The forecasting challenge of waving cold fronts: benefits of the ensemble approach

Transcription

1 The forecasting challenge of waving cold fronts: benefits of the ensemble approach Martin V. Young Tim D. Hewson Met Office, Exeter Introduction One of the major challenges confronting forecasters is predicting the behaviour of slow-moving cold fronts. Frontal waves present particular problems because numerical weather prediction (NWP) models commonly provide conflicting advice on their development, even just 24 hours in advance. The waves induce erratic movement of the frontal zone as a whole, and if the front is active (i.e. accompanied by heavy rain) then even relatively small positioning errors for the front can lead to major errors in forecasts (notably of rain and wind) for specific locations. A further complication is that, in certain active situations, a frontal wave may develop further into an intense small-scale depression. These were precisely the concerns on 28 October 2010 when an active cold front was predicted for the UK area. This paper illustrates how a range of NWP products from various deterministic and ensemble runs were utilized operationally to provide a consistent forecast, and how the forecaster managed to reconcile key differences in the forecast story that the different runs portrayed, thereby adding value to raw model output. The importance of differencetracking techniques and satellite imagery interpretation are also highlighted. Models used operationally at the Met Office Here we will first describe the NWP model runs that are routinely used by Met Office (c) forecasters, so as to provide context for the discussion in the following sections. In October 2010 the operational workhorse for short-range (1 2 day) forecasting at the Met Office was the North Atlantic European model (NAE). This was a limited-area model, with 12km resolution, that ran four times daily out to 48 hours ahead, from 0000, 0600, 1200 and 1800 UTC. NAE output was (d) 296 Figure 1. Met Office operational surface analysis at 0600 UTC on 28 October Arrows highlight the parent low (top) and the formative, diminutive, wave. Figure 2. Output from various numerical models along with verifying imagery, all valid at 1800 UTC on 29 October hour NAE forecast of sea-level pressure and rainfall rate from data time 0600 UTC on 28 October As but for the GM. (c) MSG infra-red image with superimposed radar rainfall image, the analysed sea-level pressure field, and mesoscale low centres (L). (d) 42-hour ECMWF forecast of sea-level pressure, and accumulated rainfall from 1500 to 2100 UTC. In and symbols represent rainfall type: circles for dynamic rain, Vvs and triangles for convective rain, with larger symbols for greater intensities. Symbol colour represents rainfall rate (mmh 1 ): blue is 0 1, green 1 2, yellow 2 4, orange 4 8 and red In (c) rates are: blue < 0.5, yellow 0.5 1, green 1 2, pink 2 4, red 4 8 and white 8 16.

2 The forecasting challenge of waving cold fronts Figure 3. Cyclonic feature points (the so-called dalmation plot, after Hewson, 2009b) from the ECMWF Integrated Forecasting System, from data time 0000 UTC on 28 October 2010, valid at 1200 UTC on 29 October (main Figure) and 0000 UTC 30 October 2010 (inset). Spots denote cyclonic feature positions (barotropic lows, frontal waves, and diminutive frontal waves) in 52 model runs. Colours show the wind maximum (kn) at 1km altitude found within a 600km radius of each feature point in each respective model run the legend gives values: in certain circumstances gusts realised at the surface could mirror these values. Pressure field is from the control run. Black dots denote barotropic low centres. Yellow circles/crosses denote respectively features from the control/deterministic model runs. Small black crosses show verifying feature points i.e. where the manual analyst eventually placed the parent low (large) and the tips of cold front waves (small) discussed in the text. complemented by output from the lower resolution (25km) global model (GM) that ran essentially to the same schedule, but with output not available until a little later than for the NAE. The GM runs out to 144 hours ahead at 0000 and 1200 UTC and 48 hours ahead at 0600 and 1800 UTC. The third most used model was probably the ECMWF (European Centre for Medium Range Weather Forecasts) deterministic global model, which had 16km resolution and ran twice daily (from 0000 and 1200 UTC). The forecaster also had access to many other deterministic models from around the world, in a variety of formats. This poor man s ensemble (i.e. the use of different deterministic models as a proxy for an ensemble) was complemented by genuine ensemble output, from two systems in particular. One system was MOGREPS (the Met Office Global and Regional Ensemble Prediction System), which in regional format ran from data times 0600 