Newsletter No. 149 banner

Predicting heavy rainfall in China

Linus Magnusson
Thomas Haiden


On 19 and 20 July, severe rainfall hit central and north-eastern China. ECMWF’s forecasts three to five days ahead of the event performed reasonably well, but the quality of earlier forecasts was geographically uneven.

Observations and short-range forecasts
Observations and short-range forecasts. Forty-eight-hour observed precipitation 19 July 00 UTC to 21 July 00 UTC from SYNOP observations (left) and predicted 48-hour precipitation from the HRES issued 19 July 00 UTC (shading) together with mean sea-level pressure (contours) valid 20 July 00 UTC (right).

Ensemble and high-resolution forecasts
Ensemble and high-resolution forecasts. ENS and HRES 48-hour precipitation at a grid point in Beijing valid 19 July 00 UTC to 21 July 00 UTC for a range of starting dates. Black dots in the box-and-whisker plot represent the ENS median, the wide boxes represent the 25th and 75th percentile, the narrower boxes represent the 10th and 90th percentile, and the vertical lines show minimum and maximum values.

The rainfall was connected to a low pressure system that formed over southern China and moved northward. The cyclone generally resulted in more than 50 mm of rain in 48 hours along its way, with some stations in central China receiving more than 200 mm. Further north, the precipitation in Beijing reached almost 300 mm over the two days, but local variations were large. ECMWF’s high-resolution forecast (HRES) from 19 July 00 UTC predicted rainfall in excess of 300 mm locally south-west of Beijing in the first two forecast days, which shows that the forecast system is capable of simulating such extreme rainfall, although one should not expect it to capture the exact location of the extremes.

In the last forecast before the start of the accumulation period, the HRES and ensemble forecast (ENS) median gave around 150 mm for a grid point in Beijing. Ensemble members ranged from 50 to 300 mm, which indicates a large uncertainty in the local severity even in the shortest forecasts. Looking at earlier forecasts, a risk of 25% or higher for more than 100 mm was predicted by ENS from 16 July onwards, which corresponds to a forecast range of 3–5 days.

In forecasts produced before 16 July, the southern part of the rainfall was captured well, but the extension to the north, where the most severe rainfall occurred, was missed. This is apparent when we compare EFI (Extreme Forecast Index) and SOT (Shift of Tails) values for 3-day accumulated rainfall (19–21 July) in the forecasts from 15 and 19 July.

This event was one of several episodes of extreme rainfall in China this summer. At the beginning of July, central China was hit by severe rainfall that resulted in flooding of the Yangtze River. In this case, the short-range (1–2 days) ECMWF forecasts placed the rainfall somewhat too far north. Although the location error was only in the order of 100 km, it was big enough to have presented a substantial challenge to forecasters trying to predict rainfall levels for specific river catchments.

Extreme Forecast Index and Shift of Tails
Extreme Forecast Index and Shift of Tails. EFI (shading) and SOT (contours) for 3-day precipitation (19–22 July) from 15 July (left) and 19 July (right).

Cooperation agreement

In 2014 ECMWF concluded a cooperation agreement with the China Meteorological Administration (CMA). Their local knowledge and the availability of high-density observations will help us to better assess and understand the performance of our forecasts. One of the areas of cooperation will be the evaluation of ECMWF’s forecasts in China using high-density observational datasets. This will enable ECMWF to obtain more detailed results on model performance in this area.

If insights gained from these studies lead to improved forecasts in the region, then that will not only be of importance for forecasters in China but may also be beneficial for Europe. Forecast errors can propagate with the group velocity of Rossby waves, so that initial and short-range errors originating in the area of China could reach Europe eight to ten days into the forecast. Therefore, one of many ingredients for achieving ECMWF’s strategic goal of predicting risks of extreme weather over Europe two weeks in advance may be improved analysis and forecast performance over South-East Asia.

Other recent events

On 6 August, Skopje, the capital of the former Yugoslav Republic of Macedonia, was hit by severe flash floods that killed at least 21 people. The location of this event was not well predicted by ECMWF’s forecasts.

In the second week of August, the US state of Louisiana was hit by severe rainfall over a period of three to four days. The large-scale features of this event were well predicted around a week in advance, while capturing the local details was a challenge even in the shortest-range forecasts due to its convective nature.

Evaluations of all the events mentioned in this article can be found in the ECMWF Severe Event Catalogue at