Grid-Oriented Probabilistic Load Forecasting for AI Data Centers

Qingyi Yao; Ding Lin; Tianqiao Zhao; Meng Yue; Jianhui Wang

eess.SYNov 6, 2026

Grid-Oriented Probabilistic Load Forecasting for AI Data Centers

Authors: Qingyi Yao, Ding Lin, Tianqiao Zhao, Meng Yue, Jianhui Wang

Organizations: Southern Methodist University, Dallas, TX, USA · The University of Texas at Arlington, Arlington, TX, USA · Brookhaven National Laboratory, Upton, NY, USA

Abstract

Probabilistic load forecasts for AI data centers need to represent both uncertainty under changing workloads and temporal dependence to support grid operations. This paper develops an open, grid-oriented benchmark using measured GPU power from a 448-GPU AI cluster. Forecasts are evaluated through four stages: probabilistic prediction, online uncertainty calibration, temporal scenario generation, and horizon-wide reserve sizing. Following a change in cluster utilization, a quantile LightGBM forecast achieves only 63.4% coverage for a nominal 90% interval, while online regime-conditioned calibration increases coverage to 87.3%. Even with well-calibrated marginal forecasts, independent sampling produces unrealistic ramp distributions and excessive reserve procurement. Reconstructing temporal dependence from historical forecast errors reduces the discrepancy in ramp statistics and lowers realized reserve cost by 7.4-10.6% across three forecasting tasks. These findings show that marginal forecast quality alone is insufficient for operational decisions and motivate joint evaluation of calibration, temporal structure, and decision outcomes.

Figures & tables

Fig. 1: MIT Supercloud TX-Gaia cluster, March–September 2021, hourly means. (a) GPU IT power, including an assumed 40 W idle draw for unallocated GPUs. (b) GPUs allocated by the scheduler. Shading marks the training, validation and test periods.

Fig. 2: Grid-oriented evaluation framework. Mini-plots show each stage’s output for the test origin 2021-08-02 11:00 UTC (1-h resolution, 24-h horizon). Each stage passes its output to the next and is scored with its own metrics (ES: energy score, VS: variogram score, KS: Kolmogorov–Smirnov distance).

5-min / 1-h
Forecaster	CRPS	MAE	Cov.	Min cov.	Winkler
	(kW)	(kW)	(%)	(%)
Persistence	0.504	0.738	82.9	68.0	8.01
LightGBM	0.481	0.792	82.4	73.1	7.07
LightGBM + OC	0.469	0.783	87.9	77.2	6.80
TCN	0.438	0.747	88.3	86.8	6.11

TABLE I: Forecast skill and calibration on the test period (Aug.–Sep. 2021)

Fig. 3: Empirical coverage of the central 90% interval by workload regime at the forecast origin, 1-h resolution, 24-h horizon, test period.

Fig. 4: Scenarios from the quantile TCN, 1-h resolution, 24-h horizon. (a), (b) Twenty scenarios for the test origin 2021-08-02 11:00 UTC, selected as the largest observed 1-h ramp-up among origins outside the low regime whose observed window stays above 20 kW. (c) Empirical CDF of the maximum 1-h ramp-up within the horizon over all 1440 test origins.

5-min / 1-h
Generator	ES	VS	KS	KS	Reserve	Short-
			peak	ramp	cost	fall (%)
Independent	2.38	17.7	0.52	0.70	0.353	4.7
Gaussian copula	2.30	13.0	0.14	0.35	0.333	7.1
Residual bootstrap	2.43	13.7	0.06	0.07	0.390	11.0
Schaake shuffle	2.30	12.5	0.08	0.07	0.326	10.5

TABLE II: Scenario quality and horizon-wide reserve, quantile TCN, test period

Electricity Load Forecasting

Electricity load forecasting predicts future power demand to support grid operation, planning, and energy management.

See all 50 papers

A Physics-Aware Framework for Short-Term GPU Power Forecasting of AI Data Centers
Apr 14, 2026Mohammad AlShaikh Saleh, Sanjay Chawla, Sertac Bayhan +2Physics-Informed Forecasting Time Series Forecasting
College of Science and Engineering, Hamad Bin Khalifa University, Doha, Qatar · Qatar Computing Research Institute, Hamad Bin Khalifa University, Doha, Qatar · Qatar Environment and Energy Research Institute, Hamad Bin Khalifa University, Doha, Qatar +1
Bayesian Transformer for Probabilistic Load Forecasting in Smart Grids
Mar 9, 2026Sajib Debnath, Md. Uzzal MiaElectricity Load Forecasting Probabilistic Time Series Forecasting
O&M Analytics, AES Clean Energ; The AES Corporation; Louisville, CO 80027, USA · Department of Information and Communication Engineering; Pabna University of Science and Technology; Rajapur, Pabna, Bangladesh
Power-Flexible AI Data Centers: A New Paradigm for Grid-Responsive Compute
Jun 23, 2026Chris Williams, Philip Colangelo, Ayse Coskun +15Demand Response Carbon-Aware Computing
Emerald AI · Electric Power Research Institute (EPRI) · National Grid +2

Apr 14, 2026J/K move · Enter open · S save

A Physics-Aware Framework for Short-Term GPU Power Forecasting of AI Data Centers

Mohammad AlShaikh Saleh, Sanjay Chawla, Sertac Bayhan, Haitham Abu-Rub+1

College of Science and Engineering, Hamad Bin Khalifa University, Doha, Qatar · Qatar Computing Research Institute, Hamad Bin Khalifa University, Doha, Qatar · Qatar Environment and Energy Research Institute, Hamad Bin Khalifa University, Doha, Qatar+1

Physics-Informed Forecasting Time Series Forecasting Electricity Load Forecasting Data Center Energy Management Physics-Informed ML

Open paper AskPDF arXiv

AI data centers experience rapid fluctuations in power demand due to the heterogeneity of computational tasks that they have to support. For example, the power profile of inference and training of large language models (LLMs) is quite distinct and big divergences can result in the instability of the underlying electricity grid. In this paper we propose, to the best of our knowledge, the first physics-informed DLinear time-series model that can accurately forecast power utilization of an AI data center 5-80 minutes (short-term forecasting) into the future. The physics, based on a multi-node lumped thermal resistance-capacitance (RC) network consistent with Newton's law of cooling, is captured using newly derived time-dependent ordinary differential equations (ODE) that separately models and interlinks power consumption with the GPU compute and memory utilization and temperature. The resulting model, that we refer to as PI-DLinear, trained and evaluated on a real AI data center dataset and is not only more accurate than the state-of-the-art (SOTA) models tested, but the forecast profile respects the underlying physics under power throttling and load transient events. Relative to the SOTA transformer-based and non-transformer-based models, improvements in forecasting accuracy (averaged across all look-back and prediction windows) range from 0.782%-39.08% for MSE, 0.993%-51.82% for MAE, and 0.370%-22.28% for RMSE.

Abstract

Electricity Load Forecasting

Explore similar work

A Physics-Aware Framework for Short-Term GPU Power Forecasting of AI Data Centers

Bayesian Transformer for Probabilistic Load Forecasting in Smart Grids

Power-Flexible AI Data Centers: A New Paradigm for Grid-Responsive Compute