Prediction Of Summer Precipitation In China Based On Machine Learning

The machine learning algorithm based on big data analysis has made some achievements in different fields,but has not been widely used for the prediction of seasonal precipitation.Therefore,based on the Beijing Climate Center(BCC)seasonal climate prediction model data and the monthly mean precipitation data from the National Meteorological Information Center of China,this research explores the application of machine learning in predicting summer precipitation over China.The selection of precipitation prediction factors affects the accuracy of machine learning prediction.After extensive literature review,60 meteorological factors with definite physical significance for prediction of summer precipitation were selected,including 33 atmospheric factors,7 land surface factors,13 marine factors and 7 sea ice factors.In this research,the Long Short-Term Memory(LSTM)algorithm is optimized.The training and generalization error in the LSTM network are optimized by Adaptive Moment Estimation(Adam)and Dropout,respectively,which speeds up the training and improves the generalization ability of the model.Because the prediction accuracy of the LSTM network is greatly affected by the network parameters,and the precipitation climate characteristics and prediction factors of each meteorological site are different,the hidden layer node number,training times and learning rate of 160 sites are optimized in this research.In addition,the prediction accuracy of the LSTM Network,stepwise regression,Back Propagation Neural Network and BCC model are compared.The results show that the prediction ability of the LSTM network is better than that of the stepwise regression,Back Propagation Neural Network and BCC models in considering mean value,anomaly correlation coefficients and root mean square error.The prediction accuracy of the LSTM network precipitation anomaly using April data is better than the prediction results using March and May data.The sea ice factors make a positive contribution to the seasonal precipitation prediction results.The effect of the full-factor input LSTM network prediction is better than the LSTM network that eliminates the sea ice factors.Based on the results of parameter tuning and research on the prediction scheme,the BCC model data from April is selected and the LSTM network with all precipitation prediction factors is used to predict the summer precipitation over China.The results show that the LSTM network is able to predict the overall situation of precipitation.The Ps scores of the summer precipitation return test in 2014 and 2015 are 74 and 71 points and the anomaly sign consistency rates are 55.63% and 55.25%,respectively.The average Ps score is higher than the prediction result of the national climate trend prediction meeting in flood season and operational model.This research provides a possible reference for summer precipitation prediction over China.