Research And Application Of Rainfall Prediction Algorithm Based On Deep Learning

Rainfall is the most important meteorological factor affecting the lives of most people.The main rainy season in China is from June to September,supporting 85-95%of my country’s food crops.Because agricultural activities and crop production depend on the amount and distribution of rainfall throughout the year,accurate seasonal rainfall forecasts are critical for agricultural planning and disaster mitigation.Rainfall is also an important factor in flood control.How to obtain rainfall information more quickly and accurately has attracted more and more attention of meteorological researchers.Therefore,it is of great significance to deploy a rainfall prediction model with accurate prediction and good generalization performance.Relying on the national project "Water Pollution Control and Treatment Science and Technology Major Project",this thesis conducts a series of rainfall prediction research on the basis of fully analyzing and understanding the existing machine learning algorithms.The main research results are as follows:(1)Aiming at the needs of rainfall forecasting tasks,this study proposes a set of feature engineering solutions for rainfall forecasting through visual analysis of the relationship and laws between rainfall forecasting data.This study proposes to synthesize the traditional rainfall observation index: the Precipitable Water Vapor(PWV)observation index through the feature synthesis method.By synthesizing this index,the combination of the traditional rainfall observation method and the machine learning algorithm is realized.The study also used the method of generating statistical features to generate the average rainfall for each station for each month,so that the rainfall prediction model can better capture the relationship between season and rainfall.(2)This study proposes a Rain-Attention model for the task of rainfall probability prediction.The Rain-Attention model has made a series of optimizations based on the Transformer-based Encoder structure: gated fusion mechanism,non-invasive attention mechanism.The gated fusion mechanism enables the model to adjust the weight of fusion between features according to the loss,so that the model can better analyze the feature information contained in the vector.The non-intrusive attention mechanism enables the model to better resolve the difference between the station ID vector and the side information vector.This research proves that the Rain-Attention probabilistic prediction model has better prediction accuracy and generalization performance than other models through comparative experiments.(3)This study proposes a Rain-Tab Net model for the rainfall forecasting task.The Rain-Tab Net model has made a series of optimizations based on the Encoder structure based on Tab Net: G-Softmax activation function,unsupervised pre-training.In this study,the G-Softmax activation function is used instead of the Sparsemax activation function,which can effectively improve the internal class compactness and the interclass separability.This study uses unsupervised pre-trained Rain-Tab Net model to improve the stability and performance of Rain-Tab Net model.In the final test set prediction,this study uses a strategy of jointly predicting the Rain-Attention model and the Rain-Tab Net model.Comparative experiments show that the joint prediction strategy proposed in this study is reasonable,and the Rain-Tab Net model has excellent performance.the accuracy of other models.(4)This research implements a rainfall prediction system based on the RainAttention model and the Rain-Tab Net model.The rainfall prediction system in this study is based on the B/S structure,using the Tornado framework as the web service framework,the front-end page implementation using the Vue framework,the My SQL database for data storage,and the Tensorflow for implementing the rainfall prediction algorithm.This study builds a comprehensive,easy-to-operate and stable rainfall forecasting system.