of the South African Institution of Civil Engineering Volume 57 Number 3 September 2015

of the South African Institution of Civil Engineering Volume 57 No 3 September 2015 ISSN 1021-2019

Publisher South African Institution of Civil Engineering Block 19, Thornhill Office Park, Bekker Street, Vorna Valley, Midrand, Private Bag X200, Halfway House, 1685, South Africa Tel +27 (0)11 805 5947/48, Fax +27 (0)11 805 5971 http://www.saice.org.za [email protected] Editor-in-chief Prof Gerhard Heymann University of Pretoria Tel +27 (0)12 420 3627 [email protected] Contents joint Editor-in-chief Prof Chris Clayton University of Southampton [email protected] 2 Evaluation of the effect of deteriorating riding MANAGING Editor quality on bus–pavement interaction Verelene de Koker Tel +27 (0)11 805 5947, Cell +27 (0)83 378 3996 C M W Dreyer, W J vd M Steyn [email protected] journal editorial Panel Prof G Heymann – University of Pretoria 9 Comparison of two data-driven modelling Prof CRI Clayton – University of Southampton techniques for long-term streamflow Prof Y Ballim – University of the Witwatersrand Dr P Day – Jones & Wagener (Pty) Ltd prediction using limited datasets Prof J du Plessis – University of Stellenbosch Prof GC Fanourakis – University of O K Oyebode, J A Adeyemo, F A O Otieno Prof M Gohnert – University of the Witwatersrand Prof PJ Gräbe – University of Pretoria Dr C Herold – Umfula Wempilo Consulting 18 Performance assessment of aquatic macrophytes Prof A Ilemobade – University of the Witwatersrand Prof SW Jacobsz – University of Pretoria for treatment of municipal wastewater Prof EP Kearsley – University of Pretoria Prof JV Retief – University of Stellenbosch M Shah, H N Hashmi, A R Ghumman, M Zeeshan Prof E Rust – University of Pretoria Prof W Steyn – University of Pretoria Mr M Van Dijk – University of Pretoria 26 Evaluation of the response behaviour of unconfined Prof C Venter – University of Pretoria cemented materials under dynamic loading Prof A Visser – University of Pretoria Dr E Voster – Aurecon South Africa (Pty) Ltd M J Matheba, W J vd M Steyn, R J Moloisane, T I Milne Prof J Wium – University of Stellenbosch Prof A Zingoni – University of Prof M Zuidgeest – 35 Comparison of travel time between private Peer reviewing car and public transport in Cape Town The Journal of the South African Institution of Civil Engineering is a peer-reviewed journal G Hitge, M Vanderschuren that is distributed internationally Design and reproduction 44 Finite element analyses of the structural behaviour Marketing Support Services, Ashlea Gardens, Pretoria Printing of pylons supporting an inclined coal conveyor Fishwicks, Pretoria M Perduh, J A v B Strasheim Papers for consideration should be e-mailed to the Managing Editor at: [email protected] 57 Investigating the bottom free surface nappe The South African Institution of Civil Engineering accepts no responsibility for any statement made or opinion expressed (ogee profile) across a sharp‑crested weir caused in this publication. Consequently, nobody connected with the publication of this journal, in particular the proprietor, by the flow in an asymmetrical approach channel the publisher and the editors, will be liable for any loss or damage sustained by any reader as a result of his or her action S J van Vuuren, G L Coetzee, C P R Roberts upon any statement or opinion published in this journal. © South African Institution of Civil Engineering

Journal of the South African Institution of Civil Engineering • Volume 57 Number 3 September 2015 1 TECHNICAL PAPER Evaluation of the effect of Journal of the South African Institution of Civil Engineering deteriorating riding quality Vol 57 No 3, September 2015, Pages 2–8, Paper 1116 on bus–pavement interaction

WILLEMIEN DREYER Pr Eng, an Associate C M W Dreyer, W J vd M Steyn Member of SAICE (South African Institution of Civil Engineering), works for Lidwala Consulting Engineers in Nelspruit, in the Mpumalanga Province of South Africa. She obtained a BEng Deteriorating riding quality has a negative impact not only on infrastructure, but on the (Civil Engineering) degree in 2009, a BEng (Hons) road user as well. Road users experience higher vehicle operating costs (VOCs), longer travel in Geotechnical Engineering in 2012, and an times, congestion and uncomfortable rides, to name a few. The purpose of this paper is to MEng in Transportation Engineering in 2014, all demonstrate the link between deteriorating riding quality and high vertical accelerations at the University of Pretoria. Her areas of interest are geometric and pavement engineering. (awz) and VOCs for a bi-articulated bus on an identified bus route with different responsible road authorities in rural Mpumalanga, South Africa. The link, as identified, indicated that the Contact details: speed the bus travelled played a big role in the generation of a . Recommendations for future Lidwala Consulting Engineers wz P O Box 2930 research are also proposed. Nelspruit 1201 South Africa Introduction and Background This paper focuses on a study of the T: +27 (0)861 543 9252 E: [email protected] In Mpumalanga, only about 25% of house- interaction between a bi-articulated bus holds have access to private transport, which and the pavement surface of one specific

PROF WYNAND STEYN Pr Eng, a member of makes Mpumalanga one of the provinces bus route in Mpumalanga, including the SAICE (South African Institution of Civil with the highest use of public transport (bus associated VOCs generated. The objective Engineering), is a pavement engineer with a or taxi) (DOT 2003). Even though the road of the paper is thus to demonstrate the use research interest in vehicle–pavement network in Mpumalanga is extensive, the of vehicle–pavement­ interaction data to interaction, accelerated pavement testing, and maintenance and upgrade of this network is evaluate the effect of road and operating pavement materials and instrumentation. He obtained a PhD in Civil Engineering at the a concern, especially with regard to munici- conditions on ride quality and VOCs in University of Pretoria in 2001. He spent 19 years pal and provincial roads. Deteriorating roads bi‑articulated buses on typical routes. with the CSIR (Council for Scientific and Industrial Research) in various have a direct impact (such as vehicle operat- positions, and is currently professor of civil engineering (focusing on road ing costs (VOCs)) and an indirect impact Mpumalanga road network pavement related subjects) at the University of Pretoria. His professional (such as high bus fares) on the road user. There is an estimated 5 400 km of paved activities include academic and industry research in the areas of pavement engineering, vehicle–pavement interaction, and pavement materials. He has Steyn et al (2011) indicated that, in order roads in the Mpumalanga Province, and authored and co-authored 19 journal papers, 14 book chapters and to maintain the condition of a road to ensure an estimated 8 500 km of gravel roads 68 conference papers. a safe, reliable and smooth trip for road (MDPWRT 2013). The provincial roads Contact details: users, as well as protecting the underlying of the Mpumalanga Province fall under Department of Civil Engineering materials, routine road maintenance is the responsibility of the Mpumalanga University of Pretoria required. If the road is not maintained on an Department of Public Works, Roads and Private Bag X20 ongoing basis, the structural strength of the Transport (MDPWRT). The establishment Hatfield 0028 pavement reduces over time. and maintenance of the local municipal South Africa A number of factors determine the effi- roads and streets infrastructure is the T: +27 (0)12 420 2171 cient operation of a country’s economy, such responsibility of the district and local E: [email protected] as an efficient economical system, efficient municipalities. logistics system, and an efficient transport If road maintenance is delayed for more system (Steyn et al 2011). Recently, the focus than five years, the construction cost to on a country’s logistics costs (goods transpor- repair the pavement increases by six to tation) has become more visible, as these costs eighteen times, excluding other indirect and have a direct effect on the broader economy. direct costs (SAICE 2011). Indirect costs An increase in goods transportation costs include costs experienced by the non-driver, leads to an increase in end-product costs to for example increases in food prices, as fuel the consumer and that, in turn, leads to a and time are wasted on deteriorating or con- decrease in the global competitiveness of a gested roads. Direct costs include VOCs, as country, as products become more and more well as fuel and time wasted on deteriorating expensive (SOL 2012). Various studies (Chatti or congested roads. & Zaabar 2012; Steyn & Bean 2010; SOL 2010) have proved that deteriorating road quality Road roughness and riding quality results in significant increases in repair and The definition of road roughness is the une- vehicle maintenance costs, fuel consump- venness of a pavement surface defined over Keywords: vertical accelerations, riding quality, bi-articulated bus, tion and tyre wear, which in turn, lead to an an interval between two specified points vehicle operating costs, road conditions, maintenance increase in company logistics costs. (Sayers & Karamihas 1998). Engineers use

Dreyer CMW, Steyn WJvdM. Evaluation of the effect of deteriorating riding quality on bus–pavement interaction. 2 J. S. Afr. Inst. Civ. Eng. 2015;57(3), Art. #1116, 7 pages. http://dx.doi.org/10.17159/2309-8775/2015/v57n3a1 1 13 11 2 14 12

4 8 6 3 5 10

Figure 1 Layout of the acceleration configuration

Figure 2 Accelerometer route description (Google Earth 2013) road roughness as the primary indicator of driving on a road. Furthermore, there is no responsible authorities. This research should the condition of the road (riding quality). By standard to determine human discomfort, provide an objective indication of the level determining the road roughness of a pave- or comfort expressed in physical terms of increased costs down the line due to ment, engineers can estimate what mainte- such as acceleration or amplitudes at a the perceived savings by not maintaining nance, if any, is required in order to restore given frequency. Gillespie (1992), however, a road regularly. It is envisaged that the the pavement to acceptable riding quality concluded from various investigations that results of the study upon which this paper levels (Steyn et al 2011). The International there is enough proof in the available test is based can be used to motivate for more Roughness Index (IRI) is the most widely data to outline a zone above which vibration emphasis on regular maintenance, and used statistic to indicate the riding quality of is certainly intolerable and below which it resultant lower VOCs on the South African a pavement. is irrelevant. road network (specifically at provincial and Riding quality affects VOCs directly, municipal level). since the tyres and suspension system Logistic costs The case study included a route surveyed transfer the unevenness of the road surface Transport costs, storage and port costs, with a profiler and a bus fitted with acceler- as vertical accelerations to the vehicle. These inventory-carrying costs and management ometers at the locations where the vertical vertical accelerations can lead to damage of costs, administration costs and profit are accelerations were expected to be the highest vehicle components, as well as to an increase typically incorporated into the logistics (Figure 1). The accelerometers with the in fuel consumption. Further damage can costs. By controlling and managing these highest accelerations were sensors 11 and 12, be caused to the cargo being transported, costs effectively, a country’s logistics costs located on the driver side in the rear of the depending on the suspension and speed of stay in balance with the cost of general goods bus and the rear of the trailer. The outputs the vehicle (Steyn et al 2011). (SOL 2012). received from the surveys were vertical The riding quality of a road surface influ- Extensive information is available on the accelerations, bus speed, roughness of the ences the travel experience significantly. An effects of riding quality on VOCs, including road, and GPS coordinates. uneven road surface translates the vibrations the models used to determine these effects. The road roughness data for each through the suspension and tyres to the NCHRP (National Cooperative Highway road authority (municipal, provincial and vehicle body, and from there to the occu- Research Program) Project 1-45 developed a national) was categorised in three riding pants, driver and cargo (Von Becker 1992). VOC model that reflects relevant and up to quality categories. The road roughness The judgement of the road user depends date vehicle technology. Most models relate anomalies for each section were identified largely on the ride experienced by the VOCs with oil and fuel consumption, tyre with a box plot, and the cause of these user’s vehicle (Sayers & Karamihas 1998). wear, maintenance, and repair and deprecia- anomalies were determined. Some were In addition, according to Gillespie (1992), tion (Chatti & Zaabar 2012). speed humps, potholes, stop-controlled vertical acceleration measurement is the and/or robot-controlled intersections, or most common and meaningful measure Research methodology differences in surfacing. These values were of ride vibration. It is very difficult to The identified bus route, as indicated in not discarded, but were analysed with objectively evaluate the perception of Figure 2 and Table 1, consisted of roads the data, as the anomalies formed part of comfort experienced by a road user while with different roughness levels and different the route.

Journal of the South African Institution of Civil Engineering • Volume 57 Number 3 September 2015 3 The impact of road roughness on fuel Table 1 Summary of route section lengths consumption, tyre wear, and repair and Section Distance (km) maintenance costs was analysed. The Route name Responsible authority length calibrated HDM4 model (Chatti & Zaabar, (km) From To 2012) was used to predict the fuel consump- National road National road SANRAL 12.4 0.60 13.00 tion, tyre wear, and repair and maintenance cost per km of each section of the bus route P1 Provincial road MPWDRT 3.9 13.1 17.00 under consideration. P2 Provincial road MPWDRT 1.13 27.71 28.84 The methodology followed included the following steps: Municipal road Municipal road Mbombela Local Municipality 3.33 28.88 32.21 ■■ Determine riding quality P3 Provincial road MPWDRT 12.22 48.79 61.01 ■■ Determine the VOC ■■ Simulate the impact of an increase in roughness Table 2 Road roughness categories used in study ■■ Simulate the impact of a decrease in Cantisani & Sayers et al roughness Loprencipe Combined categories (1986) ■ (2010) (m/km) ■ Measure the vertical accelerations gener- (m/km) (m/km) ated by the bus ■■ Link the deteriorating riding quality to Very good ≤ 2.0 < 1.42 ≤ 2.24 Very good to good the cost impact of the user/passenger/bus Good 1.5–3.5 1.42–2.84 company, etc. Different scenarios were sketched by simulat- Fair 2.5–6.0 2.25–4.05 Fair to mediocre ing improvements in the road roughness of Mediocre 3.8–11.0 2.84–4.06 each section of the bus route and evaluating Poor > 8.0 > 4.06 > 4.05 Poor the projected costs (HDM4) if only certain sections of the bus route could be upgraded. Table 3 Road roughness category results for each section of the selected bus route Survey location Municipal National road P1 P2 P3 The route was selected so that it consisted road of roads representing different responsible Very good to good 65.4% 55.4% 45.6% 15.0% 77.4% authorities. The accelerometers were mounted on the bus at the Nelspruit bus Fair to mediocre 30.5% 28.5% 22.8% 43.7% 13.6% terminal, where commuters board the bus Poor 4.1% 8.7% 27.2% 35.0% 3.8% to travel to Nkomeni (route indicated in Anomaly (poor) none 7.4% 4.4% 6.3% 5.2% Figure 2). Five sections were identified (see Table 1). The road sections were a national road, three provincial roads (P1, P2 and P3) Table 4 Combined awz categories and a municipal road. ISO 2631–01:1997 Cantisani & Loprencipe (2010) Category 2 2 (m/s ) speed-related awz thresholds (m/s ) Threshold values Not uncomfortable ≤ 0.63 ≤ 0.315 The road roughness categories used in this paper were developed from Sayers et al Uncomfortable 0.63–1.6 0.315–0.9 (1986) and Cantisani and Loprencipe (2010) Extremely uncomfortable > 2 > 0.9 (Table 2). The categories from Cantisani and Loprencipe (2010) were combined into three, i.e. very good to good as the The results of the provincial road sec- profiler on the bus route. The costs are first category, fair to mediocre as the tions, however, indicated that the provincial indicated in Tables 8, 9 and 10 for one bus on second, and poor as the third category. The roads were in a worse condition than each section measured per km. anomalies of each section were identified expected. According to the data, P2 and the with a box plot. These values fall in the municipal road were in the worst condition, poor category, however, because of the high with 31.6% and 41.3% respectively of the Results IRI values at specific locations, as singled roads in the poor and anomaly category. Comfort is difficult to evaluate objectively, out by the data. Section P3, however, was in a better condi- as user perception of dynamic effects plays a The roughness of each road section was tion than expected, even better than the major role. The evaluation of the root mean categorised with the limits from Table 2, tested section of the national road, as the square (RMS) acceleration (awz) is required and the results are indicated in Table 3. The roughness measured was lower. to determine the comfort of a road. ISO results of the national road section were as 2631 (1997) was applied to calculate the Vehicle operating cost model expected, with most of the section in the vertical acceleration (awz). The speed-related very good to good category and with very The calibrated HDM 4 model (Chatti & awz thresholds proposed by Cantisani and little of the road section in the poor category. Zaabar 2012) was used to determine fuel Loprencipe (2010) were adopted for the anal- However, a third of the section under consid- consumption, tyre wear, and repair and ysis of the data (Table 4). The road roughness eration was in the fair to mediocre category, maintenance cost per km. The calculations versus speed versus vertical acceleration which was higher than expected. were based on the data generated from the was plotted on three-dimensional graphs to

4 Journal of the South African Institution of Civil Engineering • Volume 57 Number 3 September 2015 indicate the impact of speed on the vertical Table 5 awz weighted averages for the national road acceleration data. IRI awz S11 awz S12 Speed (m/km) (m/s2) (m/s2) (km/h) Anomalies Very good to good/not uncomfortable 1.39 0.634 0.615 73.05 The road roughness on the national road section was generally very good, and the Fair to mediocre/uncomfortable 2.88 0.820 0.920 76.32 anomalies were analysed in the poor cat- Poor/extremely uncomfortable 4.83 0.692 0.770 67.08 egory. The anomalies identified on the P1 section were due to bridge construction joints, uneven patching and surface failures. was analysed. Uniform sections that fell into In Figure 4 the speed versus road rough- Consecutive surface failures (potholes) were one category were identified over the route, ness is presented. The thresholds, indicated the cause of the anomalies on section P2. and from these sections the weighted aver- in green, orange and red, were deduced The municipal road section was in a dete- ages over the length of the section were cal- from the research done by Cantisani and riorating state, and the highest roughness culated for each road authority. The national Loprencipe (2010). was measured on this section. The reasons road section results are shown in Table 5. Provincial road sections for the anomalies on this section were due The awz values and road roughness were to the fact that this road had quite a number expected to be quite low on this road section. The data collected on the provincial road of speed humps, as the road section was in It was also expected that the higher the road sections was analysed in a similar manner as a residential area. The community tried to roughness, the higher the vertical accelera- for the data collected on the national road. curb speeding by requesting speed humps tions would be. The weighted averages, how- P2 was in the worst state, and therefore will that were higher than design standards. The ever, indicated lower accelerations at a higher be discussed in this paper. Table 6 indicates speed humps have a negative effect on the road roughness. Another parameter that the weighted average values of the isolated riding quality, bus suspension, VOCs and should be included in the discussion is speed. data for P2. fuel consumption. The remaining identified The vertical accelerations were influenced by Compared to the national road data, anomalies on the municipal road section the road roughness, as well as by the speed of the road roughness recorded on this sec- were uneven patching and surface failures, the vehicle. Cantisani and Loprencipe (2010) tion was higher, which was expected. The such as potholes. The measured data of indicated that a section of surface failure accelerations, however, were recorded as section P3 indicated that this section was will register a high road roughness, but lower, even though this section was visibly in an average state. The reasons for the depending on the speed the vehicle travels in a worse condition than the national road. anomalies were indicated as bridge joints and over that surface failure, the accelerations The difference in values could be due to consecutive potholes. The last anomaly was could be high and low. Therefore, Cantisani variation in speed. On this section it was the result of the surface changing to concrete and Loprencipe (2010) indicated that if the not possible to drive at the posted speed of paving blocks, with higher roughness and vehicle speed was reduced, a higher level of 60 km/h, as the road was in a deteriorated also a sharp horizontal curve. roughness could be tolerated. state, and there were too many potholes and Figure 3 indicates a three-dimensional speed humps. Therefore, in order to prevent Vertical acceleration (awz) versus plot of the vertical accelerations and average further damage to the bus, the driver had roughness versus speed speed of the bus at the corresponding road to drive slowly. Hence the speeds indicated Three-dimensional graphs were plotted roughness. The vertical accelerations were in Table 6 were significantly lower than to indicate the effect speed has on vertical calculated from the accelerometers mounted those recorded on the national road section. acceleration, and therefore on users’ percep- on the bus, while the speed was calculated This is not ideal, as the state of the road, tion of ride comfort. from the GPS points recorded during the congestion, speed humps and consecutive survey, and the road roughness was obtained bus stops result in longer travel time. The National road section from the profiler data. Google Earth (2013) impact of poor riding quality on travel time The data recorded over the length of the was used to correlateVery the good GPS to Good points withFair to Mediocreand congestion Poor was deemed outside the national road section on sensors 11 and 12 the isolated sections and road roughness. scope of this case study. Very Good threshold Good threshold Mediocre threshold

14.0014

12.0012 1.6 1.4 10.0010

1.2 ) 2 1.0 8.008

(m/s 0.8 6.00

wz 6 a 0.6 IRI (m/km) IRI 0.4 5.0 4.5 4.004 0.2 4.0 IRI (M/KM) 3.5 0 3.0 2.002 70 2.5 60 50 2.0 Speed (km/h)40 30 1.5 IRI (m/km) 0.00 20 10 0 30.030 40.040 50.050 60.060 70.070 80.080 90.090 N4 regression area Speed (km/h) SPEED (KM/HR) N4 data Very good to good Fair to mediocre Poor Very good threshold Good threshold Mediocre threshold Figure 3 Three-dimensional plot awz versus roughness versus speed – national road Figure 4 Speed versus road roughness – national road

Journal of the South African Institution of Civil Engineering • Volume 57 Number 3 September 2015 5 Table 6 awz weighted averages for P2 average speeds on this section of the route were very low. The roughness values were IRI awz S11 awz S12 Speed 2 2 (m/km) (m/s ) (m/s ) (km/h) also very high on this route. The awz values, however, for the not uncomfortable and Very good to good/not uncomfortable 1.47 0.530 0.504 29 extremely uncomfortable ranges were higher Fair to mediocre/uncomfortable 3.06 0.515 0.721 25 than those recorded on the national road. Furthermore, the values indicated that the Poor/extremely uncomfortable 6.85 0.588 0.984 28 road was in such a deteriorated state that it was not possible to drive faster. Table 7 awz weighted averages for municipal road The three-dimensional plot in Figure 7

indicates high roughness and awz values for IRI awz S11 awz S12 Speed 2 2 (m/km) (m/s ) (m/s ) (km/h) low speeds, and high awz values and low roughness values at high speeds. This is Very good to good/not uncomfortable 1.76 0.441 1.047 37.882 consistent with the data on the national and Fair to mediocre/uncomfortable 3.09 0.582 0.735 27.892 provincial road sections. Speed also played a significant role in users’ perception of how Poor/extremely uncomfortable 6.32 0.511 1.042 36.624 the bus travelled on the road. Figure 8 shows the data of the municipal Figure 5 shows a three-dimensional plot conclusions with this data. The speed versus road section sorted into the thresholds as of the accelerations versus speed versus road road roughness (Figure 6) correlates with the proposed by Cantisani and Loprencipe roughness. The recorded speeds on these findings that this section might be too short (2010). The very good to good, and fair to sections were very low. This section, how­ever, to make relevant conclusions. mediocre values fell in the range as proposed was also very short (approximately 1 km) by Cantisani and Loprencipe (2010). The poor relative to the other sections. After the data Municipal road section values, however, according to this graph, actu- analysis (which included isolating uniform The data collected on the municipal road ally still fall in the good to very good category. sections) it seemed that the length of the section was analysed as shown in Table 7. The principal results of this case study section may be too short to make relevant Similar to the data recorded on the P2, the were that speed played a determining role

16

14 2.0 1.8 12 1.6

) 1.4 10 2 1.2 1.0 8

(m/s 0.8

wz a 0.6 (m/km) IRI 6 0.4 14 0.2 12 10 4 0 8 40 35 6 2 30 25 4 Speed (km/h)20 15 2 IRI (m/km) 10 5 0 30 40 50 60 70 80 90 R538(2) regression area Speed (km/h) R538(2) data Very good to good Fair to mediocre Poor Outliers poor Very good threshold Good threshold Mediocre threshold Figure 5 Three-dimensional plot awz versus roughness versus speed – P2 Figure 6 Speed versus road roughness – P2

25

2.0 20

1.5 ) 2 15

1.0

(m/s wz

a 10 0.5 (m/km) IRI 20 18 0 16 14 12 5 10 60 8 50 40 6 Speed (km/h)30 4 IRI (m/km) 0 20 10 2 30 40 50 60 70 80 90 Municipal regression area Speed (km/h) Municipal data Very good to good Fair to mediocre Poor Outliers poor Very good threshold Mediocre threshold Mediocre threshold Figure 7 Three-dimensional plot awz versus roughness versus speed – municipal road Figure 8 Speed versus roughness – municipal road

6 Journal of the South African Institution of Civil Engineering • Volume 57 Number 3 September 2015 in the generation of vertical accelerations, Table 8 Percentage improvement from current cost at 60 km/h and therefore user comfort and VOCs. The Repair and Average Fuel anomalies indicated that high roughness Average IRI Tyre wear maintenance Section speed consumption (%) (%) cost values were generated by driving over a (km/h) (%) (%) surface failure (such as a pothole), change in surfacing (from asphalt to concrete block National road 14.67 60.00 0.63 0.16 14.00 paving), uneven patching, speed humps or P1 18.94 60.00 1.07 0.28 18.46 bridge construction joints. As road roughness indicates the state P2 50.01 60.00 3.91 1.03 49.33 of the road surface, the national road sec- Municipal road 52.75 60.00 4.81 1.28 51.88 tion was mostly in a good state, with about 60% in the very good to good riding quality P3 13.85 60.00 0.57 0.14 13.32 category. Section P1 was in a good state, with Average 30.04 60.00 2.20 0.58 29.40 55% in the very good to good riding quality category. The two sections with the highest roughness values were sections P2, with 30% Table 9 Decrease in cost by improving two sections of the route at 60 km/h in the poor category, and the municipal road Repair and Average Fuel section, with 40% in the poor category. These Average IRI Tyre wear maintenance Section speed consumption (%) (%) cost values were as expected; however, section P3 (km/h) (%) (%) was in a better state than expected, with only 14% in the fair to mediocre category and 9% National road 0.00 60.00 0.00 0.00 0.00 in the poor category. P1 0.00 60.00 0.00 0.00 0.00 The vertical accelerations measured by the accelerometers indicated the comfort of P2 50.01 60.00 3.91 1.03 49.33 the ride. However, the speed the bus travelled Municipal road 52.75 60.00 4.81 1.28 51.88 played the largest role in evaluating user comfort. P2 and the municipal road were the P3 0.00 60.00 0.00 0.00 0.00 two sections with the highest roughness val- Average 20.55 60.00 1.74 0.46 20.24 ues, and on these sections the bus travelled at the lowest speeds. Table 10 Percentage increase in cost as IRI increases with 1 m/km at the average measured speed Costs associated with public transport Average Repair and Fuel In order to demonstrate the potential effect Average IRI measured Tyre wear maintenance Section consumption (%) speed (%) cost of road conditions on fuel, tyre, and repair (%) (km/h) (%) and maintenance costs, an analysis was done on each section encountered on this specific National road 47.69 69.43 1.82 0.46 48.19 bus route. The current projected average P1 36.25 51.78 2.16 0.58 26.99 cost of an articulated truck (bus) travelling P2 26.63 23.55 2.85 0.70 21.34 on this route was calculated with the HDM4 model (Chatti & Zaabar 2012). Thereafter, Municipal road 21.67 21.66 2.78 0.69 18.03 the following scenarios were analysed: P3 50.91 67.11 1.87 0.48 52.48 1. Impact on VOCs if the road roughness decreases Average 36.63 46.71 2.30 0.58 33.41 2. Impact on VOCs if the worst sections were upgraded 3. Impact on VOCs if the road roughness same for each road in order to make relevant municipal road) was determined. Indicated increases. conclusions. The fuel consumption decreased in Table 9 is the potential decrease in costs, The first scenario was to determine the with an average of about 2% over the whole determined by the HDM4 model (Chatti & associated costs of a bus route with an route. The improved roughness impacted Zaabar 2012) if the roughness of section P2 average road roughness that falls below the section P2 and the municipal road section the and the municipal road section was decreased good to very good threshold. The roughness­ most, with a decrease in fuel consumption to an average roughness below 2.24 m/km. values that were worse than 2.24 m/km of about 4% and 5% respectively. The tyre Data in Table 9 indicates that, by upgrad- (worse roughness than the good to very wear decreased with an average of about 0.6% ing the two worst sections of the road, a good threshold) was changed to 2.24 m/km over the whole route. The repair and main- saving of about 2% in fuel consumption, 0.5% in the analysis. Costs were projected with tenance costs decreased with about 30% over in tyre wear and about 20% on repair and the HDM4 model developed by Chatti and the whole length of the bus route, with the maintenance cost could be possible for the Zaabar (2012). municipal road section at a decrease of 50% bus travelling on this bus route. The fuel consumption and tyre wear and section P2 at a decrease of about 52%. In order to demonstrate the impact of no values did not change significantly, but the It could be costly to upgrade the whole maintenance, the roughness on the sections repair and maintenance costs decreased bus route to an acceptable roughness (below was increased with 1 m/km. Table 10 shows significantly. The percentage decrease in costs the very good to good roughness threshold the percentage increase in fuel consumption, is indicated in Table 8. The speed shown in of 2.24 m/km), and therefore the impact of tyre wear, and repair and maintenance cost Tables 8 and 9 is 60 km/h so that it is the upgrading the two worst sections (P2 and the if the roughness increased by 1 m/km from

Journal of the South African Institution of Civil Engineering • Volume 57 Number 3 September 2015 7 the current values. The speed indicated in There are various research opportunities Vibration. Part 1: General Requirements. Geneva: Table 10 is the average speed the bus trav- around this topic, which should be investi- International Standards Organization. elled on each road for the case study. gated further. The bus company can deter- MDPWRT (Mpumalanga Department of Public Works, mine the cost of maintaining the bus route, Roads and Transport) 2013. Annual Report 2012/13. versus the current VOCs for the buses. Available at: www.dpwrt.mpg.gov.za [accessed on Conclusions 28 January 2014]. Even though it is very difficult to quantify Recommendations SAICE (South African Institution of Civil Engineering) the level of comfort on a section of road This paper focused on a case study that 2011. SAICE Infrastructure Report Card. Midrand: for the road user, it is possible to give an did not include the impact of the bus mass, SAICE. Available at: www.saice.co.za [accessed on indication. The riding comfort on the bus suspension system or bus interior. Therefore, 3 February 2012]. route depends on more factors than the road recommendations for further research would Sayers, M W, Gillespie, T D & Paterson, W D O 1986. surface, the suspension, type of vehicle and be to include these factors as they may Guidelines for conducting and calibrating road interior of a vehicle. Even though the road influence the outcome. Furthermore, driver roughness measurements. World Bank Technical surface does play a large role, the speed the fatigue, trip duration and congestion caused Paper number 46, Washington D.C.: The World vehicle travels plays the most significant role by deteriorating riding quality could also be Bank. and has the biggest impact on VOCs. included in further studies. Sayers, M W & Karamihas, S M 1998. Little Book of The three-dimensional plots indicated Profiling. Michigan, US: The Regent of the University that high roughness and awz values were of Michigan. REFERENCES recorded at low speeds, with high awz values SOL (State of Logistics) 2010. Seventh Annual State of and low roughness values at high speed. Cantisani, G & Loprencipe, G 2010. Road roughness Logistics Survey for South Africa. Pretoria: CSIR. Therefore, it can be concluded that, due to and whole body vibration: Evaluation tools and SOL (State of Logistics) 2012. Ninth Annual State of an uneven road surface, road users tend to comfort limits. ASCE Journal of Transportation Logistics Survey for South Africa. Pretoria: CSIR. drive slower for a more comfortable ride, as Engineering, 136(9): 818–826. Steyn, W J v d M & Bean, W L 2010. The potential it decreases the vertical accelerations. Chatti, K & Zaabar, I 2012. Estimating the effects of effects of deteriorating road quality and maintenance The scenario was analysed to improve pavement conditions on vehicle operating costs. in South Africa: Exploring benefit-cost analysis. In: the riding quality of the two worst sections National Cooperative Highway Research Program Seventh Annual State of Logistics Survey for South of the bus route, and by improving the road (NCHRP) Report 720, Transportation Research Africa, Pretoria: CSIR. surface of these two sections the VOCs Board (TRB), Washington D.C. Steyn, W J v d M, Bean, W, King, D & Komba, J reduced, with 2% in fuel consumption, 0.5% DOT (Department of Transport) 2003. Key Results of 2011. Evaluation of selected effects of pavement in tyre wear and about 20% on repair and the National Households Travel Survey. Available at: riding quality on logistics costs in South Africa. maintenance costs. Lack of maintenance is www.arrivealive.co.za [accessed on 13 July 2012]. Transportation Research Record, No. 2227. often the cause of deteriorating roads, and Gillespie, T D 1992. Fundamentals of vehicle dynamics. Von Becker, P J 1992. Impacts on the road and their the impact was simulated by increasing the Warrendale, PA, US: SAE International. effects on road construction and road preservation surveyed roughness values with 1 m/km. The Google Earth 2013. Available at: www.google.com costs. In: Cebon, D (Ed.), Heavy vehicles and Roads: increases in VOCs were 2% in fuel consump- [accessed on 16 November 2012 and 30 March 2013]. Technology, Safety and Policy. London: Institution of tion, 0.6% in tyre wear and about 33% on ISO 2631-1 1997. Mechanical Vibration and Shock – Civil Engineers. repair and maintenance costs. Evaluation of Human Exposure to Whole-Body

