Bias calibration and error propagation adjustment for ML-based time series forecasting: A systematic study for PJM’s electricity load forecast amid Virginia’s data center surge

Accurate electricity load forecasting is crucial for grid stability and market efficiency. While advanced machine learning (ML) models offer marginal accuracy gains, they often introduce complexity, interpretability challenges, and high computational costs. This work systematically evaluates time series forecasting accuracy and proposes simple yet effective correction methods to enhance performance. We decompose forecast errors into three components: model bias from historical inertia, errors in dependent variables, and random noise. Based on this framework, we develop a two-stage correction method consisting of bias calibration and error propagation adjustment to improve prediction accuracy efficiently. Our empirical evaluation of PJM’s hourly load forecasts, aligned with its operational guidelines, demonstrates the method’s effectiveness across a large regional power grid spanning 13 states and Washington, D.C. A case study in Virginia’s Dominion region, where data center growth increases uncertainty, shows that these corrections nearly eliminate forecast bias. The proposed approach offers a computationally efficient, theoretically grounded alternative to complex ML models, delivering significant accuracy improvements.