Hybrid Double Seasonal ARIMA and Support Vector Regression in Short-Term Electricity Load Forecasting

Abstract

Forecasting is the main purpose of time series modelling. In short-term forecast, data can be predicted for a half hour-ahead. A half hour-ahead prediction faced with overlapping data series patterns risk. On the other hand, time series model can be analyzed with a linier or nonlinier approach. In this paper, we proposed the combination (hybrid) liner and nonlinier model for modelling the short-term electricity load in East Java. A half-hour electricity load forecasting is needed for real time controlling and short-term maintenance schedulling. However, the main problem of modelling time series data is determining linier or nonlinier time patterns. In short-term electricity load forecast, it depend on the moment of time (i.e weekdays, weekend, public holidays, joint holidays or religious holiday, etc) and the electricity load classification. In this analysis, we developed the Double Seasonal ARIMA (DSARIMA), Support Vector Regression (SVR), and hybrid DSARIMA-SVR. The DSARIMA model belong to linier model based on a well-known Box-Jenkins methodology. The SVR model belong to nonlinier model and the hybrid model is a mixing of linier and nonlinier models. The models are evaluated using Root Mean Square Error (RMSE) and Symmetric Mean Absolute Percentage Error (MAPE). The result shows that the accuracy of hybrid DSARIMA-SVR models are superior to the other individual models.