Fuel Consumption Prediction in Shovel-Truck System of Surface Mine Using Artificial Neural Network

Abstract

The operations of a mine revolve around the budget, which describes the concept of cost of operations and returns generated. The fuel expended by each mining truck varies depending on differing conditions and is therefore difficult to be estimated based solely on manufacturers' manuals. Most mining companies have challenges forecasting the anticipated fuel consumption which has an essential influence on the budget. Usually, there is an under or overestimation of fuel consumed which goes beyond the stipulated allowable limits. BCM International Limited, Nzema, faces difficulties in accurately estimating fuel consumption for planning and budgeting purposes. A review of previous studies has shown that the Backpropagation Neural Network (BPNN) is a method noted for fuel consumption prediction. However, the prediction capability of the BPNN is dependent on the type of training algorithm used for the weight adaptation and bias assignment. Because of that, this study applied and tested three training algorithms namely the Levenberg Marquardt (LMA), Bayesian Regularisation (BRA) and the Scaled Conjugate Gradient (SCGA) to build an Artificial Neural Network (ANN) model to predict truck fuel consumption at BCM International Limited. The motive here was to select the best-performing training algorithm. It was deduced that the LMA produced the best model results with the least Mean Absolute Percentage Error (MAPE) value of 10.63% followed by SCGA and BRA. The model testing also revealed that the fuel consumed can be predicted from the input elements by approximately 96% using the LMA, 87% by BRA and 83% by the SCGA. The optimum BPNN structure with the LMA was 3-48-1 describing an input of three variables, 48 neurons in the hidden layer and one output.