Research On Deep Learning Runoff Forecast Model And Optimization Based On Multi-source Data Fusion

Runoff forecasting is the basis of water resource management and regulation.With the continuous improvement of human observing ability of natural processes,it is possible to obtain multi-source data at different scales.With the increasing data representation and more comprehensive description of hydrological process,new opportunities are provided for the research of data-driven hydrological model.In this study,long-range hydrometeorological data of several catchments in the upper reaches of Hanjiang river are selected,forecasting factors are screened by correlation coefficient method,mutual information method and grey correlation analysis method,and runoff forecasting model based on long and short-term memory neural network(LSTM)is constructed by deep learning method and model structure optimization is carried out to verify its forecasting ability and applicability in different catchments.Selecting atmospheric circulation factor data set and study regional long-range hydrometeorological data,through Ensemble Kalman filter(EnKF)data assimilation algorithm,establish a deep learning runoff forecasting model based on multi-source data fusion to explore the improvement effect of multi-source data fusion on runoff forecasting accuracy.At the same time,based on the hydrological characteristics of the study area,the original runoff series is decomposed and reconstructed by various mode decomposition methods,and the combined runoff forecasting model is constructed by using the strategy of decomposition-prediction-reconstruction to explore its forecasting effect on the characteristics of runoff and flood process.The main research contents and achievements of this paper are as follows:(1)LSTM runoff forecasting model based on Deep learning method has good applicability in the upper reaches of Hanjiang river.LSTM models in Baohe,Xun he and Jiahe river basins have high forecast accuracy and good simulation effect.LSTM models in Jushui river basin can satisfiy the basic forecast requirements,but the forecast accuracy of LSTM models in Ziwu river basin is lower.With the extension of the forecasting period,the performance of the model forecast significantly decreases,and the Nash efficiency coefficient is lower than 0.5,which can not meet the requirement of runoff forecast.Bi-directional long and short-term memory neural network(BiLSTM)model is constructed by optimizing model structure and adding back-propagation LSTM hidden layer,which improves the forecast accuracy of Deep learning model.Compared with traditional LSTM model,BiLSTM model improves each forecast index and captures the extreme value of runoff sequence better in the same forecasting period.However,there are still shortcomings such as short forecasting period of model and low accuracy of flood process simulation.(2)An EnKF-BiLSTM model is built based on BiLSTM model and EnKF data assimilation method.After fusing multi-source data sets,the forecasting accuracy indexes of hydrological models in different watersheds can be further improved and change more smoothly with the extension of forecasting period.The forecast accuracy of EnKF-BiLSTM model has been improved in different forecast periods.When the forecasting period is 1d,the NSE of each research watershed model rises above 0.70,and the NSE and relative deviation increase by 2.7%～13% respectively.When the forecasting period is 2d,NSE of each watershed model will increase by 12%～16% as a whole.Although the forecast period is 3d,the Nash efficiency coefficient of the model has been lower than 0.5,and the forecast accuracy of the model is low.The improvement effect of EnKF-BiLSTM model on runoff forecast is different in each watershed.Among them,the simulation effect of Xun he river and Jiahe river is significantly improved,while that of the Jushui and Baohe river basins is relatively small.For EnKF-BiLSTM model,there are differences in runoff forecasting results in different catchments.The preliminary analysis discussed its relationship with the natural geographical characteristics and the intensity and range of atmospheric circulation influence of each watershed.(3)Based on EnKF-BiLSTM model and Empirical Mode Decomposition(EMD),Ensemble Empirical Mode Decomposition(EEMD)and Variational Mode Decomposition(VMD)methods,a Deep learning combined forecasting model is constructed.The results show that the combined forecasting model based on VMD has the best forecasting effect.In different forecasting periods,the NSE of watershed model reaches above 0.7,the mean peak relative error(MREP)is reduced,and the forecasting accuracy of peak discharge is improved.It also improves the forecast accuracy of peak discharge arrival time.It can be seen that the modal decomposition method can be used to pre-process the original runoff sequence,which can effectively overcome the non-stationarity of the flood runoff data of small watershed in mountainous area,enhance the description ability of the characteristics of runoff data with different characteristics,and effectively improve the performance of model prediction.