OCTOBER 2016 H E M I N G 1433

Met Office Unified Model Performance Following Major Changes to the Initialization Scheme and a Model Upgrade

JULIAN T. HEMING Met Office, Exeter, Devon, United Kingdom

(Manuscript received 1 March 2016, in final form 27 June 2016)

ABSTRACT

The Met Office has used various schemes to initialize tropical cyclones (TCs) in its numerical weather pre- diction models since the 1980s. The scheme introduced in 1994 was particularly successful in reducing track forecast errors in the model. Following modifications in 2007 the scheme was still beneficial, although to a lesser degree than before. In 2012 a new trial was conducted that showed that the scheme now had a detrimental impact on TC track forecasts. As a consequence of this, the scheme was switched off. The Met Office Unified Model (MetUM) underwent a major upgrade in 2014 including a new dynamical core, changes to the model physics, an increase in horizontal resolution, and changes to satellite data usage. An evaluation of the impact of this change on TC forecasts found a positive impact both on track and particularly intensity forecasts. Following implementation of the new model formulation in 2014, a new scheme for initialization of TCs in the MetUM was developed that involved the assimilation of central pressure estimates from TC warning centers. A trial showed that this had a positive impact on both track and intensity predictions from the model. Operational results from the MetUM in 2014 and 2015 showed that the combined impact of the model upgrade and new TC initialization scheme was a dramatic cut in both TC track forecast errors and intensity forecast bias.

1. Introduction In 2007, a complete reevaluation of the initialization scheme was undertaken to assess whether it was still There is a long history of initializing tropical cyclones proving beneficial to forecasts of TCs from the MetUM. (TCs) in the Met Office Unified Model (MetUM) in In the years since the scheme was first introduced there order to improve the model’s representation of TCs in had been many improvements in model formulation, both the analysis and the forecast. In the late 1980s and increases in model resolution, and the introduction of early 1990s forecasters had a tool available that inserted new observational data, particularly from satellites. so-called bogus observations of central pressure, sur- These changes were likely to have diminished the need rounded by four values of wind speed and direction at for artificial initialization of TCs. In the event, this the surface and three lower-tropospheric levels. In 1994, evaluation first found that the initialization scheme was this was superseded by a new initialization scheme that still reducing TC track forecast errors by an average of involved the insertion of bogus observations of wind 12.2%. Furthermore, a modification to the scheme that speed and direction at the surface and three lower- reduced the horizontal spread of ‘‘bogus’’ observations tropospheric levels. This technique proved extremely generated for small TCs resulted in a further reduction successful and reduced TC track forecast errors by 34% in TC track forecast errors of 4.7% (Heming 2009). This on average in trials (Heming et al. 1995). The following configuration of the initialization scheme will be known year the MetUM produced better guidance for TC track as the 2007 scheme hereafter. A diagrammatic repre- prediction to the National Hurricane Center than any sentation of the scheme is shown in Fig. 1. After a few other numerical weather prediction model for the ex- years in operation a further evaluation of the 2007 tremely active Atlantic hurricane season of that year scheme was undertaken to ensure it was still providing (Gross 1996; Heming and Radford 1998). benefit to the model forecasts. Section 2 of this paper presents the results of this evaluation. Corresponding author address: Julian T. Heming, Met Office, A major change to the MetUM was implemented in FitzRoy Road, Exeter, Devon, EX1 3PB, United Kingdom. 2014, which was the culmination of many years’ work E-mail: julian.heming@metoffice.gov.uk (Met Office 2014; Walters et al. 2016, manuscript submitted

DOI: 10.1175/WAF-D-16-0040.1

Unauthenticated | Downloaded 09/25/21 08:16 PM UTC 1434 WEATHER AND FORECASTING VOLUME 31

FIG. 1. The configuration of the 2007 scheme. to Geosci. Model Dev.). This included changes to the periods covered a significant amount of TC activity in model dynamics, physics, horizontal resolution, and both the Northern and Southern Hemispheres. The first satellite data usage. Early trials indicated that this period was from 23 August to 19 September 2010, the change would have a significant impact on TC intensity second from 20 August to 15 October 2011, and the third predictions. Section 3 of this paper evaluates the impact from 19 January to 18 March 2012. In total there were 57 of the model change on TC predictions and explains how TCs during these three periods comprising 17 in the this resulted in the development of a completely new western North Pacific, 10 in the eastern North Pacific, 19 form of TC initialization in the MetUM using estimates in the Atlantic, 4 in the western south Indian Ocean of central pressure from TC warning centers (discussed (west of 908E), and 7 in the eastern south Indian Ocean in section 4). and South Pacific. A sizeable number of forecasts were verified at each forecast lead time: for example, 460 at 24 h, 243 at 96 h, and 87 at 168 h. The control and trial 2. Evaluation of the 2007 scheme used the configuration of the MetUM operational from July 2011 to January 2012 (known as OS27), which a. Trial configuration and results had a horizontal grid spacing of approximately 25 km at In 2012 experiments were undertaken to assess the midlatitudes and 70 vertical levels. The control in- impact on TC forecasts of transplanting analysis fields cluded use of the 2007 scheme to initialize TCs, as was from another model into the MetUM. The European done in the operational model at the time, whereas in Centre for Medium-Range Forecasts (ECMWF) model the trial this scheme was switched off. Table 1 shows was chosen since it had performed better than the the various verification scores for TC track prediction MetUM for TC track prediction in the previous few for the control and trial. Scores were calculated at years. The results indicated that the performance of the 12-hourly forecast intervals, but only the 24-hourly forecast was very sensitive to the lower-tropospheric values are shown in Table 1. Details of the TC tracking winds in the analysis (Heming 2012). Given this result, it method and verification scores can be found in Heming was decided to undertake another evaluation of the 2007 (2016). scheme to ensure it was still providing benefit to MetUM For TC track the results show that despite much larger forecasts of TCs since the scheme primarily adjusted the analysis errors in the trial, which might be expected model’s lower-tropospheric wind fields. Standard con- because of the removal of the initialization, track fore- trol runs of approximately 4–8 weeks are available to cast errors were lower at all lead times. When averaged model developers at the Met Office. The 2007 scheme over all forecasts from 12 to 168 h at 12-hourly intervals, was evaluated by using three control runs available at the trial track forecast errors were 8.4% lower. The re- the time for periods during 2010, 2011, and 2012 and duction in track errors up to 96 h was significant beyond running trial forecasts without the 2007 scheme. These the 1% level. The reduction in errors at longer lead

