中文摘要：

BP神经网络模型用于水质进行评价的研究已经很多，然而，传统的BP神经网络无法考虑相邻水质级别临界处的模糊性，评价指标较多时运行速度慢，且由于训练样本少和代表性差，评价结果精度不高。建立了基于AM—MCMC算法的RAGA—BP模型，利用RAGA能够选出最优的BP网络初始结构；AM—MCMC算法模拟足够的代表性好的样本为BP网络训练所需，用于灌区的水质评价。实例研究表明，与传统的BP网络相比，基于AM—MCMC的RAGA—BP网络收敛速度提高约20％，评价结果与实际水质比较更为客观、合理。基于AM—MCMC的RAGA—BP模型能考虑相邻水质级别临界处的模糊性，克服训练样本少的缺点生成足够的代表性好的样本。快速有效地对灌区水质进行评价。此外，基于AM—MCMC的RAGA—BP模型还可用于洪灾损失评价、地震灾害评价及其他评价问题，具有广泛的实用性。

英文摘要：

Back Propagation Artificial Neural Net is widely used in evaluation of water quality, but it can not determine the fuzziness between adjacent grades of water quality and the convergence velocity and accuracy of estimation are low for lack of training samples. So the BP ANN based on Real Coded Accelerating Genetic Algorithm and Markov Chain Monte Carlo which based on Adaptive Metropolis was built and used to evaluation water quality. RAGA was used to optimize topology, initialize weights and bias of BP; AM-MCMC was adopted to produce enough simulated samples for training BP net and to determine fuzziness between adjacent grades of water quality. Adaptive Metropolis method was taken as a sampling method to improve sampling efficiency of MCMC. Results showed that RAGA-BP based on AM-MCMC can improve convergence velocity by 20%, and the evaluation results of RAGA-BP are more objective, reasonable than that of single indicator method. The model proposed in the paper, considering the fuzziness of boundary between adjacent grades, overcoming the fault of lack of training samples, can rapidly evaluate water quality for irrigation area, the RAGA-BP based on AM-MCMC can be used to evaluate the loss of flood and earthquake.