On the 30th December, 2015, new maximum peak levels were recorded on the Rivers Cree, Dee and upper Tweed due to rainfall which fell overnight from the 29th to the 30th December associated with Storm Frank; an unusually deep area of low pressure which brought weather fronts across Scotland (see figure 1). Around 370 properties were flooded in Ballater, NE Scotland, alone, but the devastation was wide spread across Scotland (see figure 2). In the previous article “Storm Frank – how was the medium range flood guidance?” we discussed the development and 2- 5 day forecast of Storm Frank and its potential impacts, whereas the aim of the current article is to discuss the variations in, and factors affecting, the forecast performance over the 24 hours prior to the event peak flows to give an overview assessment of model performance.
There had been significant snowfall over the high ground of the South West and North East over the Christmas period (see figure 3a; MODIS image for the 27th December,2015). The 27th-28th December had been relatively dry but there was some rainfall overnight on the 28th into the morning of the 29th, some of which will have fallen as snow. However, by the 29th December, much of the snow cover had melted particularly at lower elevations (figure 3b). Grid-to-Grid (G2G), the gridded hydrological model which we use to forecast river flows nationally, has a snow hydrology component, which estimates snowfall, accumulation and melt based on precipitation, air temperature and basic snow pack principles, and it had simulated the observed pattern of melt.
Storm Frank- Rainfall
The main bulk of the rainfall associated with Storm Frank occurred between 1800 hrs (+3 hours for NE Scotland) on the 29th December and 0900hrs (+3 hours for the NE) on the 30th December. Maximum recorded rainfall for the event was 130.6mm which fell in 15.5 hours at Lower Black Laggan raingauge in the southwest (Dumfries and Galloway), whereas in the Upper Dee catchment maximum rainfall was recorded at Glenmuick at 95.6mm in 15 hours. In general, 50 to 100mm of rain fell across the Southern Uplands and much of Highland Scotland (figure 4).
Sample forecast Locations
Table 1 is a sample of gauging stations located near to locations where impacts associated with flooding (see figure 2) occurred and/or new maximum recorded levels were observed (indicated by the values highlighted in red in column E). Columns F, G and H indicates the performance of the three G2G forecasts run in the 24 hours prior to the event peaks by comparing the percentage difference of the forecast peak to the observed and the difference in the estimated time of peak. For the difference in timing a negative value indicates that the forecast peak was predicted to occur earlier than it did in reality and the values are fractions of an hour (therefore a value of -0.25 indicates the forecast peak was estimated to occur 15 minutes earlier than in reality). Forecasts 1 and 2 (F1 and F2) were run at 13:45 and 18:30 on the 29/12/ 2015, and F3 at 07:30 on 30/12/2015. These forecasts use deterministic rainfall forecasts from the Met Office UKV runs at 0900hr, 1200hr and 0000hr.
The values highlighted in yellow in columns F-H show at which forecast the previous maximum flow (column D) was forecast to be exceeded. For the 8 sites (red highlight) in table 1 where new peaks were recorded, F1 predicted the previous maximum flow being exceeded for 6 of the sites, and by F3 all of the 8 occurrences were predicted.
In general, the forecasts performed well for predicting the timing of the peak flows, and for indicating the severity of the event at least 24 hours ahead (refer to previous blog article for longer lead time identification of the severity of Storm Frank). For most sites the forecasted peak occurred ahead of the observed time (red text in table 1), which is the preferable direction of timing errors, and the forecast which performed the worst in terms of timing was forecast 2. For the upper Dee gauges which are used for assessing risk for Ballater, the estimate of time for the peaks were in the range of 3 hours too early to 15 minutes late.
In regards to forecast performance for the peak flows, the severity and spatial extent of the impacts were generally predicted with sufficient lead time (+24 hours) to co-ordinate emergency response. However, in terms of whether the peak flows were over- or under- estimated there is an approximate geographical split across Scotland. For the gauges located in the NE, such as those in the River Dee catchment upstream of Ballater, Forecasts 1 and 2 underestimated the peak discharge of the fluvial event, with forecast 2 performing the most poorly in terms of timings and underestimation of the peak flows (see example in figure 5). The performance of forecast 2 can in part be attributed to the underestimation of rainfall in the forecasts for these catchments, as forecast 2 generally predicted the least amount of precipitation. However, forecast 1 overpredicted rainfall for these smaller and higher elevation catchments, but the flow at the gauge locations still underestimated peak flows.
Forecast 1 generally overpredicted the amount of rainfall across Scotland, but whereas forecast 2 underpredicted the rainfall for the Dee catchments, it overpredicted for the more southern gauges (and the River Dee at Park). The rainfall totals were however, still less than for forecast 1. The estimates of peak flows were generally improved over the southern catchments compared to those in the NE. Figure 6 shows the variations in predicted precipitation and river flow for the different forecasts for the River Cree at Newton Stewart in Dumfries and Galloway.
Based on the weather and national river forecasting results on the morning of the 29th December, additional simulations of G2G were run as new NWP model runs became available to the forecaster. In general the combined rainfall and river forecasts at gauge locations across Scotland indicated that this event would cause significant, and possibly severe, impacts at several locations across Scotland with sufficient lead times to enable co-ordination and planning of emergency response. For 6 of the 8 sites analysed in this blog which recorded new maximum peak river levels, the forecast run at 1345hrs on the 29th December estimated that the previous maximum recorded levels would be exceeded. However, the performance of the forecasts varies across Scotland, and is shown to be very sensitive to the rainfall forecast and model structural uncertainties particularly for the smaller catchments. It’s important to remember that the models analysed here are national scale models being used to predict local impacts and the detailing of rainfall intensity at the scale of small catchments is challenging, despite the relatively high resolution of the hydrological model (1km). The spatial variation in rainfall can be significant even in a situation of widespread (or dynamically forced) rainfall, and the highest resolution of the NWP models is 1.5km. Local rainfall totals and intensity can be focussed in certain catchments due to marginal changes in low-level airflow, with the potential of slow moving ‘mountain-waves’.
As science and technology improves, the predictions of rainfall and river flows will better represent local effects and thus the spatial variations. The evolution of such science will require careful assessment of the capability and application of such data to hydrological modelling in near real time as is required for flood forecasting. However, in the mean time, the performance of the forecasting tools and models which are available to us have been tested and proven this winter, during some of the worst events on record, to be invaluable in providing good and timely indications of the likely flood impacts.