CREST: Covariate-Regularized Emission Forecasting with Seasonal-Temporal Decomposition for Carbon Emission Prediction

Accurate forecasting of industrial carbon emissions is essential for translating corporate sustainability commitments into measurable outcomes. However, ESG data is often fragmented, sparse, and temporally heterogeneous, making reliable forecasting challenging. We propose CREST, Covariate-Regularized Emission Forecasting with Seasonal-Temporal Decomposition, a hybrid forecasting architecture combining a rule-regularized neural network with residual temporal modeling via Prophet. The network captures nonlinear emission-driver relationships under domain-enforced constraints, while Prophet recovers residual seasonal and trend dynamics. To address data scarcity, we introduce structure-preserving time-series augmentation via seasonal decomposition with calibrated residual jitter. Evaluated on an open-source dataset (2.28M records, 10 Indian industrial sectors) and a proprietary industrial dataset (2,080 records, 45 plants, 15 countries), the augmented CREST model achieves MAPE of \textbf{0.13\%} and \textbf{0.45\%} respectively, outperforming all baselines by a substantial margin.