Carbon Emission Forecasting in Combined Gas-Vapor Cycle Power Plants via TimeGPT: A Large Time Series Model Approach

Accurate and interpretable real-time prediction of carbon emissions is essential for improving energy efficiency, ensuring adherence to regulatory standards, and fostering environmental sustainability in power plants. Nevertheless, existing methodologies frequently encounter challenges related to accuracy and interpretability, particularly when managing complex temporal dependencies and data scarce under various operational conditions. This study addresses these limitations by introducing a predictive framework that employs TimeGPT, a large time series generative pre-trained transformer model. The proposed approach integrates the mechanistic knowledge of power plant operations with advanced machine learning techniques to capture both physical and temporal patterns in carbon emissions. The model was validated using historical and real-time operational datasets from combined gas-vapor cycle (CGVC) power plants, demonstrating superior predictive performance across various plant conditions and fuel types compared to conventional methods. The study demonstrates that TimeGPT significantly outperforms AutoLSTM, AutoTCN, and AutoGRU in carbon emission prediction, as evidenced by the lower MAE, RMSE, and MAPE values across all experiments. In the data-rich scenario, TimeGPT achieved an MAE of 2.56, an improvement ranging from 46.54% to 57.39% over competing models, and surpassed them in RMSE and MAPE. The superior performance of TimeGPT, particularly under limited data conditions, underscores its promising applicability for short-term forecasting and operational decision making in power plant environments. Explainability analyses using SHapley Additive exPlanations (SHAP) values revealed key factors affecting emissions. The turbine exhaust duct pressure has the largest negative impact on predictions, reducing the estimated emission rate.