Homotopy-based hyper-heuristic searching approach for reciprocal feedback inversion of groundwater contamination source and aquifer parameters

Abstract

Groundwater contamination source identification is critical for taking effective measures to design remediation strategies, assess contamination risks, and confirm contamination responsibilities. To resolve the “equifinality” problem resulting from simultaneous inversion of contamination source characteristics and aquifer parameters at dense non-aqueous phase liquid-contaminated sites, two reciprocal optimization frames for separately identifying the contamination sources and aquifer parameters were designed and connected. The two sets of identification results were corrected stepwise by means of a feedback correction iteration process, thereby sufficiently improving the identification accuracy. The ensemble learning machine (ESLM) incorporating Kriging, radical basis function neural network, support vector regression, and wavelet kernel extreme learning machine with swarm intelligence (SI) algorithm was embedded into the reciprocal inversion iterations to replace the multiphase flow simulation model for significantly improving the computational efficiency. To improve the optimization efficiency, a hyper-heuristic homotopy algorithm was constructed for segmentally searching the global optimum in wider areas with low dependence on initial values. Results showed that the combined application of SI-based ensemble learning machine (SI-ESLM) and hyper-heuristic homotopy algorithm effectively accomplished the simultaneous identification of contamination sources and aquifer parameters with high efficiency, while maintaining high accuracy. The SI-ESLM sufficiently approximated the outputs of the multiphase flow simulation model with increased certainty (R2=0.9977), while the mean relative error was limited to 1.5388%. Compared to traditional heuristic algorithms, this application of reciprocal inversion iterations and the hyper-heuristic homotopy algorithm significantly reduced the mean relative error of identification results from 6.51% to 1.03%.