This paper compares the performance of two single classifier paradigms (k-nearest neighbor, and radial basis function) and two multiple classifier (ensemble) techniques (random forest, and stacking). These machine learning techniques are used to predict, and monitor underground pumps energy usage consumption and water dam levels for an underground single-pump station in a gold mine in South Africa. Introducing machine learning intelligent and predictive systems to mining industry may result into better safety and reduce the consumed electrical power. The results show that random forest (RF) is more capable of predicting the underground pumps energy consumption than the k-nearest neighbor, stacking and the RBF methods. With respect to the underground dam level prediction results, RBF performed better than the other three machine learning methods with the highest overall performance when other performance measures other than prediction accuracy are measured.