Very Long Term Load Forecasting at Grid Supply Points with Limited Historical Load Data Availability

Very long-term load forecasts (VLTLFs) spanning 15 to 25 years are indispensable for strategic electricity infrastructure planning. These forecasts are even more challenging when the historical load data is scarce in addition to uncertain future load growth. This paper presents a methodology to generate seasonal peak load forecasts for grid supply points (GSPs) until 2050, despite having limited years of available recorded data. This is demonstrated with two real UK GSPs with 5 years of available data. First, macroeconomic and demographic variables (regional GDP and household statistics) are mined and interpolated to match seasonal resolution. A transfer-learning backcasting step employs ridge regression, which is compared and selected for its superior performance in reconstructing historical GSP loads. The exogenous variables are first forecasted using ARIMA, and the extended dataset then feeds into seasonal ARIMAX models to project future load peaks. Finally, projected additional demand from electric vehicles and heat pumps is incorporated based on DSO Future Energy Scenarios. For the first GSP considered, the 2035 and 2050 peak forecasts differ from DFES’ electric engagement scenario by 15% and 17%, respectively, while for the second one, the analogous differences are 12% and 8%. The methodology delivers robust, explainable VLTLFs with quantified uncertainty, demonstrating its suitability where load data availability is limited.