Research on carbon emission prediction of coal-fired units integrating multi-dimensional feature selection and stacking model

Global warming has become an urgent challenge for human society. To enhance the accuracy of carbon emission prediction in coal-fired power plants and support energy conservation and emission reduction efforts, this study develops a carbon emission prediction model based on two 330 MW units in a coal-fired power plant in Hainan. Fourteen features are initially selected from unit operation, coal quality, environmental, and temporal dimensions. Feature importance ranking and screening are performed using Random Forest. A stacking ensemble learning model is then constructed with five distinct algorithms serving as base learners. After dynamic weighting, their predictions are used as meta-features and fed into a meta learner composed of ridge regression and GBDT, with hyperparameters optimized using a random search strategy. The results indicate that the model achieves optimal performance when six key features are input: unit power generation, coal consumption for power generation, unit electricity consumption, temperature, wind speed, and load coefficient. Under these conditions, the R2 reaches 0.9440, and compared with other models, the MAE and RMSE are reduced by 11.68% and 10.84%, respectively. The proposed model can serve as a high-precision carbon accounting tool for coal-fired power plants and provide data support for the formulation of emission reduction strategies.