Research On Lightning Prediction Using Deep Spatiotemporal Neural Network Models

Lightning is a serious natural disaster and its harm should not be underestimated.Accurate lightning forecasts can help avoid risks and reduce losses.Traditional lightning forecasting includes extrapolation based methods and numerical model based methods,but there is still much room for improvement in the performance of these methods based on physical models.In this paper,the temporal and spatial attributes of lightning activity and other meteorological elements are fully considered,and the lightning forecasting problem is modeled as a spatiotemporal sequential prediction problem.An effective lightning forecasting model and system are constructed by using deep neural networks.The main work of this paper includes the following two aspects.First,this paper proposes an attentive deep learning model named ADSNet to realize hourly high resolution lightning forecast.ADSNet leverages the advantages of dualsource data,and the encoder-decoder framework naturally combines the historical observations with the numerical simulations.The channel-wise attention mechanism adjusts the information ratio of various simulated meteorological factors to enhance the long-term prediction ability of the model.Second,this paper proposes a multi-source heterogeneous data fusion model based on ADSNet,named AHSNet.It introduces automatic weather station data as an additional data source,increasing the amount of information available for prediction.Two solutions are designed and implemented to eliminate the difficulty in merging discrete data and grid data.These two solutions are not only suitable for lightning prediction,but also provide references for similar data fusion problems in other tasks.This paper validates the effectiveness of the proposed methods on Beijing-TianjinHebei real datasets.Compared with the three commonly used lightning parameterization schemes,the ADSNet model can increase the grid-to-grid cumulative scores of the 12-hour lightning forecast by 2 to 3 times and the hit rate by 1 to 2 times.The two versions of AHSNet further improve the 6-hour lightning forecast.An ablation study also verifies the effectiveness of the proposed data fusion method and attention mechanism.