Machine Learning-Based Long-Term Forecasting of Load Composition Considering Exogenous Variables

Long-term load forecasting (LTLF) plays a key role in the planning and operation of power systems, enabling utilities to anticipate demand patterns and ensure grid stability. This paper presents a novel approach to LTLF by exploring the influence of various exogenous variables not only on load forecasting accuracy at grid supply point (GSP) but also on the accuracy of forecasting the composition of the load. Using a bidirectional gated-recurrent unit (GRU) neural network, the study integrates macroeconomic factors, weather conditions, and social dynamics as key exogenous factors in the 1-year-ahead active and reactive load forecasts. Then, with a feed-forward neural network, the forecasted load is disaggregated into load categories for every day of the year. Three GSPs from a real distribution network with varying gross domestic product (GDP) and mean flow characteristics serve as case studies to validate the proposed model. Results of the first stage indicate that incorporating exogenous variables enhances forecasting performance, with population demonstrating the strongest correlation for these particular datasets. The results of the second stage demonstrate that the forecasted load can be decomposed into constituent categories accurately enough and that load decomposition is not very sensitive to errors in the forecast. The decomposed load forecast results were validated with a Household Electricity Survey study performed by the UK government.