, Saurabh KUMAR 2
, Uruya WEESAKUL 3
, Apin WORAWIWAT 3
, Anshuman SINGH 4
, Dayanand SHARMA 5
2Research Unit in Climate Change and Sustainability, Department of Civil Engineering, Faculty of Engineering, Thammasat School of Engineering, Thammasat University, Pathumthani 12120, Thailand
3Department of Civil Engineering, Faculty of Engineering, Thammasat School of Engineering, Thammasat University, Pathum Thani, 12120, Thailand
4Department of Civil Engineering, National Institute of Technology Patna, Bihar, 800005, India
5Department of Civil Engineering, Sharda University, Greater Noida, Uttar Pradesh, 201310, India
Abstract
Groundwater contamination can affect people’s lives seriously as well as have detrimental effects on ecosystems. Therefore, it is essential to use accurate predictive tools in order to predict and control groundwater contamination. This research aims at predicting solute concentrations in groundwater applying two advanced methods – genetic programming (GP) and artificial neural networks (ANN). These two methods are critical in dealing with the problems associated with risk assessment, contaminants remediation and water resources management. Advection-dispersion equation is used as an important tool in solute transport analysis and applied in combination with ANN and GP in order to optimize numerical solutions. In this paper 1,361 points (1,040 points for training and 321 points for testing) were used as data input. Analytical solute concentrations were considered in the development of ANN and GP models. The ANN models include five different learning algorithms which are Conjugate Gradient with Powell-Beale Re-starts (CGB), Bayesian Regularization (BR), Conjugate Gradient with Feedback (CGF), Cartesian Genetic Programming (CGP) and Levenberg-Marquardt (LM). The analysis of the obtained results has proven that ANN-LM was the most successful model due to the highest correlation coefficient (~0.99) and the smallest root mean square error (RMSE) during both stages (training and testing). GP model correlated ~0.75 and provided 12% RMSE in the training process and 17% RMSE in the testing phase. The lowest values of AOC (area over the curve) were attained by ANN-BR (0.077 training/0.102 testing). ANN methods performed better when the amount of available hydrological data was limited. In conclusion, both ANN and GP prove themselves to be rather powerful tools for modelling groundwater contamination. It is noteworthy that the best results were obtained in the case of ANN-LM. Thus, it is reasonable to claim that ANN-based methods were more effective than others in this research. Novelty of this paper consists in com-bining artificial intelligence-based methods with advection-dispersion equation.
