PHYSICS-INFORMED HYBRID LEARNING FOR HOURLY POWER-PLANT EMISSIONS: A MULTI-POLLUTANT BENCHMARK WITH UNCERTAINTY AND TRANSFER

Abstract

This study examines whether the same modelling strategy can be used for hourly CO₂, NOx, and SO₂ emission estimation at power plants, or whether the choice of model should depend on the pollutant and deployment setting. We assembled 10,144,680 unit-hours from 583 generating units in Georgia, Kentucky, and Texas for 2022 and 2024 by combining U.S. EPA Clean Air Markets Program Data with EIA-860 plant metadata through the EPA/EIA Power Sector Crosswalk. Emission-factor, statistical, machine-learning, and physics-informed hybrid models were tested using blocked-time, plant-holdout, and seasonal-holdout splits to limit temporal and unit-level leakage. PHASE, the proposed hybrid, uses a scaled emission-factor estimate as a physical prior and learns the remaining residual with a data-driven model while also estimating predictive uncertainty. The results differ sharply by pollutant. For CO₂, the emission-factor baseline already explained most of the variance (R² = 0.992), whereas its performance was much weaker for NOx (R² = 0.394) and SO₂ (R² = 0.413). XGBoost reached R² = 0.999, 0.973, and 0.980 for CO₂, NOx, and SO₂, respectively. For NOx, removing recent historical features caused the largest deterioration in accuracy, while removing the simple physical prior slightly improved average in-domain MAE. At the same time, PHASE-lite reduced some large-error episodes according to the Diebold-Mariano comparison (p < 10⁻³), achieved 89.2% empirical coverage for a nominal 90% interval, and retained R² = 0.965 on previously unseen plants. The evidence therefore supports a differentiated strategy: a transparent factor model is already strong for CO₂, while NOx and SO₂ require learned representations of changing operating conditions. The hybrid is most useful when uncertainty, large-error robustness, and cross-plant transfer matter more than the lowest average in-domain error. A future corroborative comparison with Sentinel-5P TROPOMI NO₂ observations over Kazakhstan thermal plants is also defined.