Research And Implementation Of Thunderstorm Prediction Algorithm Based On Neural Network

The study of thunderstorm prediction is one of the most important topics in modern weather forecasting.The general method of thunderstorm prediction is to obtain thunderstorm-related meteorological data through weather monitoring equipment for analysis.In recent years,artificial intelligence technology has also been introduced in the field of meteorology,compared with traditional methods based on expert experience,the calculation performance and prediction capabilities of new technologies are better.However,the methods of using machine learning and deep learning in thunderstorm prediction are not mature enough.Based on a variety of meteorological data,this paper studies different aspects of thunderstorm prediction,including thunderstorm path prediction,thunderstorm approach prediction and thunderstorm image prediction.The main work content is as follows:(1)A fusion method of Kalman filter and thunderstorm path prediction is proposed.The algorithm starts with lightning location data,and uses a clustering algorithm to identify thunderstorm cells based on the mapping relationship between lightning and thunderstorms.On this basis,it tracks the displaced thunderstorm cells based on changes in the preceding and following periods,and introduces Kalman filtering from them to reduce the system's tolerance;finally,extrapolate its future movement path based on the movement trend information of the thunderstorm cells to realize the prediction of the thunderstorm path.Experimental results show that the accuracy of the algorithm reaches 78.04% when the prediction step in 6 minutes,and it is higher than the current commonly used algorithm after the extrapolation time exceeds 18 minutes.(2)This thesis presents a prediction model of the approaching thunderstorm within the detection range of the electric field instrument.For the detection area of the atmospheric electric field instrument,the model first determines the data characteristics in one time slice plus the thunderstorm event label of the next time slice as a sample;the next step is to use multiple methods to separately analyze the time series of electric field data and lightning data features,which can reflect the approach of thunderstorms;Finally,the BP neural network is used to establish the binary classification model of whether thunderstorms will occur in the future.Experiments show that for the data of Guangzhou Electric Field Instruments in 2020,the model has a maximum POD of 86%,a FAR as low as 6.31%,and a CSI of 80.36% after adding quantitative lightning data features.(3)A thunderstorm image prediction method based on radar echo is proposed.The algorithm use the echo value matrix of the radar echo data to generate the grayscale image of the thunderstorm,and construct a series of thunderstorm image time series according to the weather radar scanning time interval;further,use the sliding window mechanism to expand the third dimension of the sample to construct a three-dimensional tensor.Finally,this thesis proposes to use Conv LSTM neural network to construct a thunderstorm image prediction model.For the radar echo data taken from Guangzhou for dozens of days,when the size of the convolution kernel is 3*3 and the reflectivity threshold is 60 d BZ,the CSI of the model is the highest,reaching 60.65%.