Ensemble Probabilistic Prediction Method Of The Runoff Based On Multiple Models

Accurate and reliable hydrological forecasting is the basis of water resources development and utilization.How to further improve the forecast accuracy and at the same time accurately quantify or characterize the forecast uncertainty is a hot and difficult problem in the field of hydrological forecasting at this stage.Ensemble probability forecasting,which characterizes forecast uncertainty in the form of probability or interval,is a key development direction for future hydrological forecasting.Therefore,the research on hydrological ensemble probability forecasting is of great theoretical significance and practical application value for the scientific scheduling of reservoirs.In this thesis,we propose a multi-model stochastic combination-based runoff ensemble probability forecasting method(SCMM)based on the construction of several single forecasting models for the inlet runoff of the upper reaches of the Han River Golden Gorge Reservoir.The SCMM method is compared with the Bayesian model averaging(BMA)method to demonstrate the rationality and effectiveness of the proposed method.The main work and conclusions of the paper are as follows.(1)Various runoff forecasting models covering linear and nonlinear,white box and black box,physical concept and data driven are built.Multi-method screening of forecast factors was used to construct seven single deterministic forecast models,artificial neural network model(BP),multiple linear regression model(MLR),support vector machine model(SVM),random forest model(RF),extreme learning machine model(ELM),radial basis neural network model(RBF),and two-parameter monthly water balance model(TPWB),using mean absolute error(MAE,Nash efficiency coefficient(NSE),root mean square error(RMSE),mean percentage error(MAPE),and goodness-of-fit coefficient(R~2)were used as deterministic forecast evaluation indicators to assess the forecast results.The results show that the NSE of all seven models is above 0.65 in the test period;the coefficient of goodness of fit R~2 is above0.80,among which the NSE of SVM and ELM is the best and the NSE of TPWB is the worst.were 102m~3/s,72.2 m~3/s,and 0.8%,respectively,with the largest combined error.(2)A runoff ensemble probabilistic forecasting model based on the BMA method was constructed.The MCMC-DREAM sampling algorithm was used to calculate the posterior distribution of the model parameters,and the optimal parameters of the model were estimated by the DREAM algorithm.The mean relative interval width,interval coverage and continuous ranking probability score are used as evaluation indexes of the ensemble probability forecast results,and the BMA models based on three distributions,namely,the normal distribution with homogeneous parametric variance,the normal distribution with heterogeneous parametric variance and the gamma distribution,are compared as posterior distributions.The results show that the BMA model based on the normal distribution with homogeneous variance has the highest RMSE;the BMA model with normal distribution with heterogeneous variance has the lowest RMSE among the three models.Compared with the ELM model,which is the best deterministic forecast model,the RMSE and MAE of the BMA model are slightly lower than those of the ELM,but the goodness-of-fit coefficient R~2 is better than that of the ELM.(3)An ensemble probabilistic forecast model for runoff based on the SCMM method is constructed.A single deterministic model is combined and an ensemble forecast model is established with random weighting,and a multi-objective genetic algorithm is used to rate the upper and lower limits of the ensemble forecast member model weights to finally obtain the forecast samples and the forecast intervals that can quantify the forecast uncertainty.The results show that the interval coverage of the SCMM model is 95%,which is higher than that of the BMA model(80%),and the forecast interval of the SCMM model is smaller at low flows and larger at high flows,which can reflect the forecast uncertainty more realistically,and the interval forecast performance of the SCMM model is better.The RMSE of SCMM mean forecast is 22.4 m~3/s,and that of BMA mean forecast is 85.7 m~3/s.The mean forecast error of SCMM model is much smaller than that of BMA model.Although the mean interval width of the SCMM model forecasts is higher than that of the BMA model,the interval and probability forecasting performance is better considering the simple structure and fewer parameters of this ensemble forecasting method.Therefore,the proposed SCMM method is still competitive in practical operational forecasting.