Prediction of Groundwater Quality Index and Identification of Key Variables Using Bayesian Neural Network

Abstract

Prediction of groundwater quality index (GWQI) and quantification of the influence of each key water quality variable is vital for water resource management. This paper shows the potential of the automatic relevance determination-based Bayesian neural network(ARD-BNN) technique for GWQI prediction and compares the predictive performance of ARD-BNN with artificial neural network (ANN), long short-term memory network (LSTM), convolutional neural network (CNN) and hybrid CNN-LSTM. What discriminates the present approach from the previous approach is that GWQI is simulated from the global criteria of WHO and the artificial intelligence (AI) models are trained and cross-validated in a synthetically simulated data set of 1470 examples and applied to predict GWQI using newly collected groundwater samples from an area surrounding Amarpur dolerite dyke, Dhanbad, Jharkhand (INDIA). The present analysis suggests that ARD-BNN is relatively skilful (MSE_ARD-BNN = 0.01) relative to other AI models investigated (MSE_ANN = 0.06;MSE_CNN = 1.60; MSE_LSTM = 3.85;MSE_CNN-LSTM = 8.21). The efficacy of the method and stability of results is also tested in the presence of different levels of correlated noise which suggests that the ARD-BNN model is considerably unwavering for up to 20% correlated noise; however, adding more noise (∼50% or more) degrades the results. Sensitivity analysis via ARD-BNN-based soft-pruning strategy identifies that [NO₃⁻], [SO₄²⁻],[pH],[F⁻]and [EC], are key water quality variables for predicting GWQI in the study area.