A scalable explainable deep learning framework for predictive analysis and interpretability of CO₂ emissions patterns

For environmental monitoring and making decisions that are good for the environment, it is important to be able to accurately anticipate CO \(_2\) emissions. This paper presents an interpretable deep learning framework utilizing a BiLSTM network combined with SHAP for CO \(_2\) prediction. Using a sliding window technique, CO \(_2\) time-series data are reformed to show how things change over time. To stop data leaking, chronological data separation and min-max normalization are used. The proposed BiLSTM model achieved the best result. It has an RMSE of \(8.58 × 10⁵\) , an MAE of \(1.19 × 10⁵\) , a \(R^2\) score of 0.99, and a MAPE of 0.0073% on an unseen dataset. Global and local explanations based on SHAP give complementing feature-level insights using a distinct explainability model trained on engineered country, sector and temporal attributes.