The prediction models are the basis for scientific formulation of emergency disposal and rescue measures. In order to quickly and accurately construct the forecasting model of sudden water pollution accidents, this paper regards the problem as the Bayesian estimation problem and obtains the posterior probability density function of the model parameters according to the finite difference method and Bayesian inference. Then, by using the improved Metropolis-Hastings sampling method, more reasonable prediction model parameters are obtained. Taking the sudden water pollution event in a certain open channel as an example, the effects of different observation noises on the parameters calibration results are discussed under the two scenarios of non-uniform flow control with non-uniform flow and non-equal capacity control, and compared with the parameter and true value of the Bayesian-Markov chain Monte Carlo method. The results show that, the improved Bayesian-Markov chain Monte Carlo method can give a better parameter value, more application and has a good anti-noise, which can provide a new approach to build the forecast model of sudden water pollution accidents.