Time Series Forecasting Based on Support Vector Machine Using Particle Swarm Optimization

Abstract

In recent years, due to the non-linear nature, complexity, and irregularity of time series, especially in energy consumption and climate, studying this field has become very important. Therefore, this study aims to provide a high accuracy and efficiency hybrid approach to time series forecasting. The proposed model is called EDFPSO-SVR (Empirical Mode Decomposition- Discrete Wavelet Transform- Feature selection with Particle Swarm Optimization-Support Vector Regression). In the proposed hybrid approach, the first step is to decompose the signal into the Intrinsic Mode Functions (IMF) component using the Empirical Mode Decomposition (EMD) algorithm. In the second step, each component is transformed into subsequences of approximation properties and details by converting the Wavelets. In the third step, the best feature is extracted by the PSO algorithm. The purpose of using the PSO algorithm is feature extraction and error minimization of the proposed approach. The fourth step, using time vector regression, has dealt with time series forecasting. Four data sets in two different fields have been used to evaluate the proposed method. The two datasets are electric load of England and Poland, and the other two datasets are related to the temperature of Australia and Belgium. Evaluation criteria include MSE, RMSE, MAPE, and MAE. The evaluation results of the proposed method with other Principal component analysis (PCA) feature extraction algorithms, and comparisons with methods and studies in this field, indicate the proper performance of the proposed approach.