基于自适应LSTM代理模型和MCMC方法的地下水污染源反演识别研究OA

To efficiently and accurately conduct groundwater pollution source inversion,this paper employs deep learning methods,specifically the Long Short-Term Memory(LSTM)model and Multilayer Perceptron(MLP),to establish proxy models for pollution transport simulation.The DREAM-MCMC algorithm is then applied,using an adaptive update strategy to identify the groundwater pollution source inversion results.Finally,sensitivity analysis is used to discuss the inversion outcomes,thereby constructing a comprehensive groundwater pollution source inversion system.Two numerical examples are used to validate the proposed system.The results show that the LSTM-based proxy model achieves higher accuracy in simulating the model.Specifically,in Example 1,the three evaluation metrics—coefficient of determination(R²),Mean Squared Error(MSE),and Mean Relative Error(MRE)—reach 0.999 9,0.03 and 0.001,respectively,while in Example 2,the values are 0.883 4,333.65 and 0.362.In comparison,the proxy model built using the MLP method in Example 1 has values of 0.999 1,0.76 and 0.005,and in Example 2,the values are 0.810 3,665.42 and 0.262.These results demonstrate that the proxy model built using LSTM achieves higher approximation accuracy for the simulation model.By combining the DREAM-MCMC algorithm with the adaptive update strategy,and comparing it to the inversion method without the adaptive update strategy,the results indicate that the method with the adaptive update strategy generally exhibits lower relative errors in the inversion outcomes,confirming that this strategy significantly improves the inversion accuracy.Finally,the sensitivity analysis of the inversion results further elucidated the relationship between them.The results from the two numerical examples prove that this system can efficiently and accurately solve groundwater pollution source inversion problems,providing new insights into addressing groundwater pollution.