Application of neural network techniques to predict the heavy metals in acid mine drainage from South African mines

Abstract

Acid mine drainage (AMD) refers to acidic water generated during mining activities and is characterised by a low pH, high salt content, and the presence of heavy metals. To treat water sources contaminated with AMD, sampling and laboratory analysis will have to be done for each water source to determine the concentrations of heavy metals. This process is time-consuming, high in cost and may involve human error or negligence. The application of neural network (NN) techniques to predict the heavy metals in AMD from South African mines has been presented. Four specific objectives were pursued in this dissertation. The first one was to identify AMD and analyse for heavy metals in the AMD. Heavy metals that were identified and found to be in high concentrations in the AMD sample from Sibanye Western Basin AMD Treatment Plant are Zn, Fe, Mn, Si, and Ni. The other objectives of the study were to determine the input, output, and hidden layers of the NN structure (application of NN); (2) to find the appropriate algorithm to train the NN, and to compare the NN results (outputs) with the measured concentrations of major heavy metals sampled (targets). The Backpropagation Neural Network (BPNN) model had three layers which included the input layer (pH, SO42−, and TDS), the hidden layer (five neurons) with a tangent sigmoid transfer function (tansig) and the output layer (Cu, Fe, Mn, and Zn) with linear transfer function (purelin). The predictions for heavy metals (Zn, Fe, Mn, Si, and Ni) using the NN method focusing on a BP forward pass (feed-forward backpropagation NN) with ten different algorithms were presented and compared with the measured data. The mean square error (MSE) value was calculated for ten algorithms and compared to identify the one that is most appropriate for the prediction process and the model by having the lowest value. It was determined that the Levenberg-Marquardt back-propagation (trainlm) algorithm resulted in the best fitting during training because it resulted in an MSE value of 0.00041, meaning the error was very low when this algorithm was used.