Unauthenticated | Downloaded 09/25/21 08:16 PM UTC OCTOBER 2016 H E M I N G 1435

TABLE 1. 2007 scheme control versus trial results. Mean track forecast statistics, where boldface indicates the better score. The 12-hourly statistics are calculated, but only 24-hourly results are shown. The t values and significance levels (%) of track error differences are shown. The trial had the 2007 scheme switched off.

0 h 24 h 48 h 72 h 96 h 120 h 144 h 168 h No. of cases 571 460 386 310 243 182 132 87 Control track errors superior 346 198 159 130 103 76 47 28 Trial track errors superior 191 254 219 179 140 104 85 59 Track errors equal 34 8 8 1 0 2 0 0 Control track error (km) 33 103 184 285 403 557 723 871 Trial track error (km) 47 92 160 258 372 523 665 797 Reduction in track error of trial relative to 240.8 11.4 11.9 9.8 5.8 3.8 6.6 6.0 control (%) t value of track error differences 8.103 23.585 24.421 22.908 22.414 21.628 21.599 21.503 Significance level of track error differences (%) ,0.1 ,0.1 ,0.1 0.2 0.8 5.3 5.6 6.8 Control track skill score (%) — 41 51 49 — — — — Trial track skill score (%) — 48 58 55 ————

times was significant beyond the 7% level. Model skill in forecast tracks for individual forecasts for Hurricanes predicting the track of TCs against climatology and Igor, Katia, and Maria. For the control persistence (CLIPER; Neumann 1972) was calculated persisted with a westward track for too long and was for the first 72 h of the forecast. The trial skill scores were also a little too fast. The trial was slower and turned on average 6.0% higher than the control results. The Igor northwestward sooner, resulting in a much better frequency of superior performance for TC track fore- forecast. For the control showed an casts shows that the trial was superior in 57% of all erroneous turn toward the west, whereas the trial cor- forecasts compared to 42% for the control. This latter rectly persisted with a northwestward track. Finally, for measure indicates the percentage of TC track forecasts the control again did not turn the that were improved or worsened as a result of switching system northwest and then north soon enough, result- off the 2007 scheme. Results show that switching off the ing in an equatorward bias. The trial was much closer to scheme was beneficial in 15% more forecasts than it was the observed track. detrimental. In another forecast for Hurricane Igor, the trial For intensity forecasts (not shown in Table 1) the showed a much better prediction of recurvature and differences between the control and trial were fairly acceleration into the midlatitudes at longer lead times. modest. Trial forecasts on average had a lower central Figure 3 shows the 168-h forecast from data at 1200 UTC pressure by 0.8 hPa and a higher maximum 10-m wind 14 September 2010 for the control and trial alongside the 2 speed by 0.7 kt (1 kt 5 0.51 m s 1). The trial reduced the verifying analysis. The control did not engage the hur- mean absolute error in maximum 10-m wind speed by ricane with the midlatitude jet, resulting in it remaining just 0.6 kt. Intensity tendency skill (the skill of the model in the subtropics close to . Although the trial to predict strengthening or weakening) was 1.8% higher forecast was a little to the southeast of the verifying in the control than the trial. position for Hurricane Igor, it was a much better forecast. b. Individual cases In some cases the lack of initialization of the TC in the These results indicate overall that switching off the trial resulted in a poorer analysis position, which in turn 2007 scheme reduces TC track forecast errors and has resulted in a poorer forecast. For example, Fig. 4a shows little impact on TC intensity forecasts. However, it is forecast tracks for Hurricane Danielle. The trial had a worth looking at a few individual cases to see how larger initial error than the control, which resulted in a switching off the scheme affects some characteristics of persistent westward bias in the forecast compared to the the MetUM’s predictions. control, which had a better forecast track. The problem Cross-track errors (not shown in Table 1) indicate of poorer analyses without initialization was particularly that control forecasts had an equatorward bias relative an issue during the formative stages of TCs in the to the observed track, particularly in the North At- western North Pacific. Figure 4b shows forecast tracks lantic. Several cases bear this out and show that with- for Tropical Storm Nanmadol at the time when it was a out the initialization scheme the equatorward bias in relatively weak tropical storm. There was a large anal- forecasts is eradicated. For example, Fig. 2 shows ysis error in the position of the storm in the trial, and the

