CREST: Structure-Preserving Learning for Sparse Industrial Emission Forecasting

Accurate industrial carbon-emission forecasting is difficult when measurements are sparse, operational conditions vary across facilities, and future operational covariates are unavailable at deployment. We introduce \textbf{CREST}, a lightweight hybrid forecasting framework that combines \emph{rule-regularised neural learning} of emission--operation relationships with \emph{residual temporal decomposition} using Prophet. CREST is backbone-agnostic and separates operational prediction from systematic temporal variation, allowing future operational covariates to be unnecessary at inference. We further introduce structure-preserving augmentation that perturbs only the residual component of an STL decomposition, preserving trend and seasonality while increasing training diversity. We evaluate CREST on an open-source benchmark containing 2.29M records from 10 Indian industrial sectors and a real industrial dataset covering 2,080 observations from 45 plants across 15 countries, using strictly temporal train/test splits. CREST-FFN achieves 1.59\% MAPE on the open-source dataset and 2.68\% on the industrial dataset without augmentation, outperforming conventional machine-learning baselines under heterogeneous industrial conditions. Structure-preserving augmentation further reduces MAPE to \textbf{0.13\% and 0.45\%}, respectively, corresponding to 92\% and 83\% reductions in error. Notably, the lightweight FFN achieves performance comparable to recurrent CREST variants while using substantially less computational complexity. These results show that combining domain constraints, temporal decomposition, and structure-preserving data augmentation can provide a robust and efficient alternative to scaling neural forecasting models for sparse industrial settings.