Research On Mid-long-term Runoff Forecasting Based On Long-short-term Memory Recurrent Networks And Its Structural Reduction Variant

The runoff process is affected by many factors such as geographical environment,climate changes,and human activities.It is characterized by randomness,fuzziness,grey,and chaos.The research on future runoff forecasting of the river basin is of great significance for obtaining economic and social benefits such as flood prevention,power generation,agricultural irrigation,industrial residential water use,river navigation,and landscape tourism.Therefore,in order to better plan water resources,it is particularly important to improve the accuracy of mid-long-term runoff forecasting.This paper uses the annual runoff and monthly runoff data from the Yichang station in the Three Gorges Basin,and uses artificial neural network to predict the mid-long-term runoff.To avoid ANN's failure to memorize past data and the problem of gradient descent,we use the LSTM and GRU of recurrent neural network to create an annual and monthly runoff forecasting model,and propose a Fastr,more computationally efficient gate structure--Simple-LSTM as a predictive network,which has got very good scores in training and testing.Compared with the Back Propagation(BP)neural network and support vector machine(SVR)prediction results,the LSTM series models have a higher prediction.Based on the existing research,long-term runoff forecasting system for the Three Gorges Basin have been designed and completed.This paper takes the Yichang station in the upper reaches of the Yangtze River as the research object,uses the annual runoff and monthly runoff data of the Yichang station in the Three Gorges watershed,and uses the artificial neural network to predict the mid-long-term runoff.The resea rch focuses on the problems that the traditional artificial neural network prediction model can not memorize past data,easily converge to local optimum and gradient descent,etc.,and introduce the long-short-term memory networks(LSTM)and gated-recurrent-unit(GRU)variants in the recurrent neural network.In order to establish a forecast model for annual and monthly runoff,a more simple and computationally efficient gate structure Simple-LSTM,is proposed as the forecasting network.Case studies show that the Simple-LSTM network shows good performance in both training and testing sets.Compared with BP neural network and Support Vector Machine(SVR)prediction results,Simple-LSTM has a higher prediction qualification rate and higher computing efficiency.Based on the above researches,a set of mid-long-term runoff forecasting systems for the Three Gorges Watershed has been designed.The main achievements of the research work are as follows:(1)In order to overcome the problems of slow convergence,large fluctuations in oscillation,convergence to local points,and gradient attenuation in traditional artificial neural networks,this paper introduces LSTM and GRU networks and builds a mid-long-term runoff forecasting models based on LSTM and GRU.Simplifying LSTM' structure,a more efficient and Fastr forecasting network,which is called Simple-LSTM,has been obtained.Based on this,an annual runoff and monthly runoff forecasting system has been established to obtain better prediction results.(2)Tensor Flow framework has been used to construct LSTM,GRU,and Simple-LSTM training and forecastion models.The influence of different time steps on GRU runoff forecastion accuracy and computational efficiency has also been researched.The forecastion results of LSTM series forecasting systems are compared with BP neural network and SVR prediction results.In terms of annual runoff forecasting,on the test set,the prediction pass rate of the LSTM series model are better than those of the BP neural network and SVR.Calculated of the relative error rate of qualification rate by 20%,t he GRU pass rate is the highest,which is 9.01% and 6.79% higher than BP network and SVR respectively.Simple-LSTM has higher computational efficiency,which is 24.5% and 16.9% higher than LSTM and GRU,respectively,and its pass rate is very close to GRU's(less in 0.79%).In terms of monthly runoff forecast,LSTM has the best prediction accuracy,which is 0.5% and 0.65% higher than GRU and Simple-LSTM,respectively,and compared with SVR and BP networks,the passing rates increased by 2.34% and 6.17%,respectively.Simple-LSTM has higher computational efficiency,which is 38.85% and 25.84% higher than LSTM and GRU respectively,and its pass rate is very close to LSTM's(less in 0.65%).(3)A set of mid-long-term runoff forecasting and demonstration systems for the Three Gorges Watershed has been designed and implemented,which improves the existing mid-long-term runoff forecasting system for the Three Gorges River Basin and provided new ideas for runoff forecasting methods.