Grid-Oriented Probabilistic Load Forecasting for AI Data Centers

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

AI summary · arXiv: 2611.08929v1 · eess.SY

Grid-Oriented Probabilistic Load Forecasting for AI Data Centers

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

Original paper Submitted to arXiv on 06 Nov. 2026 License: NONEXCLUSIVE-DISTRIB 1.0 Also in: fr

How to read this summary? Illustrated guide

The essentials

An open benchmark shows that calibrating forecast intervals and preserving temporal dependence both matter for realistic AI data-center load scenarios and reserve sizing.

Question

How well do probabilistic forecasts of AI-cluster power capture changing workload uncertainty and temporal dependence, and how do these properties affect grid-oriented reserve decisions? The paper also asks whether accessible measured data and a staged evaluation can reveal where forecast value is lost.

Method

The authors use GPU telemetry and scheduler logs from a 448-GPU cluster, covering 5 February to 1 October 2021, to define three tasks: 5-minute resolution over 1 hour, 15-minute resolution over 4 hours, and hourly resolution over 24 hours. They compare four forecasters, apply a regime-conditioned online calibration method, and evaluate four scenario generators using 200 trajectories; reserve decisions are assessed across three reserve-to-shortfall cost ratios.

Findings

Result At the 24-hour horizon, LightGBM's nominal 90% interval covered 63.4% of test outcomes, while online calibration raised coverage to 87.3%.
See the passage in the paper
The online correction raises LightGBM coverage to 87.9%, 87.9% and 87.3%, and the lowest regime coverage at the 24-h horizon from 57.3% to 82.2%.
Result Independent scenario sampling produced the largest energy and variogram scores in all three tasks and substantially distorted ramp and peak statistics.
See the passage in the paper
Independent sampling has the largest ES and VS in every task, and its KS distances for the peak (0.52–0.66) and the largest ramp-up (0.70–0.88) show that the observed statistics are concentrated at one end of the simulated distribution.
Result The Schaake shuffle had the lowest energy and variogram scores in all three tasks and reproduced ramp-up and peak statistics more closely than independent sampling.
See the passage in the paper
The Schaake shuffle combines the conditional marginals of the TCN with the empirical rank structure. It has the lowest ES and VS in all three tasks and KS distances of 0.07–0.13 for ramp-up and 0.08–0.09 for the peak.
Result Using the Schaake shuffle reduced realized reserve cost by 7.4–10.6% across the three tasks compared with independent scenario sampling.
See the passage in the paper
Across the three tasks the cost falls by 7.4%, 9.1% and 10.6%, and the shortfall rate is 10.5%, 9.3% and 7.6%.
Authors’ interpretation The authors conclude that reserve sizing benefits from scenarios that preserve both forecast-conditioned uncertainty and temporal dependence.
See the passage in the paper
These results underscore the need for load scenarios that preserve both forecast-conditioned uncertainty and temporal dependence when sizing reserves over a forecast horizon.

Limitations

The benchmark targets cluster GPU IT power rather than total data-center demand, so its results do not directly characterize CPU, cooling, or facility losses.
See the passage in the paper
The target is therefore cluster GPU IT power, which excludes CPU, cooling and facility losses.
The evaluation uses data from a single cluster during 2021, limiting the scope of the reported empirical findings.

Why it matters

The paper provides an open, staged benchmark that connects probabilistic forecasts to scenario quality and reserve outcomes using measured AI-cluster GPU power. The authors argue that marginal forecast accuracy or calibration alone is not enough: grid-oriented evaluation also needs realistic temporal dependence and decision-level assessment.

Written by AI from the paper. Passages are checked word for word against the text we analysed; the rest may contain errors.

