An analysis of the feasibility of SARIMAX-GARCH through load forecasting

Load forecasting is critical for energy sector planners and generation companies to predict the level of energy that should be generated to maximize energy security. Accurate predictions of future energy consumption and supply capacity increase public confidence in energy sector planners, and help make informed decisions on power generation infrastructure to reduce capital costs. To forecast load, energy sector planners have used variations of the AutoRegressive Integrated Moving Average (ARIMA) model, such as Seasonal AutoRegressive Integrated Moving Average with eXogenous factors (SARIMAX). However, the accuracy of these models is limited due to heteroskedasticity, where the variance of data is not constant. Consequently, ARIMA can be combined with General AutoRegressive Conditional Heteroskedasticity (GARCH), a model that forecasts variance, in the SARIMAX-GARCH model. We hypothesized that SARIMAX-GARCH will be more accurate in predicting load than SARIMAX, a variant of ARIMA. We trained SARIMAX-GARCH and SARIMAX and selected the best models using Akaike Information Criterion (AIC) and Bayes Information Criterion (BIC), resulting in a Mean Absolute Percentage Error (MAPE) of 13.2% for SARIMAX and 11.0% for SARIMAX-GARCH. This shows that SARIMAX-GARCH is more accurate than SARIMAX for load forecasting data. The results indicate that SARIMAX-GARCH could potentially be improved through ensemble techniques and other exogenous variables. The results of our study will help energy sector planners and generation companies in forecasting energy consumption with more accurate predictions.