Unauthenticated | Downloaded 09/25/21 08:16 PM UTC 1436 WEATHER AND FORECASTING VOLUME 31

FIG. 2. Control (red circles) and trial (green squares) forecast tracks (24-h steps) for (a) Hurricane Igor from 0000 UTC 11 Sep 2010, (b) Hurricane Katia from 0000 UTC 4 Sep 2011, and (c) Hurricane Maria from 1200 UTC 8 Sep 2011. Corresponding analysis positions are shown as triangles and best-track observed positions as TC symbols (24-h steps). The trial had the 2007 scheme switched off. forecast took Nanmadol toward the north and then c. Conclusion from evaluation of the 2007 scheme northeast, whereas the storm actually moved west and then northwest. The control forecast was much better in The evidence from the trial presented above was this case. However, it should be noted that by 2 days carefully considered in mid-2012 in order to make a later the trial forecast track was slightly better than the decision about the future of the 2007 scheme for control (not shown). initializing TCs. Cyclone Funso developed in the Mozambique Chan- The mean error statistics for TC tracks give a clear nel and was initially slow moving but later accelerated indication that switching off the scheme was beneficial, southward into the open Indian Ocean. Control fore- with an overall reduction of error of 8.4%. This re- casts persistently tracked Funso westward toward the duction in track error was achieved without any sig- coast of Mozambique at a fairly slow pace. The trial nificant detriment to TC intensity forecasts. Individual forecasts were much better both in terms of direction cases show that with the scheme switched off the model and speed of movement. This is illustrated in Fig. 5, performed much better for the recurvature of strong which shows all forecast tracks from 0000 UTC runs of TCs—particularly in the Atlantic. There was benefit the control and trial configurations of the model for seen in both Northern and Southern Hemisphere TC Cyclone Funso. predictions. On the negative side, there was evidence

Unauthenticated | Downloaded 09/25/21 08:16 PM UTC OCTOBER 2016 H E M I N G 1437

FIG. 3. (a) Control (solid red lines) and trial (dashed green lines) 168-h forecast of mean sea level pressure from data time 1200 UTC 14 Sep 2010 for Hurricane Igor. (b) Verifying analysis. Trial had the 2007 scheme switched off. that switching off the scheme sometimes produced following the implementation of the 2007 scheme worse forecasts during the formative stages of TCs, there were further changes to the MetUM, including most notably in the western North Pacific. increases in model horizontal and vertical resolution, As discussed in the introduction, the 2007 scheme increases in data assimilation resolution, the intro- was found to be beneficial to MetUM forecasts of TCs, duction of hybrid four-dimensional variational data but to a lesser degree than when the first configuration assimilation (4DVAR; Clayton et al. 2013), and fur- of the initialization scheme was introduced in 1994. ther improvements in observational coverage and The reduced benefit was likely due to the many im- usage, particularly satellite data. Some of these provements in model formulation, increases in model changes would have acted to constrain the model’s resolution, and the introduction of new observational analysis of TCs and reduce the requirement for arti- data, particularly from satellites, that occurred in the ficial initialization of the low-level wind field to the intervening period. In the few years immediately point that this form of initialization actually became

FIG. 4. Control (red circles) and trial (green squares) forecast tracks (24-h steps) for (a) Hurricane Danielle from 0000 UTC 26 Aug 2010 and (b) Tropical Storm Nanmadol from 1200 UTC 23 Aug 2011. Corresponding analysis positions are shown as triangles and best-track observed positions as TC symbols (24-h steps). The trial had the 2007 scheme switched off.

Unauthenticated | Downloaded 09/25/21 08:16 PM UTC 1438 WEATHER AND FORECASTING VOLUME 31

FIG. 5. (a) Control and (b) trial forecast tracks (colored symbols 24 h apart all valid at 0000 UTC) for Cyclone Funso in January 2012. Best-track observed positions are shown as TC symbols (24-h steps). The trial had the 2007 scheme switched off. detrimental to the model’s analysis and forecast Office (2014) (P. Earnshaw 2015, personal communication) capabilities. illustrates the improved results for ENDGame against its Alongside evidence from the trials of the 2007 predecessor New Dynamics when compared to inde- scheme, consideration was also given to the fact that a pendent analyses from ECMWF for a set of twelve 3-day major model change, which had been in development forecasts. for several years and was planned for implementa- In addition to the new dynamical core, a number of tion within the following two years, was showing other changes were also implemented. There was an signs of performing much better for TC prediction in increase in the model’s horizontal resolution from a initial tests (Heming 2014). In conjunction with this grid spacing at midlatitudes of approximately 25 km change there were also plans to test a completely new (N512) to 17 km (N768). The resolution of the data technique for TC initialization that would be tar- assimilation component was also changed from ap- geted at improving the analysis of TC intensity (dis- proximately 60 km (N216) to 40 km (N320). A package cussed later in this paper). Having considered all of of changes to the satellite data assimilation included these factors, a decision was made to switch off the 2007 scheme for initialization of TCs beginning 17 July 2012.

