Genetic Algorithm and Particle Swarm Optimization for Parameter Optimization of Least-Square Support Vector Regression Model in Electricity Load Demand Forecasting

Abstract

Abstract Accuracy is very important in time series forecasting where the model obtained depends on historical, linear, or nonlinear data patterns. This study aims to forecast short term electricity load in East Java, Indonesia, whereas electricity load is one of major and vital need in our daily life. The linear approach is carried out using the ARIMA method, while the nonlinear approach used in this study consist of the SVR and LSSVR models. The parameter selection greatly affects the results of accuracy, so an optimization method is needed. The usual grid search optimization method does not guarantee an optimum solution and more importantly, it is not efficient in some practical applications. Therefore metaheuristic optimization methods are needed, including GA and PSO which can find the entire possible space for the search for solutions. PSO is easier to apply but is prone to have premature convergence due to trapped in the minimum locale. To overcome this, modified PSO is developed to seek the optimal position according to the specified criteria. The result of this research is that the accuracy of the nonlinear approach is much better than the linear approach. The addition of an optimization method to the SVR provides a significant change in accuracy compared to LSSVR. Meanwhile, MPSO is the best optimization method because it produces a lowest RMSE value.