Very-Short-Term Load Forecasting for AI-Data Centers Using Gaussian Process Regression

Data centers form the critical backbone of the digital infrastructure, underpinning the explosive growth of cloud computing and Artificial Intelligence (AI) workloads. The substantial energy consumption and rapidly fluctuating workload patterns observed in the AI-focused data centers create significant challenges for operational efficiency, cooling management, and integration with sustainable energy systems. As a result, accurate load forecasting is essential for effective resource allocation, energy optimization, and ensuring reliable, uninterrupted supply of electricity. In this paper, we employ kernel-based Gaussian Process Regression (GPR) to forecast the load demand of an AIdata center over a very short-term horizon - specifically, minuteahead forecasts with a$\mathbf{1 5}$-second resolution. In particular, we develop three GPR models, each using a different kernel, and train them on a set of previously observed AI-data center load measurements. The models are then evaluated on their ability to forecast demand over a 30-day period, and their performance is compared against a Light Gradient Boosting Machine (LightGBM) model. The results clearly indicate that all GPR models outperform LightGBM, with the GPR model using the Squared Exponential kernel achieving the highest forecasting accuracy.