Short-Term Load Forecasting with Advanced Machine Learning Approaches: Spatial Variability and Model Performance

The increasing adoption of residential solar and the electrification of transport and heating brings unprecedented variability and randomness into load profiles, challenging the reliable and economic operation of the power grid. This variability is further compounded by unmonitored behind-the-meter (BTM) photovoltaic (PV) devices and weather fluctuations that affect solar generation and load demand. To address these complexities, this study explores a range of advanced forecasting techniques-including statistical methods, machine learning, and deep learning-to improve the accuracy of short-term load forecasting while assessing computational complexity. We applied these methodologies across different load zones in the New York power system and evaluated their performance against standard engineering practices. This research emphasizes the necessity for zone-specific forecasting models and suggests a cautious approach to adopting complex machine learning techniques in grid management. The most effective model achieved a mean absolute percentage error (MAPE) of 0.41% for day-ahead load forecasting, substantially outperforming the current industry practice of around 2%.