3. Major upgrade to the MetUM a. Background to the model upgrade Upgrades to the dynamical core of the MetUM typ- ically take many years to develop. In 2002 the ‘‘New Dynamics’’ upgrade was implemented (Davies et al. 2005). The successor to New Dynamics was named ENDGame and was implemented in the MetUM op- erationally in July 2014 (Wood et al. 2014; Met Office 2014). A major benefit from ENDGame was an in- crease in atmospheric variability. This manifested itself in improved detail and intensity of large-scale storms, which arose from the use of less artificial damping in FIG. 6. Globally integrated eddy kinetic energy (EKE) for various resolutions from a set of twelve 3-day ENDGame and the ENDGame formulation. A simple measure of the New Dynamics forecasts initialized from ECMWF analyses. atmospheric variability is provided by the globally in- Comparable EKE results from ECMWF analyses are also tegrated eddy kinetic energy. Figure 6, taken from Met shown.

Unauthenticated | Downloaded 09/25/21 08:16 PM UTC OCTOBER 2016 H E M I N G 1439

TABLE 2. Summary of differences between the control and trials during the assessment of GA6. Positive indicates trial values higher than control. Negative indicates trial values lower than control.

Track error Central pressure 10-m wind Central pressure mean absolute 10-m wind mean absolute change (%) change (hPa) change (kt) error change (hPa) error change (kt) Trial25 27.3 27.1 18.9 23.0 26.7 Trial17 28.6 211.1 113.4 23.6 29.0 reduced thinning of some types of data and improved and satellite data usage changes and with a small addi- usage of some other types. Changes to model physics tional benefit from the resolution increase. included an increase in entrainment rate for deep Results for TC intensity prediction showed that in convection and improvements to several other physical both trials TCs were much stronger and absolute errors parameterization schemes. The complete package was and biases in forecast intensities were reduced com- known as Global Atmosphere 6 (GA6) and is described pared to the control. In Trial25 the mean forecast in detail in work by Walters et al. 2016, manuscript central pressure was 7.1 hPa lower and 10-m winds submitted to Geosci. Model Dev.). 8.9 kt higher than for the control. The mean absolute error in central pressure forecasts was reduced by b. Trial of the model upgrade 3.0 hPa, and the mean absolute error in 10-m wind Prior to implementation of this package of changes, forecasts was reduced by 6.7 kt. In Trial17 the mean the model was trialed using periods from June to forecast central pressure was 11.1 hPa lower and 10-m September and November to December 2012. There winds 13.4 kt higher than in the control. The mean were 36 TCs during these trial periods. The package absolute error in central pressure forecasts was re- of changes was tested with and without the increase in ducedby3.6hPaandthemeanabsoluteerrorin10-m horizontal resolution. The two trials were known as wind forecasts reduced by 9.0 kt. Trial25 (25-km grid spacing) and Trial17 (17-km grid For TC intensity prediction it is also worth examining spacing). Details of the impact of the trial configu- the variation in bias with forecast lead time. Figure 7 is rations on TC forecasts from the MetUM can be taken from Heming (2014) and shows the mean bias in found in Heming (2014) and are summarized in central pressure for the trials. Figure 7 first shows that Table 2. the control (New Dynamics with 25-km grid spacing) For TC track prediction the key result was that Trial25 had a weak bias in the analysis of approximately 15 hPa. reduced track errors by 7.3% and Trial17 reduced track This bias initially increased with lead time, as a result errors by 8.6% relative to the control. Thus, there was a of a tendency to spin up TCs at a slower rate than reality, clear benefit to TC track forecasting from GA6, with before reducing again at longer lead times to a value most of the benefit coming from the dynamics, physics close to 15 hPa. Trial25 had a similar weak bias in the

FIG. 7. Mean TC central pressure bias during the trial of GA6: Trial25, model upgrade without horizontal resolution increase; Trial17, model upgrade with horizontal resolution increase.

Unauthenticated | Downloaded 09/25/21 08:16 PM UTC 1440 WEATHER AND FORECASTING VOLUME 31

TABLE 3. Results of the 2015 scheme control versus trial. Mean track and intensity forecast statistics, where boldface indicates the better score. The 6-hourly statistics are calculated, but only the 24-hourly results are shown. The t values and significance levels (%) of track error differences are shown. The trial included use of the 2015 scheme.

0 h 24 h 48 h 72 h 96 h 120 h 144 h No. of cases 417 324 241 83 58 36 24 Control track errors superior 166 143 96 43 27 15 8 Trial track errors superior 186 155 134 39 29 19 16 Track errors equal 65 26 11 1200 Control track error (km) 38 72 125 185 270 439 637 Trial track error (km) 35 68 116 179 257 394 572 Reduction in track error of trial relative to 7.9 4.8 7.0 3.3 5.1 10.3 10.2 control (%) t value of track error differences 21.927 20.816 22.165 20.550 20.935 20.949 21.267 Significance level of track error differences (%) 2.7 20.8 1.6 29.2 17.7 17.4 10.8 Control track skill score (%) — 55 68 70 — — — Trial track skill score (%) — 57 71 72 ——— Control central pressure bias (hPa) 11.5 12.8 12.1 9.4 3.7 0.9 2.3 Trial central pressure bias (hPa) 6.3 9.5 10.2 8.2 2.9 1.2 1.4 analysis, but the bias reduced with increasing forecast increasing lead time than for Trial25. In fact, the bias lead time to a value of just below 5 hPa by 144 h. For actually became negative at lead times of 120 h and Trial17 there was again a large weak bias in the above (i.e., the forecast TCs were too intense on analysis, but the bias reduced at a faster rate with average).

