Research On A Short-imminent Precipitation Forecast Method Based On Deep Learning

Precipitation has a great impact on people’s life and work,accurate prediction of precipitation can help people take preventive measures in time to avoid or reduce losses caused by floods and other disasters.Therefore,accurate prediction of precipitation is of great significance.In this thesis,the Doppler radar image sequence data is used,and the deep learning method is used to build a short-term precipitation forecasting model,aiming to improve the accuracy of precipitation forecasting.The main work is as follows:(1)In the radar data preprocessing,this thesis mainly considers the relationship between the radar echo reflectivity factor and precipitation,and adopts the percentile method and the idea of convolutional neural network,that is,the convolutional layer is used to select the sliding area,and the pooling layer is used to perform percentile pooling.Taking the selected radar reflectivity factor as the main feature information,finally,the original radar image of15×4×101×101 can be processed into feature information of 15×4×10×4.Such processing not only improves the quality of data,but also meets the needs of subsequent models.(2)In the construction of precipitation prediction models,combining the strong feature learning ability of convolutional neural networks and the ability of recurrent neural networks that are good at dealing with time-series problems such as precipitation,a precipitation prediction model based on a three-dimensional convolutional long and short-term memory network(Conv3D-LSTM)was first constructed based on the previous work.Then the model is further improved,and the main elements of the improvement are the adoption of a bidirectional gated recurrent unit network module instead of a long and short-term memory network module,which allows the model to learn key information on the temporal sequence before and after the radar feature data,and the introduction of an attention mechanism,which allows the model to focus on regions with larger reflectivity factors,and finally an improved model of a three-dimensional convolutional bidirectional gated recurrent unit network based on the attention mechanism(Att-Conv3D-Bi GRU)is constructed.(3)In order to verify the effectiveness of the improved model,the results of the two models are analyzed experimentally and compared with the commonly used precipitation prediction models.The experimental results show that the improved model based on Att-Conv3D-Bi GRU performs better in predicting the total precipitation at the target site in the next 1-2 hours and in the evaluation indexes of the model.The coefficient of determination(R~2)of the improved model in this thesis reaches 0.63,which is higher than that of the Conv3D-LSTM model at 0.59.Meanwhile,its mean absolute error(MAE)and root mean square error(RMSE)are 6.42 mm and 9.36 mm,which are lower than that of the Conv3D-LSTM model at 6.85 mm and 10.18mm.Moreover,the improved model outperforms the commonly used precipitation prediction model in all the prediction performance evaluations,which indicates that the improved model effectively improves the accuracy of precipitation prediction.