Research On Mid-long Term Runoff Forecasting Model Based On Decomposition Method And Machine Learning Combination

The ’14th Five-Year Plan’ puts forward clear requirements for digital twin construction and intelligent water conservancy system construction.As an important part of the ’ four pre ’ system,medium and long-term runoff forecast also plays a vital role in water resources planning management and efficient utilization.However,with the influence of climate change and human activities,the problems and complexity of medium and long-term runoff forecasting are becoming more and more significant.Due to the runoff time series is highly misalignment and non-smooth characteristics,it is often difficult to obtain satisfactory results for forecasting monthly runoff time series using traditional,single intelligent methods.In view of this situation,hybrid model becomes an important and effective way to improve forecast performance,and it is important to combine the advantages of various algorithms and construct more highefficiency modeling organically to increase the quality of hydrological forecasting.Therefore,this paper uses the combination of decomposition method and machine learning model to study the medium and long term runoff forecast.At first,This areticle describes the research status of time series trend change characteristics and medium and long-term runoff forecasting,and then adopts the measured monthly runoff data from two stations,Hongjiadu hydrological station in Wujiang River basin and Yingluoxia hydrological station in Heihe River basin,as the research object,and combines various artificial intelligence algorithms such as traditional machine learning models,deep learning models,decomposition methods,etc.,to investigate the application of various algorithms and their combined models in medium and long-term runoff forecasting.The main research results are as follows.In this areticle the major research contents and results were as bellow:(1)In order to study the variation characteristics of hydrological time series,this paper uses Mann-Kendall mutation test,Mann-Kendall and wavelet analysis means to analyze the variation characteristics of runoff data at two hydrological stations,Hongjiadu and Yingluoxia.The results showed that the runoff time series of Hongjiadu hydrological station showed an increasing trend,while the runoff time series of Yingluoxia hydrological station showed a decreasing trend.The monthly runoff time series of Hongjiadu hydrological station has multiple oscillation cycles,while the monthly runoff measured time series of Yingluoxia hydrological station has only one main cycle,and the first main cycle of both hydrological stations is 10 years.The comprehensive analysis shows that the monthly runoff time series of both Hongjiadu and Yingluoxia hydrological stations have a certain degree of non-stationarity,this also shows that the decomposition method is needed to preprocess the time series,reduce its non-stationarity,and improve the prediction accuracy of the time series.(2)The BP,SVM,RF,LSTM,and ELM in medium and long term runoff forecasting was comparatively studied,and the grey wolf optimization algorithm(GWO)with outstanding global optimization capability was selected for parameter selection in order to further improve the parameter optimization of the single model,the improvement studies of Hongjiadu Station and Yingluoxia Station have been carried out for 50 years.The combination of GWO method improves the effect of parameter selection in the model and applies it to the monthly runoff forecasts for 50 years at Hongjiadu hydropower station and Yingluoxia hydrological station,and RMSE、MSPE、R、NSE four standard indicators were used to evaluate the model performance.Through comprehensive analysis of the index results,it is found that the prediction performance of the LSTM model is the best,the prediction accuracy of the SVM and ELM models is not much different,which is slightly lower than the LSTM model,and the prediction ability of the RF and BP models is the last two in the five models.This indicates that the LSTM model,which has excellent generalization performance and robustness,in comparison to the classical runoff forecasting method based on neural network,this method has higher forecasting precision and can forecast runoff quickly and accurately.(3)Due to the complexity and volatility of the runoff time series,a single prediction model cannot achieve comprehensive and accurate prediction results.Time series pre-processing techniques can reduce the noise in the time series and smooth the data,the results show that the forecasting precision is high and the stability is high.According to this article,the five selected machine learning models are coupled with EEMD,WD and EMD decomposition methods,and the runoff series are preprocessed by the decomposition method to form the monthly runoff sequence combination model based on one-time decomposition model,and the results show that the prediction accuracy of EEMD-GWO-LSTM,WD-GWO-LSTM,EMD-GWO The prediction accuracy of the other four machine learning models coupled with the decomposition method is also improved,this shows that the performance of the one-time decomposition combination model after time series preprocessing is better than that of a single model.The decomposition performance of the three decomposition methods can be ranked from high to low as follows: EEMD>WD>EMD.(4)In order to solve the problem of high complexity of high-frequency subsequence from EEMD primary decomposition,this paper use three kinds of entropy to evaluate the subsequences with high complexity and more effective information after decomposition of the measured runoff series of two stations.On this basis,through the secondary smoothing treatment,the monthly runoff forecast combination model based on secondary decomposition is constructed,and then the forecasting precision of the primary model is improved.in which the prediction errors of RMSE,MAPE,R,NSE values of the WD-EEMD-GWO-LSTM model of Hongjiadu hydropower station compared with the EEMD-GWO-LSTM model increased by 37.89%,49.05%,2.35%,4.90%,respectively The prediction errors of RMSE,MAPE,R,and NSE values of the WD-EEMD-GWO-LSTM model at Yingluoxia hydrological station were improved by130.41%,103.92%,2.00%,and 4.17%,respectively,compared with its primary decomposition model EEMD-GWO-LSTM.Therefore,the hybrid prediction method of quadratic decomposition and machine learning model used in this paper can effectively improve the prediction accuracy of medium-and long-term runoff forecasting and provide an effective basis for runoff prediction.