Forecasting Carbon Emissions in the AI Industry: Integrating ESG Semantics with Large Language Models

Abstract

This study introduces an explainable framework for forecasting carbon emissions, specifically designed for firms in the AI industry. It integrates structured financial indicators with semantic features sourced from corporate sustainability reports. Leveraging Large Language Models (LLMs) alongside a hybrid LSTM-Attention architecture, our proposed model effectively captures both temporal dependencies and textual insights related to environmental, social, and governance (ESG) factors. Empirical results demonstrate that our approach significantly outperforms traditional baselines across standard evaluation metrics. Furthermore, SHAP analysis enhances the model's interpretability, revealing firm size and capital intensity as key predictors. Our research contributes to the AI industry by operationalizing unstructured ESG narratives into interpretable, scalable inputs for environmental risk modeling and sustainable AI governance.