Daily Reservoir Inflow Forecasting Based On Deep Neural Network

Daily reservoir inflow forecasting plays an important role in reservoir management and operation.It can be used for reservoir drought management,flood control dispatch,irrigation water,hydroelectric power generation,industrial and domestic water,etc.Therefore,accurate daily reservoir inflow forecasting is very necessary for reservoir water resources planning.Due to human activities,climate change,water conservancy project construction and other multi-source factors,the daily reservoir inflow has the characteristics of randomness,non-stationarity,non-linearity and time-varying,which brings great challenges to forecasting.In this paper,the daily inflow data of Ankang Reservoir is taken as the research object.According to the characteristics of the data and the analysis of the forecasting model,the current forecasting method of daily reservoir inflow based on decomposition and artificial intelligence model is improved.The research contents of this paper mainly includes the following two aspects:(1)The current decomposition method combined with artificial intelligence model has achieved good prediction accuracy in daily reservoir inflow forecasting,but it still has some shortcomings.Directly using decomposition method to decompose the daily reservoirs inflow data,without analyzing the characteristics of the data and stationary pretreatment,for the daily reservoir inflow data with large fluctuations can not achieve good forecasting results.The time-varying information of time series data can not be fully utilized by the current forecasting model.To solve the above problems,a daily inflow forecasting model based on decomposition ensemble learning and LSTM(DEL-LSTM)is proposed.By analyzing the characteristics of daily reservoir inflow data,the data stationary processing method is studied.The daily reservoir inflow data are pretreated by logarithmic transformation,which solves the fluctuation problem of daily reservoir inflow data and improves the forecasting performance of the forecasting model.Comparing with the commonly used time series decomposition and reconstruction methods,EEMD and FT are used to extract the characteristic items of the daily inflow data of Ankang Reservoir.In order to achieve more accurate forecasting of daily reservoir inflow,LSTM network model which can utilize time-varying information of time series data is constructed for each feature item after decomposition and reconstruction.Experiments show that the DEL-LSTM forecasting method proposed in this paper effectively improves the accuracy of daily reservoir inflow forecasting.Compared with other forecasting models currently used,the validity of the DELLSTM forecasting method is verified.(2)The forecasting method of decomposition combined with artificial intelligence model mostly uses the same forecasting model to predict the decomposed subsequence or the reconstructed feature items,without analyzing the characteristics of the decomposed data and the applicable forecasting model.In order to solve the above problems and improve the forecasting accuracy of daily reservoir inflow,a forecasting method of daily reservoir inflow based on data feature recognition(DFR-RIF)is proposed.In order to effectively represent the data characteristics of each feature item after decomposition and reconstruction of daily reservoir inflow data,the characteristics of time series data are analyzed,and the stationarity,non-linearity and complexity of each feature item are tested.The advantages and disadvantages of different forecasting models and the types of data they are good at processing are studied.According to the test results of the feature items and the theoretical analysis of the forecasting model,the matching relationship between the feature items and the forecasting model is established,and the corresponding forecasting model of the feature items is determined by the experimental method.In this paper,the validity of the matching relationship between the feature items and the forecasting model is verified by experiments,and the structure of the forecasting model for each feature item is determined.The experimental results show that the proposed DFR-RIF forecasting method improves the performance of daily reservoir inflow forecasting.