Modelling the Precipitation Nowcasting ZR Relationship Based on Deep Learning
Authors: Ma, J., Cui, X. and Jiang, N.
Journal: Communications in Computer and Information Science
Volume: 1587 CCIS
Sudden precipitations, especially heavy ones, usually bring troubles or even huge harm to people’s daily lives. Hence a timely and accurate precipitation nowcasting is expected to be an indispensable part of our modern life. Given that the current precipitation nowcasting methods are based on radar echo maps, the ZR relation that transforms radar echoes into precipitation amounts is crucial. However, traditionally the ZR relation was typically estimated by location-dependent experiential formula which is not satisfactory in both accuracy and universality. Therefore, in this paper, we propose a deep learning based method to model the ZR relation. To evaluate, we conducted our experiment with the Shenzhen precipitation data as the dataset. We introduced and compared several deep learning models, such as CNN, LSTM, and Transformer. The experimental results show that Transformer + CNN has a higher prediction accuracy. Furthermore, to deal with the unbalanced datasets and emphasize on heavy precipitation, we tried to use the SMOTE algorithm to expand heavy precipitation samples, and it shows that it can effectively improve the prediction accuracy of heavy precipitation. Similarly, we also tried to use a customized loss function to enhance the weight of heavy precipitation samples during model training, and it also demonstrate that it can achieve a better accuracy of heavy precipitations. Both approaches can improve the prediction of heavy Precipitation samples by more than 30% on average.