The Research And Application Of Combined Forecasting Model For Daily Runoff

Affected by climate,geography,human activities and other factors,runoff series are periodic,random and trendy.River plays an important role in navigation,power generation and irrigation.Accurate runoff forecast is important for water conservancy project scheduling,flood control and drought relief decision-making.Due to the short prediction period of the current daily runoff forecasting model,this paper will study the hydrological forecasting method based on machine learning.Firstly,build five simple single model based on five kinds of machine learning algorithms including Bayesian Ridge Regression(BRR),Support Vector Regression(SVR),Random Forest Regression(RFR),K-Nearest Neighbor(KNN)and Multiple Linear Regression(MLR).Secondly,build five improved single models based on the simple single model.Thirdly,the combination model is constructed according to the simple average method,weighted average method and wavelet combination method.Mean absolute percentage error,relative root mean square error,coefficient of determination and qualified rate,are selected for the optimization of calibration process parameters and the comparison of model prediction effect.The main conclusions are as follows:(1)The five simple single models are suitable for runoff forecast when the prediction period is less than 5 days.The models are applied to the daily runoff prediction of Cuntan station,Yichang station,Hankou station and Datong station in the Yangtze River Basin.It's found that the coefficient of determination of the five models are higher than 0.9 when the prediction period is shorter than 5 days.The order of prediction effect is KNN > RFR?MLR?BRR >SVR.The coefficient of determination are lower than 0.9 when the prediction period is more than 10 days,and the order of accuracy of each model is: RFR > BRR?MLR > SVR > KNN.(2)The five improved single models can improve the prediction speed and the forecast accuracy.The five improved single model are: IS-BRR,IS-SVR,IS-MLR,IS-KNN and ISRFR.The improved single model reduces the dimension of samples by adding the column vector method,and reduces the training and prediction time.The prediction speed of IS-BRR and IS-MLR models is increased by 2 times,and the prediction speed of IS-KNN,IS-RFR and IS-SVR models is increased by 30 times.The coefficient of certainty of each model increased by 5% ? 30%.Among them,the prediction characteristics of IS-BRR model,IS-SVR model and IS-MLR model are similar,and the prediction curve is relatively smooth.The prediction characteristics of IS-KNN and IS-RFR are similar,and the prediction curve is more serrated.(3)The prediction accuracy of simple average combination model,weighted average combination model and wavelet combination model are improved in turn,which are better than the improved single models.The longer the prediction period is,the more significant the improvement effect of each combination model is.Each combination method corresponds to six combination models: BRR-RFR,BRR-KNN,MLR-RFR,MLR-KNN,SVR-RFR and SVRKNN.Compared with the improved single model,the coefficient of determination of the 18 combination models corresponding to the three combination methods are increased by 5% ?40%.For each combination method,BRR-RFR and MLR-RFR models are the best.(4)The prediction results of 18 combined models for Datong and Hankou station are better than Cuntan and Yichang station.When the prediction period is 30 days,the mean absolute percentage error of Cuntan station and Yichang station is 0.18?0.25,which is higher than Hankou station and Datong station(0.13?0.15),and the relative root mean square error of Cuntan station and Yichang station is 0.4?0.5,which is higher than Hankou station and Datong station(0.18 ? 0.23).For Cuntan station and Yichang station,the models need to be optimized.