FIG. 8. (a) TC track forecast errors and (b) TC central pressure forecast bias during the trial of the 2015 scheme. The control and trial used the GA6 configuration of the MetUM. The trial included the 2015 scheme. The operational results were for an earlier model configuration.

Unauthenticated | Downloaded 09/25/21 08:16 PM UTC OCTOBER 2016 H E M I N G 1441

FIG. 9. Control (red circles) and trial (green squares) forecast tracks (24-h steps) from data time 1200 UTC 23 Oct 2013 plotted against best-track observed positions (24-h steps) for Hurricane Raymond. Corresponding analysis positions are shown as triangles. The trial in- cluded the 2015 scheme.

GA6 was operationally implemented in the MetUM have the potential to reduce the model’s weak bias while on 15 July 2014. The impact of GA6 on operational TC allowing the model to make its own balanced adjust- forecasts is discussed in section 5. ments to the wind structure. Thus, a new form of TC initialization was developed and tested, which is here- after referred to as the 2015 scheme. 4. TC initialization using central pressure estimates b. Formulation of the 2015 scheme a. Background to the new initialization technique TC warning centers around the globe (e.g., In section 2, we saw that switching off the 2007 scheme Meteorological Agency, National Hurricane Center) in the MetUM was beneficial to TC track forecasting produce estimates of the position and structure of all overall with no detriment to TC intensity. Two years active TCs every 3 or 6 h. These include estimates of after switching off the initialization scheme, a major and central pressure. While in model upgrade (GA6) resulted in a further reduction in most cases these are estimates (i.e., not directly mea- TC track forecast errors as well as a significant reduction sured) based on a combination of techniques such as in TC intensity forecast bias—particularly at longer lead Dvorak (1975, 1984) and Knaff and Zehr (2007), they times, as explained in section 3 above. However, it was provide information that is potentially of value to nu- notable that GA6 had virtually no impact on the model’s merical models. While the 2007 scheme made use of the weak bias in the analysis for TCs. While the intensity estimates of maximum sustained wind and radii of 34-, bias at short lead times was reduced by GA6, TC pre- 50-, and 64-kt winds provided by TC warning centers, dictions were handicapped by having to spin up from a the estimates of central pressure had remained unused relatively weak analysis. Thus, attention was turned to by the MetUM. how this weak bias could be reduced, possibly through The 2015 scheme was designed to ingest estimates of the development of a new form of TC initialization. the central pressures of all active TCs from a variety of Given that there was clear evidence that the 2007 TC warning centers around the globe. These are avail- scheme, which involved the generation of lower- able at 6-hourly intervals and sometimes at 3-hourly tropospheric wind observations, was now detrimental intervals. It was considered that assimilating a single to TC track forecasts in the MetUM, it was decided not central pressure observation every 6 h may have a lim- to return to a form of initialization that involved direct ited impact on the MetUM analysis or forecast. Hence, adjustments to the wind structure of the model. How- the scheme was designed to produce hourly values of TC ever, usage of central pressure estimates appeared to central pressure. These are based on a combination of

Unauthenticated | Downloaded 09/25/21 08:16 PM UTC 1442 WEATHER AND FORECASTING VOLUME 31

FIG. 10. (a) Control and (b) trial vertical cross sections of the analyzed u component of the wind for Hurricane Raymond valid at 1200 UTC 23 Oct 2013. The trial included the 2015 scheme. interpolation and extrapolation of the estimates from observations, including passing through quality control TC warning centers. For example, the 1200 UTC run of procedures. the MetUM, which uses a hybrid 4DVAR data assimi- c. Trial of the 2015 scheme lation system (Clayton et al. 2013), has an observational time window from 0900 to 1459 UTC. If estimates of The trial period chosen was 26 September–12 November position and central pressure for an active TC are 2013. In total there were 23 TCs during the trial period, available at 0600 and 1200 UTC, these are used to derive comprising 13 in the western North Pacific, 5 in the estimates of central pressure at 0900, 1000, and 1100 UTC eastern North Pacific, 3 in the Atlantic, and 2 in the north by linear interpolation. The 1200 UTC value from the Indian Ocean. Forecasts were run to 144 h at 0000 and warning center is used directly and values for 1300 and 1200 UTC and in addition forecasts were run to 48 h at 1400 UTC are derived by linear extrapolation. Hence, 0600 and 1800 UTC. The numbers of forecasts verified at six values of TC central pressure and position are each forecast lead time were 324 at 24 h, 83 at 72 h, and 24 presented for data assimilation during the 6-h time at 144 h. Although the trial period was from 2013, the window and are used in the production of a model control and trial simulations used the configuration of analysis, being treated in a similar way to other surface the MetUM operational from July 2014, which included

Unauthenticated | Downloaded 09/25/21 08:16 PM UTC OCTOBER 2016 H E M I N G 1443

FIG. 11. (a) Control and (b) trial vertical cross sections of the 96-h forecast u component of the wind for Hurricane Raymond valid at 1200 UTC 27 Oct 2013. The trial included the 2015 scheme. the major model upgrade described in section 3 (GA6). the reduction in track errors was significant beyond the The GA6 configuration of the model was used since it 2% level, but at other lead times the reduction was was vital to assessing the 2015 scheme against this significant only between the 10% and 30% levels. The baseline, which was already in operation at the time the trial track forecast skill scores were on average 2.7% trials were conducted. Table 3 shows the various verifi- higher than the control. The frequency of superior per- cation scores for TC track and intensity prediction for formance for TC track forecasts shows that the trial was the control and trial runs. Scores were calculated at superior in over 50% of all forecasts compared to less 6-hourly forecast intervals, but only the 24-hourly values than 43% for the control. Track forecast errors for the are shown in Table 3. control and trial are shown in Fig. 8a and also include the The results for TC track show that forecast errors values for the configuration of the model that was op- were lower in the trial at all lead times. When averaged erational during the trial period (i.e., the version before over all forecasts from 6 to 144 h at 6-hourly intervals, the major model upgrade described in section 3). This the trial track forecast errors were 6.2% lower. How- illustrates the combined impact of GA6 and the ever, statistical significance was mostly not high. At 48 h 2015 scheme.

