Research On Medium And Long Term Runoff Forecasting Method Coupled With Climate Forecast System And Deep Learning

Medium and long-term streamflow prediction is the main content of hydrological forecasting,streamflow prediction is helpful to optimize reservoir dispatching,prevent floods and drought,rational formulation of power generation plan of hydropower stations.However,the process of streamflow formation is a highly complex dynamic system.Climate characteristics,geographical environment and human activities make it difficult to accurately predict the process of streamflow.The use of meteorological data for streamflow prediction has become the focus and hotspot of current hydrological forecasting research.The interaction between atmospheric temperature,natural precipitation,evaporation,soil infiltration and other factors affecting river streamflow is complex,and it is difficult to directly apply to streamflow prediction.In this paper,Wunonglong Hydropower Station is taken as the research object.The research includes the acquisition and analysis of ensemble forecasting data,streamflow prediction by deep learning method,coupled ensemble forecasting and deep learning method for medium and long term streamflow prediction.The main work and research results are as follows:(1)In this paper,a method of screening meteorological factors based on grey relational analysis is proposed.Considering the characteristics of CFS ensemble forecasting,a complete set of ensemble forecasting data extraction technology is proposed from three aspects of data download,analysis and screening,which can obtain atmospheric,precipitation and temperature data.In Wunonglong Hydropower Station watershed,the inverse distance weighting method is used to process the ensemble forecast data in the area in space downscaling,so as to calculate the meteorological forecast data in the watershed.Meteorological data show that the CFS ensemble forecast makes monthly scale prediction of 77 meteorological factors in the basin.In this paper,grey correlation analysis is used to judge the close relationship between meteorological data and streamflow.There are many meteorological factors in CFS forecast,not all meteorological factors are related to streamflow.Grey correlation analysis method is used to calculate the correlation between all meteorological factors and streamflow process.Eighteen meteorological factors with high correlation were screened,including atmospheric movement,rainfall,humidity,temperature and other meteorological characteristics.(2)This paper presents a streamflow prediction model based on deep learning method.Considering the difficulty of establishing streamflow prediction model by traditional methods,this paper constructs a deep neural network to let the network model learn the nature of the data by itself,so as to establish a streamflow prediction model with higher accuracy and longer forecasting period,thus avoiding the complex parameter selection problem of the traditional streamflow prediction model.LSTM and seq2 seq were used to establish streamflow prediction models with forecast periods of 1 month and 12 months respectively.When LSTM model is used to predict the next monthly streamflow,the streamflow data of all previous months are used.Considering the regularity,periodicity and trend of streamflow in each year,the qualified rate of the prediction results of the deep neural network streamflow prediction model reaches 65.5%.In this paper,a streamflow prediction model based on seq2 seq is constructed.The forecasting period of the model is 12 months,which can forecast the whole year's streamflow process.(3)This paper presents a streamflow prediction model coupled with meteorological forecast data.Although the streamflow prediction method based on LSTM and seq2 seq proposed in the previous paper can predict the streamflow in a long forecasting period,its research data only include streamflow data,and there is no model to couple meteorological ensemble forecasting data into streamflow prediction.In this paper,a ridge regression method is proposed to couple the meteorological data provided by CFS ensemble forecast with the prediction data of depth learning method for 1-9 months,and to establish a ridge regression model and forecast streamflow.The accuracy of forecasting results has been greatly improved.Finally,this paper summarizes the whole paper and points out the next research directions.