A methodology for long-term hourly electricity demand forecasting driven by GDP and invariant load profiles

Long-term electricity demand forecasting is necessary for grid and operations planning, as well as for the analysis and planning of energy transition strategies. However, accurate long-term load forecasting with high temporal resolution remains a challenge as most existing approaches focus on aggregated forecasts, which require accurate prediction of many other variables. This study proposes a parsimonious methodology that uses t-tests to verify the temporal invariance of historical load profiles and the correlation of load with gross domestic product (GDP) to arrive at a long-term hourly load forecast. Applying this methodology first to a controlled benchmark analysing Singapore’s isolated grid, and of multi-year historical data (2004–2022) reveals that its relative hourly load has remained statistically invariant, with an overall percentage deviation of 4.24% across seasonality indices. Utilizing these invariant historical load profiles, five-year-ahead total yearly forecasts were generated using GDP as a predictor, and the hourly loads were then forecasted using hourly seasonality index fractions. The maximum Mean Absolute Percentage Error (MAPE) across multiple experiments of six years ahead forecasts was only 6.87%. To systematically evaluate the generalizability and macroeconomic limitations of this top-down methodology beyond Singapore, the methodology was subsequently tested across six European countries. Results show that the methodology achieves high accuracy (MAPE < 8%) for economies with strong GDP-load correlations (Belgium, Bulgaria, Croatia, Spain) but exhibits some performance degradation in decoupled economies where macroeconomic growth no longer drives electricity demand expansion linearly (Germany, UK). Additionally, invariant seasonality metrics were incorporated into a short-term forecasting model based on exponential smoothing (ES), demonstrating comparable or improved accuracy relative to existing machine learning-based methods. The findings show that parsimonious approaches can be effective in arriving at long-term, high resolution forecasts.