Dynamic Electricity Carbon Emission Factor Estimation for Gas-Fired Power Units Using Temporal Encoding and Bayesian Calibration

To address the critical challenge of insufficient accuracy in dynamic carbon emission factor (CEF) calculation for thermal power units due to the lack of hourly fuel consumption data, this study innovatively proposes a carbon emission inversion model (TS-BEFNet) that integrates temporal feature encoding with Bayesian physical constraints. The model constructs a bidirectional LSTM encoder-decoder network to extract latent generation efficiency coefficients in an unsupervised manner from hourly power generation and environmental features. A variational Bayesian correction framework is then established to dynamically adjust efficiency deviations, enabling the derivation of real-time carbon emission data and dynamic CEFs. Case study validation demonstrates that the model achieves over 96.5% accuracy in daily carbon emission estimation. Compared to default CEF values for gas-fired units, the model improves real-time carbon emission monitoring accuracy by more than 30.96% during heating seasons and 4.83% in non-heating seasons, while effectively capturing real-time emission trends of gas-fired units.