Application Of Machine Learning Method In Lightning Forecast And Early Warning

When encountering catastrophic weather,accurate forecasting and warning of upcoming lightning events can effectively prevent or reduce the casualties and operational losses of various industries due to lightning disasters.Traditionally,researchers have used the acquired meteorological data to carry out a series of lightning forecast and early warning work.In recent years,data has become an important production factor in various industries.As the best method technology combined with the application of "big data",machine learning has made breakthroughs in various modeling and prediction tasks compared with traditional methods.Progress and performance.In this situation,based on the inherent advantages and urgent needs of the development of big data applications by the meteorological department and industry,this paper introduces machine learning methods into the existing lightning forecasting and early warning technology.First of all,using meteorological feature data selected in ERA5 that has a good correlation with lightning occurrence and lightning frequency data with grid statistics,a high-performance lightning prediction model is constructed based on the integrated algorithm XGBoost.In the basic forecast,the predicted hit rate POD is 90.41%,the false alarm rate FAR is 7.46%,the critical success index CSI is84.27%,the skill score TSS is 0.83,and its performance is compared with the model method proposed in traditional research.Obviously improved,and on this basis,a multi-class forecast on lightning frequency interval was further realized.Subsequently,using the same ERA5 data and the average lightning amplitude data of gridding statistics,a prediction model about lightning amplitude was constructed based on the fusion method of Stacking model.In the research,the five single models for lightning amplitude prediction constructed earlier were fused to obtain the final prediction model.After the fusion model,the average RMSE of the lightning amplitude prediction in the range of 1-100 k A is 10.96296,and the predicted RMSE in the range of 0-50 k A amplitude reaches 6.47134.Under the conditions of sufficient data,the model has initially achieved the Predict performance.Finally,a strong convection early warning model with spatiotemporal characteristics in the output is constructed.The research is based on recurrent neural network,using improved neurons,double-layer network structure and weighted loss function,to predict the echo shape and movement trajectorywithin the next 1h through the historical echo data of the first 2h to achieve the relevant warning work.After training,the model can effectively predict the generation and dissipation of subsequent echoes and the distribution change of the reflectance value.Within 30 minutes,it can better predict the echo characteristics of the area above 40 d BZ.According to the above research,the feasibility and effectiveness of the application of machine learning method technology in lightning forecast and early warning have been confirmed.