Towards CLEAN: Cycle Learning Energy-Emission Assessment Network (Papers Track)
Yanming Guo (University of Sydney); Jin Ma (University of Sydney); Kevin Credit (Maynooth University); Qian Xiao (Maynooth University)
Abstract
Formulating effective climate policies is challenging due to the limitations of current decision-making models, such as Computable General Equilibrium models, which often rely on static, linear assumptions that oversimplify complex, real-world scenarios. To address these limitations, we introduce the Cycle Learning Energy-Emission Assessment Network, a scalable deep-learning framework trained on the Carbon Monitor Energy-Emission time series dataset. CLEAN captures global patterns between energy structures and sectoral carbon emissions, enabling accurate energy-emission predictions at national, regional, city, and sectoral levels. By framing the problem as a bidirectional supervised regression task, our model achieves a Mean Absolute Percentage Error of approximately 5%, significantly outperforming traditional models like XGBoost, especially with our novel data augmentation method, Densify. This demonstrates CLEAN’s superior performance and generalization capabilities, making it a powerful tool for climate policy analysis. The CMEE dataset, CLEAN model, and Densify augmentation methods are open-sourced at \url{https://anonymous.4open.science/r/CLEAN-D66A}.