8 Journal of the South African Institution of Civil Engineering • Volume 57 Number 3 September 2015 Comparison of two TECHNICAL PAPER Journal of the South African data-driven modelling Institution of Civil Engineering techniques for long-term Vol 57 No 3, September 2015, Pages 9–17, Paper 1043 streamflow prediction OLUWASEUN OYEBODE (MWISA, MIWA, MIAHS) is a Master’s (research) candidate in Civil Engineering at the Durban University of using limited datasets Technology. He graduated with a BSc Upper Second Class Honours in Environmental Engineering from the University of Ibadan, O K Oyebode, J A Adeyemo, F A O Otieno Nigeria. His focus is in the fields of hydrological modelling and climate change impacts on water resources, with particular interest in the development of models using evolutionary computation and artificial intelligence techniques. His current This paper presents an investigation into the efficacy of two data-driven modelling techniques research relates to the use of genetic programming and differential in predicting streamflow response to local meteorological variables on a long-term basis and evolution-trained neural networks to model streamflow response to local under limited availability of datasets. Genetic programming (GP), an evolutionary algorithm hydro-climatic variables in the upper uMkhomazi River. approach and differential evolution (DE)-trained artificial neural networks (ANNs) were applied Contact details: for flow prediction in the upper uMkhomazi River, South Africa. Historical records of streamflow, Department of Civil Engineering and Surveying rainfall and temperature for a 19-year period (1994–2012) were used for model design, and Durban University of Technology PO Box 1334, Durban 4000, South Africa also in the selection of predictor variables into the input vector space of the model. In both T: +27 (0)84 807 3576 approaches, individual monthly predictive models were developed for each month of the year E: [email protected] using a one-year lead time. The performances of the predictive models were evaluated using three standard model evaluation criteria, namely mean absolute percentage error (MAPE), root DR JOSIAH ADEYEMO (MASABE, MASCE, MIWA, mean-square error (RMSE) and coefficient of determination (R2). Results showed better predictive MWISA) is a Senior Lecturer in the Department performance by the GP models (MAPE: 3.64%; RMSE: 0.52: R2: 0.99) during the validation phase of Civil Engineering and Surveying at the when compared to the ANNs (MAPE: 93.99%; RMSE: 11.17; R2: 0.35). Generally, the GP models Durban University of Technology. He obtained his BSc (Honours) at the University of Ilorin, were found to be superior to the ANNs, as they showed better performance based on the three Nigeria, his MSc at the University of Ibadan, evaluation measures, and were found capable of giving a good representation of non-linear Nigeria, and his doctorate at the Tshwane hydro-meteorological variations despite the use of minimal datasets. University of Technology. He focuses on developing and applying evolutionary optimisation techniques to solve real-world science and engineering design problems at minimum cost and INTRODUCTION for maximum benefit. He is renowned for the development of a Numerous researchers have applied vari- multi-objective evolutionary algorithm called multi-objective differential The need to manage water resources in ous approaches to predicting streamflow evolution algorithm (MDEA) which is used by many researchers worldwide. arid and semi-arid regions has always been – from the use of traditional auto-regressive Contact details: of high importance to water managers (AR) models (Jayawardena & Lai 1994; Wang Department of Civil Engineering and Surveying and decision-makers, especially in this et al 2009; Wu et al 2009), to the use of con- Durban University of Technology era of increased climate variability. Water ceptual, process-based and physically-based PO Box 1334, Durban 4000, South Africa ­resources engineers and other stakeholders models (also referred to as the “knowledge T: +27 (0)31 313 2985 E: [email protected] have developed various approaches to man- driven models”) (Limbrick et al 2000; Butts aging the relatively little amount of water et al 2004; Chiew 2006; Jiang et al 2007; PROF FRED OTIENO Pr Eng (FSAICE, SFWISA) is a in these regions in order to ensure constant Leander & Buishand 2007), to the data- C-rated researcher with the National Research availability of water for domestic, indus- driven models (DDMs) (Cannon & Whitfield Foundation (NRF), South Africa. He is a Fellow of trial, ecological and agricultural purposes. 2002; Maity & Kashid 2010; Zakaria & the South African Institution of Civil Streamflow remains a fundamental compo- Shabri 2012; Galelli & Castelletti 2013; Engineering, a Senior Fellow of the Water nent of the water cycle and a major source Kasiviswanathan & Sudheer 2013). In recent Institute of Southern Africa (WISA), and was the WISA President in 2007/2008. He has over 30 of freshwater availability for human, animal, years, the application of DDMs have gained active years of consulting, lecturing and plant and natural ecosystems (Makkeasorn et more popularity due to their good perfor- research experience in a number of disciplines, more recently focusing on al 2008). Therefore, prediction of streamflow mance when applied to complex hydrological water resources management, water and wastewater treatment, solid waste both on a short-term and long-term basis is modelling problems. DDMs are models management and general environmental management. of crucial importance to water managers as that give representation of system state Contact details: it forms the basis upon which their decisions variables such as input, and internal and Department of Civil Engineering and Surveying are made. While short-term predictions are output variables, while characterising the Durban University of Technology PO Box 1334, Durban 4000, South Africa made to provide signals about flood dangers nature of hydrological processes within the T: +27 (0)31 373 2375 and drought, long-term predictions help in system. They do this by taking into account E: [email protected] providing information for long-term water only a few assumptions about the physics of supply strategies (Kisi & Cigizoglu 2007). the system being modelled. DDMs are now Such information is needed, for example, being considered as an approach that could when making decisions on the location and complement or replace the knowledge-driven Keywords: data-driven models, artificial neural networks, genetic programming, sizing of reservoirs on a river. models (Solomatine & Ostfeld 2008; Londhe streamflow prediction, upper uMkhomazi River

Oyebode OK, Adeyemo JA, Otieno FAO. Comparison of two data-driven modelling techniques for long-term streamflow prediction using limited datasets. J. S. Afr. Inst. Civ. Eng. 2015;57(3), Art. #1043, 9 pages. http://dx.doi.org/10.17159/2309-8775/2015/v57n3a2 9 & Charhate 2010). A major reason is that results from the latter have been found to exhibit higher-degree uncertainties in their structural makeup and parameterisations when compared to DDMs (Poulin et al 2011; Il-Won et al 2012; Montanari & Di Baldassarre 2013). Hence, the use of DDMs is seen as a promising technique for solving these sensitivity and uncertainty issues, as well as other hydrological modelling-related problems. The genetic programming (GP) approach is a prominent DDM that has proven appli- cability to hydrological modelling. GP is a member of the evolutionary algorithm (EA) family and has been applied in a wide range of science-related and engineering analyses. GP has performed well in various water- Figure 1 Location of the uMkhomazi River and gauging stations around the catchment related studies, such as sediment transport modelling, streamflow prediction, rainfall- streamflow prediction in the study area. METHODOLOGY runoff modelling, ecological modelling, This study is unique, as comparative studies uncertainty assessment studies, etc (Liong between two evolutionary-inspired tech- Genetic programming et al 2002; Muttil & Lee 2005; Ni et al 2010; niques (GP and a DE-trained ANN) are very Genetic programming (GP) (Koza 1992) is a Garg 2011; Selle & Muttil 2011; Sirdari et al uncommon, especially when employed under population-based search which is inspired 2011; Kisi et al 2012). limited availability of datasets. by the Darwinian principle of natural selec- Another extensively used data-driven tion (survival of the fittest). GP is a member modelling technique is artificial neural of the evolutionary algorithm (EA) family networks (ANN). ANN is inspired by STUDY AREA AND DATASETS which performs its operations by genetically neuroscience and uses its adaptive learning The upper uMkhomazi River is located breeding a population of computer programs feature to solve problems in domains with within the province of KwaZulu-Natal in to solve problems. GP initialises by randomly little or no prior knowledge of the system South Africa, and is the third largest river in generating programs that are perceived being modelled. Over the last two decades, the province. The river is of high importance to be candidate solutions to the problem. ANN has been successfully applied to due to its role as a major source of water Programs are then chosen from the pool, various fields of water resources, including supply to the densely populated urban and evaluated based on a “fitness function” function approximation, classification and areas of Durban and Pietermaritzburg. The which describes how well they solve the forecasting studies (Coulibaly et al 2001; uMkhomazi River is approximately 160 km given problem. The selected best programs Moradkhani et al 2004; Cigizoglu 2005; long and is elevated at about 3 300 m above are then transformed into a new generation Dibike & Coulibaly 2006; Kisi & Cigizoglu sea level. The river derives its source from of computer programs using genetic opera- 2007; Heng & Suetsugi 2013). With the the upper Drakensberg Mountains and tors which apply slight modifications to the incorporation of ANNs, ensembles of discharges into the Indian Ocean, drain- structure of the selected programs to achieve models are being built to form modular ing an area of 4 400 km2. The climate is better solutions/programs. These succes- or hybrid models in order to increase the characterised by wet summers which occur sive iterations continue until a termination confidence level of predictive assessment between November and March, and dry win- criterion is met. The program returned at studies and to reduce model uncertainty ters which extend over the months of June the end of the run is finally chosen as the (Abrahart et al 2012). to September. Mean annual precipitation best program and the model that best solves However, in order to achieve accurate and varies between 700 to 1 200 mm year-1, with the given problem. The principal operators reliable predictions in hydrological studies, highly intra- and inter-seasonal streamflows employed in GP are: large datasets are often required for the estimated to produce an average annual yield 1. Selection: Parent programs are chosen purpose of model training (Babovic & Keijzer of 568 million (Flugel & Marker 2003). probabilistically based on their fitness 2002). These huge numbers of datasets are Past records of mean streamflow on values for the purpose of reproduction. limited to certain regions and are often a monthly basis were obtained from the 2. Crossover: A modification to the unavailable in some areas, especially in Department of Water Affairs (DWA). structure of the parent programs which developing countries (Ni et al 2012). Nineteen-year data from gauging station­ involves swapping some sections to pro- The main objective of this study is U1H005 (uMkhomazi River @ Lot 93 1821) duce offspring programs. to develop models capable of long-term with geographical coordinates between 3. Mutation: The creation of an offspring streamflow prediction in response to non- 29΄ 44΄ 37.3΄ south longitudes and program by randomly altering a struc- linear fluctuations of meteorological vari- 29΄ 54΄ 17.8΄ east latitudes were applied in tural member or node of a selected parent ables in the upper uMkhomazi River. The this study (Figure 1). The South African program. potentials of the GP and ANN approach in Weather Service (SAWS) provided the corre- The GP representation consists of numerical providing models, using the few available sponding climatic data from three independ- constants and variables generally referred to datasets, were subjected to test and their ent weather data stations (Pietermaritzburg, as “terminals”, T, and arithmetic, relational ­performances evaluated comparatively so Shaleburn and Giant’s Castle) located within and trigonometric operations which are as to determine the suitable approach for the study area. internal nodes called “functions”, F. The

10 Journal of the South African Institution of Civil Engineering • Volume 57 Number 3 September 2015 selected terminals and functions constitute and non-linear input-output interactions rainfall values from the three independent the primitive set of the GP algorithm. The makes it suitable for predictive studies in the weather stations for the same month in five major preparatory steps that should be field of water resources. the preceding year (R1t, R2t, R3t) and their adopted before applying GP to a problem corresponding temperature values (T1t, T2t, Selection of input variables involve (i) selecting the set of terminals, T3t). Weather stations 1, 2 and 3 represent (ii) selecting the set of primitive functions, The modelling strategy employed in this Pietermaritzburg (PMB), Shaleburn, and (iii) determining the fitness function, study was to subject the two approaches (GP Giant’s Castle weather stations respectively. (iv) determining the parameters for control- and ANN) to the same set of datasets to The approach employed for long-term ling the run, and (v) defining the criterion avoid any form of bias. Hence, the same set streamflow prediction in this study was for terminating the run (Maity & Kashid of input variables were used for both models. to adopt a one-year lead time. Therefore, 2009). The reader is referred to Koza (1992), However, the choice of input variables was the streamflow being modelled for a given

Babovic & Keijzer (2000) and Poli et al (2008) dependent on the few available datasets. month in the next year (Qt+1) is designated as for more in-depth discussion on GP. Although there are several processes that the target output. The mathematical repre­ influence streamflow generation in river sentation of the one-year lead time model Artificial neural networks (ANNs) hydrology, such as precipitation, tempera- adopted can be expressed as: ANN is a computational intelligence (CI) ture, evaporation, soil moisture, vegetation method inspired by the neurological process- cover, land use, etc (Loucks & van Beek Qt+1 = f (Qt, Qt–1, Qt–2, R1t, R2t, R3t, T1t, ing ability of the human brain. ANN models 2005; Raghunath 2007), only the available T2t, T3t) (1) consist of a pool of simple processing units datasets of rainfall and temperature were called neurons which communicate by send- used alongside that of streamflow for input GP models ing signals to one another over a large num- variable selection. The streamflow, rainfall The GP predictive models for long-term ber of weighted connections (Kröse & van and temperature data made available by streamflow prediction in this study were der Smagt 1996). The operating principles of the DWA and SAWS cover a 19-year period developed using an objective function – to ANNs is based on parallel distributed infor- (1994–2012). The rainfall and temperature minimise the mean-square error that can mation processing that is capable of storing datasets were collected from three inde- be obtained between the predicted and the experiential knowledge gained through the pendent weather stations located within observed values of streamflow. The mean- process of learning, and making it available the study area. Results of serial correlation squared error function which measures the for future use (Elshorbagy et al 2010). The analysis show high correlation between the fitness of evolved programs is calculated by processing units function by receiving inputs values of streamflow for the past three years taking the average of the squared raw errors from external sources or other neurons and that of any pre-selected year. The results, over the values in the training dataset. This in the network, and computing output however, revealed that streamflow for the can be expressed mathematically as: signals which is transmitted to other units. pre-selected year had close relationship n 2 These processing units are found in layers with rainfall and temperature values of the (Qoi – Qpi) F = Min ∑ (2) commonly categorised as input, hidden or preceding year across the three independent i=1 n output layers. The use of an activation func- weather stations. The datasets were split tion in the hidden node of ANNs helps in randomly into two subsets, with two thirds Qo and Qp are observed and predicted values transforming the non-linearity in the inputs of the datasets used for model training and of streamflow respectively, n is the number into a linear space. The commonly used the remaining third for validation. The of data points, and i is the counter from 1 to activation functions are sigmoidal functions random splitting was done in a manner in the number of data points. such as the logistic and hyperbolic tangent which the validation datasets were within The ability of GP to screen and prioritise functions (Maier & Dandy 2000). The major the range of the training datasets, thereby input variables during its run contributes network topologies that characterise the making the datasets representative of the to the fitness of the evolved programs, thus architecture of ANNs are the feed-forward same population. ensuring the accuracy of its predictions. This neural networks (FFNN) and re-current is achieved by expressing the contribution neural networks (RNN); with multilayer of each input variable as a function of its perceptron (MLP), radial basis function MODEL DEVELOPMENT frequency of occurrence. The primitive set of (RBF) networks, Kohonen’s self-organising Both the GP and ANN models that were the GP was supplied with arithmetic, com- feature maps (SOFM) and Elman-type RNN investigated for long-term streamflow parison, logistic and trigonometric functions as the most popular ANNs (Coulibaly & prediction in this study were developed by in order to capture details of the relationship Evora 2007; Jha 2007). Numerous specialised adopting a monthly approach. It has been between the input variables and the target learning algorithms have been employed found that the use of individual monthly output. A distributed population structure, for the purpose of training and subjecting models in high-lead-time prediction pro- which involves the subdivision of the popula- ANNs to adaptive learning. The earliest and duces better predictions when compared to tion space into multiple subpopulation or most popular method that has been used the adoption of a single model, which often demes, was employed in this study. This to train ANNs is the back propagation (BP) produces poor predictions (Sivapragasam et subdivision allows for occasional migration algorithm. However, in recent times research al 2011). Hence, a total of twelve individual of individuals among demes for exchange of has produced improved algorithms for ANN. monthly models (one for each month of the genetic material, in order to achieve evolu- These include methods based on gradient year) were developed using both modelling tion of the entire population, quicken the descent like quick propagation (QP) and the approaches. The input spaces of the GP and evolution process and also to prevent prema- Levenberg-Marquardt (LM) algorithm, and ANN models were populated with a total of ture convergence. heuristic methods such as genetic algorithm nine input variables. These input variables The implementation of GP in this study (GA) and differential evolution (DE). Hence, comprised streamflow values for a given was done by using a program-based GP tool the ability of ANNs to assimilate complex month in the last three years (Qt, Qt–1, Qt–2), called Discipulus (Francone 2011). Discipulus

Journal of the South African Institution of Civil Engineering • Volume 57 Number 3 September 2015 11 Table 1 Summary of parameter settings used Input layer Hidden layer Output layer to control the GP algorithm run

Parameters Value I1 Initial 80, Program size maximum 512

Mutation frequency 95% I2 Crossover frequency 50%

Block mutation rate 30% O1

Instruction mutation rate 30% Linear activation Instruction data mutation rate 40% function

Population size 500 Sigmoidal-type Maximum number of run 300 activation function I9 Maximum number of 10 000 generations since start of run

Homologous crossover 95% Figure 2 Architecture of the three-layer feed-forward neural network (FFNN) Number of demes 10 layer nodes from 2 to 10 using a single (one) 3. Coefficient of determination (R2): Migration rate 1% stepping approach. indicates a better model as its value The ANN was trained using a differential approaches 1. is a linear genetic programming (LGP) soft- evolution (DE) algorithm. A total run of 2 (Qo – Qo )(Qp – Qp ) ware that evolves models in the form of com- 10 000 generations was adopted for optimal R2 = ∑ 2 2 (5) puter programs based on the least sum of training after a number of trial runs. The ∑(Qo – Qo ) ∑(Qp – Qp ) squared errors. The goodness-of-fit is mea- population size, NP, crossover constant, CR, sured using R-square and F-score statistics and mutation scale factor, F, were used to Qo and Qp represent observed and predicted against observed values of the training and control amplification of differential variation streamflows respectively, Qo and Qp repre- validation datasets. The default parameter during the run. Following the suggestion sent their corresponding mean values, n is the settings recommended by Francone (2011) of Price and Storn (2013), NP, CR and F number of data points, and i is the counter were used to control the GP run (Table 1). were set at “D multiplied by 10”, 0.9 and from one to the number of data points. Francone (2011) states that the default set- 0.4 respectively, (where D is the number of Considering that the maximum number of tings for a Discipulus project work quite well weights and biases in the selected architec- lags needed to predict the next year’s flow is for most projects, and that Discipulus auto- ture). In the hidden layer of the FFNN, a three, the 19-year datasets constituted 16 data matically sets, randomises, and optimises the logistic sigmoidal-type activation function points for each monthly model. Lower values GP parameters for project runs. of between 0 and 1 was used to scale the of MAPE and RMSE would indicate better The GP algorithm for each computa- inputs in the range 0.1–0.9. A linear activa- predictive accuracy of the model, while higher tion was run on an Intel Core i7 PC with tion function was, however, employed in the values of R2 (close to 1.0) would indicate bet- 3.40 GHz and 4 GB RAM. The maximum output layer. ter predictive accuracy of the models. size of each evolved program was restricted to 512, initialising with 80 instructions Performance evaluation per program. This was done to prevent The performance of the models developed in RESULTS AND DISCUSSIONS the phenomenon of bloating, which means this study was evaluated using three stand- The performance evaluation results of the over-growing of programs without limits and ard statistical measures, namely mean abso- two DDM approaches (GP and ANN) on without any improvement in the fitness of lute percent error (MAPE), root mean-square long-term streamflow prediction in the the population (Bleuler et al 2001). error (RMSE) and coefficient of determina- upper uMkhomazi River are presented in tion (R2). The three performance evaluation Tables 2a and 2b, for the training and valida- ANN models criteria can be computed using the following tion datasets respectively. It can be observed The multi-layer feed-forward neural network mathematical expressions: from Table 2a that both the GP and ANN (FFNN), one of the most widely used net- 1. The mean absolute percent error (MAPE): models provided very competitive perfor- work architecture in hydrological modelling indicates a better model as its value mances during the training phase, with the systems, was employed for the purpose approaches zero. ANN models having a slight edge over the of comparison. The architectural design GP models. The maximum values of MAPE n of the FFNN models developed comprise 1 Qp – Qo and RMSE recorded by the ANN models MAPE = ∑ × 100 (3) three layers – one input, one hidden and n i=1 Qo were 5.31% and 0.72 respectively, while their an output layer (Figure 2). The input layer corresponding values in the GP models were consists of nine input nodes representing 2. Root mean-square error (RMSE): computed to be 11.15% and 1.50. However, in the nine selected input variables, while the indicates a better model as its value both approaches, R-squared values showed output node consists of only one neuron approaches zero. a high correlation between observed and (target output). The optimal architecture of predicted streamflows. The R-squared values n 2 each individual model was determined by (Qo – Qp) ranged between 0.9918–1.0000 in the ANN RMSE = ∑ (4) incrementally varying the number of hidden i=1 n models, and 0.9891–0.9994 in the GP models.

12 Journal of the South African Institution of Civil Engineering • Volume 57 Number 3 September 2015 Table 2(a) Comparison of MAPE, RMSE and R2 values between GP and ANN models during training Table 3 Network architecture showing optimal number of hidden layer nodes in the Training phase ANN models MAPE RMSE R2 Month Optimal network Month GP ANN GP ANN GP ANN architecture

January 3.9401 1.2865 1.4968 0.6778 0.9964 0.9992 January 9-10-1

February 0.9966 0.4008 0.4974 0.2689 0.9994 0.9997 February 9-7-1

March 3.6570 3.1207 1.2469 1.1404 0.9982 0.9989 March 9-10-1

April 5.6798 3.4220 1.0710 0.7181 0.9891 0.9949 April 9-7-1

May 2.6864 0.6667 0.1982 0.0694 0.9970 0.9991 May 9-10-1

June 1.1854 1.4E-08 0.0607 7.8E-10 0.9986 1.0000 June 9-5-1

July 1.2691 0.2050 0.0558 0.0150 0.9985 1.0000 July 9-7-1

August 4.2479 4.1465 0.1047 0.1799 0.9972 0.9918 August 9-8-1

September 11.1474 5.3100 0.0607 0.2615 0.9988 0.9984 September 9-7-1

October 3.1860 4.2952 0.1607 0.4268 0.9994 0.9953 October 9-9-1

November 4.2854 2.2617 0.4855 0.2183 0.9975 0.9995 November 9-10-1 December 6.0007 1.1E-10 0.8642 3.6E-11 0.9972 1.0000 December 9-4-1 Average 4.0235 2.0929 0.5252 0.3313 0.9973 0.9981

could prevent over-training. Furthermore, 2 Table 2(b) Comparison of MAPE, RMSE and R values between GP and ANN models during validation it was noticed that the optimisation of the network architecture, as determined using the Validation phase DE-algorithm, resulted in a slow convergence 2 Month MAPE RMSE R rate, and hence increased computational GP ANN GP ANN GP ANN time. This can be understood better from Table 3, which presents the optimal network January 3.4179 65.4074 1.3535 29.1835 0.9969 0.0992 architecture of the individual ANN models February 3.0490 22.7069 1.0126 18.3385 0.9923 0.9651 as returned at the end of each run. It was March 4.3077 35.3504 1.3797 20.2504 0.9932 0.3559 observed during the runs that the training speed becomes slower as the number of April 6.1725 21.9832 1.1868 6.3037 0.9741 0.5174 hidden layer nodes increases. This implies May 1.8900 57.2785 0.1245 3.8866 0.9954 0.2272 that increase in the number of synaptic con- June 0.5932 40.5837 0.0325 1.6614 0.9990 0.6659 nections between units imposes a greater number of weights on the network. This is in July 0.7575 87.6803 0.0257 2.5899 0.9999 0.5788 line with the submission of Karthikeyan et al August 7.5730 58.1474 0.3401 3.9421 0.9881 0.0080 (2013) that the number of hidden layer nodes September 5.1472 100.6726 0.1983 1.8091 0.9740 0.2979 in­fluence computational time, and conse- quently, that ANNs require an ample period October 6.1092 125.0948 0.2653 3.3390 0.9905 0.0959 of time for continuous training in order to November 2.1735 255.0309 0.2918 24.8077 0.9978 0.1231 achieve better convergence when used on December 2.4802 89.3958 0.5242 16.0652 0.9978 0.0917 small datasets. Some researchers (Ilonen et al 2003; Ghaffari et al 2006; Corzo & Average 3.6392 93.9918 0.5612 11.1708 0.9916 0.3486 Solomatine 2007) have, however, opined that the idea of increasing the computational time It can also be noted in both approaches correlations recorded during training, with should not be seen as a guarantee to achieving that low flows, which occur during the R-squared estimates of 0.9740–0.9999 during better generalisation, as this effort may yield months of May to July, produced smaller validation. It may be inferred that the ability no practical improvement in the results. error estimates compared to the high flows of the GP approach to screen and prioritise In contrast, GP exhibited better generali- which occur between December and March, input variables contributed to the fitness of sation ability at a faster learning rate, a prod- with February as an exception. its models (Makkeasorn et al 2008). uct of its ability to distribute the population On the other hand, the GP models However, the error estimates in the into demes (Brameier & Banzhaf 2001). The performed considerably better than the DE-ANN models increased marginally, yield- distribution of the population into demes ANN models during validation, as the ANN ing higher MAPE and RMSE values while allowed for occasional migration of individu- models produced higher errors. The errors generating lower R-square values. Although als between sub-populations for exchange produced in the GP models during validation the DE-ANN learning process was satisfac- of genetic materials, thereby leading to the were better converged towards zero and are tory, the generalisation was poor. The poor occurrence of parallel evolution. This evolu- estimated to be 0.6%–7.6% and 0.03–1.38 for generalisation may be attributed to the use tion further minimised the chances of the MAPE and RMSE respectively. Also, all the of small datasets (Zhang et al 2010), and also GP algorithm converging to local optima, GP models maintained the highly positive to the non-existence of an approach that and also ensured faster convergence.

Journal of the South African Institution of Civil Engineering • Volume 57 Number 3 September 2015 13 100 100 January February /sec) /sec)

3 80 3 80 60 60 40 40 20 20 Streamflow (m Streamflow (m 0 0 1 2 3 4 5 6 7 8 9 10 11 1 2 3 4 5 6 7 8 9 10 11 Data points Data points

140 50 March April 120 /sec) /sec)

3 3 40 100 80 30 60 20 40 10 20 Streamflow (m Streamflow (m 0 0 1 2 3 4 5 6 7 8 9 10 11 1 2 3 4 5 6 7 8 9 10 11 Data points Data points

14 10 May June 12 /sec) /sec)

3 3 8 10 8 6 6 4 4 2 2 Streamflow (m Streamflow (m 0 0 1 2 3 4 5 6 7 8 9 10 11 1 2 3 4 5 6 7 8 9 10 11 Data points Data points

8 10 7 July August /sec) /sec) 3 6 3 8 5 6 4 3 4 2 2 1 Streamflow (m Streamflow (m 0 0 1 2 3 4 5 6 7 8 9 10 11 1 2 3 4 5 6 7 8 9 10 11 Data points Data points

25 25 September October /sec) /sec)

3 20 3 20 15 15 10 10 5 5 Streamflow (m Streamflow (m 0 0 1 2 3 4 5 6 7 8 9 10 11 1 2 3 4 5 6 7 8 9 10 11 Data points Data points

40 60 35 November December

/sec) /sec) 50 3 30 3 25 40 20 30 15 20 10 5 10 Streamflow (m Streamflow (m 0 0 1 2 3 4 5 6 7 8 9 10 11 1 2 3 4 5 6 7 8 9 10 11 Data points Data points

Observed ANN GP

Figure 3 Observed and predicted streamflows during training

14 Journal of the South African Institution of Civil Engineering • Volume 57 Number 3 September 2015 120 100 January February

/sec) 100 /sec)

3 3 80 80 60 60 40 40 20 20 Streamflow (m Streamflow (m 0 0 1 2 3 4 5 1 2 3 4 5 Data points Data points

80 40 70 March 35 April /sec) /sec) 3 60 3 30 50 25 40 20 30 15 20 10 10 5 Streamflow (m Streamflow (m 0 0 1 2 3 4 5 1 2 3 4 5 Data points Data points

16 6 14 May June

/sec) /sec) 5 3 12 3 10 4 8 3 6 2 4 2 1 Streamflow (m Streamflow (m 0 0 1 2 3 4 5 1 2 3 4 5 Data points Data points

10 12 July August

/sec) /sec) 10

3 8 3 8 6 6 4 4 2 2 Streamflow (m Streamflow (m 0 0 1 2 3 4 5 1 2 3 4 5 Data points Data points

7 10 September October 6 /sec) /sec)

3 3 8 5 4 6 3 4 2 2 1 Streamflow (m Streamflow (m 0 0 1 2 3 4 5 1 2 3 4 5 Data points Data points

70 50 November December 60 /sec) /sec)

3 3 40 50 40 30 30 20 20 10 10 Streamflow (m Streamflow (m 0 0 1 2 3 4 5 1 2 3 4 5 Data points Data points

Observed ANN GP

Figure 4 Observed and predicted streamflows during validation

Journal of the South African Institution of Civil Engineering • Volume 57 Number 3 September 2015 15 Figures 3 and 4 present plots of observed (GP) and artificial neural networks (ANN), of anarchy? Emerging themes and outstanding values against predicted values as simulated were employed comparatively for long-term challenges for neural network river forecasting. by the two approaches for both training and streamflow prediction. Results clearly showed Progress in Physical Geography, 36(4): 480–513. validation phases. The plots clearly reveal the efficacy of the GP approach in giving a Babovic, V & Keijzer, M 2000. Genetic programming remarkable performance of the GP models better representation of complex and non- as a model induction engine. Journal of during both training and validation, unlike the linear input-output relationships, despite the Hydroinformatics, 2: 35–60. ANN models which were not able to replicate use of limited datasets. The GP models devel- Babovic, V & Keijzer, M 2002. Rainfall runoff modelling the performances recorded during the training oped obtained better performance as average based on genetic programming. Nordic Hydrology, phase, as a result of some under- or over- values of mean absolute percent error (MAPE) 33(5): 331–346. estimation of observed values. Despite ensur- = 4.02% and 3.64%; root mean-square error Bhattacharya, M, Abraham, A & Nath, B. A 2001. ing that the validation datasets were within the (RMSE) = 0.53 and 0.56, and R2 = 0.9973 and Linear genetic programming approach for range of the training datasets, the differences 0.9916 during training and validation respec- modeling electricity demand prediction in Victoria. between observed and predicted streamflows tively. However, the corresponding values of Proceedings, Abraham, A & Koppen, M (Eds), during the months with high flows were more MAPE, RMSE and R2 in the ANN models International Workshop on Hybrid Intelligent pronounced during validation. were estimated to be 2.09% and 93.99%; 0.33 Systems, 2001 Adelaide. Heidelberg: Physica-Verlag, The poor performance of the ANN and 11.17, and 0.9981 and 0.3486 respectively. pp 379–393. models during validation is considered to be Though the use of ANNs remains a flexible Bleuler, S, Brack, M, Thiele, L & Zitzler, E 2001. due to the problem of over-parameterisation approach known for its prominent feature of Multiobjective genetic programming: Reducing and over-fitting, which is typical of ANNs capturing non-linearity inherent in hydro- bloat using SPEA2. Proceedings of IEEE Congress on (Bhattacharya et al 2001), especially when logical systems modelling, this study clearly Evolutionary Computation, Vol. 1: 536–543. subjected to small numbers of datasets. showed that the large number of datasets Brameier, M F & Banzhaf, W 2001. A comparison of This further indicates that the use of ANNs required to achieve accurate and reliable linear genetic programming and neural networks is problem-specific and data-dependent results serve as a major drawback to its use, in medical data mining. IEEE Transactions on (Bhattacharya et al 2001), and that high especially in areas where the availability of Evolutionary Computation, 5(1): 17–26. difficulty exists in modelling hydrologic pro- datasets is limited. Over-training could also Butts, M B, Payne, J T, Kristensen, M & Madsen, H cesses with limited datasets (Ni et al 2010). have been a problem. The convergence rate 2004. An evaluation of the impact of model On the contrary, the GP models in most of the DE-trained ANNs was found to be structure on hydrological modelling uncertainty cases simulated the streamflows closely and slower, requiring a considerable amount of for streamflow simulation. Journal of Hydrology, achieved better convergence than they did time for model training. A potential solution 298(1): 242–266. during training. Both the low and high flows in this regard is the hybridisation of learning Cannon, A J & Whitfield, P H 2002. Downscaling were substantially reproduced by the GP algorithms, which is a combination of two recent streamflow conditions in British Columbia, models, including the spikes that character- or more learning algorithms for ANNs to Canada, using ensemble neural network models. ised the streamflows in some months. This achieve better adaptive learning. Journal of Hydrology, 259(1): 136–151. further affirms the ability of GP in capturing In contrast to ANNs, the GP models Chiew, F H 2006. Estimation of rainfall elasticity of normal events, as pointed out in Londhe & trained faster and achieved better conver- streamflow in Australia. Hydrological Sciences Charhate’s (2010) river flow predictive study. gence, thereby producing close agreement Journal, 51(4): 613–625. Also, the overfitting problems often associ- between observed and predicted values, Cigizoglu, H K 2005. Application of generalized ated with the ANNs were greatly minimised with highly positive correlations during regression neural networks to intermittent flow in the GP models, the reason being that both training and validation. Generally forecasting and estimation. Journal of Hydrologic GP ranks its potential candidate solutions it can be concluded from this study that Engineering, 10(4): 336–341. (program models) in terms of their fitness, genetic programming can be employed for Corzo, G & Solomatine, D 2007. Baseflow separation and often discard those with poor fitness. long-term streamflow prediction in the techniques for modular artificial neural network The ability of the Discipulus GP model to upper uMkhomazi River despite the limited modelling in flow forecasting. Hydrological Sciences produce better solutions via the combination availability of datasets. The monthly models Journal, 52(3): 491–507. of the best single program models into team developed can be deployed as predictive tools Coulibaly, P, Anctil, F & Bobee, B 2001. Multivariate models (Francone 2011), also ensured the for the purpose of planning and management reservoir inflow forecasting using temporal neural predictive accuracy of the GP models. of water resources within the uMkhomazi networks. Journal of Hydrologic Engineering, The consistency of the GP approach in region. In addition, this study further dem- 6(5): 367–376. simulating the hydro-climatological pro- onstrates GP as a powerful predictive tool Coulibaly, P & Evora, N 2007. Comparison of neural cesses in the study area is evident, as the GP in hydrologic modelling studies, which can network methods for infilling missing daily weather models were able to accurately capture the be considered as an alternative approach to records. Journal of Hydrology, 341(1): 27–41. rainfall-temperature-streamflow relationship the ANNs, especially in data-sparse regions. Dibike, Y B & Coulibaly, P 2006. Temporal neural in each month of the year. The results agree Future work will focus on the conjunctive networks for downscaling climate variability and with those in similar studies (Makkeasorn use of GP and other evolutionary computa- extremes. Neural Networks, 19(2): 135–144. et al 2008; Guven 2009; Londhe & Charhate tion techniques, such as screening and Elshorbagy, A, Corzo, G, Srinivasulu, S & Solomatine, 2010), in which GP-derived models have also optimisation tools, for improving the perfor- D 2010. Experimental investigation of the been found to showcase better predictive mance of ANNs. predictive capabilities of data-driven modeling accuracy than the ANNs. techniques in hydrology. Part 1: Concepts and methodology. Hydrology and Earth System Sciences, REFERENCES 14(10): 1931–1941. CONCLUSIONS Abrahart, R J, Anctil, F, Coulibaly, P, Dawson, C W, Flugel, W & Marker, M 2003. The response units In this study, two data-driven modelling Mount, N J, See, L M, Shamseldin, A Y, Solomatine, concept and its application for the assessment of techniques, namely genetic programming D P, Toth, E & Wilby, R L 2012. Two decades hydrologically related erosion processes in semiarid

