Forecasting seasonal to sub-seasonal rainfall in Great Britain using convolutional-neural networks

Abstract

Traditional weather forecasting approaches use various numerical simulations and empirical models to produce a gridded estimate of rainfall, often spanning multiple regions but struggling to capture extreme events. The approach presented here combines modern meteorological forecasts from the ECMWF SEAS5 seasonal forecasts with convolutional neural networks (CNNs) to improve the forecasting of total monthly regional rainfall across Great Britain. The CNN is trained using mean sea-level pressure and 2-m air temperature forecasts from the ECMWF C3S service using three lead-times: 1 month, 3 months and 6 months. The training is supervised using the equivalent benchmark rainfall data provided by the CEH-GEAR (Centre for Ecology and Hydrology, gridded estimates of areal rainfall). Comparing the CNN to the ECMWF predictions shows the CNN out-performs the ECMWF across all three lead times. This is done using an unseen validation dataset and based on the root mean square error (RMSE) between the predicted rainfall values for each region and benchmark values from the CEH-GEAR dataset. The largest improvement is at a 1-month lead time where the CNN model scores a RMSE 6.89 mm lower than the ECMWF. However, these differences are exacerbated at the extremes with the CNN producing, at a 1-month lead time, RMSEs which are 28.19 mm lower than the corresponding predictions from the ECMWF. Following this, a sensitivity analysis shows the CNN model predicts increased rainfall values in the presence of a low sea-level pressure anomaly around Iceland, followed by a high sea-level pressure anomaly south of Greenland.