AI-generated keywords: AI data centers probabilistic forecasting electricity demand reserve sizing temporal dependence

Read further

Comprehensive Summary
Layman's Summary
Blog article
Original abstract

The paper examines whether probabilistic forecasts of AI data-center electricity demand can support grid operations when they capture both uncertainty at each forecast step and how demand changes over time. This matters because reserve requirements depend on ramps, peaks, and sustained high demand—not just average forecast accuracy. To investigate, the authors create an open benchmark using power telemetry and scheduler logs from MIT’s 448-GPU TX-Gaia cluster, covering February to October 2021. The forecast target is GPU information-technology power, including an assumed idle draw of 40 watts for each unallocated GPU. It excludes CPU use, cooling, and other facility loads. The benchmark includes three forecasting tasks: 5-minute forecasts over one hour, 15-minute forecasts over four hours, and hourly forecasts over 24 hours. The authors compare persistence and seasonal-naive baselines with quantile LightGBM and a temporal convolutional network (TCN). They also evaluate a four-stage process: probabilistic forecasting, online uncertainty calibration, full-horizon scenario generation, and reserve sizing. The online calibrator adjusts forecast quantiles based on recent errors and workload regime. For scenario generation, the paper compares independent sampling, a Gaussian copula, bootstrapped historical error trajectories, and a Schaake shuffle, which combines forecast-specific marginal distributions with historical error-rank patterns. Reserve decisions are assessed by comparing the realized cost of holding reserve with the cost of shortfall. The TCN achieves the best continuous ranked probability score (CRPS) across all three tasks: 0.438, 0.741, and 1.462 kW, respectively. However, persistence has lower mean absolute error than the TCN at the two shorter resolutions, showing that model performance depends on the metric. A key challenge is a mid-May increase in cluster utilization, after which demand remains at a higher level throughout the test period. At the 24-hour horizon, LightGBM’s nominal 90% interval covers only 63.4% of outcomes. Online, regime-conditioned calibration raises coverage to 87.3%, improves CRPS from 1.746 to 1.551 kW, and increases the lowest regime-specific coverage from 57.3% to 82.2%. The TCN is already closer to nominal coverage at this horizon, with 87.6%. Calibration makes little difference to its coverage and worsens its CRPS to 1.535 kW. These results suggest that calibration can help models that do not adequately track changing conditions, but may add unnecessary interval width to better-conditioned forecasts. Accurate marginal calibration alone does not guarantee realistic trajectories. TCN forecast errors have strong lag-one correlations, ranging from 0.84 to 0.90. Independent sampling removes this dependence and distorts ramp and peak distributions. The Schaake shuffle performs best overall on scenario scores and closely reproduces key trajectory statistics. At a reserve-cost ratio of 0.1, its realized reserve cost is 7.4%, 9.1%, and 10.6% lower than that of independent sampling across the three tasks. Its shortfall rates are 10.5%, 9.3%, and 7.6%, compared with 4.7%, 4.2%, and 3.3% for independent sampling. The latter procures too much reserve relative to the 10% target. The authors conclude that grid-oriented forecasting should evaluate calibration, temporal dependence, and decision outcomes together. The benchmark covers only one cluster and a historical period, and its target excludes non-GPU facility demand. The idle-power contribution is also assumed rather than measured. Broader testing is therefore needed before generalizing the reserve-cost benefits to other data centers or grid settings.

Why AI Data-Center Forecasts Need to Look Beyond Averages

AI data centers can place new demands on electricity grids, but predicting their power use is not simply a matter of estimating an average. Grid operators need to know when demand might surge, how quickly it could ramp up, and whether high usage is likely to persist. Those patterns influence how much reserve capacity to hold—and how much it may cost to prepare for uncertainty.

In “Grid-Oriented Probabilistic Load Forecasting for AI Data Centers,” Qingyi Yao, Ding Lin, Tianqiao Zhao, Meng Yue, and Jianhui Wang investigate whether probabilistic forecasts can represent both uncertainty at individual forecast steps and the way demand evolves over time. Their results point to an important distinction: a forecast can describe uncertainty reasonably well at each moment and still produce unrealistic demand trajectories across a full day.

An Open Benchmark Based on GPU Power

The researchers build an open benchmark from power telemetry and scheduler logs for MIT’s 448-GPU TX-Gaia cluster. The data covers February through October 2021. The forecast target is GPU information-technology power. It includes an assumed idle draw of 40 watts for every unallocated GPU, but excludes CPU use, cooling, and other facility loads.

The benchmark tests forecasts at three operational time scales:

Five-minute forecasts extending one hour ahead.
Fifteen-minute forecasts extending four hours ahead.
Hourly forecasts extending 24 hours ahead.

