A domain-informed forecasting framework for short-horizon carbon emissions prediction in power plants under reduced-history conditions

Accurate short-horizon forecasting of CO₂ emissions from thermal power plants is increasingly important for unit-level operational forecasting and emission-aware operation support. In practice, emission forecasting is hindered by limited plant-specific data, operating regime shifts, and the limited interpretability of purely data-driven models. This study proposes a domain-informed predictive framework that uses thermodynamic analysis to construct physically meaningful exogenous variables for a pre-trained time-series forecasting model in combined gas-vapor cycle (CGVC) power plants. By incorporating physically meaningful variables, the framework improves feature relevance and supports stable short-horizon forecasting within the evaluated reduced-history operating conditions. The approach is evaluated on real plant operational data under data-rich, reduced-history, and short-horizon settings. The corresponding observed MAE ranges for TimeGPT and TimeGPT-EF are 2.565–2.584, 2.784–2.792, and 2.115–2.283 t/h, respectively. Lightweight plant-specific fine-tuning further supports rapid adaptation using limited same-day operating data. SHAP-based analysis identifies turbine exhaust duct pressure, steam flow, and temperature-related variables as major contributors to the model predictions. Rolling-origin evaluation provides preliminary within-unit evidence of short-horizon predictive stability, while split-conformal calibration improves the empirical coverage of the original under-calibrated prediction intervals on the May 24 test set. This study provides a practical and interpretable framework for short-horizon carbon emission forecasting in power plants, with potential applicability to emission monitoring after further validation across broader operating conditions.