A0233
Title: Optimization of electricity demand forecasting using machine learning ensemble methods for Israel's energy grid
Authors: Moshe Kelner - University of Haifa and Noga - Israel System Operator (Israel) [presenting]
Abstract: Accurate electricity demand forecasting is critical for Israel's isolated energy grid, producing 80 billion kWh annually from domestic sources. Daily generation unit scheduling relies on demand forecasts, making accuracy essential for economic and environmental optimization. Forecast errors have significant consequences. Underestimation forces expensive backup unit activation, while overestimation causes unnecessary generation costs. An ensemble system is developed, combining five machine learning algorithms to predict daily electricity demand and peak demand for three-day horizons. The approach incorporates 1,200 meteorological features, including hourly temperatures across regions, humidity, and historical patterns. Feature selection used statistical significance testing, with training on 2000-2022 data and 2023 testing. Research addresses three methodological challenges - establishing performance-based ensemble weighting criteria, improving rare extreme event prediction despite limited examples, and developing adaptive algorithms for consistent performance during seasonal transitions with high weather variability. Model performance was evaluated using MAPE between forecasted and actual values. The ensemble demonstrates improved accuracy over individual models, directly supporting energy sector decisions through optimized scheduling, cleaner source prioritization, reduced costs, and enhanced grid reliability.
