Identifying groundwater contamination sources based on the hybrid grey wolf gradient algorithm and deep belief neural network

• Published in: Stochastic environmental research and risk assessment Vol. 37; no. 5; pp. 1697 - 1715
• Main Authors: Li, Jiuhui, Wu, Zhengfang, He, Hongshi, Lu, Wenxi
• Format: Journal Article
• Language: English
• Published: Berlin/Heidelberg Springer Berlin Heidelberg 01.05.2023; Springer Nature B.V
• Subjects: Accuracy; Algorithms; Aquatic Pollution; Chemistry and Earth Sciences; Computational efficiency; Computational Intelligence; Computer applications; Computer Science; Computer simulation; Computing costs; Computing time; Contamination; Earth and Environmental Science; Earth Sciences; Environment; Groundwater pollution; Iterative methods; Math. Appl. in Environmental Science; Neural networks; Nonlinear systems; Optimization algorithms; Optimization models; Original Paper; Physics; Probability Theory and Stochastic Processes; Simulated annealing; Simulation; Simulation models; Statistics for Engineering; Waste Water Technology; Water Management; Water Pollution Control; Deep learning; Groundwater contamination; IGCSs; Hybrid grey wolf gradient optimization algorithm; Simulation optimization
• ISSN: 1436-3240; 1436-3259
• DOI: 10.1007/s00477-022-02360-6

The simulation optimization (S/O) method is widely used in the identification of groundwater contamination sources (IGCSs). However, in most cases, the IGCSs has the characteristics of many variables to be identified and a high degree of nonlinearity. When the grey wolf optimization algorithm (GWO) is used to solve the optimization model for this kind of problem, due to the relatively weak local search ability it has the disadvantage of premature convergence. To improve the GWO, the GWO and gradient descent algorithm were integrated to construct a hybrid grey wolf gradient optimization algorithm (HGWGO) with little dependence on the initial value and a strong local search ability. The HGWGO was then applied to solve the optimization model and improve the accuracy of the IGCSs results. In addition, when solving the optimization model, calling the simulation model hundreds of times will generate a large calculation load and consume a massive amount of computing time, which would seriously hinder the IGCSs. Thus, a surrogate model of the simulation model was established by applying a deep belief neural network (DBNN) to participate in the iterative calculation. The results showed that compared with the simulated annealing algorithm and GWO, the HGWGO had a higher calculation accuracy and could improve the accuracy of IGCSs. Although the HGWGO requires a high computational cost, the improvement in solution accuracy was sufficient to compensate for this shortcoming. The DBNN surrogate model had a high accuracy and could save 99% of the computing time by participating in the iterative calculation instead of the simulation model.