On Efficient Tuning of LS-SVM Hyper-Parameters in Short-Term Load Forecasting: A Comparative Study

Power load forecasting is essential in the task scheduling of every electricity production and distribution facility. This paper studies the application of a variety of tuning techniques for optimizing the least squares support vector machines (LS-SVM) hyper-parameters in a short-term load forecasting problem. Clearly, the construction of any effective and accurate LS-SVM model depends on carefully setting the associated hyper-parameters. As a result, available optimization techniques including genetic algorithms (GA), simulated annealing (SA), Bayesian evidence framework and cross validation (CV) are applied and then compared for performance time, accuracy and computational cost. As a measure of effectiveness, the introduced algorithms are trained and tested on historical data obtained from Ontario´s Independent Electricity System Operator (IESO) for the Canadian city, Toronto. Experimental results show that optimized LS-SVM by Bayesian framework can achieve greater accuracy and faster speed than other techniques including LS- SVM tuned with genetic algorithm, simulated annealing and 10- fold cross validation.