Hierarchical forecasting of monthly electricity demand

Abstract

Energy demand forecasting is a vital tool for energy management, maintenance planning, environmental security, and investment decision-making in liberalised energy markets. The mini-dissertation investigates ways to anticipate power usage using hierarchical time series and South African data. Approaches such as topdown, bottom-up, and optimal combination are applied. Top-down forecasting is based on disaggregating total series projections and spreading them down the hierarchy based on historical data proportions. The bottom-up strategy aggregates individual projections at lower levels, whereas the optimal combination methodology optimally combines bottom forecasts. An out-of-sample prediction performance evaluation was performed to assess the models’ predicting ability. The best model was chosen using mean absolute percentage error. The top-down technique based on predicted proportions (Top-down forecasted proportions) was superior to the optimal combination and bottom-up approach. To integrate forecasts and build prediction ranges for the proposed models, linear quantile regression, linear regression, simple average, and median were used. The best set of forecasts was picked based on the prediction interval normalised average width. At 95%, the best model based on the prediction interval normalised average width was a simple average.