In this study, a heuristic search strategy based on stochastic-simulation statistic (S–S) approach was developed for groundwater contaminant source characterization (GCSC) with simulation model parameter estimation. First, single kernel extreme learning machine (KELM) was built as surrogate system of the numerical simulation model to reduce huge computational load while evaluating the likelihood. However, compared with single KELM, multi-kernel extreme learning machine (MK-ELM) is more flexible for large amounts of data. To improve the approximation accuracy of the surrogate system to numerical simulation model, the MK-ELM surrogate system was first developed. Then, a heuristic search iterative process was first designed for GCSC with simulation model parameter estimation. The self-adaptive sampling method was proved to be more efficient than one-time sampling. Based on this idea, a self-adaptive feedback correction step was inserted into the heuristic search iterative process to ameliorate the training samples of the surrogate system in the posterior region, which further improved accuracy of simultaneous identification results. Finally, the identification results were obtained when the iteration terminated. The proposed approaches were tested in a hypothetical case study. It was shown that the heuristic search strategy can be used to assist in groundwater contaminant source characterization with simulation model parameter estimation.

在这项研究中，开发了一种基于随机模拟统计量（SS）方法的启发式搜索策略，用于通过模拟模型参数估计来进行地下水污染物源特征（GCSC）。首先，建立了单核极限学习机（KELM）作为数值模拟模型的替代系统，以减少巨大的计算量，同时评估可能性。但是，与单个KELM相比，多内核极限学习机（MK-ELM）在处理大量数据时更为灵活。为了提高替代系统对数值模拟模型的逼近精度，首先开发了MK-ELM替代系统。然后，首先设计了具有仿真模型参数估计的启发式搜索迭代过程。事实证明，自适应采样方法比一次性采样更有效。基于此思想，将自适应反馈校正步骤插入启发式搜索迭代过程中，以改善后部替代系统的训练样本，从而进一步提高了同时识别结果的准确性。最终，当迭代终止时获得识别结果。在假设的案例研究中对提出的方法进行了测试。结果表明，启发式搜索策略可用于模拟模型参数估计，以辅助地下水污染物源的表征。将自适应反馈校正步骤插入启发式搜索迭代过程中，以改善后部区域中替代系统的训练样本，从而进一步提高了同时识别结果的准确性。最终，当迭代终止时获得识别结果。在假设的案例研究中对提出的方法进行了测试。结果表明，启发式搜索策略可用于模拟模型参数估计，以辅助地下水污染物源的表征。将自适应反馈校正步骤插入启发式搜索迭代过程中，以改善后部区域中替代系统的训练样本，从而进一步提高了同时识别结果的准确性。最后，当迭代终止时获得识别结果。在假设的案例研究中对提出的方法进行了测试。结果表明，启发式搜索策略可用于模拟模型参数估计，以辅助地下水污染物源的表征。