Unauthenticated | Downloaded 09/25/21 08:16 PM UTC 1444 WEATHER AND FORECASTING VOLUME 31

FIG. 12. Control (solid red lines) and trial (dashed green lines) central pressure forecasts plotted against the best-track observed data (solid blue line) for (September 2013). The trial included the 2015 scheme.

For TC intensity, the statistics indicate that the weak these two changes now ranged between approximately bias in the analysis was markedly reduced in the trial 0 and 10 hPa. with the central pressure bias of 11.5 hPa cut to 6.3 hPa. d. Case studies With increasing lead time the difference between the control and trial narrowed and beyond 108 h both had a Examination of some individual cases illustrates some bias very close to zero. Thus, the introduction of as- of the characteristics of the 2015 scheme. similation of central pressure estimates has reduced the 1) HURRICANE RAYMOND weak bias in the analysis and short-lead-time forecasts without resulting in an overdeepening at longer lead The reduction in track forecast errors is exemplified times when the control already had a very small bias. in a forecast for Hurricane Raymond in the eastern Figure 8b shows these results together with the central North Pacific from data at 1200 UTC 23 October 2013, pressure bias for the configuration of the model, which shown in Fig. 9. The control forecast had a fast westward was operational during the trial period. This shows that movement for the hurricane whereas the trial had a the combination of GA6 and the 2015 scheme slashes slower movement with a gradual curve toward the north. the bias in forecast central pressure. Biases that The latter matches the observed track far better and thus ranged between approximately 17 and 28 hPa before produced much lower track forecast errors. At the

FIG. 13. Control (solid red lines) and trial (dashed green lines) central pressure forecasts plotted against the best-track observed data (solid blue line) for Typhoon Wipha (October 2013). The trial included the 2015 scheme.

Unauthenticated | Downloaded 09/25/21 08:16 PM UTC OCTOBER 2016 H E M I N G 1445

FIG. 14. Control (solid red lines) and trial (dashed green lines) central pressure forecasts plotted against the best-track observed data (solid blue line) for (November 2013). The trial included the 2015 scheme. analysis time in this case the observed central pressure the observed intensity and even being too deep in a was 995 hPa. The control analysis had a central pressure couple of runs. The resulting forecasts had central of 1003 hPa whereas the trial had a central pressure of pressures much closer to the observed values in most 996 hPa. The lower central pressure in the trial was ac- cases, as seen in Fig. 12. companied by a stronger and vertically deeper vortex, as For Typhoon Wipha (Fig. 13) the control analysis was seen in the cross sections of the u component of the wind unable to represent the intensity of the TC and by shown in Fig. 10. The difference in vortex depth between 1200 UTC 13 October 2013 had a central pressure of the control and trial resulted in differences in the 969 hPa compared to an observed value of 930 hPa. In steering level. Thus, the control and trial forecast tracks contrast the trial analysis central pressure was 935 hPa. started to diverge. By 96 h into the forecast the control Consequently, trial forecasts of central pressure were was erroneously tracking Raymond westward whereas much better in most cases. However, this case illustrates the trial was correctly turning it to the northwest and a known characteristic of the MetUM in that it tends to then north. The cross sections of the u component of the continue deepening the TCs beyond the point at which wind at this time (Fig. 11) show that the control had a they reach their peak intensity in reality as they move weak, shallow, and sheared vortex whereas the trial into the subtropics. This resulted in an overdeepening in had a strong and vertically deep vortex. In reality, some forecasts, particularly for the trial. There were Hurricane Raymond was going through a period of in- also a couple of trial analyses that had large negative tensification at this time and had developed a strong and central pressure biases (i.e., low centers too deep). This deep vortex, as in the trial forecast. Clearly, in this case can happen as a result of assimilating central pressure the assimilation of central pressure estimates helped observations that have observation minus background develop a stronger vortex and a more accurate vertical values larger (in absolute terms) than the difference structure and steering level, which in turn resulted in a between the observed and background central pressure better forecast track. because of a positional error in the location of the TC in the background field. 2) TYPHOONS WUTIP AND WIPHA 3) TYPHOON HAIYAN The impact of the 2015 scheme on TC intensity can be seen in the central pressure predictions for Typhoons Typhoon Haiyan devastated parts of the central Wutip and Wipha. During the period from 27 to Philippines and was likely the most intense TC recorded 29 September 2013 Typhoon Wutip deepened from 1000 at landfall (Lander 2014). It occurred in November 2013, to 965 hPa. The control analysis did not keep pace with which fell during the latter part of the trial period for the the rate of deepening and the subsequent forecasts were 2015 scheme. The control and trial forecasts of the track consequently unable to predict the intensity of the ty- of Typhoon Haiyan were both very good with 72–96-h phoon. However, the assimilation of central pressure forecasts of landfall having an error around 100 km. estimates resulted in the analysis being much closer to Typhoon Haiyan went through rapid intensification with

