Study On Medium And Long Term Runoff Forecast Of Qingjiang River Basin Based On Neural Network And Attention Mechanism

There is a large amount of liquid water on the surface of the earth,but most of it exists in the marine environment,and fresh water resources that can be effectively used account for 2.53% of the total.Even so,nearly 70% of fresh water exists in the form of glaciers,snow and frozen soil.And the available fresh water only accounts for 30% of the world's total fresh water.It seems that fresh water is scarce.Due to the huge population and unreasonable spatial distribution of water resources,my country has become a water-scarce country.Therefore,effective and reasonable runoff forecasting can provide reference value for relevant departments and play a pivotal role in flood control and national economic development.This paper mainly analyzes the characteristics of multi-temporal and spatial scales of the Qingjiang River Basin,the study of runoff forecasting based on multiple factor identification methods,and the intelligent forecasting method based on the attention mechanism.Based on the measured runoff data of Qingjiang Geheyan,the medium and long-term runoff forecast is carried out.,The main innovative results of the thesis are as follows:(1)The distribution of runoff during the year is represented by the percentage and unevenness of the average monthly runoff in the years.The results show that the distribution of runoff during the year is uneven.The flood season is mainly concentrated in April to September,and the dry season is mainly concentrated in October.From January to February,the runoff subsided during the dry season is obvious.Analyzing the evolution trend of runoff,the results show that the overall trend of the monthly runoff sequence of Geheyan is not significant,and the monthly runoff has not increased significantly over the years.Analyzing the mutation point,the result shows that there is a mutation point.The periodic calculation of the runoff time series shows that there are multiple time scale cycles in the runoff series,including the first major cycle of ten months,the second major cycle of eighteen months,and the third major cycle of twenty-seven months..Preliminary analysis of the stability of Geheyan flow sequence,preliminary conclusion that Geheyan runoff sequence is stable,and the randomness is small.(2)Combine four different predictor screening methods with three models to predict Geheyan runoff,and analyze the qualification rate of the prediction results.The horizontal comparison shows that the prediction effect of BP neural network and Elman neural network is better than that of multiple linear regression.The longitudinal comparison shows that the model effect of the predictor screening scheme based on stepwise regression and principal component analysis is better than the effect of correlation coefficient method and rank correlation analysis method.Among them,the Elman neural network prediction model based on principal component analysis has the relatively best effect.(3)Using the long-term memory capability of LSTM can solve the problem of gradient explosion and gradient disappearance to a certain extent.On this basis,the attention weight distribution ability is used to perform weight distribution on the importance of factors to extract more effective information.The forecast of the 10-day runoff measured in Geheyan was compared with the result of a single LSMT linear forecast.The result showed that the LSMT forecast model with the attention mechanism was improved compared with the original single model.