Groundwater Pollution Source Identification via an Integrated Surrogate Model and Multiobjective Heuristic Optimization Algorithms

Abstract Simulation‐optimization methods are commonly used in groundwater pollution source identification. Traditional simulation‐optimization methods require multiple calls to the numerical model, which leads to a considerable computational burden. Surrogate models based on machine learning can replace numerical models while maintaining accuracy. Previous studies have focused on the fitting accuracy of surrogate models, this study emphasizes the importance of the precision of surrogate models for the inversion process. We use the analytic hierarchy process to integrate ConvLSTM, convolutional neural network, and BiLSTM to improve the precision of the surrogate model. GMS is used to construct numerical models of two hypothetical cases and a practical case. Compared with the best results of the single deep learning methods, the integrated surrogate model improves the precision of the solution of the two hypothetical cases by 90% and 26%, respectively. In addition, the accuracy of the pollution source information obtained by incorporating the integrated surrogate model into the optimization model is higher than that obtained by ConvLSTM as the surrogate model. The inversion results of 7 metaheuristic optimization algorithms are compared through two hypothetical cases, and then the optimization algorithm with higher accuracy is applied to the solution of the practical case. To obtain more accurate results, we reobtain a batch of training data by resampling and training the integrated surrogate model. The results show that constructing an integrated surrogate model and selecting an optimization algorithm can improve the solution accuracy of the simulation‐optimization method. This research provides a new perspective for the construction of simulation‐optimization methods.