2401 © IWA Publishing 2017 Water Science & Technology | 76.9 | 2017 Flood forecasting within urban drainage systems using NARX neural network Yves Abou Rjeily, Oras Abbas, Marwan Sadek, Isam Shahrour and Fadi Hage Chehade ABSTRACT Urbanization activity and climate change increase the runoff volumes, and consequently the Yves Abou Rjeily (corresponding author) Oras Abbas surcharge of the urban drainage systems (UDS). In addition, age and structural failures of these Marwan Sadek Isam Shahrour utilities limit their capacities, and thus generate hydraulic operation shortages, leading to flooding Lille University of Science and Technology, fl Laboratoire de Génie Civil et Géo-Environnement, events. The large increase in oods within urban areas requires rapid actions from the UDS Villeneuve d’Ascq, operators. The proactivity in taking the appropriate actions is a key element in applying efficient France E-mail: [email protected] management and flood mitigation. Therefore, this work focuses on developing a flooding forecast Yves Abou Rjeily system (FFS), able to alert in advance the UDS managers for possible flooding. For a forecasted storm Marwan Sadek Fadi Hage Chehade event, a quick estimation of the water depth variation within critical manholes allows a reliable Lebanese University, fl Modeling Center, evaluation of the ood risk. The Nonlinear Auto Regressive with eXogenous inputs (NARX) neural Beirut, network was chosen to develop the FFS as due to its calculation nature it is capable of relating water Lebanon depth variation in manholes to rainfall intensities. The campus of the University of Lille is used as an experimental site to test and evaluate the FFS proposed in this paper. Key words | case study, flooding forecast, NARX neural network, proactivity, urban drainage systems INTRODUCTION The fast growth of cities and the aging of their infrastruc- ). In recent years, dynamic management has shown a tures are the main reasons behind the overstressing of huge potential in increasing the capacities of these infrastruc- urban drainage systems (UDS). Cities’ expansion, not being tures, by optimizing their operations (Beeneken et al. ; accompanied by suitable drainage infrastructure upgrades, Rocha et al. ; García et al. ). The efficiency of the results in generating frequent flooding events, environment dynamic management is highly dependent on the proactivity degradation and decreasing of groundwater recharge of the system managers and operators, which could be signifi- (Konrad & Booth ; Wang et al. ; Brandes et al. cantly enhanced by a flooding forecast system (FFS). An FFS ). Furthermore, climate change affects rainfall intensity consists of a model capable of predicting flooding occurrence and patterns, and consequently the runoff volumes (Ponce and warning in advance the managers of the system. et al. ; May ). This is leading to shortages in the Offering sufficient lead time to system operators to take capacity of the UDS (Berggren et al. ), and hence is caus- preventive measures and apply optimal management strat- ing frequent flooding appearances. The floods are difficult to egies, an FFS is considered as an essential tool in mitigating predict and result in significant damage, economic conse- flooding impacts. It is important to note that the complexity quences and casualties (Kenyon et al. ). Therefore, and dynamicity of UDS operations, together with temporal flood risks have become a major challenge for cities and a and spatial loading variability (Ocampo-Martinez ), major concern for researchers and practitioners. make the construction of the FFS a demanding task, taking The first step in mitigating flooding impacts consists of into consideration the time delays, constraints and nonlinea- understanding the actual system operation during the flooding rities. Traditionally, flooding forecasts are based on historical events. Implementing a monitoring system combined with a data and mathematical models or graphs that concern pat- hydraulic simulation model has proved efficient in analysing tern recognition (Rajendra Acharya et al. ). Recently, the UDS operation and flooding origins (Abou Rjeily et al. black box models are used to allow the prediction of urban doi: 10.2166/wst.2017.409 Downloaded from https://iwaponline.com/wst/article-pdf/76/9/2401/208868/wst076092401.pdf by guest on 27 July 2020 2402 Y. Abou Rjeily et al. | Flood forecasting using NARX neural network Water Science & Technology | 76.9 | 2017 flooding occurrences. In order to accomplish this prediction, rainfall intensities. The use of a black box model, instead of the models are trained on historical data and combined with conducting hydrologic–hydraulic simulation for the entire rainfall radar forecasts (Duncan et al. ). FFSs developed UDS operation, allows