Research On Medium And Long-Term Runoff Prediction Model Based On Hybrid Intelligence

Smart water management is one of the important paths to achieve a new stage of highquality water development.Forecasting is a key part of the "Four Forecasts" business service application system and an important part of building smart water resources.Therefore,accurate runoff forecasting is particularly important.In this paper,monthly runoff data from Guangzhao Reservoir and Xinfengjiang Reservoir in the Pearl River Basin of China are used as the target of this research to study medium and long-term runoff forecasting.By comparing conventional machine learning models and deep learning models,a suitable forecast model is selected,and based on this,both data pre-processing and model parameter optimization are used to improve the monthly runoff forecast accuracy and provide a new option for the medium and long-term runoff forecasting.The main contents are as follows:(1)In order to reveal the trends and patterns of medium and long-term runoff series,the Mann-Kendall mutation test,Mann-Kendall trend test and wavelet analysis were used to analyze the monthly runoff data of the two reservoirs for mutation,trend and periodicity,respectively.The results show that the monthly runoff series of the two reservoirs tend to be stable in general,with a slight increase in the monthly runoff of Guangzhao reservoir and a slight decrease in the monthly runoff of Xinfengjiang reservoir;in terms of abrupt variability,the abrupt variation points appear several times in the two reservoirs,but they all fluctuate little;in terms of periodicity,the first main cycle of the two reservoirs is ten months.Comprehensive analysis of the three runoff variation characteristics shows that the overall variation of the two reservoirs is not significant,which provides a reasonable comparison for the next runoff forecast.(2)To solve the problems of high data requirements and many parameters of physical models,conventional machine learning models BP neural network,ELMAN neural network,support vector machine(SVM),random forest model(RF)and deep learning model long short-term memory neural network(LSTM)are constructed for monthly runoff forecasting.The monthly runoff data of Guangzhao and Xinfengjiang reservoirs in the Pearl River basin in China were used for validation and the prediction performance of these five models were compared.The results showed that the LSTM model had the best prediction performance,the SVM model was second only to the LSTM model in terms of prediction accuracy,while the BP model,ELMAN model and RF model had similar overall prediction performance,and the ELMAN model was slightly better than the BP model and RF model.(3)To address the problems of difficult parameter selection of machine learning models,which leads to low prediction accuracy,the Sparrow Search Algorithm(SSA),which has the advantages of fast search speed and strong robustness,is first used to optimize the parameters of the five models constructed,and then a new prediction model(Attention-LSTM)is constructed by introducing the attention mechanism in the LSTM model,thus trying to further improve the prediction accuracy.The results show that the SSA algorithm can effectively improve the parameter selection efficiency and improve the prediction accuracy of the model,while the Attention-LSTM model with the introduction of the attention mechanism is compared with the LSTM model,BP model,ELMAN model,SVM model and RF model to verify that the Attention LSTM model has better prediction performance.(4)To address the problems that runoff sequences are nonlinear and complex,a single machine learning model cannot reveal the underlying multiscale phenomena,resulting in limited prediction accuracy.In this paper,four decomposition methods,namely empirical modal decomposition(EMD),ensemble empirical modal decomposition(EEMD),complete ensemble empirical modal decomposition(CEEMDAN)and variational modal decomposition(VMD),are firstly used to decompose the monthly runoff series into several subseries,and the decomposed subseries are input to the model and predicted separately,so as to reduce the modeling difficulty and improve the accuracy of monthly runoff prediction.The results show that the prediction accuracy of the decomposed model is significantly higher than that of the undecomposed model,which proves that the decomposition algorithm can reduce the complexity of the input sequences and effectively extract the effective information implied by the runoff sequences,so that the prediction accuracy can be improved substantially,The decomposition methods of EMD,EEMD,CEEMDAN and VMD are compared,and it is concluded that the subseries decomposed by VMD is smoother and has the best decomposition effect compared with other decomposition methods,followed by CEEMDAN and EEMD decomposition,and the worst is EMD decomposition.Finally,a summary of the proposed hybrid intelligent prediction model is presented and an outlook is given.