Predicting Carbon Emissions via Deep Learning Time Series Models

Reducing carbon emissions is a major global challenge, and is crucial for China to achieve its dual-carbon goal. Accurate carbon emissions predictions can provide valid data support for governments and businesses to better estimate the effect of emission reduction measures and adjust corresponding policies flexibly. With the increase of intelligence in industry, the sampling interval of carbon emissions data is gradually coming from quarterly to days or even hours. The statistical time series prediction method is no longer suitable for the huge amount of new data. To overcome the limitations of existing studies, this paper explores the effectiveness of the PatchTST model, a deep learning approach, across five key carbon emissions sectors in 31 provinces in Mainland China. We compare PatchTST with traditional statistical models such as ARIMA and machine learning models like Prophet in terms of prediction accuracy and robustness. Our results indicate that PatchTST not only outperforms these models but also exhibits superior generalizability across different regions and sectors. To the best of our knowledge, this study marks the first comprehensive evaluation of PatchTST's potential in carbon emissions forecasting, highlighting its suitability for data-driven time series prediction tasks in environmental sciences.