Short-term electrical load prediction has always played an important role in power system planning and operation. Support vector machines (SVMs) have been successfully employed in last decade on various prediction problems from different areas, and showed a high rate of success in load forecasting. Least squares support vector machines (LS-SVMs) as reformulation of standard SVMs bring significant improvements in terms of model simplification and computing time reduction. In this paper a forecasting model is proposed for the hourly electrical power load demand, using LS-SVMs. Model is based on historical daily load demands in combination with calendar and climate features. The historical daily loads are chosen to have the similar calendar category as the forecasting period, in order to compose LS-SVM training set with similar meteorological features. The presented model was tested by applying it on real-life load data obtained from electric distribution utility “ED Jugoistok” for the territory of city Niš and its surroundings. Experimental results show that the proposed approach gives acceptable errors both hourly and daily for the entire period of prediction and provides an interesting alternative for electricity load forecasting.
load forecasting, least squares support vector machines, time series.