Ensemble Kalman filter (EnKF) and optimization methods are two mainstream methods in groundwater contamination source identification, but most researchers usually only use or improve on one method. However, each method has strengths and weaknesses, and using one method alone may cause inaccuracy of the identification results for complex situations. Therefore, it is necessary to explore the combination of methods. In this paper, for the optimization method, to further enhance the solution precision of the optimization model for groundwater dense non-aqueous phase liquid (DNAPL) contamination source identification (GDCSI), we first constructed an improved butterfly optimization algorithm by introducing the dynamic switching probability mechanism in the previous butterfly optimization algorithm. Next, the EnKF method and optimization method (based on the improved butterfly optimization algorithm) were used respectively for GDCSI, to assess the strengths and weaknesses of the methods. Then, according to the strengths and weaknesses of the EnKF and optimization methods, the two methods were merged to build a more robust and practical combined search method (CSM) for GDCSI, to further enhance the identification accuracy and effectiveness. In addition, we innovatively applied the EnKF method to combine two single deep learning surrogate models: the deep belief neural network surrogate model and the deep residual neural network surrogate model. And a deep learning combined surrogate model was established, which further enhanced the approximation precision and applicability to the groundwater DNAPL contamination multiphase flow numerical simulation model. Finally, the results of CSM were compared with those of two single methods to analyze the former's effectiveness. The results showed that the CSM significantly improved the identification accuracy and effectiveness of GDCSI compared with any single method.
Read full abstract