Data-centric strategies for carbon-efficient carbon-intensity forecasting

Conference item

Carbon-intensity forecasting and carbon-aware scheduling are vital for decarbonizing flexible loads in data centers, yet existing approaches neglect the carbon cost of the forecasting models themselves and lack a unified performance–carbon metric. We introduce the EcoAdjusted Accuracy Score (EAAS), a novel metric that penalizes CO2 emissions per unit of accuracy. We train five representative time-series learners (ARIMA, linear regression, Prophet, LightGBM, XGBoost) on identical UK grid carbon intensity traces and rank them by EAAS. To improve EAAS score, and building on an XGBoost baseline, we propose three new data-centric optimizations: Data Input Reduction, Vectorized Operations, Cache-Friendly Processing, both individually and in a combined hybrid form. These strategies show promising potential in cutting runtime and carbon emissions while matching or improving predictive accuracy, boosting EAAS by up to 14.7% over the baseline. These results highlight the potential of targeted data manipulations, without hardware or scheduler changes, in balancing between forecasting performance and environmental impact.

Authors: Wang, X; Zilberman, N

• Publication status: Published
• Peer review status: Reviewed (other)
• Publisher: Figshare
• Host title: Proceedings of the International Workshop on AI Systems for the Environment (AISE-2025)
• Publication date: 2025-10-12
• Acceptance date: 2025-07-25
• Event title: International Workshop on AI Systems for the Environment (AISE 2025), presented at the 28th European Conference on Artificial intelligence (ECAI 2025)
• Event location: Bologna, Italy
• Event website: https://sites.google.com/view/aise25
• Event start date: 2025-10-25
• Event end date: 2025-10-25
• DOI: 10.6084/m9.figshare.30339046
• Language: English
• Keywords: carbon-intensity forecasting; Eco-Adjusted Accuracy Score (EAAS); data-centric optimization