Groundwater contamination source identification based on Rat Swarm Optimizer algorithm with a back-propagation neural network surrogate model

ABSTRACT In groundwater contamination remediation, accurate portrayal of the hydraulic parameters of the contaminated site with accurate identification of the contaminant release history is crucial. In this paper, the Karhunen–Loeve (KL) expansion was used to describe the non-homogeneity of the hydraulic conductivity while modeling the simulation model. In the solution process of the optimization model, typical algorithms such as genetic algorithms (GAs) might encounter problems such as low search efficiency. To address the challenges, this paper adopted the Rat Swarm Optimizer (RSO) algorithm to solve the optimization model. Compared with GA, the RSO's average relative error of identify the source position was 1.55% and the source strength was 2.61%, both of which were better than 7.91 and 97.9%, respectively, in GA. To solve the problem of a shortage of computational resources caused by repeated and massive calls to the simulation model, this paper considered the use of a back-propagation neural network (BPNN) to establish a surrogate model of the simulation model. According to the results, the BPNN surrogate model had a high fitting accuracy to the simulation model, with R2 of 0.9906 and RMSE of 9.2446 g/day. The identification of groundwater contamination sources based on BPNN and RSO achieved excellent results.