16 Journal of the South African Institution of Civil Engineering • Volume 57 Number 3 September 2015 catchments of southern Africa. ASTM Special Kisi, O, Cimen, M & Shiri, J 2012. Suspended sediment with minimal datasets: A case study in the West Technical Publication, 1420: 163–177. modeling using genetic programming and soft Malian River, China. Environmental Engineering Francone, F D 2011. Register Machine Learning computing techniques. Journal of Hydrology, Science, 27(5): 377–385. Technologies, Inc. Discipulus Users Manual, 450–451: 48–58. Ni, Q, Wang, L, Zheng, B & Sivakumar, M 2012. Version 5. Available at: www.rmltech.com. Koza, J R 1992. Genetic Programming: On the Programming Evolutionary algorithm for water storage forecasting Galelli, S & Castelletti, A 2013. Assessing the predictive of Computers by Means of Natural Selection. response to climate change with small data sets: capability of randomized tree-based ensembles in Cambridge, MA, US & London, UK: The MIT Press. The Wolonghu Wetland, China. Environmental streamflow modelling. Hydrology and Earth System Kröse, B & Van Der Smagt, P 1996. An Introduction Engineering Science, 29(8): 814–820. Science Discussions, 10: 1617–1655. to Neural Networks. Amsterdam: University of Poli, R, Langdon, W W B & McPhee, N F 2008. Garg, V 2011. Modeling catchment sediment yield: A Amsterdam. A Field Guide to Genetic Programming. London: genetic programming approach. Natural Hazards: Leander, R & Buishand, T A 2007. Resampling of regional Lulu Enterprises UK Limited. 70(1): 39–50. climate model output for the simulation of extreme Poulin, A, Brissette, F, Leconte, R, Arsenault, R & Malo, Ghaffari, A, Abdollahi, H, Khoshayand, M, river flows. Journal of Hydrology, 332(3): 487–496. J-S 2011. Uncertainty of hydrological modelling Bozchalooi, I S, Dadgar, A & Rafiee-Tehrani, M Limbrick, K J, Whitehead, P, Butterfield, D & Reynard, in climate change impact studies in a Canadian, 2006. Performance comparison of neural network N 2000. Assessing the potential impacts of various snow-dominated river basin. Journal of Hydrology, training algorithms in modeling of bimodal drug climate change scenarios on the hydrological regime 409(3): 626–636. delivery. International Journal of Pharmaceutics, of the River Kennet at Theale, Berkshire, south- Price, K & Storn, R. 2013. Differential evolution (DE) 327(1): 126–138. central England, UK: An application and evaluation for continuous function approximation. Available at: Guven, A 2009. Linear genetic programming for time- of the new semi-distributed model, INCA. Science of www1.icsi.berkeley.edu/~storn/code.html [accessed series modelling of daily flow rate. Journal of Earth the Total Environment, 251: 539–555. on 19 July 2013]. System Science, 118(2): 137–146. Liong, S Y, Gautam, T R, Khu, S T, Babovic, V, Keijzer, M Raghunath, H M 2007. Hydrology: Principles, Analysis Heng, S & Suetsugi, T 2013. Using artificial neural & Muttil, N 2002. Genetic programming: A new and Design. New Delhi: New Age International network to estimate sediment load in ungauged paradigm in rainfall runoff modeling. Journal of the (Pvt) Ltd. catchments of the Tonle Sap River Basin, Cambodia. American Water Resources Association, 38(3): 705–718. Selle, B & Muttil, N 2011. Testing the structure of a Journal of Water Resource and Protection, Londhe, S & Charhate, S 2010. Comparison of data- hydrological model using genetic programming. 5(2): 111–123. driven modelling techniques for river flow forecasting. Journal of Hydrology, 397(1): 1–9. Il-Won, J, Moradkhani, H & Chang, H 2012. Hydrological Sciences Journal, 55(7): 1163–1174. Sirdari, Z Z, Ghani, A A & Hasan, Z A 2011. Prediction Uncertainty assessment of climate change impacts Loucks, D & Van Beek, E 2005. Water Resources of bed load transport in Kurau River based on for hydrologically distinct river basins. Journal of Systems Planning and Management: An Introduction genetic programming. Paper presented at the 3rd Hydrology, 466–467: 73–87. to Methods, Models and Applications. Delft International Conference on Managing Rivers in Ilonen, J, Kamarainen, J-K & Lampinen, J 2003. Hydraulics, the Netherlands: UNESCO Publishers. the 21st century: Sustainable Solutions for Global Differential evolution training algorithm for feed- Maier, H R & Dandy, G C 2000. Neural networks for Crisis of Flooding, Pollution and Water Scarcity, forward neural networks. Neural Processing Letters, the prediction and forecasting of water resources 6–9 December 2011, Penang, Malaysia. 17(1): 93–105. variables: A review of modelling issues and applications. Sivapragasam, C, Muttil, N & Arun, V. Long-term Jayawardena, A & Lai, F 1994. Analysis and prediction Environmental Modelling & Software, 15(1): 101–124. flow forecasting for water resources planning in a of chaos in rainfall and stream flow time series. Maity, R & Kashid, S S 2009. Hydroclimatological river basin. Proceedings, International Congress on Journal of Hydrology, 153(1): 23–52. approach for monthly streamflow prediction using Modelling and Simulation, 12–16 December 2011, Jha, G K. 2007. Artificial neural networks and its genetic programming. ISH Journal of Hydraulic Perth, Australia, pp 4078–4084. applications. Indian Agricultural Statistics Research Engineering, 15(2): 89–107. Solomatine, D & Ostfeld, A 2008. Data-driven Institute (IASRI), New Delhi. Available at: www. Maity, R & Kashid, S S 2010. Short-term basin-scale modelling: Some past experiences and new iasri.res.in/ebook/ebadat/5-Modeling%20and%20 streamflow forecasting using large-scale coupled approaches. Journal of Hydroinformatics, 10(1): 3–22. Forecasting%20Techniques%20in%20Agriculture/5- atmospheric–oceanic circulation and local outgoing Wang, W-C, Chau, K-W, Cheng, C-T & Qiu, L 2009. ANN_GKJHA_2007.pdf. longwave radiation. Journal of Hydrometeorology, A comparison of performance of several artificial Jiang, T, Chen, Y D, Xu, C-Y, Chen, X, Chen, X & Singh, 11(2): 370–387. intelligence methods for forecasting monthly V P 2007. Comparison of hydrological impacts of Makkeasorn, A, Chang, N B & Zhou, X 2008. Short- discharge time series. Journal of Hydrology, climate change simulated by six hydrological models term streamflow forecasting with global climate 374(3): 294–306. in the Dongjiang Basin, South China. Journal of change implications – A comparative study between Wu, C, Chau, K & Li, Y 2009. Predicting monthly Hydrology, 336(3): 316–333. genetic programming and neural network models. streamflow using data-driven models coupled with Karthikeyan, L, Kumar, D N, Graillot, D & Gaur, S Journal of Hydrology, 352(3–4): 336–354. data-preprocessing techniques. Water Resources 2013. Prediction of ground water levels in the Montanari, A & Di Baldassarre, G 2013. Data errors and Research, 45(8): W08432. uplands of a tropical coastal riparian wetland hydrological modelling: The role of model structure Zakaria, Z A & Shabri, A 2012. Streamflow forecasting using artificial neural networks. Water Resources to propagate observation uncertainty. Advances in at ungauged sites using support vector machines. Management, 27(3): 871–883. Water Resources, (51): 498–504. Applied Mathematical Sciences, 6(60): 3003–3014. Kasiviswanathan, K & Sudheer, K 2013. Quantification Moradkhani, H, Hsu, K-L, Gupta, H V & Sorooshian, S Zhang, L, Merényi, E, Grundy, W M & Young, E F 2010. of the predictive uncertainty of artificial neural 2004. Improved streamflow forecasting using self- Inference of surface parameters from near-infrared

network based river flow forecast models. Stochastic organizing radial basis function artificial neural spectra of crystalline H2OH2O ice with neural Environmental Research and Risk Assessment, networks. Journal of Hydrology, 295(1): 246–262. learning. Publications of the Astronomical Society of 27(1): 137–146. Muttil, N & Lee, J H W 2005. Genetic programming the Pacific, 122(893): 839–852. Kisi, O & Cigizoglu, H K 2007. Comparison of for analysis and real-time prediction of coastal algal different ANN techniques in river flow prediction. blooms. Ecological Modelling, 189(3–4): 363–376. Civil Engineering and Environmental Systems, Ni, Q, Wang, L, Ye, R, Yang, F & Sivakumar, M 2010. 24(3): 211–231. Evolutionary modeling for streamflow forecasting

Journal of the South African Institution of Civil Engineering • Volume 57 Number 3 September 2015 17 TECHNICAL PAPER Performance assessment Journal of the South African Institution of Civil Engineering of aquatic macrophytes Vol 57 No 3, September 2015, Pages 18–25, Paper 1011 for treatment of municipal wastewater MUMTAZ SHAH is a PhD candidate in the Civil and Environmental Engineering Department at the University of Engineering and Technology, Taxila, Pakistan. He has 11 years of experience in infrastructure M Shah, H N Hashmi, A R Ghumman, M Zeeshan design in various foul sewer projects, and water and treated sewage effluent projects, where he was involved in the preparation of concept design, preliminary design, detail design, and the preparation of engineering design reports. He is well versed in The objective of the study was to evaluate the performance of three different aquatic local and international design standards/codes of practice related to water, macrophytes for treatment of municipal wastewater collected from Taxila in Pakistan. A physical wastewater and surface water. model of a typical treatment plant was constructed and was operated for six experimental Contact details: runs with each species of macrophyte. Every experimental run consisted of a thirty-day c/o Dr Hashim Nisar Hashmi (Prof CED) Faculty of Civil and Environmental Engineering period. Regular monitoring of influent and effluent concentrations were made during each University of Engineering and Technology experimental run. Locally available macrophyte species (water hyacinth, duckweed and water Taxila, 47080, Pakistan. T: +92 341 440 0046 lettuce) were selected for testing. To evaluate the treatment performance of each macrophyte, E: [email protected] BOD5, COD, and nutrients (nitrogen and phosphorus) were monitored in the effluent from the PROF DR HASHIM NISAR HASHMI currently works in model at different detention times for each experimental run, after having ensured steady state the Civil and Environmental Engineering Department conditions. The average reduction of effluent value for each parameter, using water hyacinth, at the University of Engineering and Technology, Taxila, was 50.61% for BOD5, 46.38% for COD, 40.34% for nitrogen and 18.76% for phosphorus. For Pakistan. He obtained his BSc in Civil Engineering from the Engineering University, Lahore, Pakistan, in 1984 duckweed, the average removal efficiency for the selected parameters was 33.43% for BOD5, (with distinction), and his PhD from the Queens 26.37% for COD, 17.59% for nitrogen and 15.25% for phosphorus, and for water lettuce the University, United Kingdom, in 1993. He has published many research papers in average removal efficiency was 33.43% for BOD5, 26.37% for COD, 17.59% for nitrogen and leading international scientific journals, and in the proceedings of international 15.25% for phosphorus. The mechanisms of pollutant removal in this system include both conferences. He has successfully supervised ten PhD candidates, and has also received the Best Teacher Award from the Higher Education Commission aerobic and anaerobic microbiological conversions, sorption, sedimentation, volatilisation of Pakistan. and chemical transformations. The rapid growth of the biomass was measured within the Contact details: first ten days of the detention time. It was also observed that performance of macrophytes Faculty of Civil and Environmental Engineering is influenced by variations in pH and temperature. A pH of 6–9 and temperature of 15–38oC University of Engineering and Technology is most favourable for treatment of wastewater by macrophytes. The option of macrophytes Taxila, 47080, Pakistan. T: +92 321 557 0450 E: [email protected] OR [email protected] for treatment of municipal sewage under local environmental conditions can be explored by further verifying the removal efficiency under different environmental conditions. PROF DR ABDUL RAZZAQ GHUMMAN currently works at the University of Engineering and Technology, Taxila, Pakistan. He obtained his BSc in Civil Engineering from the Engineering University, Lahore, Pakistan, in 1980, an INTRODUCTION Although some communities treat their MPhil in Water Resources Engineering from the same Wastewater is any liquid that has been wastewater in a suitable way, others lack university, and a PhD from the University of London in 1995. He has published about 113 research papers in leading scientific journals adversely affected in quality by anthropogen- adequate treatment systems, thus discharg- and in the proceedings of international conferences. He has successfully ic influence. It comprises liquid waste dis- ing untreated wastewater into the natural supervised six PhD candidates, and he is also a recipient of the Best Teacher charged by domestic residences, commercial environment. Pollutants (e.g. heavy metals) Award from the Higher Education Commission of Pakistan. properties, industry or agriculture, and can enter aquatic systems via numerous path- Contact details: encompass a wide range of potential con- ways, including effluent discharge, and urban Faculty of Civil and Environmental Engineering University of Engineering and Technology taminants and concentrations. In the most and agricultural run-off. Contaminants Taxila, 47080, Pakistan, T: +92 300 522 3338 common usage it refers to the municipal present in sewage commonly include a wide E: [email protected] wastewater that contains a broad spectrum range of metallic and organic compounds of contaminants resulting from the mixing (Montaigne 2002). MUHAMMAD ZEESHAN is currently working as lecturer and laboratory engineer at the University of of wastewater from different sources. Urban Wastewater treatment technology needs Engineering and Technology, Taxila, Pakistan. He wastewater contains 99% water, and other to be appropriate and sustainable. It also obtained a BSc in Environmental Engineering from the materials make up the remaining portion. needs to be less costly, easy to operate and Engineering University, Lahore, Pakistan, and is now studying towards a Masters in Environmental The potential pollutants include pathogens, maintain, and very efficient in removing Engineering. He has a broad interest in water and wastewater treatment, health oil and grease, metals, organic matter (OM), both organic matter and heavy metals. In and safety, and experimental work. He is currently working on industrial solids and nutrients such as nitrogen (N) developing countries, natural treatment wastewater treatment, focusing on reuse and recycling, using technology that involves a membrane bioreactor and bio-fouling reduction. and phosphorous (P). The actual proportion systems are more suitable. Natural treat- of each constituent within any given waste- ment systems are considered one of the best Contact details: Faculty of Civil and Environmental Engineering water varies depending on the spatial and treatment options, particularly in warm University of Engineering and Technology temporal differences (IWMI 2004). climates (Duenas et al 2003). Wetlands Taxila, 47080, Pakistan. T: +92 346 442 2191 In recent years the amount of waste- with macrophytes are one of the many E: [email protected] water produced from several activities has types of natural systems that can be used

Keywords: aquatic macrophytes, BOD5, duckweed, municipal sewage, increased as a result of the rapid improve- for the treatment of municipal wastewater. wastewater treatment ment of living standards (Trepanier 2002). According to Trepanier (2002), a wetland

Shah M, Hashmi HN, Ghumman AR, Zeeshan M. Performance assessment of aquatic macrophytes for treatment of municipal wastewater. 18 J. S. Afr. Inst. Civ. Eng. 2015;57(3), Art. #1011, 8 pages. http://dx.doi.org/10.17159/2309-8775/2015/v57n3a3 decrease wastewater quality indicators Raw influent input still remains unanswered. It is important Pre-settling tank to determine the lower bounds of pollut- ant content that can be reached because of their removal by aquatic plants, and Raw influent sampling under what conditions removal will occur. This is reflected in the range of applica- Raw influent tion of aquatic plants (macrophytes) for sampling wastewater treatment. Therefore, a physical macrophyte-based treatment plant model Lagoon containing was constructed to treat municipal wastewa- macrophyte ter from the University of Engineering and Technology (UET), Taxila. Wastewater was discharged into a physical model containing Duckweed Water lettuce

Water hyacinthWater macrophytes. The objective of the study was to evaluate the removal performance of pol- lutants COD, BOD5 and nutrients nitrogen Treated effluent and phosphorus (N and P) by different mac- sampling Treated effluent sampling rophytes species. To disposal To disposal To disposal

MATERIALS AND METHODS Disposal body Treatment system A circular storage tank (with a volume of Figure 1 Experimental setup for the study 2 280 litres) was used for the collection of municipal wastewater from the uni- specifically constructed for the purpose of Pakistan, untreated municipal wastewater is versity sewer. From this storage tank, raw pollution control and waste management, indiscriminately discharged into water bod- wastewater was distributed to individual at a location other than existing natural ies. Rapid urbanisation and industrialisation model compartments (5’(L) x 6’(W) x 3’(D)) wetlands, is known as a constructed wetland. have resulted in increased pollution loads containing different macrophyte species. Wetlands have many unique benefits as a in the rivers and streams. In large cities, The study was conducted as a continuous wastewater treatment process, including the municipal wastewater, along with commer- flow system. Figure 1 shows the layout of the ability to operate on ambient solar energy, to cial/industrial effluent, is being discharged experimental setup. self-organise and increase treatment capacity into water bodies in the immediate vicinity over time, being rich in biodiversity, and (i.e. rivers, surface drains and canals). As a Municipal wastewater the ability to produce oxygen and consume result, the pollution level in the water bodies The municipal wastewater for the study was carbon dioxide, thereby achieving high levels is increasing all the time. There are several collected from the municipal sewer of UET of treatment with minimum maintenance sophisticated treatment systems available, Taxila containing wastewater of both the (Wiley 2005). Macrophytes have been such as the activated sludge process, rotating university and the university colony. used effectively to treat different types of biological contactor, and aerated lagoon. wastewaters, and this is mainly due to their However, these require high capital, opera- Selection of species of macrophytes nutrient absorbing capacity, simplicity, low tional and maintenance costs. Accordingly, a The selection of macrophytes for the construction/operation and maintenance biological treatment system with low opera- study was based on the local availability of cost, low energy demand, process stability, tional and capital costs is preferred, especially macrophytes, as well as the environmental and potential benefits of the harvested mate- for developing countries like Pakistan, which conditions of the area. Keeping these factors rials (Metacalf & Eddy 1991). has a warm climate all year round. Pakistan in mind, three macrophytes were selected – Macrophytes have several properties in also has sufficient land for natural wastewa- water hyacinth, duckweed and water lettuce. relation to the treatment processes. The most ter treatment technology on the outskirts of important effects of the macrophytes in rela- cities. Maximum advantage should therefore Analytical procedure tion to the wastewater treatment processes be taken of the climate and land availability The study of the physical model was car- are the physical effects of the plant tissues, for wastewater treatment purposes. The ried out over 18 months. The BOD5, COD, which give rise to a filtration effect and pro- most attractive method is a combination ammonia and phosphorus were analysed vide a surface area for attached microorgan- of a waste stabilisation pond along with according to the Standard Methods for isms. The pollutants’ removal of macrophytes, macrophytes. The mechanism of treatment in the Examination of Water and Wastewater by plant uptake and oxygen release, affects the such a system is the same as in a constructed (APHA 2012). wastewater treatment processes to different wetland, and is an effective low-cost technol- extents. The macrophytes also provide habitat ogy method that requires minimum energy for wildlife (Reed et al 1988). to operate and which is suitable for urban as RESULTS AND DISCUSSIONS Water pollution is becoming a serious well as rural areas in Pakistan. The removal of pollutants by macrophytes issue throughout the world due to rapid Although treatment of wastewater by may occur through a number of processes, population growth, unsuitable treatment macrophyte plants started long ago, the including sedimentation, filtration, plant technologies and inadequate management. In question of how low aquatic plants can uptake/removal efficiency, adsorption,

Journal of the South African Institution of Civil Engineering • Volume 57 Number 3 September 2015 19 formation of solid compounds, and microbi- 180 al-mediated reaction (Watson et al 1989).

BOD 5 130 Biochemical oxygen demand (BOD5) is (mg/ℓ) a measure of the oxygen consumption of 5 NEQS limit microorganisms in the oxidation of organic BOD 80 matter. Settleable organics are rapidly removed in the experimental system by qui- escent conditions, deposition and filtration. 30 Attached and suspended microbial growth is 0 3 5 7 10 15 20 30 responsible for the removal of soluble BOD5. Duration (days) The influent concentrations had a range of Run 1 Run 2 Run 3 Run 4 Run 5 Run 6 95–160 mg/ℓ, showing a medium strength Figure 2 Reduction of BOD with water hyacinth sewage. The BOD5 effluent concentrations at 5 the 30th day of each experiment run for the water hyacinth system were 70 mg/ℓ (Run 1), 180 76 mg/ℓ (Run 2), 45 mg/ℓ (Run 3), 59 mg/ℓ (Run 4), 63 mg/ℓ (Run 5) and 79 mg/ℓ (Run 6). The duckweed-based system, however, 130 showed 81 mg/ℓ (Run 1), 105 mg/ℓ (Run 2), (mg/ℓ) 60 mg/ℓ (Run 3), 84 mg/ℓ (Run 4), 84 mg/ℓ 5 th

(Run 5) and 110 mg/ℓ (Run 6) at the 30 day BOD 80 of each experimental run. Similarly, water NEQS limit lettuce showed 91 mg/ℓ (Run 1), 110 mg/ℓ (Run 2), 65 mg/ℓ (Run 3), 89 mg/ℓ (Run 4), 30 93 mg/ℓ (Run 5) and 119 mg/ℓ (Run 6) at the 0 3 5 7 10 15 20 30 30th day of each experimental run. Duration (days) The results showed that water hyacinth Run 1 Run 2 Run 3 Run 4 Run 5 Run 6 showed the maximum removal (50.61% Figure 3 Reduction of BOD with duckweed average reduction) of BOD5 compared to 5 duckweed and water lettuce. Figures 2, 3 and 4 show the removal of BOD5 by water hya- 180 cinth, duckweed and water lettuce respec- tively. It was observed that removal of BOD5 occurs mainly in the first ten days of each 130 experimental run; after that the removal is (mg/ℓ) at a slower rate. This can be attributed to 5

a higher plant uptake in the first ten days. BOD 80 Similarly, significant plant growth was NEQS limit observed during this period as well. In the study it was also found that the desirable 30 concentration of BOD5 (i.e. < 80 mg/ℓ), as 0 3 5 7 10 15 20 30 prescribed by the National Environmental Duration (days) Quality Standards of Pakistan, was achieved Run 1 Run 2 Run 3 Run 4 Run 5 Run 6 in the water hyacinth system at the tenth day in each run. Figure 4 Reduction of BOD5 with water lettuce

BOD5 removal efficiencies were also observed against the organic loading rate 60 (OLR). The removal efficiencies, resulting in different OLRs, are presented in Figure 5. It 50 shows an increase in removal efficiency with 40 an increase in OLR, as well as the optimum 30 OLR of 112-113 kg BOD5/ha-d, and a reduc- tion of system removal efficiency by a further

Reduction (%) 20 increase in OLR results. 10 COD 0 COD is the amount of oxygen necessary 122 113 92 73 to oxidise the organic compound (OC) Organic loading rate (kg/ha-d) completely to CO2, H2O and NH3. COD Water hyacinth Duckweed Water lettuce is measured via oxidation with potassium Figure 5 BOD removal efficiency versus OLR dichromate (K2Cr2O7) in the presence of 5

20 Journal of the South African Institution of Civil Engineering • Volume 57 Number 3 September 2015 320 sulfuric acid and silver, and is expressed in mg/ℓ. Thus, COD is a measure of the O2 270 equivalent of the organic matter, as well as micro-organisms in the wastewater. If 220 the COD value is much higher than the

BOD5 value, the sample contains large 170

COD (mg/ℓ) amounts of organic compounds that are NEQS limit not easily biodegraded. Water hyacinth 120 was capable of decreasing COD from its initial value to a final value below the 70 0 3 5 7 10 15 20 30 National Environmental Quality Standards Duration (days) of Pakistan, i.e. < 150 mg/ℓ (50.61% average Run 1 Run 2 Run 3 Run 4 Run 5 Run 6 reduction). Similarly, with duckweed it was observed that a reduction of COD occurred Figure 6 Reduction of COD with water hyacinth from an initial value of 130 mg/ℓ to a final of 87 mg/ℓ (33.43% average reduction) and, with 320 water lettuce, the COD was reduced from an initial 131 mg/ℓ to a final 94 mg/ℓ (28.59% 270 average reduction). Around 30–40% of the decrease in the parameters occurred within 220 the first ten days of the experiment. The results confirmed the growth of 170

COD (mg/ℓ) macrophytes and showed a high performance NEQS limit in terms of the removal of COD, mainly due 120 to a well-developed root system. Similarly, it was observed that a major part of the 70 0 3 5 7 10 15 20 30 degradation of COD in the wastewater could Duration (days) be attributed to micro-organisms possibly Run 1 Run 2 Run 3 Run 4 Run 5 Run 6 establishing a symbiotic relationship with the plants. Figures 6, 7 and 8 show removal of Figure 7 Reduction of COD with duckweed COD by water hyacinth, duckweed and water lettuce respectively. 320 Ammonia Nitrogen (NH3-N) 270 Urban wastewater contains a high concen- tration of nutrients, in addition to other 220 pollutants. The major nutrients found in wastewater are N and P, which, if not treated, 170

COD (mg/ℓ) would cause a number of problems to the NEQS limit environment, especially to receiving water 120 bodies. An excess of nutrients in a water body causes overproduction of phytoplank- 70 0 3 5 7 10 15 20 30 ton and results in O2 depletion. N is an Duration (days) essential nutrient and it enters a wetland Run 1 Run 2 Run 3 Run 4 Run 5 Run 6 in particulate and dissolved inorganic and organic forms. Particulate forms of N are Figure 8 Reduction of COD with water lettuce removed through a series of processes including ammonification, nitrification, 4 denitrification and ammonia volatilisation. In fresh sewage, about 25% of the N is in the organic form and 75% in the ammonium 3 form. The organic nitrogen fraction is con-

verted almost entirely to NH3-N, and further 2 -N (mg/ℓ) converted to nitrate-N (NO3-N) via micro- 3 bial oxidation. Like for COD, the N removal NH 1 was measured by a different set of experi- ments. In the water hyacinth system, the reduction of N ranged from 2.42 to 1.45 mg/ℓ 0 0 3 5 7 10 15 20 30 (40.34% average reduction), whereas the Duration (days) reduction in duckweed ranged from 2.37 to Run 1 Run 2 Run 3 Run 4 Run 5 Run 6 1.95 mg/ℓ (17.59% average reduction), while in water lettuce N was reduced from 2.42 to Figure 9 Reduction of ammonia nitrogen with water hyacinth 2.09 mg/ℓ (14.45% average reduction). The

Journal of the South African Institution of Civil Engineering • Volume 57 Number 3 September 2015 21 water hyacinth showed the highest level 4 of N removal compared to duckweed and water lettuce. 3 The results of the present study showed that plants are significantly more efficient 2 in terms of N removal. In addition to plant -N (mg/ℓ) 3 uptake, N removal can occur by NH3 NH volatilisation (favoured by high pH), nitrifi- 1 cation and denitrification (under anaerobic conditions), and formation of organic films. 0 In the present study, N removal occurred 0 3 5 7 10 15 20 30 by volatilisation because the pH was higher Duration (days) than 6.5. It has been observed that a greater Run 1 Run 2 Run 3 Run 4 Run 5 Run 6 ratio of plant biomass to model volume can enhance the contact between plant Figure 10 Reduction of ammonia nitrogen with duckweed roots and wastewater, resulting in a greater nutrient removal. 4 Figures 9, 10 and 11 show the removal of N achieved in the study. At the end of the 3 experiment the presence of plants significant- ly reduced the ammonia-N from their initial 2 levels. A general decline of ammonia-nitrogen -N (mg/ℓ) 3 was found in all the experimental setups up NH until day 10. Thereafter the reduction was 1 much less. Reduction by water hyacinth was greater than for duckweed and water lettuce. 0 0 3 5 7 10 15 20 30 -3 Phosphorus (PO4 ) Duration (days) Phosphorus (P) is an essential nutrient for all Run 1 Run 2 Run 3 Run 4 Run 5 Run 6 life forms, and is the eleventh most abundant mineral in the earth’s crust. P is needed Figure 11 Reduction of ammonia nitrogen with water lettuce for plant growth and is required for many metabolic reactions in plants and animals. 3.5 Organic phosphorus (OP) is a part of living plants and animals, their by-products, and 3.0 their remains. Generally P is the limiting 2.5 nutrient in freshwater aquatic systems. (mg/ℓ) 2.0

Therefore, if all P is used, plant growth will –3 4 cease, no matter how much N is available. P PO 1.5 typically functions as the ‘growth-limiting’ factor because it is usually present in very 1.0 low concentrations. The natural scarcity of 0.5 phosphorus can be explained by its attrac- 0 3 5 7 10 15 20 30 tion to organic matter and soil particles. Any Duration (days) unattached or ‘free’ P is quickly removed Run 1 Run 2 Run 3 Run 4 Run 5 Run 6 from the aquatic system by aquatic plants. Excessive concentrations of P can quickly Figure 12 Reduction of phosphorus with water hyacinth cause extensive growth of aquatic plants. A normal adult excretes 1.3–1.5 g of P per 3.5 day. Primary treatment removes only 10% of the P in the waste stream; secondary treat- 3.0 ment removes only 30%. Tertiary treatment 2.5 is required to remove additional P from (mg/ℓ) 2.0 the water. The amount of additional P that –3 4 can be removed varies with the success of PO 1.5 the treatment technologies used. Available technologies include biological removal and 1.0 chemical precipitation (Polprasert 2007). 0.5 Figures 12, 13 and 14 show the removal of 0 3 5 7 10 15 20 30 P in water hyacinth-, duckweed- and water Duration (days) lettuce-based systems respectively. Water Run 1 Run 2 Run 3 Run 4 Run 5 Run 6 hyacinth showed a maximum removal of P (18.76% average reduction), whereas duckweed Figure 13 Reduction of phosphorus with duckweed

22 Journal of the South African Institution of Civil Engineering • Volume 57 Number 3 September 2015 3.5 showed a 15.25% average reduction and water lettuce a 10.69% average reduction within 3.0 the 30-day experimental period. The highest 2.5 removal was observed in water hyacinth. This was due to the synergistic effect of aquatic (mg/ℓ) 2.0

–3 plants. Plants and micro-organisms all utilise 4 P as an essential nutrient and contain P in PO 1.5 their tissues, though the portion of tissue P 1.0 is very small compared with C and N. The reduction of TP may be due to the uptake of 0.5 0 3 5 7 10 15 20 30 soluble P, the filtration of particulate matter Duration (days) through the roots, and settling. Although the Run 1 Run 2 Run 3 Run 4 Run 5 Run 6 initial level of P was low, a plant like water hyacinth, with its well-developed root system, Figure 14 Reduction of phosphorus with water lettuce can further purify wastewater.

