Lightning Nowcasting Based on Gated Depthwise Separable Convolution with Dual-source Meteorological Spatio-temporal Data

Lightning prediction is the task of predicting the future lightning area through historical meteorological observation data. At present, the mainstream methods of lightning prediction include thunder storm recognition extrapolation, numerical model prediction and machine learning lightning prediction methods. In the traditional lightning forecast, it is difficult to find the internal relationship between meteorological data and the generation, development and extinction of lightning through artificially designed equations. There are relatively few machine learning lightning forecasting methods, and the prediction results have the problems of long-time interval and low prediction accuracy. In order to solve the above problems, this paper proposes a gated depthwise separable convolution structure (GDSC) that can realize multi-scale feature fusion. By establishing the global connection of local features, the same effect of attention mechanism is achieved with less calculation. The gated memory unit is used to adaptively determine the importance of each scale feature. On the basis of SimVP network, GDSC is used to improve it, and the SimVP-GDSC model for lightning nowcast is constructed. Experiments are carried out on real lightning data sets, and the results prove the effectiveness of SimVP-GDSC model.