CarbonOracle: Automating Energy Mix & Renewable Energy Source Forecast Modeling for Carbon-Aware Micro Data Centers

Geo-distributed data centers (DCs) have a substantial impact on global electricity consumption and carbon emissions, with their energy demands expected to increase alongside emerging technologies such as Generative Artificial Intelligence (GenAI) and Natural Language Understanding (NLU). In response to environmental and operational concerns, major cloud providers are investing in DC infrastructures powered by renewable energy sources (RES). However, the design and management of energy-efficient data centers present new challenges. Current forecasting models for RES production and electricity grid energy mix are often limited in accuracy and forecasting horizon, hindering carbon-aware service management. To tackle these challenges, we introduce CarbonOracle, a Machine Learning (ML) service that automates data extraction from self-hosted RES, energy grids, and weather APIs, while also simplifying the ML training and forecast of RES production and electricity grid carbon emissions. Its application programming interface serves ML-based forecasts for RES production (e.g., solar, wind) and energy mix metrics, designed to support carbon-aware deployments, enabling integration with container schedulers and other applications. Through a comprehensive evaluation over a real data center testbed, our results show that CarbonOracle has an error rate of approximately 9% for forecasts related to self-hosted photovoltaic (PV) panels, while its forecasts for electricity grid carbon emissions have an error rate of less than 4%.