Factors affecting the performance 40 35 Temperature 30 In order to check the performance of mac- 25 rophytes under various temperature condi- 20 tions, additional experiments were conduct- 15 ed with pH 7.5. Figure 15 shows the effect of 10 temperature variation on the performance of Removal efficiency (%) 5 water hyacinth, duckweed and water lettuce regarding BOD removal. During the experi- 0 5 10 15 25 35 ments it was observed that macrophytes Temperature (°C) are sensitive to temperature and show no Water hyacinth Duckweed Water lettuce growth and pollutant (BOD5) removal at a temperature below 10oC. Almost all of the Figure 15 Effect of temperature variation on performance of macrophytes plants of these three species cease to survive at this temperature. As the growth of species o 50 is negligible at temperature below 10 C, 45 there was no uptake of nutrients (N and P) by 40 the plants. A temperature between 15–38oC 35 is suitable for treatment of municipal waste- 30 water by macrophytes, as high growth was 25 observed at this temperature. 20 15 pH

Removal efficiency (%) 10 In order to monitor the effect of the variation 5 of pH on the performance of macrophytes, 0 5 6.5 8.5 10 experiments in the laboratory were conduct- pH ed at a temperature of 25oC and at different Water hyacinth Duckweed Water lettuce levels of pH. At a pH below 5, macrophyte

performance (BOD5 removal) is almost zero. Figure 16 Effect of pH variation on performance of macrophytes This is mainly due to the highly acidic nature of the wastewater. On the other hand, when Table 1 Plants (macrophytes) height measurement the pH gradually increases, performance improves up to a pH of 7.5, and, at further Plant height (ft) increase in pH again starts retarding mac- Time Run 1 Run 2 Run 3 Run 4 Run 5 Run 6 (days) rophyte performance (BOD5 removal). At a pH of 10 (at high alkalinity) the performance WH WL WH WL WH WL WH WL WH WL WH WL of macrophytes was again decreased to zero. 0 0.65 0.51 0.74 0.53 0.69 0.54 0.70 0.55 0.62 0.61 0.62 0.67 Therefore, a pH of 6–9 is most suitable for the performance of macrophytes. Figure 16 5 0.70 0.55 0.79 0.57 0.74 0.60 0.75 0.59 0.67 0.65 0.68 0.72 shows the effect of pH variation on the per- 10 0.78 0.60 0.87 0.61 0.81 0.66 0.82 0.65 0.75 0.72 0.76 0.79 formance of water hyacinth, duckweed and water lettuce regarding BOD removal. 20 0.81 0.64 0.91 0.63 0.85 0.68 0.84 0.68 0.80 0.75 0.81 0.81 5

30 0.83 0.69 0.94 0.65 0.87 0.70 0.86 0.70 0.84 0.77 0.83 0.84 Plant growth The plant stems filtered and reduced some (WH = water hyacinth; WL = water lettuce) particles in the wastewater. When they died,

Journal of the South African Institution of Civil Engineering • Volume 57 Number 3 September 2015 23 they acted as a net that filtered some of the 1 300 pollutants. They reduced inflow and induced particles accumulation or precipitation in 1 100 ) the systems. Plant heights were observed in 2 900 order to investigate plant growth during each trial. The measured plant heights are given 700 in Table 1. 500 Biomass (g/m

Plant biomass productivity 300 Primary productivity and biomass are 100 the important parameters. In general, –5 0 5 10 15 20 25 30 35 the productivity of macrophytes is higher Time (days) than that of terrestrial communities and Run 1 Run 2 Run 3 Run 4 Run 5 Run 6 agricultural crops because they do not suf- fer from a shortage of water. Macrophytes Figure 17 Growth of water hyacinth in study model have a high tolerance for the fluctuations in environmental conditions and show 800 high photosynthetic efficiencies. Uptake of 700 nutrients by macrophytes is essential for ) their growth and reproduction. The high 2 600 productivity of macrophytes enables sub- 500 stantial amounts of nutrients to be stored 400 in plant biomass. Measurements of biomass were made after the 5th, 10th, 20th and 30th Biomass (g/m 300 day of each experimental run for water 200 hyacinth, water lettuce and duckweed. The 100 plant biomass growth in the model was plot- –5 0 5 10 15 20 25 30 35 ted and is shown in Figures 17, 18 and 19. Time (days) It is quite clear from the results that there Run 1 Run 2 Run 3 Run 4 Run 5 Run 6 was a major rise in plant biomass in the first ten days. However, in the remaining twenty Figure 18 Growth of water lettuce in study model days it was much less when compared to the initial growth. 600 The results showed that the optimum 550 period for harvesting was found to be 8–10

) 500 days. At the optimum point, the growth rate 2 450 of the plant is lowest. 400 350 Biomass (g/m CONCLUSION 300 As far as removal efficiencies are concerned, 250 water hyacinth was found to be the most 200 effective macrophyte, while considerable –5 0 5 10 15 20 25 30 35 removals of pollutants were also found with Time (days) duckweed and water lettuce. The perfor- Run 1 Run 2 Run 3 Run 4 Run 5 Run 6 mance of the water hyacinth-based system Figure 19 Growth of duckweed in study model was found to be 50.61% for BOD5, 46.38% for COD, 40.34% for nitrogen and 18.76% for phosphorus. For the duckweed-based system severely affects macrophytes performance be explored by further verifying the removal the efficiencies were 33.43% for BOD5, and it was found that, at a temperature efficiency under different environmental 26.37% for COD, 17.59% for nitrogen and below 10oC, macrophytes were unable to conditions. There is a need for a macrophyte 15.25% for phosphorus. Similarly, for water perform treatment and that the favourable system to be used over time for the treat- o lettuce these values were 33.43% for BOD5, temperature for treatment is 15–38 C. Pre- ment of wastewater, because its performance 26.37% for COD, 17.59% for nitrogen and treatment of wastewater before plant accli- is comparable to that of conventional waste- 15.25% for phosphorus. The mechanisms matisation could be potentially effective. water treatment plants and, in addition, of pollutant removal in the system include The harvested macrophytes can be used, such a system has very low operation and both aerobic and anaerobic microbiologi- directly or after processing, as soil additives, maintenance costs. cal conversions, sorption, sedimentation, mulch, fertiliser, green manure, pulp and volatilisation and chemical transformations. fibre for paper-making, animal feed, human The pH of the wastewater affects the perfor- food, organic malts for biogas production, REFERENCES mance of macrophytes and it was found that and for composting. The option of macro- APHA (American Public Health Association) 2012. macrophytes gave optimum performance at phytes for treatment of municipal sewage Standard Methods for the Examination of Water and pH 6–9. Temperature is another factor that under local environmental conditions can Wastewater, 22nd ed. Washington DC: APHA.

24 Journal of the South African Institution of Civil Engineering • Volume 57 Number 3 September 2015 Duenas, J F, Alonso, J R, Rey, A F & Ferrer, A S Metcalf & Eddy Inc. 1991. Wastewater Engineering. Reed, S C, Middlebrooks, E J & Crites, R W 1988. 2003. Characterization of phosphorous forms in New York: McGraw-Hill. Natural Systems for Waste Management and wastewater treatment plants. Journal of Hazardous Montaigne, F & Essick, P 2002. Water pressure. Treatment. New York: McGraw-Hill. Materials, 97: 1–3. National Geographic, 202: 2–23. Watson, J T, Reed, S C, Kadlec, R H, Knight, R L & IWMI (International Water Management Institute) Trepanier, C, Parent, S, Comeau, Y & Bouvrette, J 2002. Whitehouse, A E 1989. Performance expectations and 2004. A framework for global assessment of the Phosphorous budget as a water quality management loading rates in constructed wetlands. In: Hammer, D A extent of wastewater irrigation. The need for tool for closed aquatic mesocosms. Journal of Water (Ed.), Constructed Wetlands for Wastewater Treatment, a common wastewater typology. Netherlands: Research, 36: 1007–1017. Chelsea, MI, US: Lewis Publishing. International Water Management Institute, Polprasert, C 2007. Organic Waste Recycling: Technology Wiley, J 2005. Wastewater Microbiology, 3rd ed. p 958. and Management, 3rd ed. London: IWA Publishing. New York: McGraw-Hill, p 746.

Journal of the South African Institution of Civil Engineering • Volume 57 Number 3 September 2015 25 TECHNICAL PAPER Evaluation of the response Journal of the South African Institution of Civil Engineering behaviour of unconfined Vol 57 No 3, September 2015, Pages 26–34, Paper 1156 cemented materials under dynamic loading MOKGELE MATHEBA (Pr Tech Eng) focuses on pavement engineering and materials investigation, implementation and monitoring of labour-based road projects, as well as project management related to M J Matheba, W J vd M Steyn, R J Moloisane, T I Milne laboratory testing. He holds the qualifications BTech (Transportation Engineering) and MTech (Civil) obtained from the former Technikon Pretoria and Tshwane University of Technology respectively. He is currently the operations manager for Africa at There is a significant increase in traffic loading on most roads in the developing African Geostrada Engineering Materials Laboratory, a business unit of Aurecon. He is also an external moderator for the BTech courses at the University of countries, and South Africa is one of them. Often this increased traffic loading results in the South Africa. premature failure of pavement structures. Mechanistic-empirical (M-E) design methods based Contact details: on fundamental principles are better able to accommodate changes in the design environment. Geostrada Materials Engineering Laboratory The successful use of design methods depends on the accuracy of the input material PO Box 11126, Hatfield, 0028, South Africa parameters. Therefore, as design is moving towards M-E design methods, there is a need for the T: +27 (0)12 432 0500, E: [email protected] material parameters to reflect the actual pavement response to dynamic loads. The objective of PROF WYNAND STEYN (Pr Eng, MSAICE) researches this paper is to report on the investigation of the response of cement-stabilised sub-base layers vehicle–pavement interaction, accelerated pavement to dynamic load by evaluating stiffness at a known strain level. This stiffness was compared testing, and pavement materials and instrumentation. He obtained a PhD in Civil Engineering from the with those derived from unconfined compressive strength (UCS) tests performed at static load. University of Pretoria in 2001. He spent 19 years with The strain and stiffness values were also evaluated against compacted density, cement content, the CSIR (Council for Scientific and Industrial Research) moisture content and material type. It was found that stiffness of some of the cement-stabilised in various positions, and is currently professor of civil engineering (focusing on road pavement-related subjects) at the University of Pretoria. Professional sub-base layers may possibly be overestimated through the use of static loads. activities include academic and industry research in the areas of pavement engineering, vehicle–pavement interaction and pavement materials. He has authored and co-authored 20 journal papers, 17 book chapters and 74 conference papers. He is Associate Editor of the International Journal for BACKGROUND on the accuracy of the input parameters and Pavement Engineering and has a B3 NRF rating. the way in which the material parameters are Contact details: Introduction determined (Theyse 2001). Therefore there Department of Civil Engineering, University of Pretoria Private Bag X20, Hatfield, 0002, South Africa Trends in the road transport sector indicate is a need for material parameters to reflect T: +27 (0)12 420 2171, E: [email protected] significant increases in the growth of traffic the actual pavement response under dynamic loading, due to an increase in the number tyre loads. The main parameters affecting JONES MOLOISANE (Pr Tech Eng, MSAICE) has a research interest in low-volume roads, road of vehicles on all road types in developing material behaviour include stress, strain construction materials and soil stabilisation. He holds countries, such as South Africa (Kekana and stiffness. the qualifications MTech (Civil) cum laude and MSc (Appl Sci) obtained from the former Technikon Pretoria 2006). The increase in the traffic loading Traffic stress can cause a decrease in the and University of Pretoria respectively. He is currently a on road pavements leads to increased dam- strength of the pavement, as the pavement lecturer (focusing on transportation engineering technology, and construction age, high maintenance and on-going repair layers undergo different phases of distress contract management and administration-related subjects) in the Department of Civil Engineering at the Tshwane University of Technology, South Africa. He is costs, as well as the shortening of the life of (De Beer 1985; Theyse et al 1996). In real also currently the chairman of the board of directors and shareholder of Virtual the roads. Often traffic loading results in life these pavements and material layers are Consulting Engineers (Pty) Ltd, and a director and shareholder of Delta Built premature failure of roads, with expensive responding to dynamic traffic loading (De Environment Consultants (Pty) Ltd and Science Ignite. maintenance having to be carried out, and Beer 1990; Theyse 2001). Relatively small Contact details: Department of Civil Engineering, Tshwane University of Technology in many cases, even complete reconstruction decreases in layer strength would reduce Private Bag X680, Pretoria, 0001, South Africa of failed roads being undertaken (Cottrell the expected structural capacity. This raises T: +27 (0)12 382 5221, E: [email protected] et al 2003). The South African road design uncertainty about the actual structural method is being reviewed by the South capacity of pavements, since the stiffness of DR TERENCE MILNE (Pr Eng, MSAICE) has extensive experience in road and rehabilitation design, African National Roads Agency (SANRAL) the sub-base will vary considerably during documentation and supervision, contract (SANRAL 2007), partly to accommodate the construction and completion phases and administration, implementation of rural water and roads upliftment projects in developing areas, and the the increase in traffic loading and traffic there would be no clear basis to determine management of large infrastructure projects, including volume, so that roads will be able to support what the ‘correct’ stiffness of the sub-base engineering, procurement, and construction management (EPCM). He these increased loads. The revision of the layer should be (Jordaan 1994). Therefore graduated from the University of Witwatersrand with a BSc (Eng) in 1985, after which he commenced employment with Van Wyk & Louw (later Africon and South African design method for flexible there is a need to look at the materials’ now Aurecon), where he is currently technical director and EPCM project pavements, which focuses on the effect of the response to dynamic repetitive loading. manager in capital projects delivery. He obtained his MSc (Eng) in 1993 in the field of Structural Concrete, and his PhD in Engineering in 2004 from the increase of traffic loading in the pavement Different vehicle configurations cause University of Stellenbosch. structure, is a long-term project with the different load functions in terms of load Contact details: aim of improving the existing Mechanistic- level and frequency on a pavement structure Aurecon Empirical (M-E) design methods of flexible (Steyn et al 1999). The following terminology PO Box 74381, Lynnwood Ridge, 0040, South Africa T: +27 (0)12 427 2000, E: [email protected] pavements (SANRAL 2007). M-E design for load case description has been proposed methods are better able to accommodate by Steyn et al (1999): ■ Keywords: cement-stabilised sub-base layer, dry and wet conditions, changes in the design environment. The suc- ■ Static load – load that is independent on dynamic loading, stiffness cess of these M-E design methods depends time and location

Matheba MJ, Steyn WJvdM, Moloisane RJ, Milne TI. Evaluation of the response behaviour of unconfined cemented materials under dynamic loading. 26 J. S. Afr. Inst. Civ. Eng. 2015;57(3), Art. #1156, 9 pages. http://dx.doi.org/10.17159/2309-8775/2015/v57n3a4 ■■ Dynamic load – load that is dependent on Table 1 Determination of layer stiffness (Van Wijk et al 2007) time and independent of position (load Stiffness Calculation/estimation method magnitude changes according to some ■ Tri-axial testing* time-based function and the position Existing unbound layers ■ Back-calculation from deflections** (i.e. deflection bowl parameters) is static) ■ ■ Moving dynamic load – load that is ■ Primary reliance on tri-axial testing* New unbound layers ■ dependent on both the time and position Soil classification and simple tests* (load magnitude changes according to ■ Estimate from unconfined compressive strength (UCS)* Existing cemented soils some time-based function, and position ■ Back-calculation from deflections** (i.e. deflection bowl parameters) also changes). ■ Estimate from unconfined compressive strength (UCS)* Real traffic causes moving dynamic loads, New cemented soils ■ Use tri-axial test* while dynamic load is mainly used in the ■ Estimate flexural strength from beam test* search to simplify the understanding of pave- * Laboratory-related test that is used to destructively assess the specimen’s failure, in order to understand ment response. The response analysis of the its material behaviour under different loads. pavement layers’ stiffness should be dictated ** Field-related test that is used to non-destructively assess the structural properties of a flexible pavement by the type of load applied. Consequently, dynamic pavement response analysis should be performed for time-dependent loading. certain limits and micro cracks develop static loading. The maximum compressive The objective of this paper is to report between coarse particles, and the layer fails strength value is used to predict the effective on the investigation of the response of the to recover to its original position, resulting strength of cemented layers. cement-stabilised sub-base layers to dynamic in permanent deformation. Previous studies loading. The investigation was carried out by have found that micro cracks start to occur Flexural strength (beam test) test method evaluating the changes in stiffness at known at 35% of actual strength and 25% of strain- The composite of tensile strength and tensile strain level, and these were compared to the at-break under repeated loading (CSRA 1986 strain capacity of a material sample are stiffness from dynamic loads derived from TRH 13; De Beer 1990; Jones et al 2001). determined through its flexural strength the unconfined compressive stress (UCS) The strain-at-break is deduced from flexural (Natt & Joshi 1984). The method is reliable test. This work is based on an extensive labo- strength tests in the laboratory. The practice for determining tensile strains, as it can be ratory study where two materials (ferricrete of using the strain-at-break of the material measured from a bending level of a material and norite) were stabilised with cement and to normalise the working strain in the stiff- sample. The strain at break is measured tested in both wet and dry conditions using ness reduction is in doubt since significant under static load, but the elasticity of the a repetitive loading test method, as reported discrepancies in the strain-at-break results material cannot be determined, because by Matheba (2013). from different research laboratories have the sample is continuously under load. In been noted (Theyse et al 2007). The response considering a field sub-base layer response to Analysis of cement-stabilised of a cement-stabilised layer to an applied repetitive load, there is a need to incorporate sub-base layers load is highly affected by moisture content, it for determination of elastic strain of the The response of each layer in a pavement compaction and cement content. material response to load. structure to traffic loading or environ- Repetitive loading increases fatigue mental conditions affects the response of cracking of the cement-stabilised layer which Falling weight deflectometer (FWD) the other layers. A cement-stabilised layer results in the crushing of the layer over time test method normally becomes stiffer or gains more (De Beer 1990). Once the layer is crushed The FWD test method produces a dynamic strength mainly when the cement content the stiffness is reduced to a natural material impulse load that simulates a moving wheel in the material is increased. Normally the layer. At this stage the layer is more sensi- load. Complete deflection bowls are used cement-stabilised layer develops shrinkage tive to environmental conditions, such as in an iterative procedure, known as back- cracks resulting from drying and thermal moisture. calculation, to estimate the pavement layer stress of the cement-stabilised layer, which effective stiffness. The goal of the back- is not related to traffic loading but to the Design material tests calculation process is to estimate a set of hydration caused by the cement in the Different types of material design test meth- layer effective stiffnesses that best match the material. With increased traffic loading the ods are used to determine input parameters measured deflections at all locations where effective stiffness, which is defined as the in the M-E design method. These material deflections are measured. ratio of the rate change of stress with the design test methods are used to obtain the elastic strain of the layer, will decrease (De materials’ strength for the different layers Beer 1990). The maximum tensile strain at in the pavement structure (Van Wijk et al EXPERIMENTAL PROCEDURE the bottom of the layer (which controls the 2007). The stiffness of various pavement The dynamic loading was developed to effective fatigue life of the pavement) and the layers is estimated by means of destructive simulate field conditions where the cement- vertical compressive stress (σ) on top of the or non-destructive methods as indicated in stabilised pavement structures and layers cemented layer (controlling its crushing life) Table 1. are exposed to similar circumstances. The should be monitored to evaluate the life of aim is to have a better understanding of the the layer. In practice the maximum tensile Unconfined compressive strength performance of the cemented layers under strain is referred to as the strain-at-break of (UCS) method repeated loading. A dynamic loading method the material to normalise the working strain The UCS measures the strength of the was developed on the basis of the UCS with in the stiffness. stabilised material. The UCS test method regard to the preparation of the cement- Strain-at-break refers to the strain when determines only the maximum compressive stabilised specimen only. For the testing of a cement-stabilised layer is loaded beyond strength of any stabilised material under samples, a repetitive load was applied on the

Journal of the South African Institution of Civil Engineering • Volume 57 Number 3 September 2015 27 was determined using the expression in Africa, and one cement agent, i.e. CEM II B Equation 1: A-L 32.5N were used to evaluate the strength characteristics of cement-stabilised materi- Applied load als. The name ferricrete is derived from the A Stress (σ) = (1) Area which load is applied combination of ferruginous and concrete to on the surface (Brink 1985). Weathered norite material is a sub-division of igneous rock. It is largely com- Where: posed of calcium-rich plagioclase laboradorite Applied load = kN and pyroxene (Brink 1985). These materials Area (m²) = r2 x 3.41 (radius of a are commonly used in the construction of ­specimen multiplied by roads, especially in the sub-base layers in the Plate 1 Cemented specimen under dynamic 3.41) Gauteng Province and surrounding areas loading Radius (r) = radius of a specimen (m) (Sadzik 2004), and their performance has been extensively studied. The yellowish- specimen of the cement-stabilised material A permanent displacement of the cemented brown ferricrete material was sampled from to measure both the stress and micro-strain specimen under repetitive load was mea- the borrow-pit situated along Road K109 for deducing an effective stiffness of the sured for every load applied. For every phase, near Brakpan and Benoni in the Gauteng stabilised material layer. the average displacement was determined Province. This material was used in 2001 as a and used to calculate the material strain by sub-base material in the upgrading construc- The use of dynamic load means of the expression in Equation 2: tion of Road K109. Ferricrete material is an and its specifications iron-oxide cemented regolith material. When h Dynamic loading equipment consists of a Microstrain = (2) ferricrete material is mixed with cement and 50 mm diameter piston that applies a vertical H water, it absorbs water very quickly during load on the specimen. A spacer of 127 mm the hydration process and produces calcium in diameter and 3.8 mm in thickness, that Where: hydroxide, which tends to cause carbonation. weighs 4.25 kg, is placed on top of the speci- h = average displacement (mm) This leads to a reduction in strength in the men during the testing process. The plate H = height of the specimen which is cement-stabilised layer. The weathered norite allows for equal distribution of the load at 127 mm material was sampled from a borrow-pit situ- the point where it reaches the specimen, as ated east of the Bakwena Platinum freeway, shown in Plate 1. The behaviour of the cemented specimen near to the Bakwena Platinum freeway/Bela There are two electronic sensors – sen- was observed by noting the change in strain Bela intersection on National Route 1 () sor A measures the specimen’s micro-strain caused by loads applied dynamically until in Pretoria, Gauteng Province. This material as a function of a displacement, and sensor B­ it failed, while stiffness was calculated at was used in 2001 as a sub-base material in the measures the actual applied stress as a func- each phase. A recovered displacement of construction of the Bakwena Platinum free- tion of the load on a specimen. The dynamic the cemented specimen under repetitive way (National Route 4 (N4)), which runs from load testing equipment is programmed to load was measured at every load applied. the N1/Sefako Makgatho Drive intersection to apply an initial load of 1.8 kN per 1.5 second For every phase, the average displacement the Botswana border. on the specimen. This load induces a stress was determined and used to calculate the Common cement, CEM II A-L 32.5N, equivalent of 100 kPa. The stress of 100 kPa ­material strain. which complies with the South African is applied for 2 500 load repetitions before Environmental influences, such as Bureau of Standards (SABS) EN197 (now it increases the stress in 100 kPa incre- regional moisture conditions, play an impor- South African National Standards (SANS) ments. De Beer et al (1999) found that a tant part in the performance of the cement- 50197-1) was used as a cementing stabiliser combination of traffic, such as buses, light stabilised layer. This factor may not have full (SABS 2001). This type of cement was cho- commercial vehicles and trucks, contrib- impact to cement-stabilised layers without sen as it is a product that has been modified utes to pavement deterioration differently a load application since the environmental to accommodate environmental effects such because of their differing axle loads, and factor is a secondary variable. The soil sam- as temperature (Paige-Green & Netterberg the load from slow-moving traffic induces ples were tested under two environmental 2004). CEM II A-L 32.5N cement has been more stress levels as more loads are being conditions: used successfully in many recent projects in applied to the pavement. Therefore a rate of ■■ Dry condition – the cemented specimen South Africa. It consists of fly ash (to develop 1.5 second intervals was established based on was not covered for moisture loss during strength in the material) and a small amount the concept that if vehicles travel at 60 km/h repetitive load testing, irrespective of how of lime (to reduce plasticity in the material). at intervals of 25 m, it will take 1.5 seconds long the test took to complete. (Moisture Ordinary tap water without dissolved for the second vehicle to reach the point loss from a specimen was only monitored salts was used to mix the samples in prepara- where the first vehicle had been. A 0.008 by the material’s response to load). tion for determining the natural properties second tyre/pavement contact stress used in ■■ Wet condition – the specimen was of materials and the preparation of cemented the pavement analysis (De Beer et al 1999) wrapped in impermeable plastic to pre- specimens. Contaminated water was avoided, in the upper layer up to 300 mm (Steyn & vent any evaporation occurring from the because mineral salts and impurities in Sadzik 2007) was for the vehicle travelling at cemented specimen when the repetitive the water affect the chemical properties of a higher speed. This assumption was based loading test was applied. cement-stabilised materials. on the research conducted for the measure- The laboratory test on grain-size dis- ment of tyre-to-pavement contact stress, and Selected materials tribution, Atterberg limits and California is generally described as Stress-in-Motion Two types of materials, i.e. ferricrete and bearing ratio (CBR) were conducted on these (SIM) technology (De Beer et al 1999). Stress norite from the Gauteng Province in South ferricrete and norite materials in accordance

28 Journal of the South African Institution of Civil Engineering • Volume 57 Number 3 September 2015 100 cement-stabilised specimens were prepared 90 according to the conventional preparation 80 of the UCS specimen method to determine 70 their behaviour as per TMH 1 (CSRA 60 1986). Both the 95% and the 100% Modified 50 AASHTO density compaction efforts were

% Passing % 40 used in the study to relate the conventional 30 compaction densities of between 95% and 20 100% Modified AASHTO density, as 10 required in the COLTO (COLTO 1998) 0 0.001 0.01 0.1 1 10 100 specifications and various project specifica- Grain size (sieve size (mm)) tions in South Africa. Ferricrete material (unstabilised) Ferricrete material with 2% cement Ferricrete material with 3% cement Ferricrete material with 4% cement EXPERIMENTAL RESULTS Figure 1 Grain size distribution of the unstabilised ferricrete material and the ferricrete material AND DISCUSSION stabilised with 2%, 3% and 4% cement content The effect of compaction and moisture content (monitored only) on stiffness of 100 a cement-stabilised layer was evaluated 90 by comparing the resulting stiffness from 80 dynamic load against the effective stiffness 70 from static load. Grain-size distribution of 60 the unstabilised ferricrete material and the 50 ferricrete material stabilised with 2%, 3% and

% Passing % 40 4% cement content are depicted in Figure 1. 30 More grain fractions of the ferricrete 20 material stabilised with 2%, 3% and 4% 10 cement content were retained above the 0 0.001 0.01 0.1 1 10 100 0.075 mm sieve compared to the unstabilised Grain size (sieve size (mm)) material. An increase in cement content Norite material with 3% cement Norite material with 2% cement showed improvement in the ferricrete mate- Norite material (unstabilised) Norite material with 4% cement rial. The bond between the grain fractions of the stabilised ferricrete material lower than Figure 2 Grain size distribution of the unstabilised norite material and the norite material the 0.475 mm sieve was improved with the stabilised with 2%, 3% and 4% cement content addition of the cement content. Material properties of the ferricrete Table 2 Material properties of the ferricrete material material (both the unstabilised and stabilised with 2%, 3% and 4% cement content) are Ferricrete material presented in Table 2. Properties Stabilised Stabilised Stabilised Unstabilised The unstabilised ferricrete mate- with with with material 2% cement 3% cement 4% cement rial showed higher plasticity because the disintegration of weathered fine grain Grading modulus (GM) 1.91 1.64 2.06 2.00 of the ferricrete material contributed to Initial consumption of cement (ICC) (%) 4 – – – higher colloidal clay content of the mate-

Liquid limit (LL) (%) 28 26 19 14 rial. The contribution to higher colloidal clay content by the weathered fine grain of Plasticity index (PI) (%) 11 8 3 1.5 the ferricrete material was also reported by Maximum dry density (MDD) (kg/m³) 2 016 – – – Paige-Green (1989), and Paige-Green and Bam (1995). A grading modulus (GM) of the Optimum moisture content (OMC) (%) 12.1 – – – ferricrete material stabilised with 2% cement California bearing ratio (CBR) values (%) content was unusual to be lower than the At 90% Modified AASHTO density 11 unstabilised. This could possibly have been CBR was done only on the unstabilised influenced by the higher amount of finer At 93% Modified AASHTO density 26 material, and the unconfined compressive fractions in the sample representation and strength (UCS) was done on the cement- At 95% Modified AASHTO density 32 rapid initial setting due to the high zinc con- stabilised materials tent in the ferricrete material. In theory, the At 97% Modified AASHTO density 39 stabilised ferricrete material increases the bonding between the material particles and with the TMH 1 (CSRA 1986) standard Two sets of specimens were prepared for reduces the fraction of the finer particles. methods to categorise their natural proper- every mix of 2%, 3% and 4%, compacted at This ferricrete material had a CBR of 32% at ties, and were classified in accordance with 95% and 100% Modified AASHTO density 95% Modified AASHTO density, which is the TRH 14 (CSRA 1985 TRH 14) classifica- respectively, and a set of three specimens lower than the G5 but above the G6 soil clas- tion guidelines. was prepared to serve as a control test. The sification as per TRH 14 (CSRA 1985 TRH

Journal of the South African Institution of Civil Engineering • Volume 57 Number 3 September 2015 29 14) classification guidelines. This could have 7 000 been caused by the higher optimum mois- 6 000 ture content (OMC) in achieving maximum dry density (MDD). This ferricrete material 5 000 may be sensitive to a high moisture climate 4 000 when used in pavement layers. 3 000 Grain-size distribution of the unstabilised norite material and the norite material stabi- 2 000 lised with 2%, 3% and 4% cement content are 1 000 depicted in Figure 2. (MPa) stiffness effective Static 0 The bond between the finer fractions 94 95 96 97 98 99 100 101 smaller than the 0.475 mm sieve was Compaction effort (%) improved with increases in the cement Norite material Ferricrete material Average of norite and ferricrete content. However, the grain fraction at 3% cement content could have had a poor bond Figure 3 Static effective stiffness of the norite, ferricrete and the average of both the norite and and this could have been caused by the ferricrete materials replicating of the quartering of the materials where finer factions were compromised. respectively. Both materials would be regard- strength of the ‘lot’. These results are pre- Material properties of the norite material ed as suitable to be used for stabilisation to sented in Table 4. (both the unstabilised and stabilised with a C3 and C4 cemented material. Cemented The compressive strength of the stabi- 2%, 3% and 4% cement content) are presented materials are classified in accordance with lised ferricrete material was less than the in Table 3. the TRH 13 (CSRA 1986 TRH 13) classifica- stabilised norite material using conventional The unstabilised norite material showed tion guidelines. Following this, specimens of UCS. These results are applicable for all lower plasticity because the material was the stabilised materials were prepared. cement contents (i.e. 2%, 3% and 4%). only slightly weathered and more particles Three samples were crushed at seven-day were still in a fresh stage. This was con- age using the conventional UCS method for Effect of compaction effort firmed by lesser fraction of the finer material control tests. An average of the three results The static effective stiffness of the norite, passing the 0.425 mm sieve. Also, when was taken as the result of the compressive ferricrete and the average of both the norite a small percentage of the cement content of 2% was added, the finer grains became Table 3 Material properties of the norite material less sensitive to water and developed into Norite material non-plastic. Similar results were observed by Properties Stabilised Stabilised Stabilised Paige-Green and Netterberg (2004) during Unstabilised with with with material their earlier work on the use of different 2% cement 3% cement 4% cement types of cement and their effect in strength Grading modulus (GM) 2.32 2.42 2.35 2.52 generation. This confirms that this type of norite material has a low natural plasticity Initial consumption of cement (ICC) (%) 2 – – – index (PI). This norite material had a CBR of Liquid limit (LL) (%) 14 – – – 49% at 95% Modified AASHTO density after soaking for four days, which is higher than Plasticity index (PI) (%) 3 Non-plastic Non-plastic Non-plastic the minimum requirement of 45% at 95% Maximum dry density (MDD) (kg/m³) 2 231 – – – Modified AASHTO density for the G5 mate- Optimum moisture content (OMC) (%) 8.9 – – – rial classification as per TRH 14 (CSRA 1985 TRH 14) classification guidelines. This was California bearing ratio (CBR) values (%) found to be realistic due to the higher per- At 90% Modified AASHTO density 17 centage of stone particles which contribute CBR was done only on the unstabilised At 93% Modified AASHTO density 36 to high resistance in bearing capacity. material, and the unconfined compressive strength (UCS) was done on the cement- The norite and ferricrete materials were At 95% Modified AASHTO density 49 stabilised materials classified as per TRH 14 (CSRA 1985 TRH At 97% Modified AASHTO density 55 14) as G5 and G6 natural gravel materials

Table 4 Unconfined compressive strength (UCS) results of the ferricrete and norite materials

Unconfined compressive strength (UCS) (kPa)

Ferricrete material (average of three results) Norite material (average of three results) Compaction effort Cement content and standard deviation Cement content and standard deviation (%) 2% Stdev 3% Stdev 4% Stdev 2% Stdev 3% Stdev 4% Stdev

100 760 10.5 850 7.4 890 8.1 960 4.8 1 800 5.5 2 300 6.7

95 450 5.3 550 6.4 610 8.2 760 5.2 1 460 4.5 1 660 8.2

Stdev: Standard deviation

30 Journal of the South African Institution of Civil Engineering • Volume 57 Number 3 September 2015 and ferricrete materials determined at 6 000 95% and 100% Modified AASHTO density R2 = 0.8318 compaction effort are depicted in Figure 3. 5 000 The test results were obtained from samples that had the same compaction, but con- R2 = 0.6843 tained different cement contents, i.e. 2%, 3% 4 000 and 4%. 2 The static effective stiffness of the norite, 3 000 R = 0.142 ferricrete and the average of both the norite and ferricrete materials improved with an (MPa) stiffness effective Static 2 000 increase in the densification of the mate- 1.5 2.0 2.5 3.0 3.5 4.0 4.5 rial. The static stiffness increased by 18.8% Cement content (%) (625 MPa) from 95% to 100% Modified Norite material Ferricrete material Average of norite and ferricrete AASHTO density compaction effort. It is Poly. (ferricrete material) Poly. (norite material) Poly. (average of norite and ferricrete) evident that densification keeps soil particles together by reducing voids in the layer which Figure 4 Static effective stiffness of the norite, ferricrete and the average of both the norite and eventually increases the angle of material ferricrete materials friction and cohesion. The effective stiffness increases with the increase in cohesion of 180 the material to resist the applied stress of the static load. 150 120 Influence of cement content R2 = 0.4123 on stiffness response 90 R2 = 0.3734 The static effective stiffness of all the 60 2 norite, all the ferricrete and the average R = 0.0599 of both the norite and ferricrete materials 30

against cement content are depicted in Dynamic effective stiffness (MPa) 0 Figure 4. 1.5 2.0 2.5 3.0 3.5 4.0 4.5 The static effective stiffness of the norite Cement content (%) material increased by 48.4% from the 2% to Norite material Ferricrete material Average of norite and ferricrete 3% of the cement content, and by 14.1% from Poly. (ferricrete material) Poly. (norite material) Poly. (average of norite and ferricrete) the 3% to 4% of the cement content, while the static effective stiffness of the ferricrete Figure 5 Dynamic effective stiffness of the norite, ferricrete and the average of both the norite material increased by 7.8% and 1.5% from and ferricrete materials the 2% to 3% and 3% to 4% of the cement content respectively. Static effective stiffness 70 improved with the increase in cement con- 60 tent. An increase in cement content in both the ferricrete and norite materials improved 50 material strength and increased resistance 40 to cracking and crushing when the material 30 was under load. The dynamic effective stiffness of the 20 dynamic load (MPa) Effective stiffness to norite, ferricrete and the average of both the 10 norite and ferricrete materials in all wet and 0 dry conditions against cement content are 1 2 3 4 5 depicted in Figure 5. Cement content (%) The dynamic effective stiffness increased Effective stiffness of ferricrete material at dry conditions by 41% and 48% with an increase from the Effective stiffness of ferricrete material at wet conditions 2% to 3% and the 3% to 4% of the cement content, while the static effective stiffness Figure 6 Dynamic effective stiffness at 100% Modified AASHTO density compaction effort in the increased by 48% and 14.1% with an increase variations of the conditions of the stabilised ferricrete material from the 2% to 3% and the 3% to 4% of the cement content respectively. It can be It was noticed that an increase of 1% of the Effect of moisture on the material confirmed that the stiffness response to cement content on the sub-base layer may The sensitivity to variation in moisture both dynamic and static loading was directly increase the effective stiffness response to of the ferricrete and norite materials was improved with an increase in cement con- dynamic load proportionally, irrespective determined by comparing the differences tent, because the cement particles surround- of the type of material or material proper- in stiffness in both wet and dry conditions. ing the granules act as bridges, resulting in ties, because the layer damage induced by This was uncontrolled and not monitored, bonding and increasing the strength. The dynamic load is primarily in the form of and would depend on the laboratory condi- rate of increase in shear strength becomes extensive micro cracks during the effective tions (temperature and humidity), as well as greater as the cement content is increased. fatigue phase. on the duration of testing. The stiffness of

