Power Load Forecasting with Least Squares Support Vector Machines and Chaos Theory

Abstract

In this paper, a novel approach to power load forecasting based on least squares support vector machines (LS-SVM) and chaos theory is presented. First, with the data from EUNITE network, we find the chaotic characteristics of the daily peak load series by analyzing the largest Lyapunov exponent and power spectrum. Average mutual information (AMI) method is used to find the optimal time lag. Then the time series is decomposed by wavelet transform. Cao's method is adopted to find the optimal embedding dimension of the decomposed series of each level. At last, with the optimal time lag and embedding dimension, LS-SVM is used to predict future load series of each level. The reconstruction of predicted time series is used as the final forecasting result. The mean absolute percentage error (MAPE) is 1.1013% and the maximum error is 25.1378 MW, which show this approach is applicable for power load forecasting