A novel carbon emission estimation method based on electricity-carbon nexus and non-intrusive load monitoring

Accurate carbon accounting is foundational for power enterprises' participation in the carbon market. Current research on estimation of carbon dioxide emissions through electricity-carbon index analysis primarily relies on an enterprise's total electricity consumption, which often leads to uncertainty and poor interpretability. In reality, carbon dioxide emissions within an enterprise are predominantly generated in specific key processes, indicating a strong correlation between the electricity consumption of key equipment and carbon dioxide emissions. In this context, to enhance both estimation accuracy and interpretability, a two-stage deep learning structure is proposed. This structure leverages non-intrusive load monitoring and employs deep learning algorithms to first disaggregate an enterprise's total electricity consumption into the consumption of key equipment and then use these data to estimate carbon dioxide emissions. Utilizing real-time data from a power plant in China, the proposed two-stage deep learning structure as well as three representative deep learning networks (Recurrent Neural Network, Long Short-Term Memory network, and Gated Recurrent Unit network) are applied to construct electricity-carbon models for carbon dioxide emission estimation. The experimental results highlight that when employing the two-stage structure, models across three deep learning networks demonstrate a marked enhancement in estimation accuracy compared to traditional models. The two-stage structure reduces the mean squared error (MSE) for models across these three networks by 55.1%, 47.9%, and 46.9%, respectively, compared to their baseline values. This research aims to enhance the precision of carbon dioxide emission estimation and serves as a valuable reference for electricity-carbon research in various sectors.