Journal of the South African Institution of Civil Engineering • Volume 57 Number 3 September 2015 31 the ferricrete material stabilised with 2% and 225 4% under dynamic loading at wet and dry 200 conditions when the material was compacted 175 at 100% Modified AASHTO density compac- 150 tion effort, as depicted in Figure 6. 125 The intention with the two points 100 depicted in Figure 6 was to evaluate the 75 effect of the variation in moisture, rather dynamic load (MPa) Effective stiffness to 50 than to evaluate the quantity of the effect. 25 The dynamic effective stiffness of the 0 ferricrete material stabilised with 2% 1 2 3 4 5 cement content showed an improvement Cement content (%) of 123% from 22 MPa to 49 MPa in stiff- Effective stiffness of norite material at dry conditions ness when the material deviated from wet Effective stiffness of norite material at wet conditions to dry conditions. On the other hand, the dynamic effective stiffness of the ferricrete Figure 7 Dynamic effective stiffness at 100% Modified AASHTO density compaction effort in the material stabilised with 4% cement showed variations of the conditions of the stabilised norite material no improvement when the material devi- ated from wet to dry conditions. These C4 C3 results revealed that higher compaction 120 and cementation of the material resulted in better material bonding. This means 100 R2 = 0.1091 the material could have a higher capacity 80 to resistance fatigue when the material is under stress. However, the stiffness of the 60 norite material stabilised with 4% cement 40 showed no improvement when the material to dynamic load (MPa) 20

was tested at wet to dry conditions. This Effective stiffness response indicated that this type of material becomes 0 more rigid at maximum compaction effort 2 000 2 500 3 000 3 500 4 000 4 500 5 000 (100% Modified AASHTO density) at opti- Effective stiffness response to static load (MPa) mum moisture content (OMC), and possibly Effective stiffness to dynamic load (MPa) Linear (effective stiffness to dynamic load (MPa)) at 4% cement with the resulting material being less sensitive to water and moisture. Figure 8 Relationship between effective stiffness response to dynamic load against static load At that stage, the cement-stabilised material became more cohesive and rigid irrespec- 180 y = 8E-10x2 + 0.0004x + 40.605 tive of moisture content and resulted in 160 R2 = 0.6054 C3 material being more resistant to fatigue and cracking 140 under controlled loading. Fine-grained 120 materials, such as this ferricrete material, 100 commonly display a different response to 80 C4 material coarse-grained materials as the modulus 60 decreases with increasing stress level until a dynamic load (MPa) Effective stiffness to 40 certain value is reached, after which it tends 20 to increase slightly when a conditioning 0 load stage has been passed (Theyse 2001). 0 40 000 80 000 120 000 160 000 200 000 The stiffness of the norite material stabi- Load repetitions (number) lised with 2% and 4% under dynamic loading Effective stiffness of both norite and ferricrete materials at wet and dry conditions when the material Poly. (effective stiffness of both norite and ferricrete materials was compacted at 100% Modified AASHTO density compaction effort, as depicted in Figure 9 Relationship between the effective stiffness and load repetitions to dynamic load Figure 7. The intention with the two points content showed an increase of 30% from The stiffness results of the stabilised depicted in Figure 7 was to evaluate the 170 MPa to 200 MPa in stiffness when the norite material response to dynamic load effect of the variation in moisture, rather material deviated from wet to dry condi- indicated that: than to evaluate the quantity of the effect. tions. The dynamic effective stiffness of the ■■ Compaction improved the strength of The dynamic effective stiffness of the norite norite material stabilised with 2% cement material by increasing cohesion. material stabilised with 2% cement content content showed an increase of 50% from ■■ Cement content increased rigidity of showed an increase of 97% from 52 MPa 100 MPa to 150 MPa in the stiffness, and material to loading. to 98 MPa in stiffness when the material when stabilised with 4% cement content ■■ The material still remained slightly sensi- deviated from wet to dry conditions. On the showed an increase of 100% from 150 MPa tive to moisture. other hand, the dynamic effective stiffness to 300 MPa in the stiffness when the mate- A comparison of the static and dynamic of the norite material stabilised with 4% rial deviated from wet to dry conditions. effective stiffness of both the stabilised

32 Journal of the South African Institution of Civil Engineering • Volume 57 Number 3 September 2015 Table 5 Suggested values of the effective stiffness related to the C4 and C3 material strength Noting the relationship of the material response to static and dynamic loads for the Suggested stiffness of dynamic load Static load (minimum values) (minimum values) determination of C4 and C3 material stiff- Category ness, the effective stiffness under static load, Effective stiffness Load repetition Effective stiffness (MPa) (number) (MPa) which was determined from the minimum UCS values of 750 MPa and 1 500 MPa, was C4 60 47 000 2 860 used to categorise the effective stiffness of

C3 80 83 300 4 340 the dynamic load. It was found that the min- imum effective stiffness and the load repeti- tions for the C4 and C3 materials evaluated norite and ferricrete materials in both wet The suggested modelled values of the under dynamic load could be developed and dry conditions is depicted in Figure 8. effective stiffness and number of load with a low level of confidence at that stage The depiction in Figure 8 is related to repetitions responding to dynamic load are because they correlated poorly. Although the proposed stiffness of the C4 and C3 presented in Table 5. the parameters for the C3 and C4 materials stabilised materials as proposed by Theyse The values in Table 5 relate to the effec- response were developed, the material of the et al (1996). For the determination of the C4 tive stiffness of the static load. Finally, it same properties as ferricrete material could and C3 stabilised materials stiffness, effec- has to be mentioned that even though a reach a higher compressive stiffness at an tive stiffness of static loads of 2 836 MPa sub-base layer may have an effective stiffness early stage, as it has a short working strain and 4 297 MPa which were derived from above the suggested values for C4 and C3, a and low load-carrying, while the material of the minimum UCS value of 750 kPa and number of load repetitions (stress) will be of a similar quality as the norite material could 1 500 kPa respectively, were used to catego- impact because it will determine the fatigue reach a constant working strain and still rise the effective stiffness of the dynamic phase of the sub-base layer. absorb more stress and load repetitions with- load. All the stiffness responses to dynamic out a sign of decrease in stiffness. It meant load were plotted to determine a trend that the norite material (C4) was expected line to be used for the classification of the CONCLUSIONS to have a longer life span than the ferricrete material, considering that stiffness did not A sub-base layer is fully functional when material (C3) in worst conditions. correlate well with a regression coefficient its properties are improved by stabilising of only 0.11. This poor correlation indicated the material to improve the layer’s effective that the stiffness of both the stabilised stiffness. The effect of dynamic loading on ACKNOWLEDGEMENTS ferricrete and materials could have been the response of this parameter of layer was This paper is based on the work reported influenced by the variation in moisture and of great concern, and an investigation was in the first author’s Master of Technology stabilisation, and also the potential of effec- conducted to determine the response of a (MTech) degree dissertation submitted to the tive stiffness response versus static could cement-stabilised layer by measuring the Tshwane University of Technology. Tshwane possibly not have been a good correlation in response to static and dynamic loading. University of Technology, the National terms of what it measured. Reading off from The stiffness response to both dynamic Department of Transport’s Northern the regression line, it was discovered that the and static loading was improved with the Transportation Centre of Development, minimum effective dynamic stiffness values increase in cement content and compaction. Geostrada Materials Engineering Laboratory of approximately 60 MPa and 80 MPa may The particles surrounding the granules and Aurecon are fully thanked for financial be for the C4 and C3 stabilised materials’ acted as a bond, thus resulting in an increase support. Tshwane University of Technology strength respectively. in strength. The rate of increase in shear is also thanked for permission to publish A comparison of the dynamic effective strength became greater as the cement this work. stiffness and load repetitions of both the content increased. It has been observed stabilised norite and ferricrete materials in that an increase of 1% of cement content both wet and dry conditions are depicted in on the sub-base layer could increase the REFERENCES Figure 9. effective stiffness response to dynamic load Brink, A B A 1985. Engineering Geology of Southern In relating the minimum dynamic effec- proportionally, irrespective of the type of Africa, Vol. 4. Silverton: Building Publications. tive stiffness values obtained in Figure 8, material, or the material’s properties. This COLTO (Committee of Land Transport Officials) 1998. together with the load repetitions, the stiff- was because the layer damage induced by Standard Specifications for Road and Bridge Works ness of both the norite and ferricrete materi- dynamic load primarily failed in the form of for State Road Authorities. Pretoria: Department of als in both wet and dry conditions were extensive micro cracks during the effective Transport. plotted against the number of load or stress fatigue phase. Cottrell, J R B H, Schinkel, T O & Clark, M T 2003. repetitions which caused strain at break at There was an increase in stiffness in all A traffic data for mechanistic-empirical pavement the end of working strain and are depicted the norite and ferricrete materials when designs. Virginia, US: Virginia Transport Research in Figure 9. Polynomial regression was used compaction was increased from 95% to 100% Council. to determine the relationship between the of maximum dry density. As the ferricrete CSRA (Committee of State Road Authorities) 1986. effective stiffness to dynamic load and load material had higher fine material content TMH 1 – Standard Methods of Testing Road repetitions. The dynamic effective stiffness and plasticity, it was noted that an increase Construction Materials. Technical Methods for correlated reasonably well with the regres- in compaction effort improved the effec- Highways, Pretoria: National Institute for Transport sion coefficient of 0.61. The expected load tive stiffness more than that of the norite and Road Research, CSIR. repetitions induced from dynamic load at the material. The effective stiffness response to CSRA (Committee of State Road Authorities) 1986. proposed minimum effective stiffness values dynamic and static loads could be direction- TRH 13 – Cementitious Stabilisers in Road of C4 and C3 materials were 47 400 and ally proportional to the cohesion of the Construction. Technical Recommendations for 85 500 respectively. material particles and density of the layer. Highways, Pretoria: Department of Transport.

Journal of the South African Institution of Civil Engineering • Volume 57 Number 3 September 2015 33 CSRA (Committee of State Road Authorities) 1985. dissertation (unpublished), Pretoria: Tshwane Steyn, W J vd M, De Beer, M & Du Preez, W 1999. TRH 14 – Guidelines for Road Construction University of Technology. Simulation of dynamic traffic loading for use in Materials. Technical Recommendations for Natt, G S & Joshi, R C 1984. Properties of cement and Accelerated Pavement Testing (APT). Report 99/057, Highways, Pretoria: Department of Transport. lime-fly ash stabilised aggregates. Transportation Pretoria: National Institute for Transport and Road De Beer, M 1985. Behaviour of cementitious sub- Research Record No. 998, Washington DC: Research, CSIR. base layers in bitumen base road structure. MEng Transportation Research Board. Steyn, W J vd M & Sadzik, E 2007. Application of the dissertation (unpublished), Pretoria: University of Paige-Green, P 1989. The influence of geotechnical portable pavement seismic analysis (PPSA) for Pretoria, South Africa. properties on the performance of unpaved gravel pavement analysis. Proceedings, 26th Southern De Beer, M 1990. Aspects of the design and behaviour wearing course materials. PhD thesis (unpublished), African Transport Conference (SATC 2007), of road structures incorporating light cementitious Pretoria: University of Pretoria. 9–12 July 2007, Pretoria. layers. PhD thesis (unpublished), Pretoria: University Paige-Green, P & Bam, A 1995. The hardness of gravel Theyse, H L, De Beer, M & Rust, F C 1996. Overview of Pretoria, South Africa. as an indicator of performance in unpaved roads. of the South African Mechanistic Pavement Design De Beer M, Kannemeyer, L & Fisher, C 1999. Towards Research Report RR 93/560, Pretoria: Department of Analysis Method. Divisional Publication DP-96/005, improved mechanistic design of thin asphalt layer Transport. 75th Annual Meeting of the Transportation surfacing based on actual tyre/pavement contact Paige-Green, P & Netterberg, F 2004. Cement Research Board (TRB), National Research Council, Stress In Motion (SIM) data in South Africa. stabilisation of road pavement materials Washington DC. Proceedings, 7th Conference on Asphalt Pavements laboratory testing programme – Phase 1. Theyse, H L 2001. The development of mechanistic- for Southern Africa (CAPSA), 29 August to Confidential Contract Report CR 2003/42, empirical permanent sub-grade deformation 2 September 1999, Victoria Falls, Zimbabwe. Pretoria: National Institute for Transport and Road model from simulator data. MEng dissertation Jones, D, Paige-Green, P & Prozzi, J 2001. Pozzolanic Research, CSIR. (unpublished), Johannesburg: Rand Afrikaans stabilisation of road pavement layers: A comparison SABS (South African Bureau of Standards) 2001. University. between Californian and South African practice. SABS EN-197-1:2000. Cement – Part 1: Composition, Theyse, H L, Maina, J W & Kannemeyer, L 2007. Report CR 2001/17, Pretoria: Division of Roads and Specifications and Conformity Criteria for Common Revision of the South African flexible pavement Transport Technology, CSIR. Cements. Pretoria: SABS. design method: Mechanistic-empirical component. Jordaan, G J 1994. The South African mechanistic Sadzik, E (CSIR) 2004. Personal interview, Pretoria. Proceedings, 9th Conference on Asphalt Pavements pavement design method. Research Report (Notes are in the possession of the first author.) for Southern Africa (CAPSA), 2–5 September 2007, RR 91/242, Pretoria: Department of Transport. SANRAL (South African National Roads Agency Gaborone, Botswana, pp 256–292. Kekana, P 2006. Traffic volume at the South African toll Limited) 2007. Revision of South African Pavement Van Wijk, A J, Hervey, J & Hartman, A M 2007. gates. MSc dissertation (unpublished), Cape Town: Design Method (SAPDM). Peer review of the Assessing material properties for pavement University of Cape Town. documentations prepared for the revision of rehabilitation design. Proceedings, 9th Conference Matheba, M J 2013. Behaviour of unconfined cemented the South African Mechanistic Design Method, on Asphalt Pavements for Southern Africa (CAPSA), materials under dynamic loading. MTech Publication 1998/009584/06, Pretoria: SANRAL. 2–5 September 2007, Gaborone, Botswana.

34 Journal of the South African Institution of Civil Engineering • Volume 57 Number 3 September 2015 Comparison of travel TECHNICAL PAPER Journal of the South African time between private Institution of Civil Engineering car and public transport Vol 57 No 3, September 2015, Pages 35–43, Paper 1167 in Cape Town GERHARD HITGE Pr Eng has been working as a transportation engineer in both the public and private sectors for the past 20 years, and has a G Hitge, M Vanderschuren Master’s Degree in Transportation Engineering (University of Pretoria). He was employed by Transport for Cape Town (TCT) at the time of The objective of urban transportation planning is, or should be, to optimise the access to writing this paper, and now operates as an opportunities for all people. One of the factors that defines access is to minimise the travel time independent consultant. His current focus is on the development of integrated, multi-modal transport systems that are between home and both primary and secondary activities. Optimisation refers to the balance sustained by higher-density, mixed-use urban development. between the benefits of reducing travel time with the cost of that reduction. Cost includes Contact details: operational, infrastructure and environmental costs. However, the reality in many cities is that ELUTI travel time is often minimised for some users or communities, while it remains relatively high PO Box 2072 for others. , 7551 This paper explores the core components of travel time of an integrated public transport Cape Town system, and compares that with travel time in the private transport system. This is done by T: +27 (0)82 372 2730 E: gerhardhitge@.co.za using travel time data for Cape Town to estimate the value of time spent on each component of a typical trip in Cape Town in 2013. The paper concludes that travelling by public transport takes significantly longer than by A/PROF MARIANNE VANDERSCHUREN (who is an Associate Member of SAICE) has a Bachelor’s private car for the average trip in Cape Town. It then highlights where to focus investment in the Degree in Transport Planning and Engineering public transport system to move towards an integrated, multi-modal system that can compete (Tilburg, Netherlands), a Masters Degree in with the private car, and thereby become attractive to all communities. System Engineering, Policy Analysis and Management (Delft, Netherlands) and a PhD in Modelling of Intelligent Transport Systems (Enschede, Netherlands). In 2000 she moved to INTRODUCTION and South Africa’s national governments South Africa. Here she has been researching the transferability of ‘developed Noise, pollution, global warming and (CCT 2013a; DoT 2007b) to improve public world’ knowledge in sustainable transport to South Africa and impaired liveability are just a few of the transport to provide a better service to cur- surrounding countries. negative impacts of transportation sys- rent public transport users. It simultaneously Contact details: tems around the world. These problems aims to make public transport competitive Centre for Transport Studies can largely be overcome by transforming with the private car in order to provide a via- University of Cape Town New Engineering Building car-centric cities into cities where public ble alternative mode to traditional car users. , 8000 transport and non-motorised transport are One of the key factors that define South Africa the dominant modes. This paradigm shift, accessibility is the travel time between T: +27 (0)21 650 2593 from the previous predict-and-provide for home and activities, or opportunities. E: [email protected] cars approach that has been displayed in Several frameworks have been developed the developed world, has also entered the to determine level of service indicators for developing world. South Africa confirmed the effectiveness of public transportation a commitment to the development of a systems (Hassan et al 2013; Cervero 2013). car-competitive public transport system in These systems contain a variety of criteria 2007 (DoT 2007a). How individuals choose attributes to evaluate, ranging from cost between routes, modes, departure times, measures, operational performance and utili- etc, has always been an important research sation, but virtually always consider travel question in transportation planning and time as a key indicator. management (Chen & Mahmassani 2004; Travel time was confirmed to be impor- Emmerink et al 1995). Goodwin (1995) tant to the South African travelling public as concluded that there is one simple but it was indicated as the most common deter- important proposition for travel behaviour minant of transport mode choice (32.5%) analysis that arises from past research: peo- in the 2013 National Household Travel ple differ. Therefore, planning should not Survey (StatsSA 2014). The cost of travel was just rely on analysis of averages. indicated as the second largest determinant of mode choice (26.2%). In his book The Background Transit Metropolis, Cervero (1998) states that The objective of urban transportation plan- the central premise is that transit will only ning is, or should be, to optimise access to become time-competitive with the car by opportunities for all people in a city. It is the improving the match between how services stated policy of the Cape Town metropolitan are configured and cities are designed. Keywords: travel time ratio, competitiveness of public transport, Cape Town

Hitge G, Vanderschuren M. Comparison of travel time between private car and public transport in Cape Town. J. S. Afr. Inst. Civ. Eng. 2015;57(3), Art. #1167, 9 pages. http://dx.doi.org/10.17159/2309-8775/2015/v57n3a5 35 The average travel time in Cape Town was, at about 90 minutes in 2013, at the upper end of the global range, which aver- ages around 70 minutes per person per day (Metz 2010; Schafer & Victor 1998). Of greater significance is the discrepancy between modes in Cape Town, with car users travelling at the global average of 70 minutes, but public transport users averaging around 110 minutes (CCT 2013b). This significant difference resonates with the discrepancy in the levels of spending on infrastructure for the two largely separate sub-systems of private and public transport networks, over the past three decades. Due to the legacy of apartheid, with spa- tial planning and economic suppression of non-white citizens, there is still a very strong correlation between race, income level and mode choice. The previously disadvantaged groups of the population are bound to use the much slower public transport modes (CCT 2009). It should follow that one of the key focus areas of the public transport investment strategy should be on the reduc- tion of travel time in public transport, both Figure 1 Study methodology relative to that of the private car and in real terms moving closer to the global average for metropolitan areas. Legend The interventions in the public transport IRT trunk network system discussed in this paper cannot and IRT feeder network Railway network should not compensate for the problems cre- Freeway ated by historic spatial planning. This paper Expressway Primary arterial therefore does not elaborate on the spatial Secondary arterial and land use planning interventions that Local roads and streets Central business district contribute to making public transport more Cape Town International Airport competitive with the car. Cape Town Municipal Area

Objectives and methodology Origins The primary objective of this paper is to Destinations show that travel time in public transport is not competitive with that of the private car in Cape Town. The second objective is to highlight three different areas of interven- tions to reduce public transport travel time. The methodology followed (summarised in Figure 1) was to identify the main com- ponents of a public transport trip that affect the total travel time, through a review of the literature. A survey was designed to collect relevant data for Cape Town. The literature Figure 2 Origins and destinations used for travel time survey also facilitated a comparison and interpreta- tion of the results from the data analysis, Table 1 Origins and destinations from which the approach to intervene was deducted. Origins Destinations

1. Elsies River 5. 1. Bellville DATA COLLECTION 2. 6. 2. Montague Gardens A survey was designed to collect data on travel times between seven typical origins 3. Diep River 7. Parklands 3. Wynberg and four typical destinations in Cape Town. The sample areas were geographically strati- 4. Macassar 4. Cape Town (CBD) fied to obtain a random mix of travel times,

36 Journal of the South African Institution of Civil Engineering • Volume 57 Number 3 September 2015 Table 2 Distribution of walking speeds recorded 100 as part of full journey travel time

Attribute m/s km/h 90

Average 1.01 3.62 80 Standard deviation 0.47 1.68

70 making use of a variety of transport network elements, as shown in Figure 2, and sum- Cape Town sample marised in Table 1. 60 A (single) gathering point was selected, The elderly Adults Children in each of the various districts, as a 50 departure and arrival point. These areas are, on average, less than 2 km from the nearest bus, minibus and train station, and 40 surveyors were required to walk to these stations or pick-up areas. Park-and-Ride 30 was not evaluated as an option due to the Cumulative proportion smaller than x (%) relatively low proportion of people using 20 this option at present (Wentley & Hitge 2013). The modes used between the origin and destination (O-D) pairs are as follows: 10 private vehicle, train (Metrorail), bus

(GABS), minibus-taxi (informal taxi, MBT) 0 and where possible, a limited number of 0.6 0.8 1.0 1.2 1.4 1.6 1.8 2.0 2.2 2.4 2.6 BRT (MyCiTi) trips. With certain O-D pairs Speed (m/s) not all modes were surveyed due to their unavailability in connecting these O-D Figure 3 Walking speed distribution pairs. In other cases the main mode would be used in conjunction with other modes to Table 3 Walking distance for different modes (km) complete a trip. Average time Mode Maximum Minimum Average The survey included three departures (min)* during the AM peak period (i.e. 06h00, Rail 4.22 0.61 2.13 35.9 06h15 and 06h30), two during the off-peak (OP) period (i.e. 12h00 and 12h15) and one Bus 3.05 0.38 1.43 24.1 return trip during the PM peak period (i.e. MBT 2.62 0.2 1.14 19.2 16h30). The modes for each AM, OP or PM period were conducted on the same day for All PT 1.36 22.8 each O-D pair. *At average walking speed of 1.01 m/s Surveyors were provided with a detailed route plan for each O-D pair, the travel time survey questionnaire and a GPS device. coordinates on the device itself, and this data survey is compared to similar international The questionnaire was created to docu- can be recovered after the survey has been and local values. ment the trip information. The GPS device completed. was programmed to record the position / The survey was conducted during Walking time coordinates of the surveyor throughout the November and the first week of December Walking time to public transport is a func- survey, and to save the position and time 2012. GPS and questionnaire data were anal- tion of walking speed and distance. Walking when the surveyor departs, arrives and other ysed and interpreted to compose the data set speed is influenced by personal characteris- points during the journey when congestion is used for this analysis. tics (age, fitness level, etc), gradient, ambient experienced or when the surveyor is waiting conditions like wind speed and surface qual- at a transfer station (i.e. way points). ity (Hermant 2012). The distance a person The data gathered from the paper-based ANALYSIS OF CAPE TOWN has to or is willing to walk, is influenced by surveys includes quantitative and qualitative TRAVEL TIME SURVEY DATA the proximity of the nearest public transport, data, whereas the GPS data only includes The data was analysed in terms of the three surface quality, lighting, safety and trip quantitative data. Two types of GPS devices trip components recorded during the trips, purpose. were used – active and passive trackers. An including walking time, waiting time and in- The survey data provides a sample size active tracking GPS device sends a signal vehicle travel time. While data was collected of 304 walking entries, of which Table 2 containing the longitude and latitude coor- for the BRT mode, the small data set was not provides a summary and Figure 3 illustrates dinates of the surveyor to a server which used, as (a) the coverage at the time of the the cumulative distribution relative to typi- then stores this data. This data can then be survey was very limited, and (b) the service cal ranges for various parts of Cape Town recovered at a later stage for analysis. Passive was in an early developmental stage. During (adapted from CROW 1998). The average trackers store the longitude and latitude the discussion, data obtained from the speed of just over 1.0 m/s is similar to those

Journal of the South African Institution of Civil Engineering • Volume 57 Number 3 September 2015 37 Table 5 Average speeds for the four modes in 300 different time periods 250 Speed (km/h) AM Off

200 Car 29.9 38.4

Train 22.6 28.0 150

Average Bus 21.5 20.6 100

Number of transfers MBT 27.9 33.8 50

0 efficient, especially during the AM peak 5 10 15 20 25 30 35 40 45 50 55 60 65 70 75 80 85 90 when headways are low, and (b) during off- Minutes peak periods headways can be excessive on Median value Average value all modes. The design of public transport interchanges in Cape Town often results Figure 4 Transfer time distribution in five-minute intervals for all transfers in walking distances between arriving and departing modes that exceed those of other Table 4 Transfer time distribution for all transfers in public transport journeys developing cities (Schalekamp 2007). The NHTS indicates that only about 16% Transfer 1 Transfer 2 Transfer 3 All transfers of public transport passengers in the country Number 449 288 115 852 make any transfers, and that transfers are far more prevalent for rail passengers. Given Average 13:05 16:35 18:59 15:04 the greater extent of rail coverage and usage Standard deviation 13:03 16:25 19:01 15:17 in Cape Town, this number is unsurpris- ingly higher for Cape Town, at 21% of public Median 09:00 11:00 12:00 10:00 transport passengers (CCT 2013b). The survey also indicated that the number of found by several other studies, including and more than to minibus taxi services. It further second and third transfers are, at 14% and as described by Hermant (2012). reveals that 52% of people in the Western 6% respectively, significantly higher in Cape The range of walking distances to and Cape walk to their first public transport Town than in other metropolitan areas. from main modes gives an indication of the mode in under five minutes, a further 20% In an assessment of the bus system in coverage of that mode, and the degree to reaching it in under ten minutes, and less Abu Dhabi in 2009 (Hassan et al 2013), the which it is accessible from random parts of than 15% of people walk for more than 15 waiting time for transfer of passengers was the city. Table 3 shows the average walking­ minutes. recorded as 38% from 5–10 minutes, 34% distance by passengers using the three Table 3 shows that, while the survey was from 10–20 minutes, and 13% less than five traditional main modes of public transport, not meant to be representative of the City’s minutes. The median transfer time is there- as well as the theoretical average walking travel patterns, similar walking times were fore well below the ten minutes recorded for time, using the average speed from Table 2. recorded in the three separate surveys. Cape Town. Perhaps unsurprisingly the walk to rail had the highest minimum, maximum and Transfer time In-vehicle travel time ­average distances, while minibus-taxi (MBT) Table 4 shows the transfer time distribu- The in-vehicle travel time depends both on recorded the lowest on all three. While tion for the sequential transfers in a trip, the speed and the distance travelled. Speed passengers may elect to use a feeder service while Figure 4 shows the distribution of all is influenced by the speed limit and the pre- to reach rail or even bus stations to reduce recorded transfer times graphically. The vailing level of service (LOS), which dictate their travel time, many cost-sensitive public average transfer time increases notably with the ability to maintain the desired speed for transport users are known to walk these subsequent transfers, i.e. those further from vehicles in mixed traffic. For vehicles with distances. home on the journey. Of the 856 transfers, dedicated or semi-dedicated rights-of-way, A household survey of about 2% of all only 15 were recorded as being zero minutes, the speed is not only determined by the households in Cape Town was conducted or being less than one minute from arriving design standards of the way and vehicles, but during 2012 (CCT 2013b). Its ‘stated prefer- to departing. also by the frequency and duration of stops ence’ component found that minibus taxi is The data was not collected to distinguish along the route. the most and rail the least accessible public between the time taken to walk between Table 5 shows the average in-vehicle transport modes in Cape Town. The aver- arriving and departing modes, or the actual speeds for the four different modes in the age total walking time was incidentally also time waiting for the second mode to depart, three different time periods of the survey. recorded as 22.8 minutes, which includes although this distinction is important when The table shows that the car has the highest trips to public transport, as well as about assessing interventions to reduce the transfer average speed during both the AM and off- 21% of respondents who make entire trips time. peak periods, followed by the minibus taxi, on foot. Table 4 shows that the median travel time with conventional GABS buses consistently Data from the National Household Travel is significantly below the average, and that travelling slower than rail. Survey (NHTS) (StatsSA 2014) confirms this the standard deviation is very high. An The average peak hour speeds are intui- trend, showing that the national walking dis- assessment of the factors contributing to tively expected, with MBTs being marginally tance to train and bus is about 65% and 15% this reveals that (a) some transfers are very slower than cars due to more frequent stops,

38 Journal of the South African Institution of Civil Engineering • Volume 57 Number 3 September 2015 of the extent of the road network relative to 0 10 20 30 40 50 60 the rail network in Cape Town. A significant 80 70 contribution to the length of rail trips is the fact that the rail network is radially outwards 70 from the CBD, with an absence of north- 60 south links in the eastern part of the city.

60 Total trip travel time 50 Table 6 shows the travel time of all three 50 components for the four main modes. The 40 in-vehicle travel time for the public transport modes ranges from 25% to 50% higher than 40 that of the car. However, the waiting and walking time adds substantially to all public Speed (km/h) Speed 30 30 transport modes, and even doubles the total trip travel time for rail. The result is that 20 travel time by public transport was recorded 20 in the survey as between two and three times that of the car. 10 The Household Travel Survey (CCT 10 2013b) also found that persons travelling by train had the longest travel time at 75 0 0 minutes (during the AM peak). This was 0 5 10 15 20 25 30 followed by bus travel at 71 minutes and Distance (km) minibus taxi at 51 minutes. Car drivers Train MBT Linear (train) Linear (MBT) travelled an average of 46 minutes during the Bus Car Linear (bus) Linear (car) AM peak. The NHTS found a similar trend with the average travel time by train in the Figure 5 Speed–distance relationship for four modes Western Cape, the longest at 81 minutes, followed by bus at 70 minutes, minibus taxi Table 6 Travel time per trip component and distance per mode at 49 minutes, car at 40 minutes and walking only, at 29 minutes. Travel time Car Train Bus MBT All PT The average trip distance for all modes, In-vehicle 42.1 55.5 63.8 51.8 57.1 including the car, was slightly lower than Waiting 0.0 31.0 33.4 22.1 29.0 the average for all public transport modes, at 27.02 km. Walking 0.0 35.5 23.8 18.5 27.3

Total 42.1 122.0 121.0 92.4 113.4 Travel time by mode to four

Relative to car 1.00 2.90 2.88 2.20 2.70 typical destinations The discrepancy between absolute and Avg distance 24.65 30.06 26.17 27.19 27.81 relative travel time between car and public transport is further highlighted by analysing but despite erratic driving making up speed However, the effect of congestion on travel the travel time differences between the four along the route. speed is clear when comparing the off-peak destinations. Table 7 shows that both mode The low speed of train trips is concern- speeds for car and MBT to that of the AM choice and destination play a critical role in ing, considering the fact that trains run on a peak. the competitiveness between car and public dedicated right of way, and stations are typi- Figure 5 shows the variation of speed with transport. cally not less than 1 km apart. Speeds of 30 distance travelled for the four modes and for Bellville is the most central destination to 40 km/h should be expected for this mode the forward direction (excluding PM peak (shortest trip length) and therefore intuitively (Vuchic 2007). The low speed is ascribed to data). Of interest is that the speed increases requires the shortest overall travel time. the fact that substantial speed restrictions slightly with distance for car and train, but While Cape Town CBD is probably the most exist on many parts of the rail network due that it decreases slightly for both bus and eccentric destination, the transport system to poorly maintained tracks, and station minibus-taxi (MBT) (the distance axis is not developed radially from there to make it capacity requires arriving trains to some- to scale, so the gradient of the trend line has relatively more accessible by both modes. times wait for trains that are not ready to no meaning). However, because of this radial nature, travel depart. In addition, the average age of the The data reveals that, for the 28 O-D time by public transport is only about twice rolling stock fleet is about 40 years, and over- pairs, the distance by car is about 10% that of the car for Cape Town compared to crowding results in an inability to collect all shorter than the average of the public the more than three times for Bellville. It is waiting passengers, which leads to longer transport modes, with rail being almost 20% also the only destination where the distance dwell times for trains at many stations. longer than the road-based public transport by public transport modes is lower than for The data was not collected to distinguish modes. Given the geographically stratified the car. between travelling on local roads, arterial nature of the O-D pairs, it is believed that Table 7 also reveals that, while or freeways, for which design speeds vary. the distance per mode gives a fair indication Montague Gardens is only the second

Journal of the South African Institution of Civil Engineering • Volume 57 Number 3 September 2015 39 Table 7 Average speeds per mode to the four destinations

Total trip time Total trip in-vehicle Total trip walking Total trip waiting Total trip in-vehicle including waiting, Average distance time (min) time (min) travel time walking and in-vehicle speed in-vehicle

Bellville as destination

Average 23.61 27.56 23.03 44.81 82.37 37.15

Car 21.02 29.15 29.15 43.87

All PT 24.47 27.56 23.03 49.98 99.96 34.93

Cape Town as destination

Average 27.49 34.94 19.89 53.65 91.09 34.59

Car 28.83 47.48 47.48 41.94

All PT 26.99 34.94 19.89 55.96 107.39 31.85

Montague Gardens as destination

Average 27.00 29.82 45.85 58.22 112.11 30.52

Car 24.20 45.97 45.97 32.65

All PT 28.02 29.82 45.85 62.66 136.10 29.75

Wynberg as destination

Average 26.33 28.54 28.50 57.22 99.90 30.17

Car 23.03 46.68 48.77 33.86

All PT 28.14 28.29 29.98 63.35 119.35 28.83

poorly with international benchmarks, and is Time (T) = Distance (L) / Speed (V) Legend certainly also an area that holds significant Car potential for improvement. Mode 1 The next step in the process is to inter- rogate where interventions should be made Main mode to reduce the actual travel time of public Walking Lw1 LM transport, as well as the travel time relative VM Vw1 LM to the private car. This is done by comparing the impact of different strategies along a VM Lc1 Lw2 typical trip. Vw2 Vc1 LC Lc2 VC Vc2 IMPROVING THE TRAVEL TIME RATIO IN CAPE TOWN Figure 6 Speed and distance components of a trip The competitiveness of public transport rela- tive to the private car is usually calculated longest trip by car, it requires the longest of public transport trips is currently longer by travel time ratios, which are defined as travel time by public transport. The much than that for the car. Secondly, the in-vehicle the quotient of the travel time by private longer transfer time component also speed of the car is higher than the speed car and public transport between the same confirms the effect of this destination of public transport, and the car trip is not origins and destinations in the city (Jones Jr not being situated along a trunk public subjected to transfer time. 2013). Evidence from the Stockholm region transport route, so that a second feeder trip While cost-sensitive Capetonians are shows that for a travel time ratio of up to 1.5, is typically required. As discussed before, willing to walk long distances as a trade-off the share of public transport is 50% to 70%. this second or third transfer time is longer to incurring costs, this adds to their total Therefore, public transport loses most of its than the first. trip length and reduces their accessibility. competitiveness when the travel time ratio The use of bicycles to reduce the travel time exceeds about 1.5. Discussion of findings of this component of the trip is notable, and In order to analyse this principle, the The above analysis confirms that, even for a clear area of potential operational interven- impact of reducing the travel time ratio for the relatively small sample, the analysis and tion that requires relatively minor infrastruc- Cape Town from its current level to a desired comparison show that the data falls within ture provision, at least initially. level of 1.5 will be analysed. Since Cape typical ranges and is therefore considered While the proportion of people mak- Town is moving towards a single integrated acceptable for further interpretation. ing transfers is relatively low, the length public transport system (CCT 2013a), the The data shows clearly that public of transfers poses an area for significant objective is not to improve the travel time on transport is not competitive with the car improvement in total travel time. In addition, individual modes, but rather for the system on a variety of fronts. Firstly, the distance the in-vehicle travel time compares very as a whole.

