Research On Lightning Spatiotemporal Prediction Of Multi-source Data Based On Attention Mechanism

All along,severe lightning weather not only affects people's daily production and economic life and normal operation of industries,but even poses a serious threat to national defense and military,causing a large number of citizens' lives and property losses and national financial and economic losses every year,so lightning weather has become a major concern for public safety and a series of weather-sensitive industries.There are many ways to reduce the harm caused by lightning and thunderstorms to all sectors of society.The mainstream way is to predict thunderstorms and lightning activity in the future,and then notify all sectors of society through early warning forecasts to raise the awareness of all sectors and grou ps to reduce the damage caused by lightning and thunderstorms.This paper analyzes the current situation of thunderstorm identification.This paper analyzes the current research status and shortcomings in the field of thunderstorm identification,and proposes a rasterized thunderstorm identification method based on discrete wavelet transform(DWT)and densityfast search algorithm(CFSFD)to assist in lightning activity prediction;for lightning activity prediction,this paper constructs a spatiotemporal prediction model based on Conv LSTM,CNN neural network and attention mechanism to improve the prediction accuracy.The prediction accuracy is improved,and the main research is as follows:(1)Research on thunderstorm identification based on discrete wavelet transform.To address the problems of high complexity,low computational efficiency and inability to effectively identify multi-scale and multi-resolution thunderstorms in the field of thunderstorm identification,this paper proposes a new and efficient t hunderstorm identification method based on rasterized geographic data,ADTD lightning location data,DWT algorithm and CFSFD algorithm.The method firstly rasterizes and quantifies the lightning location data and geographic data,secondly uses DWT algorithm to eliminate the discrete lightning data points in the feature space that affect the recognition effect,and finally uses CFSFD algorithm to efficiently cluster the remaining real lightning data points to find the center of the thunderstorm to achieve the effect of accurate thunderstorm recognition.This method solves the problems of high complexity and low computing efficiency of existing research methods,and achieves multi-scale,multi-quantity,multi-moment and multi-area thunderstorm identification.(2)Spatiotemporal lightning prediction research based on attention mechanism for multisource data.To address the problems of poor long-term prediction,single data source,overreliance on manual experience,and low prediction accuracy in current lightning prediction research,this paper constructs a spatiotemporal prediction model based on Conv LSTM,CNN neural network,and attention mechanism.Lightning data,radar combined reflectivity data(radar data),WRF simulation data and other multi-source data are introduced to solve the problem of single data source for lightning prediction,enrich the geographical information of lightning activity,and enhance the influence of different simulation data impact factors on lightning in different time with the help of attention mechanism,and combine thunderstorm identification and lightning activity prediction to correct the lightning prediction results and improve the prediction accuracy.