Probabilistic Hydrologic Forecasting System Based On The Bayesian Method

On the basis of reviewing of the hydrological model, flood forecasting system and hydrologic forecasting uncertainty in home and in abroad, the parameter calibration techniques for conceptual hydrological model and the probabilistic hydrologic forecasting system were studied and discussed in detail. This study was supported by the National Key Basic Research Program and the National Natural Science Foundation of China. The main results and innovation points of this thesis were summarized as follows:(1) After concisely reviewed the parameter calibration techniques for conceptual hydrological model, the parameter optimization algorithms were introduced and compared and the parameter space was analyzed deeply. The multi-objective functions and their combination as well as the solutions were also discussed in detail. It is found that automatic calibration using multiple objectives could consider of different aspects of the discharge hydrograph and has a better goodness-of-fit than that of the traditional single objective function.(2) The Bayesian processor of forecasting (BPF) and hydrologic uncertainty processor (HUP) were introduced and applied in the Baiyunshan basin and Three Gorges Reservoir (TGR) intervening basin, respectively. An autoregressive model and a linear perturbation model (LPM) were proposed to describe the Bayesian prior distribution and likelihood function of the HUP, respectively. The modified HUP is superior to current HUP that can not only avoid the normal quantile transform, but also can decrease the errors associated with the selection and calibration of the marginal distribution of the observed and simulated discharge. The application results show that the modified HUP can improve the flood forecasting accuracy.(3) The quantitative precipitation forecasting (QPF) and the coupling of the QPF with hydrologic model, and a precipitation uncertainty processor (PUP) were concisely reviewed and studied. A stochastic precipitation generator for TGR intervening basin was developed and the nearest neighbor bootstrapping regressive (NNBR) method was adopted to simulate the temporal disaggregation of the precipitation forecasts. The stochastic precipitation generator was coupled with the proposed flood forecasting model for TGR intervening basin and the analytic distribution of the quantitative precipitation forecasting uncertainty was obtained. It is shown that the coupling model could increase the flood forecasting precision and could also provide the interval estimation of the discharges to be forecasted.(4) A precipitation-dependent hydrologic uncertainty processor (PD-HUP) was studied and discussed. A probabilistic hydrologic forecasting system based on Bayesian method was developed to simulate the influence of flood forecasting accuracy corresponding to the predictive uncertainty, which is decomposed into quantitative precipitation forecasting uncertainty and hydrological model uncertainty. The PD-HUP and the proposed stochastic precipitation generator for TGR intervening basin were integrated together to get the predictive density function. By using the hydrologic and meteorological data of the TGR intervening basin, the proposed system was studied and discussed, it is shown that the proposed system not only can increase the forecasting precision, but also provide more useful information for flood control decision-making, such as the predictive distribution, interval estimation and so on.(5) A real-time flood-updating model based on the Bayesian method was proposed and developed. The ARMA model was used to describe the prior distribution of observed discharge and the AR model was adopted to simulate the likelihood function of forecasting error. Both the prior distribution and the likelihood function were assumed to be linear-normal distribution and they were integrated together to form a Bayesian posterior distribution, the mean values were treated as the finial results to issue flood warring and flood control decision-making. It is shown that the proposed model not only has a more superior forecasting precision than AR model and the recursive least-squares estimate (RLS) model, but also provides the posterior distribution of the discharge to be forecasted as well as the interval estimation, which could combine forecasting process and decision-making process together.