40 Journal of the South African Institution of Civil Engineering • Volume 57 Number 3 September 2015 Equation to evaluate Table 8 Current values of trip components and travel time ratio competitiveness of public transport Descriptor Parameter Base values Time Figure 6 illustrates the different speed components of multi-modal public trans- Lc1/2 Minor road distance travelled 7.7 15.4 port and the car between the same origin Vc1/2 Speed on minor roads 30 and destination. An equation is then LC Arterial distance travelled 18 constructed to test the impact of several 36.0 variables on different travel time ratios for VC Speed on arterials 30 travel by public transport (TPT) and travel TC Travel time by car 51.4 by car (TC). With Time (T) = Distance (L) / Speed Lw1/2 Walking distance (km) 1.36 (V), the travel time by car and public trans- 22.5 V Walking speed (km/h) 3.62 port respectively are given by: w1/2

Tr1 Transfer once (21% of passengers) (min) (13) 2.7 TPT = Time walking (w*) + Transfer time (Tr) Lm Feeder mode distance (km) 7.7 + In-vehicle time mode (m) 16.5 + In-vehicle time Main mode (M) Vm Feeder mode speed (km/h) 28

Tr1 Transfer twice (14% of passengers) (min) (17) 2.4 [* 1 for the first w, Tr or c; and 2 for the second w, Tr or c] LM Trunk distance (km) 18 49.1 VM Trunk speed (km/h) 22 or TPT Travel time by PT 93.2

Lw1 Lm Lm Lw2 TPT =  + Tr1 + + Tr2 + + (1) Travel time ratio 1.81 Vw1 Vm Vm Vw2

Table 9 Trip components reduction scenarios TC = Time on local roads (c) + Time on major roads (C) + Time walking to Scenario destination Trip component PT0 PT1* PT2 PT3 PT4 PT5 PT6 Lc1 Lc Lc2 Lw TC =  + +  + (2) Trunk time 1 67% 1 1 1 76% 83% Vc1 Vc Vc2 Vw Transfer twice 1 1 1 1 0% 1 0 The survey did not distinguish between Feeder time 1 1 2% 1 1 73% 73% cars travelling on minor or major roads, or between travelling on trunk or feeder-type Transfer once 1 1 1 1 0% 1 37% services, since the current public transport Walk time 1 1 1 28% 1 1 1 system is not configured in this way. Table 8 provides the surveyed and estimated 2013 Trip time, PT (min) 93.2 77.1 77.1 77.1 88.1 77.1 77.1 base values for the above equations, based on *  the following assumed trip distance split, as Trip time is reduced from 93.2 in PT0 to 77.1 min in PT1 by reducing trunk time in PT1 to 67% of its current value. well as the prevailing travel time ratio: ■■ average trip length = 27 km ■ ■ average walking distance = 1.36 km 100 ■ ■ 70% of remaining distance on arterial 90 roads or on trunk services = 18.0 km 80 ■ ■ 30% of remaining distance on minor 70 roads or on feeder services = 7.7 km 60 In order to achieve a travel time ratio of 50

1.5, TPT has to be reduced by 16.1 to 77.1 Minutes 40 minutes, by reducing one or several of the 30 contributing components. Table 9 shows the 20 proportion that different trip components 10 must be reduced to, relative to the base case, 0 in order to achieve the desired travel time Car 1.5 Car PT0 PT1 PT2 PT3 PT4 PT5 PT6 ratio of 1.5. Figure 7 illustrates the impact of Walk time 22.5 22.5 22.5 6.4 22.5 22.5 22.5 these scenarios in graphical format. Transfer once 2.7 2.7 2.7 2.7 0 2.7 1 A comparison of the impact of Scenarios Feeder time 16.5 16.5 0.4 16.5 16.5 12 12 PT1, PT2 and PT3 indicates that the reduc- Transfer twice 2.4 2.4 2.4 2.4 0 2.4 1 tion in trunk travel time appears to be the Trunk time 51.4 77.1 49.1 33 49.1 49.1 49.1 37.5 40.6 single most effective measure to achieve the Scenarios desired effect. It is evident from both Table 9 and Figure 7 that even removing transfers Figure 7 Travel time for car and public transport scenarios

Journal of the South African Institution of Civil Engineering • Volume 57 Number 3 September 2015 41 completely (Scenario PT4) does not, on its adequate capacity so that passengers do cycling, however, requires the provision own, result in sufficient travel time savings. not have to wait for the next vehicle. of adequate storage facilities at trunk sta- The travel time of each component can ■■ Time in the feeder vehicle can be reduced tions, as well as the ability to take bicycles be reduced by either reducing the distance through shorter trips to trunk, i.e. higher onto trains and buses. factor or increasing the speed factor, or both, trunk coverage, or limited priority above ■■ Ingress walking distance can be for each trip component. For example, while general traffic, e.g. queue jump (Chitauka reduced by increasing coverage of the keeping the trunk distance at 18 km, the & Vanderschuren 2014) or pre-emptive feeder system (short-term strategy), or speed of the trunk mode needs to increase signal activation. by increasing the land use density rela- from 22 to 32.7 km/h in order to reduce the ■■ The second transfer time can be reduced tive to the feeder or trunk access points in-vehicle travel time to 33 minutes. This through: (longer-term strategy). Not only does the trunk vehicle speed falls comfortably within ■■ Reduced walking time between former increase the overall cost of the the range of 30 to 40 km/h suggested by feeder and main modes brought about transport system, but the latter approach Vuchic (2007). Increasing the trunk speed through infrastructure design, and also eliminates much of the lower speed towards the upper end of the range would ■■ A decrease in the headway of the components of the trip, hence resulting in significantly reduce the travel time ratio, departing mode, again assuming there a greater proportion of the trip at higher thereby increasing the competitiveness of is adequate capacity so that passen- speed. public transport even further. gers do not have to wait for the next ■■ Egress walking distances can be reduced Each intervention comes at different vehicle. through higher density of end use desti- implementation costs and time frames, as ■■ In designing an integrated public trans- nations in closer proximity to the trunk well as environmental and socio-economic port system, there is an opportunity stations. drivers. The next section describes the con- in Cape Town to reduce the walking siderations of these factors in context, with- distance between arrival and departure out attempting to cost typical interventions. modes by creating more compact inter- DISCUSSION change facilities. The footprints of Cape An opportunity presents itself while plan- Town’s interchanges are typically much ning to invest billions of rand on improving INTERVENTIONS TO REDUCE larger than interchanges used by similar public transport systems in South Africa TIME OF TRIP COMPONENTS volumes of people in other developing over the next decade or two. The focus of The competitiveness of the public transport cities (Schalekamp 2007). An example improvement should arguably fall primar- system can be improved by considering is that the furthest minibus taxi bay is ily on travel time improvements, as this the following factors that affect various more than 500 m from the entrance to is affected by the capital programme. interventions: the Bellville Station, before encountering Operational quality improvements could be stairs and access control to reach a rail added to the system in a variety of ways in In-vehicle travel time platform. future. ■■ Providing dedicated or semi-dedicated ■■ Vertical integration between modes are Both the absolute and relative travel time right-of-way for trunk services is the most virtually non-existent with only some of of public transport, when compared to the common infrastructure approach to giv- the most recently constructed stations private car, must be drastically reduced in ing a time advantage to public transport offering some form of grade separation. Cape Town. Reducing the absolute travel vehicles. ■■ Transfer to minibus taxi vehicles has the time would add social benefits, by giving ■■ However, as seen by the very low speeds added disadvantage that vehicles typically individuals more time for other priorities of the rail system, the quality of the only depart when a minimum occupancy like family, leisure and further education. way and protection from not only other rate has been reached. This often leads to Reducing it relative to the car would positive- vehicles, but also pedestrians, is critical to inordinately long transfer times on lower ly influence asymmetrical churn (Chatterjee ensure that operational speeds match the frequented routes. 2001; Del Mistro & Behrens 2006) in favour infrastructure design speed. of public transport, eventually achieving ■■ Similarly, a high standard maintenance Walking time a sustained shift from private car towards regime is required to ensure that the ■■ Walking speed cannot be changed as more efficient public transport modes. quality of the way is conducive to high- it is a characteristic of the community. In the distribution of travel time ratios speed travel. However, walking time would probably for individual O-D pairs in a city, with an ■■ The number of stops along a route plays improve marginally with increased fit- average of 1.5, there would be values below a substantial role in the overall in-vehicle ness levels, as well as smooth and clean 1.0 (faster than car) while others could far travel time. It is important that station walkways. exceed 2. A particular reduction in the aver- spacing be optimised between the need ■■ Replacing walking with cycling trips age ratio could be achieved through small to provide access and maintain high would contribute significantly to a sav- improvement for large passenger volumes, travel speeds. Where station spacing is ing in current walking time, and could or large improvements for small passenger low, operational and scheduling alterna- even replace short vehicle trips. At a numbers. Strategies to attract choice pas- tives, such as skip-stop services, could speed of three to four times that of walk- sengers to public transport would probably be employed to balance these competing ing, this could substantially reduce the focus on improvements at the margin of 1.5, objectives. time to first access public transport. In while social objectives could drive a reduc- some cases cycling could even replace tion in ratios that exceed 2.0. Transfer time the need for a feeder trip, especially Capital improvements must certainly ■■ The first transfer time can be reduced where a relatively long walk and short focus on the right of way, and achieving high through a decrease in the headway of feeder trips are required to access trunk in-vehicle speeds within the appropriate tech- the departing mode, assuming there is services. Promoting a greater usage of nology. However, it is as important to reduce

42 Journal of the South African Institution of Civil Engineering • Volume 57 Number 3 September 2015 the transfer time in a multi-modal system by networks. Transportation Research Record, Goodwin, P B 1995. Car dependence. Transportation, appropriate design of transfer facilities. At 1894(1): 209–221. 2(3): 151–152. facilities where large volumes of passengers Chitauka, F & Vanderschuren, M 2014. An investigation Hassan, M N, Hawas, Y E & Ahmed, K 2013. A multi- transfer, the vertical separation of modes into the performance of full BRT and partial dimensional framework for evaluating the transit should always be considered. The higher bus priority strategies at intersections by micro- service performance. Transportation Research implementation cost can be offset by the value simulation modelling in a South African context. Part A: Policy and Practice, 50: 47–61. of both the cumulative time savings, as well as Proceedings, 33rd Southern African Transport Hermant, L F L 2012. Video data collection method for higher land development potential. Conference (SATC), 7–10 July 2014, Pretoria. pedestrian movement variables and development CCT () 2009. 2009 General of pedestrian spatial parameters simulation model Household Survey Analysis for Cape Town. Cape for railway station environments. PhD thesis, ACKNOWLEDGEMENT Town: City of Cape Town. Stellenbosch: University of Stellenbosch. This paper was produced under the MISTRA CCT (City of Cape Town) 2013a. Comprehensive Integrated Jones, S L Jr (Ed.) 2013. Proceedings, 3rd International Urban Futures programme which pro- Transport Plan. Cape Town: City of Cape Town. Conference on Urban Public Transportation motes research towards sustainable urban CCT (City of Cape Town) 2013b. Household Survey Systems, Paris, 17–20 November 2013. development, under the theme: “Green, Report – Final Draft. Cape Town: City of Cape Metz, D 2010. Saturation of demand for daily travel. Fair and Dense”. The time required for the Town. Transport Reviews, 30(5): 659–674. writing process was afforded by the City of CROW (Centre for Research and Contract Schalekamp, H V 2007. Towards a user-oriented Cape Town. Standardisation in Civil Engineering) 1998. approach in the design and planning of public Recommendations for traffic provision in built-up transport interchanges. MPhil dissertation, Cape areas (ASVV). Ede, Netherlands: CROW. Town: Department of Civil Engineering, University REFERENCES Del Mistro, R & Behrens, R 2008. How variable is of Cape Town. Cervero, R 1998. The Transit Metropolis – A Global variability in traffic? How can TDM succeed? Schafer, A & Victor, D G 1998. The future mobility Enquiry. Washington DC: Island Press. Proceedings, 27th Southern African Transport of the world population. Transportation Research Cervero, R 2013. Bus Rapid Transit (BRT) – An Conference (SATC 2008), 7–11 July 2008, Pretoria. Part A: Policy and Practice, 34(3): 171–205. efficient and competitive mode of public transport. DoT (Department of Transport) 2007a. Public StatsSA 2014. 2013 National Household Travel Survey. Proceedings, 20th ACEA, Scientific Advisory Group Transport Strategy. Pretoria: Department of Pretoria: Statistics SA. Report 20, Belgium. Transport. Vuchic, V 2007. Urban Transit Systems and Technology. Chatterjee, K 2001. Asymmetric churn – Academic DoT (Department of Transport) 2007b. Public New York: Wiley. jargon or a serious issue for transport planning? Transport Action Plan. Pretoria: Department of Wentley, O & Hitge, G 2013. Understanding the Bursary paper presented to the Transport Planning Transport. utilisation of Park and Ride facilities (Cape Town Society, UK (unpublished). Emmerink, R, Axhausen, K, Nijkamp, P & Rietveld, P 1995. – 2012). Proceedings, Southern African Transport Chen, R & Mahmassani, H 2004. Travel time Effects of information in road transport networks with Conference (SATC), 8–11 July 2013, Pretoria. perception and learning mechanisms in traffic recurrent congestion. Transportation, 22(1): 21–53.

Journal of the South African Institution of Civil Engineering • Volume 57 Number 3 September 2015 43 TECHNICAL PAPER Finite element analyses of Journal of the South African Institution of Civil Engineering the structural behaviour Vol 57 No 3, September 2015, Pages 44–56, Paper 997 of pylons supporting an inclined coal conveyor MILOŠ PERDUH graduated in 2009 from the Belgrade University with a degree in civil engineering, and he subsequently obtained a Master’s degree in structural engineering. His M Perduh, J A v B Strasheim fields of interest include finite element modelling and dynamic analyses. He has been involved in projects across Europe, the As part of the coal conveyance system at Medupi Power Station, an inclined coal conveyor Middle East and southern Africa, and is will transport coal from the stockyard to the coal transfer tower, and from there to the boilers. currently employed by AECOM SA. The conveyor is supported by concrete columns (pylons), in turn supporting the steel gantries Contact details: on which the conveyor is located. The pylons can be considered as cantilever columns during AECOM SA (Pty) Ltd the construction stage, while in the final operational stage with the steel gantries positioned PO Box 3173 Pretoria, 0001 in-between the pylons, a frame system will be formed. The gantries are connected to the pylons South Africa with custom-designed sliding joints, which allow limited movement of the gantries in the T +27 (0)12 421 3500 longitudinal direction of the conveyor. This paper describes how various finite element analyses E: [email protected] of the structural behaviour of the pylons and the overall structure of the inclined coal conveyor were undertaken to assess wind and seismic actions. It focuses on modelling the behaviour DR BREDA STRASHEIM Pr Eng is a Senior of the concrete pylons during the construction period, a comparison between finite element Lecturer in Structural Engineering at models (FEMs) with different complexities and the implications of simplifying the FEMs. It will be Stellenbosch University. He graduated in civil engineering in 1974 and subsequently obtained shown that the simplified beam element models provide adequate modelling of the structural a BSc in Computer Science, an M Eng, an MBA behaviour for this kind of structure. The modelling of non-linear connections between elements and a PhD. His professional engineering career for static and dynamic conditions was also investigated, as well as the influence of the sliding joints included a period with the Department of Water between the pylons and the gantries on the overall behaviour of the structure. It will be shown Affairs, and a decade in the consulting that the overall behaviour of the structure can be highly influenced by the action of the sliding engineering practice of Geustyn Forsyth & Joubert Inc doing municipal infrastructure design, pipeline and dam construction and management mechanism and that the force distribution between the structural members can differ significantly. projects, as well as business system projects. His research interests include Recommendations on how to approach the modelling of this type of structure are made. numerical (FEA) modelling and structural dynamics. He also lectures It is concluded that the simplified model can be used to capture the behaviour of the structure, engineering management and is involved in organising and presenting as well as the complex sliding joint mechanism, which has a major influence on the performance of course work for the Construction Management Programme (CMP). the structure and the force distribution in the structural system. Contact details: Department of Civil Engineering Stellenbosch University Private Bag X1 INTRODUCTION SCOPE OF THE PAPER MATIELAND 7602 Medupi Power Station, shown in Figure 1, is The structure of the inclined coal conveyor South Africa situated in Lephalale, in the northern part of was conceived as a combination of steel T: +27 (0)21 808 4435 South Africa. After its completion it will be gantries and concrete pylons. During the E: [email protected] the fourth largest coal-fired power station in construction stage, without the steel gan- the world, and at the same time the world’s tries in position, the concrete pylons will act largest dry-cooled coal-fired power station. as pure cantilevers in both the longitudinal The first unit should have started operation and transverse directions. The paper in 2014. The cost is estimated at R140 billion. describes the analyses of the behaviour of It will have six generating units which, at the concrete pylons during the construction completion, will deliver 4 788 MW of electri- stage. After the installation of the steel cal power to the South African electricity gantries, a new ‘hybrid’ frame system, con- distribution grid. As part of the coal convey- sisting of the concrete pylons and the steel ance system, an inclined coal conveyor (ICC) gantries, will be formed in the direction 300 metres in length, consisting of concrete of the conveyor. The paper also describes pylons and steel gantries, will be installed. the finite element (FE) analyses that were This paper analyses the structural behaviour done of the behaviour of the complete ICC of the ICC and focuses on the investigation structure. A number of FE models with of the required level of complexity of finite different levels of complexity were created element modelling for this kind of structure and their advantages and disadvantages during the construction and operational are given. The connections between the stages. The main focus of the investigation steel gantries and the concrete pylons are is pointed towards the structural behaviour designed to allow a certain amount of longi- Keywords: inclined coal conveyor, Medupi, sliding joint, finite element analysis, and force distribution in the concrete pylons tudinal movement of the steel gantries, thus force distribution, concrete pylon, steel gantry, simplified model supporting the steel gantries. avoiding the development of any additional

Perduh M, Strasheim J A van B. Finite element analyses of the structural behaviour of pylons supporting an inclined coal conveyor. 44 J. S. Afr. Inst. Civ. Eng. 2015;57(3), Art. #997, 13 pages. http://dx.doi.org/10.17159/2309-8775/2015/v57n3a6 which can be used for modal analyses was required in view of the fact that non- linear gap elements cannot be used in modal analysis. ■■ Investigation into the influence of the sliding connection on the overall behaviour of the structure: A number of FE models of the concrete pylons for the systems with and without sliding con- nections were analysed and the results are compared. Special attention was paid towards both the structural behaviour of the tallest pylons and the stiff, braced (seventh) pylon. It is shown that the slid- ing connections have a major influence on the behaviour of the pylons. ■■ Modelling the effects of cracking of the pylon box section: In addition, the Inclined coal conveyor analyses of the concrete pylons with both cracked and uncracked sections were Figure 1 Medupi Power Station (Source: Eskom) compared. It is important to determine the influence of the implementation of properties of cracked sections in FE models on the overall behaviour of the structure. Both the periods of oscillation and the force distribution will differ due to cracking of the sections.

STRUCTURAL SYSTEM 1 AND COMPONENTS 2 3 4 5 Structural concept of the 6 7 inclined coal conveyor 8–12 The ICC will transport coal from the Figure 2 Three-dimensional view of FE model of inclined coal conveyor stockyard to the coal transfer tower, and from there to the boilers. Two conveyors forces in the system due to the effects of and shells. It was necessary to determine in the power station will supply coal for changes in temperature. With the aid of whether the steel gantries, which are six generating units. The structural system finite element software, the designer and actually 3D trusses, could be modelled of the ICC consists of steel bridge gantries analyst went through a process to discover with a single beam element, and to what spanning 12 concrete pylons (Figure 2). and evaluate differences in behaviour of aspects of the modelling special attention The spans of the steel bridge gantries vary the structural system with and without needed to be paid. The results from the from 14 m to 30 m. The height of the pylons implementation of the sliding connection different FE models and the different varies from 3 m to 60 m. The cross-section between the pylons and the gantries. analyses, both static and dynamic, are of the first six pylons is a hollow box sec- The paper focuses on the following presented and compared. The aim was to tion with outside dimensions of 5 800 mm aspects of the structural analysis: investigate whether simplified FE models, × 2 000 mm, while the seventh pylon is ■■ Investigation into the behaviour of the with only beam elements modelling the stiffened with additional walls and has concrete pylons during the construc- behaviour of the 3D trusses, could fully total outside dimensions of 5 800 mm × tion period: Being tall and slender struc- describe the behaviour of this type of 5 800 mm (Figure 3). The thickness of the tures, the concrete pylons are vulnerable structure. The advantages and disadvan- walls is constant throughout at 400 mm. to possible cross-wind vortex-shedding tages of using FE models with different The pylons are founded on individual pad effects. The amplitude of the additional levels of complexity are discussed. footings, with a depth of 1 500 mm. The bending moments induced by the vortex- ■■ Modelling of the non-linear sliding steel bridge gantries supporting the coal shedding effects was quantified. The tem- connection between steel gantries and conveyors consist of rigid steel portal porary measures undertaken to secure concrete pylons: Using the implemented frames, supported on two main steel the structural stability of the concrete non-linear gap elements of the FE soft- trusses. The width of the steel gantries is pylons during the construction period ware, the behaviour of the structure was 9.5 m, while the height is 4.5 m (an isomet- were also analysed. simulated under static loading. Based on ric view of one typical steel gantry is shown ■■ Comparison of results from FE models the results of this analysis, the replace- in Figure 4). with different complexities: A number ment of the non-linear gap elements with The pylons indicated are numbered from of different FE models were created using linear elements was attempted. A model left to right from 1 to 12. This paper focuses different types of elements, e.g. beams of a sliding connection with a finite gap mainly on analyses of the behaviour of the

Journal of the South African Institution of Civil Engineering • Volume 57 Number 3 September 2015 45 first three pylons, which are the tallest, and the seventh one, which is the stiffest. For reference, the global X-direction is along the length of the conveyor and the global Y-direction is perpendicular to the direction of the conveyor, as shown in Figure 2.

Design philosophy of structural system The static structural system formed by the pylons during the construction stage is equivalent to a pure vertical cantilever in all directions. Once the steel bridge gantries Pylons 1–6 Pylon 7 have been positioned between the pylons, a new frame is formed in the longitudinal Figure 3 Plan layout and 3D view of typical pylon sections direction. In this case the steel bridge gantries not only have the function of span- ning between the pylons to support the coal conveyors, but also act as a continuous strut-tie link between the pylons, forming a key element for stability of the pylons in the longitudinal X-direction. In the transverse Y-direction the pylons will always perform as cantilevers, due to the small flexural stiffness of the gantries in that plane. The sliding connections between the pylons and the top end of each steel gantry allow limited free movement of the gantries. This means that, during operational condi- tions, the concrete pylons will be subjected to only vertical reactions from the steel Figure 4 Isometric view of typical steel gantry gantries. However, during strong winds and seismic activity, when the movement of the pylons in the longitudinal direction of the conveyor is larger than the gap provided in the sliding connection at the gantry supports, the sliding mechanism will lock, and the steel gantries with the pylons will perform as a frame system.

The gantry support sliding and locking mechanism Figure 5 shows the connection with the sliding mechanism between the pylons and steel gantries. The lower parts of each steel Pinned joint gantry, outlined in blue, are restrained from free movement and are fully pinned in all Sliding joint three global directions. The top parts of each steel gantry, outlined in red, have a sliding connection, but only in the global X-direction. The contact surfaces of the sliding connections, which have slotted holes, are made of Teflon. Taking into account that the static and kinetic fric- tion coefficients between Teflon surfaces Figure 5 Sliding mechanism at gantry supports are both 0.04, the friction forces can be neglected. This means that the top part of play the role of a strut or a tie, depending on structural analysis – the first one where the the steel gantry has a sliding support with a the loading condition. friction coefficient is equal to zero, and the finite sliding length. If the size of the move- Due to the specific environment in which second one where the friction coefficient is ment reaches the size of the allowable gap, the structure will operate, and keeping equal to one. This is justified considering the sliding mechanism will lock and neigh- in mind the strategic importance of the that the operational life and the maintenance bouring pylons will be linked. Therefore the structure, the design engineers decided to of the sliding connection cannot be explicitly steel gantry between these pylons will then consider two extreme conditions for the estimated.

46 Journal of the South African Institution of Civil Engineering • Volume 57 Number 3 September 2015 1991-1-4 (EN 1991). The effects of vortex- Vortex shedding induced force shedding do not need to be investigated if

the critical wind velocity Vcrit,i is larger than 1.25*Vm, where Vm is the 10-min mean wind velocity at the cross-section where vortex- shedding occurs. The formula for the critical wind velocity for vortex-shedding is:

Figure 6 Vortex-shedding phenomenon bni,y Vcrit,i = (3) St Table 1 Vortex-shedding results for the three tallest pylons where b is the reference width of the cross- Frequency b St V V Pylon Direction Mode b crit V /V (Hz) (m) number (m/s) (m/s) crit b section at which resonant vortex-shedding occurs, n is the natural frequency of the X Mode 1 0.33 5.8 0.06 20 31.90 1.60 i,y Pylon 1 considered flexural mode i of the cross-wind Y Mode 1 0.76 2 0.12 20 12.67 0.63 vibration and St is the Strouhal number. X Mode 1 0.38 5.8 0.06 20 36.73 1.84 To calculate the first natural frequency Pylon 2 of a uniformly loaded cantilever beam with a Y Mode 1 0.88 2 0.12 20 14.67 0.73 concentrated mass attached to the free end, X Mode 1 0.48 5.8 0.06 20 46.40 2.32 the following formula based on the Rayleigh Pylon 3 method can be used: Y Mode 1 1.1 2 0.12 20 18.33 0.92 1 k n1 = * (4) LOAD ANALYSIS gust factor approach, provided the wind flow 2π M + 0.23m is not significantly affected by the presence Operational load cases of neighbouring tall structures or the sur- 3EI where k = , l is the beam length, For the design of the structure during the rounding terrain. The method described in l3 normal operational stage, the following main the CICIND Model code for concrete chim- E is Young’s modulus of elasticity, I is the load cases were considered: neys (CICIND 2001) was used to calculate moment of inertia, m is the mass of the ■■ Self-weight load – main and secondary the along-wind effects. whole beam and M is the mass attached to structural elements, conveyors, material The mean hourly wind load at height z is: the free end. and water pipes ■ 2 ■ Imposed load – service and operating wm(z) = 0.5ρav(z) CDd(z) (1) The results for the first three pylons are personnel loads, live load from conveyors shown in Table 1. The locations where

and belt tensioning forces where ρa is the density of air, v(z) is the wind vortex-shedding is identified as a potential ■ ■ Wind load speed at height z, CD is a shape factor and problem are shown in red, implying that the ■■ Seismic load d(z) is the width of the pylon. three tallest concrete pylons may experience ■■ Loading due to temperature gradient. forces induced by vortex-shedding effects The wind load due to gusts was determined by: during the construction stage. Load cases analysed for the The pylons are designed for the opera- construction stage 3(G – 1) z h tional condition when all elements, including wg(z) = wm(z)zdz (2) During the construction stage the pylons will h2 h ∫0 the gantries, are in place. It was determined be freestanding and possibly be vulnerable to that it will not be economical to design strong winds. The SANS loading code (SANS where G is the gust factor, h is the height of all the elements for the temporary load 0160-1989) does not deal with loadings from the top of the structure above ground level conditions, but rather to provide additional the effects of gusts of wind and therefore two and z is the height above ground level. temporary supports during the construction additional internationally recognised and stage. It was thus decided to use normal accepted codes were consulted. These were Across-wind loading: Vortex- post-tensioned cables to support the pylons the CICIND 2001 Model code for concrete shedding dynamic action during construction, as shown in Figure 7. chimneys (CICIND 2001) and Eurocode Vortex-shedding occurs when wind airflow Four cables were provided on each side 1991-1-4 (EN 1991). The following two wind vortices are shed alternately from opposite of the first three pylons. Each cable was loading effects were considered: sides of the pylons, as shown in Figure 6. post-tensioned to 100 kN. It is estimated This is an additional loading of the structure that the maximum load in the cables during Along-wind loading acting in the transverse direction. This effect dynamic wind effects will increase to 150 kN The along-wind loading of a structure due can occur at relatively low wind speeds and per cable. The cables are anchored to adjacent to buffeting by wind can be assumed to gives rise to a fluctuating load perpendicular foundations and temporary stress blocks. consist of a basic component based on the to the wind direction. Structural vibrations Introduction of the cables changed the static mean hourly wind speed and a fluctuating may occur if the frequency of vortex-shed- system of the pylons from a pure cantilever component due to wind speed variations ding is close to any of the natural frequencies to that of a continuous beam with elastic sup- from the mean. The dynamic response of a of the pylons. port. Therefore, the natural frequencies of the structure in the along-wind direction can be The vortex-induced loading calcula- pylons changed and had to be recalculated. predicted with reasonable accuracy by the tions were done according to the Eurocode One elegant way to calculate the natural

Journal of the South African Institution of Civil Engineering • Volume 57 Number 3 September 2015 47 1.3 1.2 1.1 ) 2 1.0 0.9 0.8 0.7 0.6 Acceleration (m/s 0.5 0.4 0.3 0 1 2 3 4 5 Period (s)

Figure 7 Stayed pylon Figure 8 Normalised response spectrum (a = 0.05 g)

frequency is to use FE software. The problem Table 2 Summary of forces in the pylons from wind loads during construction was that the available FE software at that time Loading/Pylon Pylon 1 Pylon 2 Pylon 3 did not support modal analysis with geomet- ric non-linear cable elements. The easiest way Total axial load (SLS) on base (kN) 8 781 8 192 7 261 to address this problem was to replace the Along-wind moment My (SLS) on base (kNm) 16 010 13 878 10 085 cable element with an elastic support that had Vortex moment Mx (SLS) on base (kNm) 10 147 9 021 7 428 the same stiffness as the cable element. Table 2 contains a summary of forces in the pylons ULS axial load (unstayed) (kN) 8 781 8 192 7 261

due to the wind load during the construc- ULS bending moment My (unstayed) (kNm) 26 730 22 986 16 531 tion stage. It can be seen that the bending moments about the global Y-direction, caused ULS bending moment Mx (unstayed) (kNm) 16 235 14 434 11 885 by the wind acting in the global X-direction, Capacity factor < 1 < 1 1.14 are reduced by 30% with the implementation ULS axial load (stayed) (kN) 9 410 8 815 7 883 of the cables. The calculated axial force in the cable elements was smaller than the expected ULS bending moment My (stayed) (kNm) 17 917 15 853 11 618

maximum of 150 kN. ULS bending moment Mx (stayed) (kNm) 16 235 14 434 11 885 Seismic actions during construction Capacity factor 1 > 1 > 1 After locating the site on the seismic hazard SLS – Serviceability Limit State; ULS – Ultimate Limit State map from SANS 0160-1989 (SANS 1989), F(x, y or z) – force in X, Y or Z direction; M(x or y) – moment about X, Y or Z axis it was decided to use a ground peak accel- eration of 0.05 g. The normalised response spectrum is shown in Figure 8. The client are also vulnerable to possible seismic action Table 3 shows the results from the requested that the behaviour of the structure during the construction stage, especially seismic analyses for the first three and for should remain elastic during seismic activity in the longitudinal direction of the con- the seventh pylon. The complete quadratic and therefore the seismic behaviour factor veyor before the steel gantries are in their combination (CQC) method of modal com- used in this case is equal to one. The pylons final position. bination was used. It is important to have

Table 3 Seismic analysis results for Pylons 1, 2, 3 and 7

Model Beams Shells

Direction X Y X Y

Fx Fy Fz My Mx Fx Fy Fz My Mx Column/Cases/Info (kN) (kN) (kN) (kNm) (kNm) (kN) (kN) (kN) (kNm) (kNm)

1.0*DLL+1.0*S(X+) 319 0 9 204 10 988 0 315 0 9 039 10 787 0 Pylon 1 1.0*DLL+1.0*S(Y+) 0 445 9 204 0 17 159 0 435 9 039 0 17 135

1.0*DLL+1.0*S(X+) 308 0 8 585 9 611 0 306 0 8 419 9 405 0 Pylon 2 1.0*DLL+1.0*S(Y+) 0 434 8 585 0 16 239 0 428 8 419 0 16 325

1.0*DLL+1.0*S(X+) 292 0 7 611 8 291 0 282 0 7 776 8 359 0 Pylon 3 1.0*DLL+1.0*S(Y+) 0 438 7 611 0 14 921 0 422 7 776 0 14 739

1.0*DLL+1.0*S(X+) 445 0 4 677 6 392 0 462 0 4 919 6 899 0 Pylon 7 1.0*DLL+1.0*S(Y+) 0 474 4 677 0 6 596 0 496 4 902 0 7 107

DLL – dead load; S(X± or Y±) – seismic load in ± X or ±Y direction; F(x, y or z) – force in X, Y or Z direction; M(x, y or z) – moment about X, Y or Z axis

48 Journal of the South African Institution of Civil Engineering • Volume 57 Number 3 September 2015 Table 4 Pylon 1 forces and seismic capacity From a comparison of the bending moments, due to the wind and seismic load Loading/Pylon Pylon 1 (1.6*SX+0.48*SY) Pylon 1 (0.48*SX+1.6*SY) analyses, it can be seen that the forces due to Total axial load (SLS) on base (kN) 8 781 8 781 the wind loading are much higher than those Seismic moment My (SLS) on base (kNm) 10 988 10 988 due to the seismic loading. This means that, during the construction stage, for this type Seismic moment Mx (SLS) on base (kNm) 17 159 17 159 of structure and ground peak acceleration of ULS axial load (kN) 8 781 8 781 0.05 g, the wind loading is most probably the

ULS bending moment My (kNm) 17 596 5 274 critical loading.