a significant reduction in the calcu- using black box models have been applied largely for river lation time, which improves the proactivity of the operators. analysis and protection, and have shown a good efficiency The water depth variation within a manhole, which rep- and practicality (Elsafi ; Perera & Lahat ; Amarnath resents the output of the complex function, depends on the et al. ; Artinyan et al. ), while for UDS, such systems rainfall intensity time series and the hydrologic modifications are less developed and evaluated (Yen-Ming et al. ). Fore- that occur during a storm event. The decrease in soil infiltra- casting flooding in UDS is still based on deterministic tion and in depression storage potential are examples of the simulation model results that require a lot of calculation hydrologic modifications that occur during a storm event. time, which limits the efficiency and proactivity of the oper- ators in taking the appropriate actions. With this in mind, NARX neural network for forecasting flooding events this work focuses on developing a new methodology for implementing a feasible FFS for UDS. The black box model, which will represent the calcula- tion engine of the FFS in this study, should be able to account for rainfall intensity time series and to consider the METHODS hydrologic modifications occurring on the catchment. The NARX (Nonlinear Auto Regressive with eXogenous inputs) Proposed FFS neural network combines exogenous input with recurrent behaviour in order to calculate its output. The exogenous In this study, the proposed FFS will be focusing on quickly input of the NARX neural network can represent the rainfall predicting the water depth variation in some critical locations intensity time series. In addition, through its recurrent behav- of an UDS, instead of simulating the entire network oper- iour, a NARX neural network is able to differentiate the time ations. In order to achieve that, the engineers should first factor within a storm event. Differentiating the time and under- evaluate their system hydraulic capacity and analyse the standing what was already happening in the UDS could enable flooding origins by conducting hydraulic simulations. Based the NARX neural network to account for the hydrologic on the simulation results, the locations of critical areas and modifications occurring during a storm event. Therefore, the manholes that should be monitored can be defined. The NARX neural network and its calculation nature will be water depth variations within the defined critical manholes tested in the following sections, to evaluate its capacity to oper- will be predicted through a black box model using forecasted ate as the calculation engine for the proposed FFS. rainfall intensities. The calculation engine of the FFS will be a Due to its efficiency in representing nonlinear dynamic black box model as described in the next paragraphs. The behaviours (Hoffmeister ), the NARX model has been strategy steps for the FFS are as follows. Once the weather largely used in time series modelling and forecasting appli- forecast detects the presence of rainfall events within the fore- cations. The NARX model, presented in Equation (1), is casted period, the black box model predicts the water depth characterized by calculating an output at the actual time variations at the critical locations of the UDS. If a forecasted step, as a function of multiple precedent inputs and outputs water depth in any critical location exceeds a threshold of the previous time steps. defined by the engineers, the FFS alerts the managers of poss- ible flooding. In response to the alert, infrastructure managers y(t) ¼ f[u(t À 1), ..., u(t À nu), y(t À 1), ..., y(t À ny)] (1) evaluate the severity of the situation, identify areas likely to be inundated and take appropriate actions and precautions. where y(t): model output time series, u(t): model input time Actions and precautions could be as follows: informing series, nu and ny: time delays required by the model in order specialist operators, warning inhabitants of underground to effectively represent the dynamic behaviour of the studied basements, changing road signs for car drivers, etc. phenomenon. The response of the UDS is a complex nonlinear function NARX neural networks are highly efficient in simulating of the temporal and spatial variability of rainfall intensities. complex systems (Marsalek ). They have been used in As mentioned earlier in this work, this complex function different application types, as predictor for the next values will be replaced by a black box model, trained to predict the of a time series and nonlinear noise filtering of input signals water depth variation in some manholes according to the (Demuth et al. ). Their applicability has been proved in