Unauthenticated | Downloaded 09/25/21 08:16 PM UTC 1446 WEATHER AND FORECASTING VOLUME 31 the central pressure dropping 60 hPa in 24 h and 87 hPa in 48 h. The MetUM was unable to simulate these ex- treme rates of deepening, even over a period of 6 h—the length of forecast used as the ‘‘background’’ for the next model cycle. Thus, the observation minus background values for the central pressure observations created by the 2015 scheme were very large. These observations are subject to the same quality control procedures as other conventional observations. On this basis, once rapid intensification of Typhoon Haiyan was under way (around 0600 UTC 5 December 2013) all central pres- sure observations were flagged because of the large observation minus background values and were not as- similated into the model. The consequence of this is that the MetUM predictions of the central pressure from about 12–24 h after this point were no better in the trial than the control, as seen in Fig. 14. e. Conclusion from trial of the 2015 scheme The 2015 scheme was developed primarily with the aim of reducing the weak bias in model forecast in- tensities at short lead times. Evidence from the trial indicates that this was achieved without causing a sig- nificant overdeepening at longer lead times. The 2015 scheme was not developed with the primary aim of reducing TC track forecast errors. However, the trial track forecast errors were reduced by 6.2%, although the statistical significanceofthisresultwasmostly not high. The issue of central pressure observations being flagged by quality control when TCs rapidly intensify was seen in the case of Typhoon Haiyan and several other cases in the trial. While this is not ideal, it is con- sidered that in a case such as Typhoon Haiyan when the central pressure value was close to 900 hPa and the pressure gradient near the center of the TC was near 42 hPa over one model grid length (17 km) (Morgerman 2014), it may not be appropriate to assimilate central pressure values in the current configuration of the MetUM. This is a matter that will be the subject of further investigation and experimentation. The evidence from the trial presented above was considered in late 2014. Given the positive results overall the decision was made to implement the 2015 scheme in the MetUM on 3 February 2015.

5. Operational impact of GA6 and the 2015 scheme FIG. 15. Timeline of model changes and TC initialization in the MetUM. Figure 15 shows the sequence of changes to TC ini- tialization and the MetUM that have been discussed GA6 was implemented in the MetUM on 15 July 2014. thus far. Examination is now made of the impact of the Thus, most of the 2014 Northern Hemisphere TC season last two of these changes on operational forecasts of TCs occurred after the implementation date. The 2015 in the MetUM. scheme was implemented on 3 February 2015 meaning

Unauthenticated | Downloaded 09/25/21 08:16 PM UTC OCTOBER 2016 H E M I N G 1447

FIG. 16. The 5-yr running mean of MetUM Northern Hemisphere TC track forecast errors. the whole of the 2015 Northern Hemisphere TC season At longer lead times (beyond 96 h), the bias was very occurred after the implementation date. Hence, a time close to zero. series of Northern Hemisphere TC forecast errors from On an international level the Met Office participates the MetUM gives a good perspective on the impact of in the World Meteorological Organization’s Working these two changes after they became operational. Group on Numerical Experimentation (WGNE), which The mean TC track forecast error for Northern has undertaken an annual comparison of numerical Hemisphere TCs in 2014 was almost 25% lower than the model TC forecasts since the early 1990s. Results are mean for the previous five seasons (2009–13) and in 2015 available to the end of 2014, which includes the period was a further 3.2% below the 2014 figure. Even when immediately after implementation of GA6 in the examining the 5-yr running mean of TC track forecast MetUM. The MetUM performed well in 2014 against errors, which normally smooths out large interannual other global models, particularly in the western North variability, there was still a sharp drop in the errors in Pacific region. Figure 18 (taken from WGNE’s ‘‘Inter- 2014–15, as seen in Fig. 16. comparison of Tropical Cyclone Track Forecasts Using Figure 17 shows the Northern Hemisphere TC central Operational Global Models’’ website) is a time series of pressure bias for the years 2011–15. In 2014, although 72-h track forecast errors for this region. It shows that analyses were still too weak, the central pressure bias the MetUM (labeled UKMO in Fig. 18) outperformed dropped steadily with forecast lead time to a value close all other models in 2014 including that of the ECMWF, to zero by 168 h. In 2015 the bias in the analysis was which has been the best-performing model in recent years. more than halved compared to previous years, and These operational results support the results seen in the bias in the forecast was also markedly reduced. the separate trials of both GA6 and the 2015 scheme.

FIG. 17. Mean TC central pressure bias in the Northern Hemisphere for the MetUM.

Unauthenticated | Downloaded 09/25/21 08:16 PM UTC 1448 WEATHER AND FORECASTING VOLUME 31

FIG. 18. The 72-h TC track forecast errors in the western North Pacific for global models (taken from WGNE’s ‘‘Intercomparison of Tropical Cyclone Track Forecasts Using Opera- tional Global Models’’ website). The MetUM is labeled as UKMO.

