Long-term peak demand forecasting for solar penetrated distribution networks using gradient boosted ensembles: a 14-year case study of the Arkana–Muchea Corridor in Western Australia

For distribution networks with growing rooftop photovoltaic system penetration, it is essential to have an accurate long-term peak demand forecast to plan the distribution network. This research proposes a machine learning model to predict the daily peak apparent power demand for the distribution substations of Arkana and Muchea located in Western Australia (WA) using a 14.5-year time series (from January 2008 to June 2022) of 5,295 daily observations at each substation. Temporal variables, autoregressive lag features, rolling statistical features, and a photovoltaic penetration normalization index were used to perform feature engineering for six regression models. Overall, gradient-boosted ensemble models performed the best, with Light- GBM yielding the best accuracy of 0.937 MVA on the test set at Arkana, followed by XGBoost with 0.934 MVA and 0.9744 MVA, and neural networks with 0.945 MVA and 0.9728 MVA. Correlation analysis showed that there is a significant negative correlation between PV capacity and peak demand (correlation coefficient = −0.453, p<0.01), indicating the impact of PV distributed generation on peak demand. The framework proposed in this paper is accurate and transferable for long-term forecasting of PV-integrated distribution network peak demand, assisting in capacity planning, operational decision-making, and reliable power system management.