and 1800 UTC, at 24km resolution, with 24 members; the other was the ECMWF ensemble (0000 and 1200 UTC global runs at 32km resolution, with 51 members). Finally, there were also two high- resolution Met Office models (1.5km and 4km resolution) that could add useful detail, for example relating to convection and to coastal and topographic modulation of weather. These models nominally start from 0300, 0900, 1500 and 2100 UTC and run out to 36 hours ahead, but cover only the UK, taking boundary conditions from the NAE at the time of this case study. Meteorological analysis The synoptic analysis for 0600 UTC on 28 October 2010 is shown in Figure 1. Note the developing depression centred near 51 N 23 W with an associated cold front trailing southwestwards (both features are arrowed). The depression was undergoing rapid deepening whilst moving northnortheast. All of the available NWP models predicted that the associated cold frontal zone would move northeast into the UK on 29 October, many showing it as an active rain-bearing feature. The main differences between model predictions concerned the position and intensity of smaller-scale wave developments running along the front. Arguably the first signs of a Figure 4. ECMWF ensemble probabilities of 24-hour precipitation > 12mm and > 25mm during 29 October The run data time is 0000 UTC on 28 October

3 The forecasting challenge of waving cold fronts 298 developing wave are just apparent on the cold front nearest to the southernmost arrow, where the isobars are opening out such a feature is referred to in Hewson (2009a) as a diminutive wave. Forecasting the developments Put yourselves now in the position of the Chief Forecaster, at 0800 UTC on the 28th, which is about the time that the 0600 UTC run of the NAE becomes available. This presented a solution that was rather different to most Met Office model runs seen previously, showing a cold front wave for the following day that was much less marked and that took a (northnortheasterly oriented) track rather further east over the UK, crossing Wales and then northern England instead of Northern Ireland then Scotland see Figure 2. Many of the earlier Met Office model runs (not shown) broadly resembled the 0600 run of the GM, represented in Figure 2. The wave track of these earlier solutions also received some support from the 0000 UTC run of the ECMWF deterministic model (Figure 2(d)), though in this the wave and attendant rain were a little further east. So what was the Chief Forecaster to do? The major variations in successive NAE runs clearly presented a problem in determining the most likely outcome. When a new run steps out of line, one of four decisions can be made: follow the new run, revert to the previous run, (c) compromise between the two, or, rarely, (d) extrapolate the change, if it is thought that the new run is playing catch up. Remember also that, on average, the latest run of any model will be better than the previous run of that model, so approach is preferred most of the time. When option, (c) or (d) is preferred, various field-modification tools, as described in Carroll and Hewson (2005), are employed to generate the issued guidance products. In this case several factors suggested that the forecaster should essentially revert to a previous run (approach above). These were: 1. The 0000 UTC ECMWF and GM runs were in good agreement with each other and provided contrary guidance to the 0600 UTC NAE, as discussed above. 2. Automatically-generated cyclone database products from the ECMWF ensemble (Figure 3) and from MOGREPS (not shown but similar) displayed a strong signal for a wave just south of Ireland at 1200 UTC on 29 October. Relative to this the 0600 UTC NAE s wave position, in the Bristol Channel (not shown but near 51 N 5.5 W), was an outlier. 3. With regard to weather, the ECMWF Ensem ble Prediction System (Figure 4) provided a strong signal for a corridor of heavy rain passing across Northern Figure 5. Examples of individual ensemble member forecasts of sea-level pressure from the Met Office and ECMWF showing extreme solutions in which a small vigorous cyclonic development occurs on the frontal zone. Validity times are shown next to each figure part. also shows, after Hewson and Titley (2010), objectively identified fronts (red for warm and blue for cold) and cyclonic features (black for barotropic lows, orange for frontal waves and green for diminutive waves) and the 546 and 528dm 1000 to 500mbar thickness lines (dash-dot, green and blue respectively). Figure 6. A low probability forecaster-generated alternative solution, valid at 1200 UTC on 29 October 2010, which could potentially give rise to force 9 or 10 winds. Figure 7. Modified fields issued by the Chief Forecaster at 0910 UTC on 28 October 2010, valid at 1500 UTC, 1800 UTC and 2100 UTC on 29 October 2010 as labelled. Format as in Figure 2 (a,b).

