Short-term load forecasting (STLF) is used by building operators to make informed decisions about electricity usage and purchase. Recently, research has investigated the potential of ensemble learning for improving forecast accuracy with researchers often to combine several of the same type of learning machines into an ensemble. In this paper, a novel diversified ensemble learning method is proposed and implemented for a small library building in San Antonio, Texas. Three different machine learning models – feedforward neural network (FFNN), long short-term memory (LSTM) network, and support vector regression (SVR) – are trained separately and their outputs are combined using an ensemble FFNN using the back-propagation training method to further reduce the prediction error of the load forecast. The proposed model is tested using smart meter data including outdoor temperature and total building electrical load at 15-minute intervals. The model is tested using data gathered from four different seasons of the same year, and is shown to capture the seasonal dependencies, with a mean absolute percentage error of 7%, while it outperforms the invididual prediction models in the majority of the tested cases.