Carbon Emission Forecasting Model for Carbon Asset Management and Power Data-Driven Based on TPE-Boosting Ensemble Stacking

In the context of low-carbon development, this paper proposes a combined ensemble learning model for predicting carbon emissions in key energy-intensive industries. The model consists of several steps. Firstly, the raw data is normalized and Pearson correlation analysis is conducted to determine the correlation coefficients between carbon emissions and various features. Secondly, strong features are selected, and different algorithms such as decision tree, support vector machine regression, K-nearest neighbors regression, and ridge regression are used as base learners in the AdaBoost framework for predicting carbon emissions. Thirdly, a BP neural network acts as the meta-learner, taking the results from the base learners as input. The hyperparameters of the neural network are optimized using the Tree-structured Parzen Estimator (TPE). This process generates the final prediction of carbon emissions. Finally, the proposed model is validated using emissions data from 120 cement enterprises. The evaluation metrics, including RMSE, MAPE, and R2, demonstrate that the TPE-Boosted Ensemble Stacking model performs better than the control group. This indicates the model's superiority in predicting carbon emissions in the cement industry. The TPE optimization algorithm's rapid optimization capability, together with the high accuracy and robustness of ensemble learning, contributes to the improved accuracy and stability of the prediction model.