Application of BP Model Based on RAGA and MCMC in Probabilistic Flood Forecasting

The back-propagation artificial neural net (BP ANN) was used to describe the prior density and likelihood function of the Bayesian forecasting system (BFS) which is a general theoretical framework for probabilistic forecasting via deterministic hydrologic model. The posterior density of flood discharge was gotten by the Markov chain Monte Carlo (MCMC) simulation method based on the adaptive metropolis algorithm (AM), and then probabilistic forecasting of flood discharge was made by BFS. Real-coded accelerated genetic algorithm (RAGA) was used to optimize weights and bias of BP ANN and the initial samples in AM. The results of study case showed that BP net can catch the nonlinear functions of prior density and likelihood function well. The accuracy of forecast results by BFS was higher than that of results of Xin ´anjiang model. Not only mean of forecast discharge but also variance of forecast discharge was given by the BFS worked with BP ANN and MCMC. Then variance of predictand qualified the uncertainty of forecast, which can make decision-making of control flood more precise and reasonable.