4 Ireland and southwestern Scotland. This is consistent with most of the wave tracks that one can infer from Figure 3. Commonly, the largest rainfall totals are found just to the left (west, here) of the track. Another important feature was the markedly confluent upper trough following the front, which would help provide forcing for ascending motion and hence rainfall along the front. This configuration was expected to lead to classic rearward-sloping ascent and line convection at the cold front (Young et al., 1987). Such an upper air pattern is often only conducive to mobile small-scale frontal waves. However, a weakening of the frontogenetic deformation pattern could still allow one of the waves to develop markedly, as suggested by a minority of ensemble members (Figure 5). On Figure 3 inset note the orange and dark green spots, suggesting that gusts of 55 65kn (Beaufort force 9 or 10) could be associated with the system. Accordingly, a low probability of gale-force winds was introduced into the forecasts issued on the morning of the 28th for 29 October. Indeed, on the afternoon of the 28th supplementary guidance provided to the forecasting team included the plot shown in Figure 6, described as a lowprobability alternative solution that could potentially lead to force 9 or 10 winds south of the developed wave. The consistent theme in the vast majority of models was for heavy rain across both southwest Scotland and Northern Ireland. This contrasted markedly with the 0600 UTC NAE solution which cleared the front through Northern Ireland prior to any wave development, which would result in predominantly dry conditions there through 29 October (Figure 2). The issued guidance to forecasters produced by the Chief Forecaster (shown in Figure 7) opted for the majority solution in which marked wave development would bring heavy persistent rain across the east of Northern Ireland. Since the higher resolution models were driven by the NAE their output exhibited the same problems as it did, and so on this occasion these (0900 UTC) runs also had to be discarded. After the main morning guidance was issued by the Chief Forecaster to deadline (by 0910 UTC on 28 October), the 0600 UTC GM run output became available, and this supported the decision made, showing an evolution which took the wave further west, similar to the earlier NAE and GM runs (Figure 2). Reasons for the differing model solutions When marked differences emerge between model solutions, difference-tracking techniques are regularly invoked to identify the source at an early stage. In this case, these techniques, applied to the 0600 UTC NAE and GM runs at 500mbar and at the surface, failed to reveal any noteworthy analysis differences in the vicinity of the incipient wave (e.g. Figure 8, by the lowermost arrow) indicating that either the development was highly sensitive to imperceptibly small errors in initial conditions or that the eventual differences propagated from elsewhere. Even 12 hours later (Figure 8), differences in the vicinity of the developing wave were very small whereas much greater (and growing) differences were approaching Ireland emanating from the major cyclogenesis to the west (also arrowed: this is the parent low shown near 51 N, 23 W on Figure 1). By the time the developing wave was approaching the UK from the southwest there were major differences between the NAE and GM mean sea-level pressure (MSLP) pattern over the west and north of the British Isles (Figures 8(c) and (d)). The subsequent behaviour of the wave would then be determined largely by the flow pattern into which it propagated. In the NAE the confluent MSLP pattern and the relatively deep parent low suggest compression of the wave in the cross-front direction, quelling development (Figure 9); in the GM case, with the parent low more remote, this compression would be smaller, permitting wave development if other factors were favourable (Figure 9). These differences in compression-forcing stemmed also from differences in the life cycles of the parent low in the GM this low deepens and then fills sooner than it does in the NAE, so that by the time the wave arrives not only is the parent low in the GM more remote, it is also weakening more rapidly, which amplifies the effect. Renfrew et al. (1997) provide further discussion of the compression mechanism, which they refer to as stretching by the environmental flow. This all suggests that the major differences in the handling of the wave over the UK between the NAE and GM arose not really because of analysis vdifferences in its source region, but more from initial differences in the treatment of the major parent cyclone, which amplified with time and which defined