Operational data-driven carbon emission prediction for natural gas power plants: A hybrid feature selection framework with bootstrap stability verification

As global carbon reduction targets become increasingly stringent, power generation enterprises require high-precision carbon emission prediction models. However, traditional prediction approaches are limited by unstable feature selection and low accuracy. To address these challenges, the present study proposes a multi-level machine learning optimization strategy. The feature space is hierarchically refined through a statistical filtering layer, a multi-dimensional evaluation layer, and a recursive feature elimination layer, followed by Bootstrap stability verification and hyperparameter tuning. The framework is validated using real operational data from a natural gas power plant in China. Results show that all six optimized models achieve substantial performance improvements, with XGBoost performing best. XGBoost reaches an R2 of 0.99, an RMSE reduction from 559.69 to 55.14 tons CO₂/day, and a MAPE improvement from 8.14% to 3.45%. This result is further confirmed by rolling-window time-series cross-validation, which maintains a mean R2 of 0.96 across twelve temporal folds, indicating that the high accuracy does not result from overfitting. The feature space is also effectively reduced from 18 variables to 4 core features. The proposed hybrid optimization strategy not only enhances prediction accuracy but also significantly improves model interpretability and practical applicability.