A novel Koopman-based Assistant Features Network for long and short-term carbon emission prediction

Abstract

The growing threat of global warming has heightened the importance of accurately predicting carbon dioxide emissions, a key factor in climate change mitigation. Previous studies have developed carbon emission prediction models using aggregated yearbook statistics, which are limited by spatiotemporal constraints and data scarcity. These models often assume data stationarity, overlooking the dynamic nature of carbon dioxide emission, leading to less reliable predictions that fall short of informing effective urban emission reduction strategies. This paper presents a novel approach by introducing two first daily, city-specific, multi-source carbon dioxide emissions datasets. It integrates global carbon dioxide emissions grid data from CarbonMonitor-graced with concentration measurements from the global carbon column total observation network. To make better use of these data features, we introduce a novel deep learning model, the Koopman Assistant Feature Network (KAFNet). By incorporating carbon dioxide concentration and calendar features as additional inputs, we apply the Koopman Operator theory, which is adept at analyzing the time-varying dynamics of complex systems. This approach allows for a comprehensive consideration of the underlying dynamics within carbon emissions data, enabling end-to-end optimization of predictive targets. Our empirical analysis reveals that our model significantly outperforms the sub-optimal models it was compared against, with an average reduction of 34.5% in Mean Absolute Error and 18.8% in Mean Squared Error. This enhancement in predictive accuracy provides a robust tool for capturing the evolving trends in carbon dioxide emissions, thereby offering a solid quantitative foundation to support data-driven decision-making in urban carbon reduction policies.