the background flow pattern of the region in which the incipient wave subsequently found itself. Hence the model with the more realistic handling of the major depression would be more likely to deliver the most reliable handling of the frontal wave. The forecasting challenge of waving cold fronts (c) (d) Figure 8. Difference fields (coloured) of sea-level pressure for GM minus NAE runs from a data time of 0600 UTC on 28 October: shows the analysis differences, whilst, (c) and (d) show the forecast differences at lead times of 12, 24 and 36h respectively. Solid contours are GM and dashed contours are NAE. Yellow/orange shading represents positive differences (i.e. GM pressure higher than NAE) whilst blue shading represents negative differences (i.e. GM pressure lower than NAE), as on the legend (in mbar). Figure 9. Schematic of the lower tropospheric flow pattern around a parent low that would permit only limited development of a cold front wave (wave compressed along the flow) and greater development of such a wave. 299

5 The forecasting challenge of waving cold fronts 300 Figure 10. GM 6-hour forecast of sea-level pressure valid at 1200 UTC on 29 October 2010 along with surface observations in red, mauve and white. Bogus observations inserted by the forecaster based largely on GM fields are shown in cyan. Figure 11. MSG infra red satellite imagery with NAE fields and marine observations superimposed, all valid at 1200 UTC on 29 October Red contours are the NAE 6-hour forecast whilst cyan contours are the NAE analysis. Dashed black line denotes the axis of the hook-shaped cloud feature which usually accompanies the strongest pressure gradients in the rear quadrant of the depression. Significant differences were in fact already present in the position of the major depression centre near 50 N 23 W in the analyses at 0600 UTC on the 28th (Figure 8). Lack of surface data at 0600 UTC in the vicinity of the depression meant that the NAE and GM were largely following their predecessor (0000 UTC) solutions at the surface and, in addition, perhaps because of some upper air data available to the GM due to its later data cut-off time, the differences began to magnify, such that by 1200 UTC the GM solution had the depression centre almost 2 latitude further north than the NAE. The GM solution also fitted more closely with satellite imagery. To attempt to Figure 12. Met Office 24-hour forecast synoptic chart valid at 1200 UTC on 29 October get the 1200 UTC run of the NAE back on track, regarding the position of the depression, heavy intervention was carried out in advance of the model run. This involved creation of so-called bogus or manufactured observations using the GM s 6-hour forecasts of pressure and 10m winds as a framework (Figure 10). These manufactured observations would then be assimilated into the 1200 UTC run of the NAE. This intervention strategy was also thought to be necessary because surface data around the low, which can define its structure, was again lacking at 1200 UTC, as it had been at 0600 UTC (note the absence of genuine surface observations nearby on Figure 10). The NAE 1200 UTC surface analysis (blue lines on Figure 11) responded as expected to the intervention, and was much improved compared to its 6-hour forecast for the same time (red lines on Figure 11). Note how the blue depression centre is more distinctly located within the dry (relatively cloud-free) slot, with the tight MSLP gradient to the south and west more coincident with the hook in the lower cloud (dashed black line). Indeed image animations showed the better match of the 1200 UTC analysis even more clearly, because flow patterns can be inferred from cloud movement. The 1200 UTC NAE forecast went on to produce some modest wave developments which brought some rain back into Northern Ireland. However, these waves were considered insufficiently developed when compared with other model output (perhaps because the upper air pattern had not responded sufficiently to the intervention) so the forecast products again had to be adjusted to show a more developmental solution for the wave, similar to Figure 2, and encapsulated in the forecast synoptic chart for 1200 UTC on 29 October (Figure 12). What happened Figure 2(c) shows the satellite and radar imagery for 1800 UTC on 29 October, whilst

