CarbonGPT: Meta Causal Graph-Enhanced Large Language Models for Carbon Emission Forecasting in Large-Scale Power Distribution Networks

Carbon emission reduction has emerged as a global core development objective, where dynamic carbon flow perception serves as the critical foundation for low-carbon dispatch. However, the electrification shift of transportation carbon emissions caused by large-scale electric vehicle (EV) grid integration, coupled with the grid impact from renewable energy source (RES) volatility, poses unprecedented challenges for accurate carbon emission prediction. Although adaptive graphs and large language models (LLMs) can achieve carbon emission prediction for power distribution networks (PDNs) under data sparsity scenarios, the spurious cross-variable correlations derived from their adaptive topologies tend to be amplified by LLMs, consequently constraining model learning and reasoning capabilities. To address this, we propose a novel model named CarbonGPT. This innovative model employs a causal encoder to uncover genuine cross-variable causal relationships, while incorporating a meta causal graph dictionary and lightweight alignment to enhance the comprehension of carbon feature representations by LLMs in EV and RES grid integration scenarios. Extensive simulations in PDNs under large-scale integration of EVs and RESs have demonstrated that CarbonGPT consistently achieves state-of-the-art performance in both prediction accuracy and effectiveness. Codes are available at https://github.com/lishijie15/CarbonGPT.