6. Conclusions Davies, T., M. J. P. Cullen, A. J. Malcolm, M. H. Mawson, A. Staniforth, A. A. White, and N. Wo, 2005: A new dynamical Between 2012 and 2015 three changes were made to core for the Met Office’s global and regional modelling of the the MetUM that had large impacts on TC prediction. atmosphere. Quart. J. Roy. Meteor. Soc., 131, 1759–1782, First, the old TC initialization scheme (2007 scheme) doi:10.1256/qj.04.101. was switched off. There was then a major model upgrade Dvorak, V., 1975: Tropical cyclone intensity analysis and fore- casting from satellite imagery. Mon. Wea. Rev., 103, 420–430, to the model dynamics, physics, resolution, and satellite doi:10.1175/1520-0493(1975)103,0420:TCIAAF.2.0.CO;2. data usage (GA6). Finally, a new form of TC initializa- ——, 1984: Tropical cyclone intensity analysis using satellite data. tion was introduced involving the assimilation of central NOAA Tech. Rep. NESDIS 11, 47 pp. [Available from pressure estimates (2015 scheme). Results presented in NOAA/NESDIS, 5200 Auth Rd., Washington, DC 20233.] this paper show that each of these changes resulted in a Gross, J., 1996: 1995 National Hurricane Center forecast ver- significant reduction in TC forecast errors (for track, ification. National Hurricane Center, Miami, FL, 14 pp. [Available online at http://www.nhc.noaa.gov/verification/ intensity, or both) in forecasts from the MetUM. pdfs/Verification_1995.pdf.] Issues that require further research include how the Heming, J. T., 2009: Evaluation of and improvements to the Met 2015 scheme handles rapidly intensifying TCs and the Office tropical cyclone initialisation scheme. Meteor. Appl., MetUM’s propensity on occasions to continue deepen- 16, 339–351, doi:10.1002/met.129. ing TCs beyond their actual points of peak intensity as ——, 2012: Tropical cyclone predictions from the Met Office and ECMWF global models - Relative performance and error di- they move into the subtropics. This will be undertaken agnosis. Proc. 30th Conf. on Hurricanes and Tropical Meteo- against a backdrop of wider model development that in rology, Ponte Vedra Beach, FL, Amer. Meteor. Soc., 4A.6. the coming years will include the development of the [Available online at https://ams.confex.com/ams/30Hurricane/ convective parameterization, further increases in hori- webprogram/Paper205026.html.] zontal and vertical resolution, and coupling to the ocean, ——, 2014: The impact on tropical cyclone predictions of a major all of which are likely to have an impact upon the TC upgrade to the Met Office Global Model. Proc. 31st Conf. on Hurricanes and Tropical Meteorology, San Diego, CA, Amer. forecast performance of the MetUM. Meteor. Soc., 11A.3. [Available online at https://ams.confex.com/ ams/31Hurr/webprogram/Paper243428.html.] REFERENCES ——, 2016: The method used to verify the forecast tracks of tropical cyclones in Met Office numerical weather prediction models. Clayton, A. M., A. C. Lorenc, and D. M. Barker, 2013: Operational Meteor. Appl., in press. implementation of a hybrid ensemble/4D-Var global data as- ——, and A. M. Radford, 1998: The performance of the United similation system at the Met Office. Quart. J. Roy. Meteor. Kingdom Meteorological Office global model in predicting Soc., 139, 1445–1461, doi:10.1002/qj.2054. the tracks of Atlantic tropical cyclones in 1995. Mon. Wea.

Unauthenticated | Downloaded 09/25/21 08:16 PM UTC OCTOBER 2016 H E M I N G 1449

Rev., 126, 1323–1331, doi:10.1175/1520-0493(1998)126,1323: Met Office, 2014: ENDGame: A new dynamical core for seamless TPOTUK.2.0.CO;2. atmospheric prediction. [Available online at http://www.metoffice. ——, J. C. L. Chan, and A. M. Radford, 1995: A new scheme for the gov.uk/research/news/2014/endgame-a-new-dynamical-core.] initialisation of tropical cyclones in the UK Meteorological Morgerman, J., 2014: iCyclone chase report: Super Typhoon Office global model. Meteor. Appl., 2, 171–184, doi:10.1002/ HAIYAN in Tacloban City & Leyte, Philippines. iCyclone. met.5060020211. com Chase Rep., 47 pp. [Available online at http://www.icyclone. Knaff, J. A., and R. M. Zehr, 2007: Reexamination of tropical com/upload/chases/haiyan/iCyclone_HAIYAN_in_Tacloban_ cyclone pressure–wind relationships. Wea. Forecasting, 22, 71– City_040314.pdf.] 88, doi:10.1175/WAF965.1. Neumann, C. J., 1972: An alternative to the HURRAN tropical Lander, M. A., 2014: Super Typhoon Haiyan’s 170 kt peak in- cyclone forecast system. Mon. Wea. Rev., 100, 245–255, tensity: Has Super been dethroned? Proc. doi:10.1175/1520-0493(1972)100,0245:PAOTHT.2.3.CO;2. 31st Conf. on Hurricanes and Tropical Meteorology,San Wood, N., and Coauthors, 2014: An inherently mass-conserving Diego, CA, Amer. Meteor. Soc., 9C.8. [Available online at semi-implicit semi-Lagrangian discretization of the deep- https://ams.confex.com/ams/31Hurr/webprogram/Paper245225. atmosphere global non-hydrostatic equations. Quart. J. Roy. html.] Meteor. Soc., 140, 1505–1520, doi:10.1002/qj.2235.

Unauthenticated | Downloaded 09/25/21 08:16 PM UTC