Short-term forecasting of hierarchical time series in electricity consumption: An application using South African data

Abstract

This article presents a comprehensive framework for short-term forecasting of hierarchical electricity consumption using South African data. The study promotes the precision and validity of the predictions of various energy sources by applying Stochastic Gradient Boosting (SGB) and XGBoost with reconciliation techniques. The results demonstrate that the XGBoost model effectively predicts the electricity consumption generated from solar (PV and CSP), coal, and diesel energy sources. Conversely, the SGB approach performs more efficiently in forecasting electricity consumption met by the electricity generated from nuclear and wind energy sources and emphasises using model-specific approaches for different energy sources. This research prefers applying multiple forecasting methods to improve overall accuracy of forecasting electricity consumption met by non-renewable energy sources. At the same time, hybrid models were particularly helpful for forecasting electricity consumption met by complex energy sources like wind and nuclear. Further, the research delves into the estimation of prediction intervals under linear regression and linear quantile regression and observes that the latter is superior with narrower interval widths, reduced interval scores and similar coverage probabilities. Findings capture notable gaps in the literature and generate real-world observations for energy policy-makers through the implication of hybrid methodologies for enhancing the quality of forecasts for electricity consumption.