ConvMFP-A Convolution-based Multi-source Fusion Prediction Network for Precipitation Nowcasting

Precipitation nowcasting is very important in coping with extreme weather conditions, but the prediction accuracy of traditional precipitation prediction methods is difficult to improve, and it is impossible to predict problems such as convective initiation. Therefore, this paper proposes a convolution-based multi-source fusion precipitation prediction network (ConvMFP) based on radar echo and satellite cloud image data. The model adopts the encoder-fusion-decoder network structure, combined with the spatial channel attention mechanism, and effectively integrates multi-source features, in order to achieve accurate prediction of severe convective weather system within one hour. Experiments on meteorological datasets in the Yangtze River Delta region show that compared with the traditional optical flow-based methods and some well-known deep learning methods (ConvLSTM, TrajGRU, PredRNN++, RainFormer and SimVP), the accuracy of precipitation prediction is significantly improved, especially in dealing with the prediction problems in the initial generation stage of precipitation.