CarbonNet: Enterprise-Level Carbon Emission Prediction with Large-Scale Datasets

The precise prediction of carbon emissions is crucial to combat global climate change and foster sustainable development. Conventional carbon emissions forecasting usually relies on limited records of national or regional levels, which is coarse and compromises its accuracy. Furthermore, the data collection requires lengthy statistical cycles and high costs, making it too costly to provide timely feedback from the forecasting. Additionally, they are targeted at specific fields thus hard to construct universal models. To overcome these challenges, we propose CarbonNet, a novel firm-level carbon emission prediction scheme. To build large-scale firm-level datasets, we crawled carbon emission data and reporting data (e.g., financial statements) of 3346 companies over 31 years containing 688 data fields, and combined them together. A preprocessing scheme is proposed to aggregate data with different statistical intervals or sources, and many outliers. A factor-analysis-based features extraction scheme is proposed to build a generalized forecasting model for different types of companies. A machine learning scheme is proposed for big data mining and long-term forecasting. We evaluated CarbonNet on real-world datasets. Results show that it achieves a median relative error of 0.25, outperforming others by 22%. The corresponding carbon emissions dataset has been made publicly available to advance related research.