ULS bending moment Mx (kNm) 8 236 27 454

Capacity factor > 1 > 1 FINITE ELEMENT MODELLING

S(X or Y) – seismic load in X or Y direction; M(x or y) – moment about X or Y axis Modelling of the pylons and the gantries Table 5 Periods of oscillation the period of oscillation of the cantilever In order to investigate the behaviour of the beam with a point mass at the free end. The structure during seismic activity, finite ele- Model Beams results are identical. The modal analysis for ment models (FEM) with different levels of Direction X Y seismic loadings was done without modelling complexity were created using the Autodesk Pylon T (s) T (s) of the cable elements. The reason behind this Robot Structural Analysis Professional was that the analyst could not be sure that package (Autodesk). The sliding mechanism Pylon 1 3.05 1.3 the temporary stress blocks, used to anchor at the gantry support gaps was not incorpo- Pylon 2 2.66 1.14 the cables, would remain in position during rated in the following FEMs: ■ Pylon 3 2.1 0.91 seismic activity. Therefore, the influence of ■ Model A: FEM with only beam elements. the cables was ignored. The pylons and the steel gantries are Pylon 7 0.23 0.19 From Table 3 it can be seen that for the modelled with single beam elements. This first three (tallest) pylons, the moments model allows a quick overview of all types over 90% of the modal mass participating about the global X-axis are much higher than of results (Figure 9). to model the dynamic response realistically. those around the global Y-axis. This is to ■■ Model B: FEM in which the pylons are The results from two different models, the be expected, because the moment of inertia modelled with shell elements and the model in which the pylons are modelled with around the global X-axis is much higher. steel gantries are modelled with single beam elements, and the model in which the Therefore, the section is stiffer in that plane beam elements. This model provides pylons are modelled with shell elements, are and the period of oscillation is shorter, as is localised results for the pylons. Shear lag compared. The results of the two models shown in Table 5. As a result of the shorter effects in the pylons can be quantified. differ by from less than 1% up to 8%. Table 4 period of oscillation, the pseudo-acceleration ■■ Model C: FEM in which the pylons are shows the ultimate forces for the design of is higher and the lateral force induced by modelled with single beam elements and Pylon 1. inertia is larger, causing higher moments at the steel gantries are modelled as 3D In order to check the FE analysis results, the base. For the seventh pylon, the periods space trusses, with full complexity. This additional hand calculations were performed. of oscillation are almost identical, which model shows the effects of the 3D truss The same formula based on the Rayleigh is reflected by the very close values of the system. Because the steel gantries cannot method (Equation 4) was used to calculate bending moments at the base. displace or expand in the longitudinal

ZA1 ZB1 ZC ZD ZE ZF ZG ZH ZJ ZK ZL ZM

+58.22

–5.70

Figure 9 FE model with single-beam elements

Journal of the South African Institution of Civil Engineering • Volume 57 Number 3 September 2015 49 direction for this FEM, horizontal forces from the gantries are transferred to the pylons. ■■ Model D: FEM with full 3D modelling. The pylons are modelled with shell ele- ments and the steel gantries are modelled as space trusses. This is the most complex FE model (Figure 10). ■■ Model E: FEM with only one pylon, the tallest one. The rest of the structure is represented by a single spring support which has stiffness identical to the replaced parts of the structure. The results from this very simplified model differed by less than 10% from the results from the more complex models. This type Figure 10 Detail from FE model with shells and 3D gantries

Table 6 Modal results from four different FE models

Model C (3D truss + beams) Model Model A (beams) Model B (shells + beams) Model D (full 3D model) Gantries + Beams

Direction X Y X Y X Y X Y

Modes/ T M T M T M T M T M T M T M T M Info (s) (%) (s) (%) (s) (%) (s) (%) (s) (%) (s) (%) (s) (%) (s) (%)

Mode 1 1.52 52.91 1.27 36.42 1.58 53.93 1.34 37.14 1.56 54.63 1.4 35.38 1.51 54.83 1.33 34.23

Mode 2 0.65 54.13 0.84 42.55 0.69 54.73 0.87 43.08 0.69 55.09 0.98 40.7 0.65 55.66 0.94 39.84

Mode 3 0.56 56.42 0.6 50.51 0.6 56.2 0.62 51.38 0.6 55.91 0.76 49.29 0.56 57.18 0.73 48.68

Mode 4 0.48 69.54 0.45 52.56 0.51 69.16 0.46 51.44 0.51 69.75 0.62 50.72 0.48 69.91 0.6 50.73

T – period of oscillation; M – mass participation

Table 7 Seismic results from four different FE models

Model Model A (beams) Model B (shells + beams)

Fx Fy Fz Mx My Fx Fy Fz M My Pylon/Cases/Info (kN) (kN) (kN) (kNm) (kNm) (kN) (kN) (kN) (kNm) (kNm)

1.0*DLL+1.0*S(X+) 320 0 11 185 0 6 308 305 0 11 339 –22 6 240

1.0*DLL+1.0*S(X-) –321 0 11 065 0 –6 371 –305 0 11 218 –24 –6 242

S(X+) 321 0 60 0 6 340 305 0 60 1 6 241

S(X-) –321 0 –60 0 –6 340 –305 0 –60 –1 –6 241 Pylon 1 1.0*DLL+1.0*S(Y+) –1 444 11 125 15 088 32 0 430 11 278 14 807 33

1.0*DLL+1.0*S(Y-) –1 –444 11 125 –15 088 32 0 –430 11 278 –14 807 33

S(Y+) 0 444 0 15 088 0 0 430 0 14 807 0

S(Y-) 0 –444 0 –15 088 0 0 –430 0 –14 807 0

Model Model D (full 3D model) Model C (3D truss + beams)

Fx Fy Fz Mx My Fx Fy Fz Mx My Pylon/Cases/Info (kN) (kN) (kN) (kNm) (kNm) (kN) (kN) (kN) (kNm) (kNm)

1.0*VL+1.0*S(X+) 354 30 11 441 –612 9 381 365 0 11 000 0 9 627

1.0*VL+1.0*S(X-) –253 29 11 321 –619 –2 842 –257 0 11 115 0 –2 793

S(X+) 303 0 60 3 6 112 311 0 58 0 6 210

S(X-) –303 0 –60 –3 –6 112 –311 0 –58 0 –6 210 Pylon 1 1.0*VL+1.0*S(Y+) 33 313 11 163 16 448 3 195 54 418 11 057 17 169 3 417

1.0*VL+1.0*S(Y-) 32 –351 11 157 –16 416 3 193 54 –418 11 057 –17 169 3 417

S(Y+) 0 332 3 16 432 1 0 418 0 17 169 0

S(Y-) 0 –332 –3 –16 432 –1 0 –418 0 –17 169 0

VL – vertical load; S(X± or Y±) – seismic load in ± X- or ±Y direction; F(x, y or z) – force in X, Y or Z direction; M(x or y) – moment about X or Y axis

50 Journal of the South African Institution of Civil Engineering • Volume 57 Number 3 September 2015 the thesis that for further analyses only two FE models, i.e. the model with single beam elements (Model A) and the model with sin- gle beam elements and 3D gantries (Model C) need to be used.

Modelling of soil‑structure interaction To simulate realistic soil-structure interac- tion, the bases with elastic planar supports Fx = –18.68 were modelled. One of the advantages of models with planar supports is that the designer and the analyst can quickly deter- Fx = 71.54 mine whether some parts of the bases are Fx subject to uplifting and, if that is the case, they can perform a non-linear analysis. By performing a non-linear analysis with con- Mz tact elements between the bases and the soil, Fx it is possible to model the no-tension condi- tion between the soil and the bases. Figure 11 3D truss effects In all FE models presented the elastic stiff- ness coefficient for the discretisation model of 3 the soil kv = 200 000 kN/m was used.

Implications of simplified FE models For the FEM with single beam elements it is crucial to calculate the correct axial and flexural stiffness of the beams that represent the steel gantries. If the stiffness of the beams is not correct, the overall behaviour of the whole structure and the force distribu- 2.49 tion between the pylons will not be the same 1.70 0.85 as in the full 3D FE model. To model the 0 –0.85 effect of the space truss system correctly, –1.70 –2.55 rotation of the beams in the model around –3.40 –4.25 the global Z-axis must not be allowed. As –5.10 –5.95 shown in Figure 11, the steel gantry acts as a –6.80 –7.39 sXX, (MPa) fixed beam and can attract moment around Automatic direction the global Z-direction. This moment (Mz) Cases: 5 (10 000 kN + 11 500 kNm) can be replaced by force couple (Fx), as Figure 12 Shear lag effects shown in Figure 11. The advantages and disadvantages of of FE model can be used for preliminary different magnitudes of the moments about using a simplified FE model with only beam

design and sizing of the concrete pylons. the Y-axis (My) for the load combinations of elements are: It can, however, not be used for structures the models with gantries represented by one ■■ Computing time is faster, inputting is with sliding connections. single beam element (Models A and B), and easier, changes can be made to the model the models with full 3D gantries (Models C and the results can be checked more Selection of FE models based and D). It can be seen that the results for the rapidly. Most FE packages allow easy and on comparison of results from seismic actions in the longitudinal direction individual review of all beam element the different FE models (X) are almost identical. The results for load axial, bending, shear and torsion section The modal analysis results, the vibration combinations (VL + SX) differ between the forces. period (in seconds) and the mass participa- models in which the gantries are represented ■■ For shell elements only total resultant tion factors (in %) for the first four modes, by a single beam element (Models A and B), stress can be reviewed. It is not easy to from the above-mentioned FE models with and those in which the gantries are modelled determine what portion of the total stress elastic supports, are shown in Table 6. The as 3D trusses (Models C and D). The dif- can be attributed to bending or axial X-direction is in line with the longitudi- ferences in the tabulated results are caused stresses. Modern versions of FE packages nal direction of the conveyor, while the by the load distribution effects of the 3D provide a solution by which different Y-direction is perpendicular to it. The results trusses. The gantries transfer the horizontal shell elements can be grouped together from the modal analyses from four different reaction to the pylons. These effects should as a single element. The software is then models are very close, supporting the use of not be omitted for modelling purposes. By able to compute and present the overall simplified models for this type of analysis. studying the results for the seismic response force results for the whole element – in Table 7 contains the results for the first in the Y-direction it can be seen that they this particular case the whole rectangular pylon from the seismic analyses. These show vary by up to 10%. This exercise supported hollow section of the pylons.

Journal of the South African Institution of Civil Engineering • Volume 57 Number 3 September 2015 51 ■ ■ An FE model with beam elements can be 4.0e+04 used to compute global section results. 3.0e+04 Then analysis with another FE model 2.0e+04 with shell elements and a fine mesh can 1.0e+04 be performed, and the results from the previous model can serve as an input. 0.0e+00 This micro-modelling allows checking of –1.0e04 Force (kN) Force possible shear lag effects in stiff portions –2.0e04 of the rectangular hollow section of the –3.0e04 pylons (Figure 12). –4.0e04 –40 –30 –20 –10 0 10 20 30 40 Modelling of the sliding Displacement (mm) connection at the gantry supports using gap elements Figure 13 Non-linear model of gap element Analyses with gap elements were done to model the realistic behaviour of the structure FX+ for Diana which has special locking devices at the gantry +0.00000e+000 17.6% supports. This means that the first pylon will –2.13174e+000 10.9% deform before the allowable gap is reached and –4.26349e+000 9.5% it will then link with the second pylon. This –6.39523e+000 8.6% new system must then deform until it links –8.52698e+000 7.8% with the third pylon and so on. In the Autodesk –1.06587e+001 7.3% Robot Professional (Autodesk) structural analy- –1.27905e+001 6.8% sis program there is an option to model gap –1.49222e+001 6.6% elements directly and this can be used for static –1.70540e+001 6.2% analyses (Figure 13). However, these elements –1.91857e+001 5.7% cannot be used to perform dynamic modal –2.13174e+001 5.1% analyses because the natural frequency solver –2.34492e+001 3.4% allows only one stiffness value to be used at the –2.55809e+001 2.4% connection. The modelling of the non-linear –2.77127e+001 1.3% behaviour where the beam has the freedom –2.98444e+001 0.7% to move along the slot (very low stiffness) and –3.19762e+001 0.4% then reach the gap end (very high stiffness) is –3.41079e+001 not possible in a modal analysis. To overcome this problem, short ‘soft’ beam elements, with Figure 14: FEM of the pylon from TNO-Diana the same length as the gap, were used to model the physical gap elements. In order to model Table 8 Control calculations for the gap elements the correct stiffness of these beams, an iterative Column 1 Column 2 Column 3 process was performed. The method used was Step to create a model with automatic gap elements u (mm) M (kNm) u (mm) M (kNm) u (mm) M (kNm) and then to create another separate model with Step 2 20 21 305 n/a n/a n/a n/a soft beams. A number of iterations with dif- Step 3 21.3 22 703 20 21 330 n/a n/a ferent axial stiffnesses of the soft beams were performed until the deformation and forces in Step 4 6 6 368 25.3 5 662 5 5 320 the two models were identical. Sum 47.3 50 376 45.3 26 992 5 5 320 For FEM software packages that do not support the concept of gap elements the fol- lowing analysis can be performed to model top of the first pylon. Run the analysis frames with three columns were created. the behaviour of a frame with three pylons. and record the deflections and forces Table 8 shows the results for the analytical ■■ Step 1: Decide what is the allowable slid- in the pylons. The total deflections and hand calculation of the frame with gaps. ing length (say a). forces of the pylons will be equal to the Table 9 shows the results for the three different ■■ Step 2: Model only one pylon and apply sum of the results from all three models. frames. The first row presents results for the force P at the top of it, which will cause ■■ Step 5: Create a model with three pylons frame with automatic gap elements, the second deflection of the pylon equal to a. Record and the beams between them. Divide one is for the frame with soft beams (hand gap) the forces in the pylon. each beam into two beams and make one and the third one is for the frame with no gap ■■ Step 3: Model two pylons linked by a of them the same length as a. Apply a elements. The results for the frames with gap beam with the same stiffness as the steel force equal to the sum of the forces P, Q elements are almost identical with those for gantry and apply force Q at the top of the and R, and adjust the stiffness of the short the analytical hand calculations. The differ- first pylon, which will cause deflection of beams (length equal to a) until the results ence between the moments in the columns for the second pylon equal to a. Record the are identical with those from the model frames with and without gap elements is obvi- forces in the pylons. with gap elements. ous. The moment in the first column increases, ■■ Step 4: Model three pylons with beams A similar process has to be applied for larger while the moment in the last column decreases between them and apply force R at the frames. To test this approach a number of when using gap elements.

52 Journal of the South African Institution of Civil Engineering • Volume 57 Number 3 September 2015 Table 9 Results from the testing of the gap elements moments to model and predict the behaviour of a cracked reinforced concrete section. As Column 1 Column 2 Column 3 Model expected, the deflections of the cracked sec- u (mm) M (kNm) u (mm) M (kNm) u (mm) M (kNm) tions are between four and five times bigger than those of the uncracked sections. To pre- Auto gap 47.3 50 388 25.3 26 984 5 5 316 dict the behaviour of the structure with cracked Hand gap 47.3 50 436 25.3 26 965 5 5 287 sections, and using for the pylons an estimated effective stiffness of 22.5% of the total elastic No gap 28.5 30 351 25.3 26 984 23.8 25 353 stiffness, another seismic analysis was done, and these results are shown in Table 11. As Table 10 Summary of results for analyses of section cracking expected, the oscillation periods were longer. By comparing the values of the moments Forces/Model Linear Non-linear u, lin/u, nlin Case for the first two pylons, it can be seen that the Axial (kN) Moment (kNm) u, lin (mm) u, nlin (mm) (%) moments from the analysis with reduced stiff- Case 1 10 000 5 000 3.15 13.12 24 ness are about 30% smaller. This is because the oscillation period is longer and therefore Case 2 10 000 8 000 5.03 21.79 23 the applied lateral seismic force is smaller. Case 3 10 000 10 000 6.3 28.74 22 The pylons with reduced stiffness will deform more and therefore close the sliding gap Case 4 10 700 11 500 7.23 34.11 21 ‘faster’ than the pylons with full stiffness. If behaviour factors are introduced to In the analysis of the ICC behaviour with seen that the different groups of pylons estimate the inelastic deformation in the FE models, replacement of the automatic gap oscillate with different periods). concrete pylons, the same behaviour factor elements with soft beams was done using the In order to compare the results from Robot, should not be used for all pylons because: following steps: an identical FE model of the structure with ■■ Due to the varying heights, the pylons ■■ Step 1: Create a version of Model A, A1, soft beams was created in the Strand7 FEM do not have equal stiffness and will not an FE model with beam elements, imple- package (Strand7 Software). The results from deform identically. ment automatic gap elements and apply the modal analyses were identical. ■■ The pylons are not linked rigidly due to the seismic load as equivalent static load incorporation of the sliding mechanism. (assume some period of oscillation). Modelling of the behaviour of the ■■ The special stiff seventh pylon forms ■■ Step 2: Create a version of Model A, A2, cracked concrete pylon box section part of the structure but, because of the an FE model with beam elements and soft To calculate the effective stiffness of a cracked sliding mechanism, it takes some time to beams. Apply seismic load as in Step 1. hollow box pylon section, micro material start playing its role. Adjust the stiffness of the soft beams to modelling was performed with the Diana FE ■■ The oscillation periods are long, and have more or less the same deformations software (TNO Diana). Two models were implementation of the behaviour factor and forces in the elements as in the model created, one with linear and one with non- will not have a significant influence. with automatic gap elements. linear materials (Figure 14). Both materials, ■■ Step 3: Run modal and seismic analyses i.e. the concrete and the steel reinforcement, with the FE model with soft beams. were modelled using the material non-linear DISCUSSION OF ANALYSIS RESULTS Compare the assumed periods of oscilla- formulation. To estimate the net effective tion with the computed ones. If they are stiffness of the cracked sections, deflections Loading cases for different, repeat the process from Step 1 from these two models were compared, and operational conditions onwards with new periods of oscillation the results are presented in Table 10. (using different periods of oscillation As shown in Table 10, four different load Wind actions for each pylon, because by studying the cases were analysed. An almost identical axial For the wind loading analyses during the results from the modal analyses it can be force is combined with different bending operational stage, along-wind forces similar

Table 11 Forces in the pylons, with full and reduced E modulus, for seismic action

Model A (beams) Model A (beams) Model A (beams) Moment Model Full E Reduced E Full E Reduced E E ratio

Mode T (s) M (%) T (s) M (%) Pylon/Cases/Info My (kNm) My (kNm) (%)

Mode 1 3.0 36.0 4.9 42.8 Pylon 1 Seismic X 12 322 8 301 67

Mode 2 1.7 44.6 2.3 47.1 Pylon 2 Seismic X 11 144 8 159 73

Mode 3 1.3 47.0 1.6 50.7 Pylon 3 Seismic X 9 520 8 082 85

Mode 4 1.1 48.3 1.5 51.4 Pylon 4 Seismic X 7 465 8 446 113

Mode 5 1.0 52.7 1.3 51.4 Pylon 5 Seismic X 6 235 8 608 138

Mode 6 1.0 56.9 1.0 51.6 Pylon 6 Seismic X 5 902 3 913 66

Mode 7 0.9 56.9 1.0 55.9 Pylon 7 Seismic X 9 580 9 485 99

T – Period of oscillation; M – mass participation; Mx – moment about X axis; My – moment about Y axis

Journal of the South African Institution of Civil Engineering • Volume 57 Number 3 September 2015 53 to those in the construction stage are applied Table 12 Forces in the pylons from wind analyses during operational conditions to the pylons. Two FE models were analysed: Model Model A1 (beams) Model B1 (3D truss + beams) ■■ Model A1, the FE model with single beam elements and automatic gap elements Direction X X ■■ Model B1, the FE model with 3D trusses Fz Fx My Fz Fx My Pylon/Cases/Info and beams, and with automatic gap (kN) (kN) (kNm) (kN) (kN) (kNm) elements. Pylon 1 Wind X 30 599 15 903 30 596 15 653 A summary of the results for the wind actions in the X-direction is shown in Table 12. Pylon 2 Wind X 57 568 14 770 60 564 14 483

Seismic actions Pylon 3 Wind X 23 538 13 854 25 534 13 570 The parameters used for the seismic analyses Pylon 7 Wind X 44 374 5 679 52 405 6 325 during the operational life of the structure F(x or z) – force in X or Z direction; My – moment about Y axis are identical to those used for the analyses during the construction stage. The accelera- tion a = 0.05 g and the behaviour factor is Table 13 Forces in the pylons from seismic action in the X-direction equal to one. Two FE models were consid- Model Model A2 (beams + soft beams) Model A1 (beams + gap) ered and analysed. Model A2, the model with beam elements and soft beams, was analysed Direction X X first. From the modal analyses, the natural Fz Fx My Fz Fx My Pylon/Cases/Info frequencies of each pylon were determined. (kN) (kN) (kNm) (kN) (kN) (kNm) Using this information, the equivalent static Pylon 1 Seismic X 26 350 12 322 11 329 11 522 lateral forces for each pylon were calculated and applied in the FE beam model with Pylon 2 Seismic X 31 322 11 144 21 322 10 825 automatic gap elements, Model A1. A sum- Pylon 3 Seismic X 25 308 9 520 6 324 10 454 mary of the results for seismic action in the X-direction is given in Table 13. Some of the Pylon 7 Seismic X 9 592 9 580 13 653 8 944 results differ by up to 10% because, in order F(x or z) – force in X or Z direction; My – moment about Y axis to calculate equivalent static lateral forces by hand, it is assumed that the whole mass of each pylon is subjected to the same pseudo- Table 14 Forces in the pylons from seismic action in the Y-direction acceleration. This FE model has a more Model Model A1 (beams + soft beams) representative mass distribution, leading to more accurate results. Direction Y

To investigate the effects of seismic action Pylon/Cases/Info Fy (kN) Mz (kNm) My (kNm) Mx (kNm) in the Y-direction, analysis was done with all previously mentioned models and the Pylon 1 Seismic Y 444.49 1 615.6 –0.01 15 088 results were almost identical. This is because Pylon 2 Seismic Y 511.71 754.2 –0.01 20 619 the flexural stiffness of the gantries is much Pylon 3 Seismic Y 488.7 1 125.2 –0.02 18 574 smaller than that of the pylons, and the gantries do not restrain the pylons. In this Pylon 7 Seismic Y 660.9 1 195.1 –0.24 11 335 case even the model with the automatic gap Fy – force in Y direction; M(x,y or z) – moment about X, Y or Z axis elements, Model A2, gave very good results, because the pylons on their own act like pure vertical cantilevers during modal analyses. A the gantries are effectively tying the pylons stiff pylon. On the other hand, in the system summary of the results of only one FE model and transferring the force to the stiffest with sliding connections, the shear force is shown in Table 14. seventh pylon. induced by seismic action will be distributed to each pylon before the sliding mechanism Comparison of results from the FE FE models with automatic gap elements is locked. The force distribution depends on models with and without a gap To perform a more direct comparison, the the stiffness of each pylon. results of the modal analysis from the FEM FE models with soft beams with soft beams are used and the same Load combinations The modelling of the sliding connection equivalent lateral static load is applied to the For concrete structures it is common prac- behaviour was a major focus of this work. To FE model with automatic gap elements and tice to perform only linear static analyses, highlight the diversity of results between the the one without gap elements. The results and then linearly combine the resulting for­ FE models with and without gaps, Tables 15 are presented in Table 16. It can be seen that ces and moments of individual load cases in and 16 are presented. Table 15 compares the the moment in the first pylon is 2.5 times load combinations. In the case of models of spectral analysis results for the FE models smaller, while the moment in the seventh structures with gap elements, this approach with and without soft beams. Even though pylon is 2.25 times larger. The difference does not apply, because the behaviour of the period of oscillation in the FEM without between the force distributions is a result of sliding connections is non-linear. The effect a gap is shorter and therefore the applied the presence of the sliding joint. The shear of each load case is analysed in sequence lateral load is larger, the bending moment in force from seismic loading in the system and each load combination must be analysed the first pylon is much smaller. In this case without sliding connections will travel to the individually.

54 Journal of the South African Institution of Civil Engineering • Volume 57 Number 3 September 2015 Table 15 Comparison between models with and without gap elements (different modal analyses) response spectrum analyses using standard FE software. Model A2 Model A Model (beams + soft beams) (beams, no gap) Moment It was shown that the overall behaviour ratio of the structure can be highly influenced by Direction X X the action of the sliding mechanism and that Fz Fx My Fz Fx My Column/Cases/Info (%) the force distribution between the structural (kN) (kN) (kNm) (kN) (kN) (kNm) members can differ significantly. Pylon 1 Seismic X 26 350 12 322 60 321 6 340 51 It is recommended that, if possible, the automatic gap elements that are Pylon 7 Seismic X 9 592 9 580 178 860 15 870 166 part of FEM software should be used for

F(x or z) – force in X or Z direction; My – moment about Y axis static loading. Because these automatic gap elements are non-linear, each load combination must be analysed separately. Table 16 Comparison between models with and without gap elements (same lateral load is applied) For dynamic analyses, such as modal and response spectrum analyses, it is recom- Model Model A1 (beams + gap) Model A (beams, no gap) Bending mended that soft beams be used to calculate ratio Direction X X the periods of oscillation and force distribu- Fz Fx My Fz Fx My tion. Additional hand calculations using Column/Cases/Info (%) (kN) (kN) (kNm) (kN) (kN) (kNm) the equivalent lateral force method can Pylon 1 Seismic X 11 329 11 522 27 219 4 620 40 be performed to compare the results from the response spectrum analysis with the Pylon 7 Seismic X 13 653 8 944 159 1 238 20 110 225 FEM results.

F(x or z) – force in X or Z direction; My – moment about Y axis Effects of stiffness reduction due to cracking of the pylon box section CONCLUSION box sections, additional FE models with The effects of reduced stiffness of concrete The extensive structural analysis with FEM shell elements should be used to perform sections due to cracking were analysed. The support used as input in the design for the micro-modelling in order to obtain realistic reduction in stiffness makes the oscillation Medupi Power Station inclined coal conveyor shear lag effects and localised stress concen- periods longer and as a result the computed support structures has been outlined in trations due to the stiff corner parts of the lateral seismic loading is smaller. The reduc- this paper. box sections. tion in bending moments in the tallest Important conclusions on the use of column was about 30%. FEMs with varying complexity, considera- Construction stage loading tions for the construction stage loading, the The loading and behaviour for both the modelling of the sliding connection in the construction stage and the operational ACKNOWLEDGEMENTS structure and the modelling of the stiffness stage were considered. The concrete pylons, The permission granted by Eskom Central reduction due to cracked concrete sections being tall and slender structures, should Power Generation Office to publish this were reached. be checked for possible vortex-shedding paper is gratefully acknowledged. effects and it was shown that the additional The authors of the paper would like Finite element modelling of the forces induced by these effects could not to specially thank Prof Johan Retief inclined coal conveyor structure be neglected. The wind loading is the most from the University of Stellenbosch and Regarding the FE modelling and comparison critical loading factor during the construc- Mr Hein Barnard, Principle Specialist from between different FEMs, it is shown that the tion stage. If the ground acceleration is AECOM SA, for their support and valuable simplified beam element models provide ade- higher than 0.5 g, then seismic loading contribution. quate modelling of the structural behaviour could become the most critical loading for this kind of structure. Simplified models condition. using beam elements instead of 3D models LIST OF SYMBOLS Modelling of the gantry with shell elements for box sections must be wm(z) Mean hourly wind load sliding connection approached with caution, however, because ρa Air density of localised stress concentrations. For the The impact of detailed aspects of the v(z) Wind speed at height z modelling of the steel gantries, the effects structural system, such as the sliding and CD Shape factor of 3D space trusses have to be considered. locking mechanism linking the gantries to d(z) Width of the section

The simplest model with only one pylon and the concrete pylons, implied that special wg(z) Wind load due to gust spring support representing the rest of the modelling techniques were required. The G Gust factor structure can be used for preliminary design. non-linear mechanical aspects of the h Height of the top of the structure The differences in the results obtained are behaviour of the structure for both static Fx Force in global X direction relatively small. and dynamic loading conditions also needed Fy Force in global Y direction For the final analysis of the overall behav- to be dealt with. Using an iterative process, Fz Force in global Z direction iour of the structure, it is recommended to the sliding mechanism modelled with the ICC Inclined coal conveyor use the FE model in which the pylons are gap elements can be approximated by ‘soft’ S(X) Seismic load in global X direction modelled with beam elements and the gan- beams, which are linear elements. With S(Y) Seismic load in global Y direction tries are modelled as 3D trusses, i.e. Model implementation of soft beams the analyst ULS Ultimate limit state C. For the final design of the concrete pylon was then able to perform the modal and FEM Finite element model

Journal of the South African Institution of Civil Engineering • Volume 57 Number 3 September 2015 55 FEA Finite element analysis My Moment about global Y axis EN 1991. EN-1991-1-4. Eurocode 1: Actions on

Vcrit,i Critical wind velocity for Mz Moment about global Z axis Structures. Part 1–4: General Actions – Wind vortex-shedding SLS Serviceability limit state Conditions. Brussels: European Committee for

Vb Basic wind velocity DLL Dead load Standardization (CEN). a Allowable sliding length VL Vertical load SANS 1989. SANS 0160-1989. The General Procedure b Reference width of the cross section and Loading to be adopted in the Design of Buildings. n Natural frequency Pretoria: South African Bureau of Standards. St Strouhal number REFERENCES Strand7 2015. Finite Element Analysis Software. E Young’s modulus of elasticity Autodesk 2015. Autodesk Robot Structural Analysis Available at: http://www.strand7.com. I Moment of inertia Professional. Available at: http://www.autodesk.com TNO Diana 2015. Finite Element Analysis Software. Delft, L Length /products/robot-structural-analysis. Netherlands. Available at: http://www.tnodiana.com. z Height above ground level CICIND (International Committee on Industrial m Mass of a beam Chimneys) 2001. Model Code for Concrete Chimneys. Mx Moment about global X axis Part A. Ratingen, Germany: CICND.

56 Journal of the South African Institution of Civil Engineering • Volume 57 Number 3 September 2015 Investigating the bottom TECHNICAL PAPER Journal of the South African free surface nappe Institution of Civil Engineering (ogee profile) across a Vol 57 No 3, September 2015, Pages 57–63, Paper 980

sharp‑crested weir caused by PROF FANIE VAN VUUREN Pr Eng is a professor in the Department of Civil Engineering at the the flow in an asymmetrical University of Pretoria, and a project leader for a number of Water Research Commission projects. He has 39 years of experience in water approach channel resources engineering and holds BSc Eng, BSc Hons Eng, MEng, PhD (Eng) and MBA degrees, all from the University of Pretoria. He has worked as a specialist consultant for various consulting engineering S J van Vuuren, G L Coetzee, C P R Roberts companies, has written numerous technical reports, journal publications and chapters in books, and has presented lectures in many countries.