6 consistency in the interests of accuracy; fortunately, in the present case, that was not necessary. It is also salutary to note, from ensemble data in Figure 3 for example, how large the uncertainties can be in a relatively short-range (36h) forecast e.g. disruptive heavy rain or dry, rainfall timing differences of ±9hours, gales or modest winds. So even though verification statistics are clearly showing continual improvements in NWP performance, there remains a need for the Chief Forecaster to provide a subjective assessment which involves identifying and disentangling the uncertainties and determining the most likely outcome. This is likely to remain so over the years and probably the decades to come. A key role is to identify what action is required if following the normally reliable NWP model is likely to have a negative impact. Situations such as this arise, perhaps, a few times each month. This is of particular importance as many products and services are based on automated NWP data-feeds. Acknowledgement Thanks to Mike Bush for collecting some of the data used in this study. The forecasting challenge of waving cold fronts Figure 13. Observed rainfall totals for the 24h period ending 0900 UTC on 30 October Darker blues highlight larger totals (largest are generally plotted on top). Isohyets for 10, 30 and 50mm have been added manually (increased weight for larger amounts.) Figure 13 shows the 24-hour rainfall totals spanning the event. Note that the presence of heavy rain over Northern Ireland and southwest Scotland fitted very well with the modified fields (Figure 7 centre panel). Meanwhile there was no rain at all over northeast England, where the NAE solution in Figure 2 had shown the heaviest rain. Indeed the overall evolution was much closer to the GM runs and the ensemble products than it was to the NAE (compare Figures 2(a, b), 4 and 13), showing the benefit in this case of accepting the steer provided by those ensembles. Although the timing of the wave was faster than in the modified fields, the idea of heavy rain for Northern Ireland was clearly borne out: some disruption to travel occurred in the east as various places recorded 30 45mm in 24 hours (Figure 13), with much of this falling in less than 12 hours. Heavy rainfall warnings were issued in advance of this. An interesting sequel was the development of very strong gradients across southern Scotland and parts of northern England immediately behind the wave, which went on to become a vigorous small-scale depression These gave inland gusts of around 50kn, which supports the Chief Forecaster the previous day having highlighted a risk of strong winds developing on the wave s southern flank (note the closeness of the isobars south of the southernmost L on Figure 2(c), resembling somewhat the patterns on Figures 5 and 6). Furthermore, although there were some timing errors in the morning forecast, the position of the wave in the afternoon forecast (Figure 12) verified almost exactly (note the verifying black crosses added to Figure 3). Conclusions Despite the sophistication of current NWP models, output can often differ with regard to significant weather, even at short range, and the decision as to which (if any) model solution to follow is rarely straightforward. There is always the concern that the outlier solution may be correct. Whilst verification statistics in the past have clearly shown that the correct decision is usually made (see Carroll and Hewson, 2005), this is not always the case. However in the present study there was a sufficiently strong body of evidence from a wide variety of NWP model products, in particular the ensembles, to identify a significant error in a normally reliable NWP model in its handling of a major frontal wave some hours in advance. This, combined with evidence from satellite imagery, enabled consistent guidance to be maintained in the face of varying model solutions an important aspect when conveying the message to customers. Sometimes, of course, one has to forego References Carroll EB, Hewson TD NWP grid editing at the Met Office. Weather Forecast. 20: Hewson TD. 2009a. Diminutive frontal waves a link between fronts and cyclones. J. Atmos. Sci. 66: Hewson TD. 2009b. Tracking fronts and extra-tropical cyclones. ECMWF Newsl. 121: Hewson TD, Titley HA Objective identification, typing and tracking of the complete life-cycles of cyclonic features at high spatial resolution. Meteorol. Appl. 17: Renfrew IA, Thorpe AJ, Bishop CH The role of the environmental flow in the development of secondary frontal cyclones. Q. J. R. Meteorol. Soc. 123: Young MV, Waters AJ, Browning KA, Bader MJ Application of satellite imagery in nowcasting and very shortrange forecasting: some examples for cold fronts, Satellite and Radar Imagery Interpretation, Preprints for a Workshop held at Reading, England, July EUMETSAT: Darmstadt; pp Correspondence to: M. V. Young martin.young@metoffice.gov.uk British Crown copyright, the Met Office, 2012, published with the permission of the Controller of HMSO and the Queen s Printer for Scotland DOI: /wea