Contact details: The shape of an ogee spillway is based on the shape of the lower nappe of water flowing over Department of Civil Engineering an aerated sharp-crested weir. At the design discharge, this shape minimises the possibility University of Pretoria of sub-atmospheric pressure occurring on the spillway and maximises the discharge across Pretoria, 0001 the spillway. The formulae that are currently in use to approximate the ogee profile consider South Africa T: +27 (0)12 420 2438 only two-dimensional flow parameters, being the depth of flow over the spillway crest, the E: [email protected] inclination of the upstream wall face, and the pool depth upstream of the spillway. The current formulae for the ogee shape, does not consider the influence of three-dimensional flow. The LOUIS COETZEE obtained his BEng (Civil), BEng most significant three-dimensional flow parameters that could affect the shape of the lower (Hons) and MEng (cum laude) in Water Resources nappe are the flow velocity distribution upstream of the spillway, the orientation or angle of Engineering at the University of Pretoria, and is the water approaching the spillway, the asymmetrical cross-section of the approach channel, currently studying towards a PhD. He joined the and the curvature of the dam wall. This paper reflects the influence of asymmetrical flow in academic setting of the University of Pretoria as a junior lecturer in 2011 after completing his the approach channel. The investigation was based on a physical model constructed at the undergraduate studies. He obtained his early Department of Water and Sanitation (DWS). The inclination of the model’s sidewalls of the professional experience working for Sinotech upstream approach channel was varied to cause a change in the symmetricity, while the CC and is currently employed as an Engineer at SMEC South Africa (Pty) lower nappe profile was routinely measured. It was found that the flow in the asymmetrical Ltd. He has presented various courses on flood determination, flood lines, approach channel caused a variation from the theoretical estimated ogee profile. A comparison drainage structures, stormwater modelling, pipelines, pump stations and drainage systems in conjunction with Prof SJ van Vuuren. He has a special between the measured nappe profile and the currently used formulae was investigated. It interest in numerical and physical modelling of hydraulic structures and can be concluded that the symmetricity of the approach channel influences the shape of water distribution networks. the bottom nappe, which differs from the shape as proposed by the current ogee formulae. Contact details: It is recommended that three-dimensional flow should be examined when designing an Department of Civil Engineering ogee spillway. University of Pretoria Pretoria, 0001 South Africa List of symbols T: +27 (0)12 481 3975 Sf symmetricity factor of the approach E: [email protected] Al approach channel cross-sectional area channel measured from the centreline of the vo mean velocity in the approach channel DR PAUL ROBERTS Pr Eng, a civil engineer, 2 spillway’s crest to the left bank (m ) (m/s) retired at the end of March 2003 from the then At total approach channel cross-sectional South African Department of Water Affairs and area (m2) Forestry after a career of some 42 years in water C’ vertical displacement of the turning INTRODUCTION resources management. Since retirement he has been active in consulting work in South Africa point measured at the lower nappe of the The ogee spillway relationship (USBR 1987; and Africa. He has also served as the SANCOLD ogee profile (m) Hager 1987) is used to define the geometric (South African National Committee on Large f’ horizontal displacement of the turning profile of the spillway section of a dam or Dams) secretary since 2008.

point measured on the lower nappe of hydraulic structure. The ogee relationship Contact details: the ogee profile (m) describes the bottom nappe associated with SANCOLD Ha velocity head (m) the flow over a sharp-crested weir. The PO Box 3404 Pretoria, 0001 Hd measured water depth upstream from current relationship accommodates the South Africa the crest of the sharp-crested weir, influence of the unit discharge, the angle of T: +27 (0)12 460 9100 ­relative to the crest (m) inclination of the upstream wall face, as well E: [email protected] He total head (sum of Hd and Ha) (m) as the relationship of upstream pool depth Ho design head of ogee spillway (He - C’) (m) to the total upstream energy at the apex of P upstream wall depth of the weir (m) the structure.

Tp turning point (apex position) on lower In cases where the discharge flow rate Keywords: asymmetrical, cavitation, nappe, ogee profile, physical model, nappe of the ogee profile exceeds the design flow rate, the nappe sharp-crested weir

Van Vuuren S J, Coetzee G L, Roberts C P R. Investigating the bottom free surface nappe (ogee profile) across a sharp-crested weir caused by the flow in an asymmetrical approach channel. J. S. Afr. Inst. Civ. Eng. 2015;57(3), Art. #980, 7 pages. http://dx.doi.org/10.17159/2309-8775/2015/v57n3a7 57 Table 1 Methods for approximation of the ogee curve a H Approximation of the ogee curve based on:

The principles of projectile movement Experimental methods Tp Hd He C΄ Ven te Chow (1st principles) United States Bureau of Reclamation (Chow 1959) (USBR 1987)

United States Army Corps of Engineers Brink Velocity (Wahl et al 2008) f (USACE 1970) P Ven te Chow (modified) United States Army Corps of Engineers (Chow 1959) (USACE (b) 1987)

Hager (1987)

Creeger (Chanson 2004)

Scimeni (Chanson 2004) Figure 1 Position of the turning point of the curvature at the maximum elevation Knapp (Chanson 2004) of the ogee profile Montes (Chanson 2004) detaches from the surface of the spillway CE-05016 (Ministry of Science and Technology 2007) and a sub-atmospheric pressure region is generated that may lead to cavitation (Savage Note: The reader is referred to the associated reference for more detail about each approximation of the ogee profile & Johnson 2001; Momber 2000). Cavitation usually occurs during a unit discharge, in excess of the design head, where the surface This study evaluated the existing ogee rela- three-dimensional flow parameters which pressure can be reduced at positions along tionship, considering in particular the three- are occurring upstream of the spillway. the spillway to sub-atmospheric pressure. dimensional flow parameters resembling an Research indicated that the vertical This may cause the formation of vapour asymmetrical approach channel. displacement of a flow particle at the crest’s cavities. Vapour cavities may also be formed The physical model was set up in accor- origin is equal to the vertical distance on the spillway where an irregularity in the dance with ASTM Designation: D 5242–92 between the highest point of the nappe and surface exists. The vapour cavities (also (Standard Test Method for Open-Channel the elevation of the ogee crest (Chow 1959). referred to as miniscule air bubbles) will Flow Measurement of Water with Thin- Figure 1 depicts this position, which is rec- progress along the flow path due to the high Plate Weirs (ASTM International 2001) and ognised as the turning point of the curvature flow velocity on the spillway to a region ISO 1438: Hydrometry – Open channel (Tp). Rajaratnam et al (1968) (referenced by downstream where sufficient pressure is flow measurement using thin-plate weirs Vischer & Hager, Dam Hydraulics 1999) available that will lead to the collapse of the (ISO 2008). Froude similarity was selected indicated that the vertical value of C’ can air vacuum. This generates localised high with an undistorted scale of 1:10 for the be empirically estimated by Equation 1 and pressures. Should these vapour cavities physical model. the horizontal distance (f) to this coordinate collapse near the spillway structure, there measured from the crest can be estimated will be some superficial damage to the by Equation 2. The experiments done by spillway’s surface where the vapour bubble THEORETICAL CONSIDERATION Rajaratnam et al (1968) were actually for a has collapsed. This cavitation damage can Research on the projectile movement of a confined weir. ultimately result in substantial erosion and, free-falling jet of water revealed that the 2 if ignored, will subsequently cause failures of shape of the lower nappe of a jet of water 0.4vo C’ = 0.112∙Hd – (1) the spillway chute. Minute cracks, offsets and flowing over a ventilated sharp-crested 2g increased surface roughness intensify this weir resembled the shape of an ogee spill- 2 cavitation process. The extent of cavitation way (Chow 1959; Hager 1987; Knapp et al 0.4vo f’ = 0.250∙Hd – (2) damage is a function of the cavitation indices 1970; Melsheimer & Murphy 1970; Ministry 2g at key locations on the spillway chute and of Science and Technology 2007; Murphy the duration of flow over the spillway. This 1973; USACE (a) 1987; USACE (b) 1987; A wealth of literature sources are available emphasises the need for a geometric, accurate USBR 1987; Wahl et al 2008). However, both on the approximation of the ogee profile for and precise spillway profile to reduce the pos- the numerical relationships for the flow over spillway design, and several endeavours have sibility of sub-atmospheric pressure forma- a sharp-crested weir and the principle for been made in developing a two-dimensional tion (USACE 2009). the projectile movement of a free-falling jet relationship that would be able to math- It is hypothesised that the geometric of water describing this shape only consider ematically describe the most desirable profile of the ogee spillway will be influenced the two-dimensional characteristics of shape of the ogee curve. These relationships by the following factors: the flow, namely the available energy (i.e. exclude the asymmetricity of valleys and the a. The asymmetrical cross-sectional depth and velocity of water flowing over asymmetricity of the topographical approach approach channel upstream from the the spillway crest), the angle of inclination channels that influence the flow pattern and spillway of the upstream wall face, and the height velocity distribution upstream of the spill- b. The relative orientation of the spillway of the spillway above the natural ground way, leading to an insufficient design of the with regard to the approaching flow level (USBR 1987; USCE 1970; Vischer & ogee spillway (Van Vuuren et al 2011). c. The curvature of the spillway in relation Hager 1999). This approximation of co- Methods defining the geometry of the ogee to the depth of the structure. planar (two-dimensional) flow neglects the spillway’s crest are governed by the relative

58 Journal of the South African Institution of Civil Engineering • Volume 57 Number 3 September 2015 Table 2 Various scenarios executed with allocated notation

Description of Total discharge Unit discharge Head (H ) Equivalent ogee Symmetricity Notation d approach channel (ℓ/s) (ℓ/s/m) (mm) head (Ho) (mm) factor (Sf)

Scenario A Symmetrical (baseline) 71.929 59.890 105.10 93.35 0.500

Asymmetrical, Scenario B 72.678 60.514 116.65 105.28 0.146 sidewall inclined at 45°

Asymmetrical, Scenario C 72.609 60.457 116.05 104.74 0.084 sidewall inclined at 60°

The symmetrical approach channel, used Crest of weir: 2 mm thick stainless steel plate, laser cut to a 90° angle as baseline scenario, was compared with the Plywood sheet fixed to square asymmetrical approach layouts that were hollow section frame conducted at various sidewall inclinations. Sliding circular hollow section for adjusting the inclination of the sidewall 300 1 200 300 The various scenarios executed were labelled as indicated in Table 2. For each of the scenarios conducted a separate datasheet was set up in which the XYZ-coordinates of the lower nappe profile 1 135 were populated. The stage depth measured

Varying in stilling columns by two OTT-point inclinationthe sidewall of gauges was averaged and converted to the design head of an equivalent ogee structure. 220 3 125 220 Equation 1 as defined by Rajaratnam et al (1968) was applied to calculate the equivalent Figure 2 Generic upstream cross-sectional view of the asymmetrical approach channel with ogee design head. The measured ogee profile sidewall inclined at various angles for each of the scenarios was plotted in a three-dimensional XYZ surface plot against height of the structure, the available design ■■ provide the model with the correct the theoretical ogee profile (USACE 1970). head and upstream face slope of the spillway. dimensions and specifications in order Other factors that should also be consid- to ensure that equations provided by the ered are the approach channel symmetricity, current literature are valid and that inter- RESULTS AND DISCUSSION approach channel depth, orientation of the national credibility is achieved (ASTM The experimental results produced by the spillway relative to the flow path and the cur- International 2001; ISO 2008) asymmetrical approach channel generated vature (if applicable) of the hydraulic structure ■■ provide an undisturbed, uniform seven ogee curves for each channel configura- (Van Vuuren et al 2011). Literature reflects that approach flow pattern across the weir, tion, representing the three-dimensional shape in 1888 some of the first research was con- and vary the flow rate to enable the inves- of the asymmetrical bottom nappe. In the case ducted to investigate the shape of the ogee pro- tigation of different stages versus flow of the symmetrical base line recordings, the file. Ever since then the ogee spillway has been rate scenarios, as well as measuring the ogee curve was measured at five increments the most studied spillway geometry (Savage & stage of the approach flow accurately along the crest of the sharp-crested weir, rep- Johnson 2001; Thandaveswara 2006). ■■ measure the profile of the underlying resenting the three-dimensional shape of the The nappe-shaped profile is an ideal nappe of water flowing over the sharp- symmetrical bottom nappe. These measured profile because at the design discharge, crested weir. curves were representing the shape of the water flowing over the crest of the spill- Three configurations of the physical model nappe that extends from the left boundary, cen- way always remains in contact with the were investigated as reflected in Table 2, the tre and up to the right boundary of the nappe. surface of the spillway as it glides over it. first being the symmetrical layout with verti- The positions of the curves were chosen to Additionally, for the ogee shape, there will cal sidewall, used as baseline reference for ensure that all the scenarios could be analysed be no sub-atmospheric pressure regions that the other configuration. The other layouts individually and later compared with the other may develop on the spillway surface at the comprised one having an inclined sidewall scenarios. Each point measured was recorded design discharge. of 45° and the other an inclined sidewall of as a three-dimensional XYZ coordinate. Some of the well-regarded approxima- 60°. Six sheets of plywood were attached to These measurements were depicted in a tions of the ogee profile are grouped accord- steel frames and these were lowered into the two-dimensional XZ-plot in Figure 4, and ing to the two categories shown in Table 1. approach channel of the physical model. The were compared with the theoretical ogee plywood sidewalls could vary between angles profile by the USBR (1987), USACE (1970), that ranged from 45° to 90°. This allowed the Hager (1987) and Ministry of Science and EXPERIMENTAL SETUP symmetricity of the approach channel to be Technology (CE-05016) (2007). AND PROCEDURE changed to an asymmetrical approach chan- The measured ogee profile was plotted A physical model of a sharp-crested weir nel with one sidewall either being sloped at as a red surface in a three-dimensional was constructed at the Department of 45° or 60°. The total length of the sidewall XYZ-plot in Figure 5. Overlapping the Water Affairs and Sanitation’s Hydraulic measured 7.2 m. The cross-sectional view of measured ogee surface plot in Figure 5 is the Laboratory, South Africa. The model pro- the variable approach channel is depicted in theoretical ogee profile by the USACE (1970) vided the opportunity to: Figure 2. depicted by the green surface.

Journal of the South African Institution of Civil Engineering • Volume 57 Number 3 September 2015 59 as measured along the crest, and again over- estimated the ogee profile at the boundaries similar to the symmetrical approach channel layout. High flow velocities present in this region can be observed in Figure 3 where the flow pattern can be clearly visualised. This confirms the hypothesis that an asymmetrical approach channel influences not only the geometric shape of the ogee profile, but also the symmetricity of flow across the hydraulic structure. The cen- treline tends to be moved off-centre in the direction where the inclined sidewall had been installed. It is therefore proposed that a symmetric-

ity factor (Sf) be used to estimate the severity of the asymmetricity of the approach chan- nel. This will allow for future classification of the approach channels into different catego- ries depending on the severity of the asym- metricity. The proposed relationship for the Figure 3 High flow velocities observed at the boundary of the sharp-crested weir’s crest as symmetricity factor is given in Equation 3. caused by the steep inclination of the 60° sidewall The symmetricity factors for scenarios A to C are given in Table 2. Scenario A: Symmetrical in scenario A where the greatest ogee profile approach channel (baseline) occurred in the centre of the flow, due to the Al Sf = (3) It can be seen in the two-dimensional plot of symmetricity of the approach channel. The At scenario A that the ogee profile in the centre greatest ogee profile tends to lie between the of the nappe was greater than those measured 600 mm and 800 mm measuring location. The In the case of a symmetrical approach at the outer boundaries. This trend can be inclined sidewall was installed to the right of channel, the symmetricity factor (Sf) would visualised in more detail in the final results in the approach channel. The centreline tends to therefore be equal to 0.5, and values less than which the surface plot of the measured nappe be moved off-centre in the direction where the 0.5 would indicate that an asymmetrical was depicted in an XYZ plot. The reason for inclined sidewall had been installed. approach channel is perceptible. The clas- this outcome can be explained due to the The result was that the theoretical sification of the symmetricity factor needs fact that the ogee profile was measured for a approximations of the ogee profile tend to further investigation and will be published in symmetrical approach channel, and that an underestimate the measured ogee profile a forthcoming paper. increased flow rate exists in the centre of the between the 600 mm and 800 mm mark mea- structure, resulting in minimal contraction sured along the crest, and again overestimate and flow resistance at this location. the ogee profile at the boundaries similar to CONCLUSIONS AND The result was that the theoretical the symmetrical approach channel layout. RECOMMENDATIONS approximations of the ogee profile tend to The change in the symmetricity of the underestimate the measured ogee profile at Scenario C: Asymmetrical approach approach channel has unambiguously the centre of the weir and overestimate the channel with sidewall inclined at 60° proved to alter the shape of the ogee pro- ogee profile at the boundaries where flow Similar to scenario B the two-dimensional plot file. This alteration is appreciated in the contraction and resistance was a maximum. of scenario C reflected that the ogee profile following ways: Overestimation of the ogee profile is not as was greater towards the side of the inclined ■■ The profile of the nappe differs along the critical as underestimation. The overestima- sidewall. The greatest ogee profile tends to length of the weir. From the experiments tion of the ogee profile at the boundaries lie at the 800 mm measuring location. This conducted it was also revealed that the will simply imply that a positive hydrostatic indicated that the ogee nappe was moved off- side of the weir with the inclined sidewall pressure is present in these regions on the centre even more than in the case when the caused the profile of the nappe to be spillway. However, underestimating the ogee inclination of the sidewall was 45°. This 60° higher than at the opposite side. profile may cause a more detrimental effect, angle sidewall was steep and extended closely ■■ When comparing the asymmetrical which may lead to cavitation of the surface to the boundary of the 1 200 mm mark of the approach channel’s ogee profile with the of the spillway if significant sub-atmospheric sharp-crested weir so that this could have baseline/symmetrical nappe profile, the pressures are experienced. Long-term expo- caused the sharp-crested weir not to function profile of the asymmetrical approach sure to extensive sub-atmospheric conditions as a fully-contracted weir anymore. This may channel configuration was increased at may cause failure of the spillway structure. be the reason for the higher velocities present the vicinity of the inclined wall. in this region (less flow resistance), thus result- ■■ If an ogee spillway were to be designed Scenario B: Asymmetrical approach ing in the ogee profile to be the greatest at this considering only the two-dimensional channel with sidewall inclined at 45° location. The result was that the theoretical flow parameters, problems would arise In the two-dimensional plot of scenario B it approximations of the ogee profile tended if the cross-section is asymmetrical. The was noticed that the ogee profile was greater to underestimate the measured ogee profile profile would tend to be higher than toward the side of the inclined sidewall, unlike between the 400 mm and 1 000 mm mark expected towards the inclined sidewall

60 Journal of the South African Institution of Civil Engineering • Volume 57 Number 3 September 2015 Horizontal x-scale (mm) Scenario A: 100 mm 0 10 20 30 40 50 60 70 80 90 100 110 120 130 –15 Scenario A: 200 mm

Scenario A: 400 mm –5 Scenario A: 600 mm

Scenario A: 1 000 mm 5

Scenario A: 1 100 mm Scenario A 15 USBR (1987)

Vertical y-scale (mm) y-scale Vertical USACE compound curve (1970) 25 Hager (1987)

USBR (insignificant approach velocity, 1987) 35

CE-0516 (2007)

Horizontal x-scale (mm) Scenario B: 100 mm 0 10 20 30 40 50 60 70 80 90 100 110 120 130 140 150 160 20 Scenario B: 200 mm

10 Scenario B: 400 mm

Scenario B: 600 mm 0 Scenario B: 800 mm –10 Scenario B: 1 000 mm –20 Scenario B: 1 100 mm Scenario B –30 USBR (1987)

Vertical z-scale (mm) z-scale Vertical –40 USACE compound curve (1970)

–50 Hager (1987)

–60 USBR (insignificant approach velocity, 1987)

–70 CE-0516 (2007)

Horizontal x-scale (mm) Scenario C: 100 mm 0 10 20 30 40 50 60 70 80 90 100 110 120 130 140 150 160 20 Scenario C: 200 mm

10 Scenario C: 400 mm

Scenario C: 600 mm 0 Scenario C: 800 mm –10 Scenario C: 1 000 mm –20 Scenario C: 1 100 mm Scenario C –30 USBR (1987)

Vertical z-scale (mm) z-scale Vertical –40 USACE compound curve (1970)

–50 Hager (1987)

–60 USBR (insignificant approach velocity, 1987)

–70 CE-0516 (2007)

Figure 4 Measured ogee profile for different scenarios and compared with various other approximations

Journal of the South African Institution of Civil Engineering • Volume 57 Number 3 September 2015 61 20 20

0 0

–20 –20 Z (mm) Z (mm)

Scenario A –40 –40

–60 –60

0 0 0 0 200 20 200 20 40 40 400 60 400 60 600 80 600 80 Y (mm) 800 100 Y (mm) 800 100 1 000 120 1 000 120 1 200 140 X (mm) 1 200 140 X (mm)

20 20

0 0

–20 –20 Z (mm) Z (mm) –40 –40 Scenario B

–60 –60

–80 0 –80 0 20 20 40 400 40 400 60 600 60 600 80 80 100 Y (mm) 800 100 Y (mm) 800 120 120 140 1 000 140 X (mm) 1 000 160 X (mm)

20 20

0 0

–20 –20 Z (mm) Z (mm) –40 –40 Scenario C

–60 –60

–80 0 –80 0 20 20 40 400 40 400 60 600 60 600 80 80 100 Y (mm) 800 100 Y (mm) 800 120 120 140 1 000 140 X (mm) 1 000 160 X (mm)

Figure 5 Measured ogee profile (red profile) for different scenarios and compared with the USACE (1970) approximation (green profile)

62 Journal of the South African Institution of Civil Engineering • Volume 57 Number 3 September 2015 of the structure, yet lower than expected process, as theoretical curves do not neces- Momber, A W 2000. Short-time cavitation erosion of on the opposite side. This may cause the sarily provide optimal and safe solutions. concrete. Wear, 241: 47–52. dual problem of potential separation from A design guideline should be developed Murphy, T E 1973. Spillway Crest Design. Vicksburg, MS: the spillway structure causing cavitation that would assist design engineers to design U.S. Army Corps of Engineers Waterways Experiment damage due to sub-atmospheric pressure, suitable, efficient and safe ogee spillways by Station. and sub-optimal discharge on the oppo- considering parameters such as the asym- Rajaratnam, N, Subramanya, K & Muralidhar, D 1968. site side, respectively. metricity of the approach channel. Flow profiles over sharp-crested weirs. Journal of the The deduction of the comparison of meas- Hydraulics Division, ASCE, 94(HY3): 843–847. ured profiles with the theoretically calcu- Savage, B & Johnson, M 2001. Flow over Ogee Spillway: lated ogee profiles produced diverse results: ACKNOWLEDGEMENTS Physical and numerical model case study. Journal of i. The USBR (1987), USACE (1970), Hager We would like to thank the Water Research Hydraulic Engineering, 127(8): 640–649. (1987) and Ministry of Science and Commission for their valued contribution Thandaveswara, B 2006. National Programme on Technology (CE-05016) (2007) ogee to this study, as well as the Department of Technology Enhanced Learning (NPTEL). Adras, approximations corresponded well with Water and Sanitation for the use of their India: Indian Institute of Technology. the geometry of the measured ogee pro- hydraulic laboratory. USACE (U.S. Army Corps of Engineers) 1970. Hydraulic files, although these curves had a tendency Design Criteria. Vicksburg, MI: USACE Waterways to underestimate the actual ogee profile at Experiment Station. the position on the crest where the effec- REFERENCES USACE (U.S. Army Corps of Engineers) 1987a. Hydraulic tive resistance to flow is a minimum. ASTM International 2001. Standard Test Method for Design Criteria: Overflow Spillway Crest. Vicksburg, ii. The downstream approximation made by Open-channel Flow Measurement of Water with Thin- MI: USACE Waterways Experiment Station. Hager (1987) of the ogee curve appeared plate Weirs. US Patent No. Designation: D 5242 – 92. USACE (U.S. Army Corps of Engineers) 1987b. to overestimate the projection of the Chanson, H 2004. The Hydraulics of Open Hydraulic Design Criteria: Elliptical Crest Spillway nappe profile in the downstream region Channel Flow: An Introduction, 2nd ed. Oxford: Co-ordinates, Vicksburg, MI: USACE Waterways of the ogee. This demonstrated that the Butterworth-Heinemann. Experiment Station. formula was conservative and that the Chow, V T 1959. Open-channel Hydraulics, reprint. USACE (U.S. Army Corps of Engineers) 2009. Dam probability of sub-atmospheric pressure Illinois, US: The Blackburn Press. Safety Risk Analysis Best Practices Training Manual. occurring in this region is minimised. Hager, W H 1987. Continuous crest profile for Version 2.2 – April 2011 ed. Denver, CO: U.S. It is recommended that further studies standard spillway. Journal of Hydraulic Engineering, Department of Interior Bureau of Reclamation. be carried out (by making use of physical 113(11): 1453–1457. USBR (U.S. Department of Interior Bureau of modelling) to assess the effect of the angle ISO (International Organization for Standardization) Reclamation) 1987. Spillways, Chapter 9. In: Design of of the approach channel relative to the ogee 2008. Hydrometry – Open-channel flow measurment Small Dams, 3rd ed. Washington: U.S. Government spillway, as well as the curvature of the crest using thin-plate weirs. Switzerland, Patent No. Printing Office, pp 339–434. on the geometric shape of the ogee profile. ISO 1438. Van Vuuren, S, Steyn, G, Pilz, N & Koch, H 2011. Results obtained from this study, as well Knapp, R, Daily, W & Hammitt, F 1970. Cavitation. New Influence of 3D approaching flow on a Curved as the studies proposed above, should be York: McGraw-Hill. Ogee Spillway Section – Neckartal Dam Namibia. compared with a numerical simulation using Melsheimer, E S & Murphy, T E 1970. Hydraulic Design Stellenbosch: UP Printers. a CFD analysis. It is absolutely imperative Criteria, Sheets 111-16 to 11-16/2. U.S. Army Corps of Vischer, D & Hager, W (Eds.) 1999. Dam Hydraulics. and critical to include the effect of three- Engineers. New York: Wiley, pp. 40–44. dimensional flow when designing an ogee Ministry of Science and Technology (Myanmar) 2007. Wahl, T L, Frizell, K H & Cohen, E A 2008. Computing spillway. This is also indicative of the neces- CE-05016. Design of Hydraulic Structures, Myanmar: the trajectory of free jets. Journal of Hydraulic sity of physical modelling during the design Department of Technical and Vocational Education. Engineering, ASCE, 10.1061(2): 256–260.

Journal of the South African Institution of Civil Engineering • Volume 57 Number 3 September 2015 63 Guidelines for the preparation of papers and technical notes

Authors should comply with the following guidelines when preparing papers for publication in the journal

The Journal of the South African Institution of Civil data and other information that may be confidential or should be typed using the Times New Roman Engineering is published quarterly in March, June, or sensitive. Authors are also responsible for obtain- symbol set. September and December. Articles submitted for ing permission from copyright owners when repro- ■■ Variables in equations (x, y, z, etc, as well as lower publication are reviewed by a panel of referees under the ducing material that has been published elsewhere. case Greek letters) should be presented in italics. guidance of the SAICE Journal Editorial Panel. The jour- Proof of such permission must be supplied. Numbers (digits), upper case Greek letters, sym- nal publishes research papers covering all the disciplines bols of metric measurement units (m for metres, of civil engineering (structural, geotechnical, railway, SUBMISSION PROCEDURES s for seconds, etc) and mathematical/trigono- coastal/marine, water, construction, environmental, AND REQUIRED FORMAT metrical functions (such as sin, cos and tan) are municipal, transportation) and associated topics that ■■ Electronic submission: Manuscripts should be not written in italics, but in upright type (Roman). are relevant to the civil engineering profession, and that e-mailed to the editor at [email protected]. File Variables and symbols used in the body of the text preferably have particular relevance to civil engineering sizes should not exceed 4 MB per e-mail – figures should match the format used in the equations, in southern Africa and the African continent. may for example be sent one by one or in groups not i.e. upright or italics, whichever is applicable. When preparing articles for publication, authors ■ larger than 4 MB. Manuscripts should not be sent in ■ Metric measurement abbreviations/units should should please take note of the following and comply with PDF format as this precludes reviewing of papers per conform to international usage – the SI system of the guidelines as set out: track changes. units should be used. ■ ■ Format: Manuscripts should be prepared in MS ■■ Decimal commas may be used, but decimal points CLASSIFICATION OF ARTICLES Word and presented in double line spacing, single are preferred. CONSIDERED FOR PUBLICATION column layout with 25 mm wide margins. Line ■■ Symbols should preferably be defined in the text, ■ ■ Technical papers are well-researched, in-depth, numbers must be applied to the whole document. but if this is not feasible, a list of notations may be fully referenced technical articles not exceeding All pages should bear the authors’ names and be provided for inclusion at the end of the paper. 6 000 words in length (excluding tables, illustrations numbered at the bottom of the page. With the excep- ■■ Headings: Sections and paragraphs should not be and the list of references). Related papers that deal tion of tables and figures (see below) the document numbered. The following hierarchy of headings with ‘softer sciences’ (e.g. education, social uplift- should be typed in Times New Roman 12 pt font. should be followed: ment, etc) are accepted if they are of a technical Contributions should be accompanied by an abstract HEADING OF MAIN SECTION nature and of particular interest to the civil engineer- of not more than 200 words. Heading of subsection ing profession. The latter type of paper will be subject ■ ■ First page: The first page of the manuscript should Heading of sub-subsection not only to peer-review by civil engineers, but also include the title of the paper, the number of words of ■■ References: References should follow the Harvard to review by non-engineering specialists in the field the main text (i.e. excluding figures, tables and the system. The format of text citations should be as covered by the paper. list of references), the initials and surnames of the follows: “Jones (1999) discovered that …” or “recent ■■ Technical notes are short, fully referenced technical authors, professional status (if applicable), SAICE results (Brown & Carter 1985; Green et al 1999) articles that do not exceed 2 000 words. A typical affiliation (Member, Fellow, Visitor, etc), telephone indicated that …” technical note will have limited scope often dealing numbers (landline and mobile), and e-mail and  References cited in the text should be listed in with a single technical issue of particular importance postal addresses. The name of the corresponding alphabetical order at the end of the paper. References to civil engineering. author should be underlined. Five keywords should by the same author should be in chronological order. ■■ Review papers are considered for publication as be suggested. The following are examples of a journal article, a either technical papers or technical notes on condi- ■■ These Figures, tables, photos and illustrations: book and a conference paper: tion that they are the original work of the author should preferably be submitted in colour, as the ■■ Donald, A W 1954. Pore pressure coefficients. and will assist the reader with the understanding, journal is a full-colour publication. Géotechnique, 4(4): 143–147. interpreting or applying of the subject under review. ■■ Their positions should be clearly marked in the ■■ Terzaghi, K 1943. Practical soil mechanics. New A review paper must contain criteria by which the text as follows: [Insert Figure 1] York: Wiley. work under review was evaluated, and contribute ■■ Figures, tables, photos, illustrations and equations ■■ Smith, A W, Black, C & White, A 1999. Factors by synthesising the information and drawing new should be numbered consecutively and should affecting the strength of soils. Proceedings, ASCE conclusions from the dissemination of the previously appear in the text directly after they have been Conference on Shear Strength of Soils, Colorado, published work. referred to for the first time. ■ pp 503–532. ■ Discussion on published articles is welcomed up to ■■ Illustrations must be accompanied by appropriate Papers published previously in the Journal of the six months after publication. The length of discus- captions. Captions for tables should appear above South African Institution of Civil Engineering should sion contributions is limited to 1 500 words. Where the table. All other captions should appear below be cited if applicable. appropriate, discussion contributions will be subject the illustration (figures, graphs, photos). ■■ Footnotes, trade names, acronyms, abbreviations: to the normal reviewing process and will be for- ■■ Only those figures and photographs essential to These should be avoided. If acronyms are used, they warded to the authors of the original article for reply. the understanding of the text should be included. should be defined when they first appear in the text. All illustrations should be referred to in the text. Do not use full stops after abbreviations or acronyms. POLICY REGARDING LANGUAGE AND ■■ Figures should be produced using computer ■■ Return of amended papers: Papers requiring ORIGINALITY OF SUBMITTED ARTICLES graphics. Hand-drafted figures will not be accept- amendments will be accepted up to six months after ■■ Language: Manuscripts should preferably be ed. Lettering on figures should be equivalent to a the referee reports have been sent to authors, after presented in English, as the journal is distributed Times New Roman 9 pt font or slightly larger (up which the paper will be withdrawn from the system. internationally. Articles submitted in any of the other to 12 pt) if desired. Lettering smaller than 9 pt is official South African languages should be accompa- not acceptable. FINAL ARTICLE nied by an expanded abstract in English. ■■ Tables should be typed in Times New Roman 9 pt ■ ■■ Original work: Papers and technical notes must be font. They should not duplicate information already ■ Copyright: On acceptance of the paper or technical original contributions. Authors must affirm that sub- given in the text, nor contain material that would note, copyright must be transferred by the author/s to mitted material has not been published previously, be better presented graphically. Tabular matter the South African Institution of Civil Engineering on is not under consideration for publication elsewhere should be as simple as possible, with brief column the form that will be provided by the Institution. and will not be submitted elsewhere while under headings and a minimum number of columns. ■■ Photos of authors: The final corrected version of consideration by the SAICE Journal Editorial Panel. It ■■ Mathematical expressions and presentation of the paper should be accompanied by recent, high- is the responsibility of the authors to ensure that pub- symbols: resolution head and shoulders colour photographs lication of any paper in the journal will not constitute ■■ Equations should be presented in a clear form and a profile not exceeding 100 words for each of the a breach of any agreement or the transgression of any which can easily be read by non-­mathematicians. authors. law. The corresponding author should confirm that Each equation should appear on a separate line ■■ Proofs: First proofs of papers will be sent to authors all co-authors have read and approved the manuscript and should be numbered consecutively. in PDF format for verification before publication. No and accept these conditions. Authors are responsible ■■ Symbols should preferably reflect those used in major rewrites will be allowed, only essential minor for obtaining permission to publish experimental Microsoft Word Equation Editor or Mathtype, corrections.