The comparison includes straightforward persistence and seasonal-naive baselines, as well as quantile LightGBM and a temporal convolutional network (TCN). The researchers also evaluate a wider process: first forecasting demand probabilistically, then calibrating uncertainty online, generating full-horizon scenarios, and finally sizing reserves based on the potential costs of holding reserve versus experiencing a shortfall.

Forecast Accuracy Depends on the Metric

The TCN achieves the best continuous ranked probability score (CRPS) across all three tasks: 0.438 kW for the one-hour task, 0.741 kW for the four-hour task, and 1.462 kW for the 24-hour task. Yet it does not win on every measure. Persistence has lower mean absolute error than the TCN at the two shorter resolutions.

This contrast matters for grid applications. A model that performs well on one summary metric may not be the best choice for every operational need. Evaluations should consider the kind of uncertainty and error that affects decisions, rather than relying on a single score.

Changing Workloads Can Break Forecast Calibration

A key challenge appears in mid-May, when cluster utilization rises and demand remains at a higher level throughout the test period. That shift tests whether forecast uncertainty remains trustworthy when conditions change.

At the 24-hour horizon, LightGBM’s nominal 90% prediction interval contains only 63.4% of outcomes. The online, regime-conditioned calibrator adjusts forecast quantiles using recent errors and workload regime. With calibration, coverage rises to 87.3%, CRPS improves from 1.746 to 1.551 kW, and coverage in the worst regime increases from 57.3% to 82.2%.

The TCN behaves differently. Its 24-hour coverage is already 87.6%, closer to the nominal 90%. Calibration makes little difference to coverage and worsens its CRPS to 1.535 kW. The findings suggest that calibration can help a model that struggles to track changing conditions, but may unnecessarily widen intervals when a model is already reasonably well conditioned.

Realistic Scenarios Must Preserve Demand Over Time

For reserve planning, knowing the distribution at each forecast step is not enough. Scenarios also need to reflect how errors at nearby times move together. The TCN’s forecast errors have strong lag-one correlations, ranging from 0.84 to 0.90. Independent sampling breaks that relationship, potentially distorting the ramps and peaks that operators need to anticipate.

The researchers compare independent sampling with a Gaussian copula, bootstrapped historical error trajectories, and a Schaake shuffle. The shuffle combines forecast-specific marginal distributions with historical patterns in the ranking of errors across time. It performs best overall on the scenario scores and closely reproduces important trajectory statistics.

Those differences carry through to reserve decisions. At a reserve-cost ratio of 0.1, the Schaake shuffle’s realized reserve cost is 7.4%, 9.1%, and 10.6% lower than independent sampling across the three forecasting tasks. Its shortfall rates are 10.5%, 9.3%, and 7.6%. Independent sampling produces lower shortfall rates—4.7%, 4.2%, and 3.3%—but procures too much reserve relative to the 10% target. In other words, the cheapest-looking shortfall profile is not necessarily the most cost-effective reserve strategy.

What Grid Operators Can Take Away

The central message is that grid-oriented forecasting should be judged as a chain of decisions, not just as a set of point predictions or uncertainty intervals. Calibration, temporal dependence, and the resulting reserve costs all matter. A forecast may have well-calibrated marginal distributions yet still generate implausible trajectories if it ignores how demand changes from one step to the next.

The benchmark also has clear limits. It represents one cluster and a historical period, and its target does not include non-GPU facility demand. The idle-power contribution is assumed rather than measured. Broader testing is needed before the reported reserve-cost benefits can be generalized to other data centers or grid settings. Even so, the study offers a practical framework for asking a more useful question: not only whether a forecast is accurate, but whether its uncertainty and time patterns support sound operational decisions.

Created on 09 Nov. 2026

Was this synthesis useful?

Score: 0

Ask questions about this paper to our AI assistant

How to chat with this paper? Illustrated guide

Compare with the PDF

You can also chat with multiple papers at once here.

AI assistant instructions?

Tools for this paper

Explain it differently 1 free rewrite left today

Save to a collection Group papers in collections to compare them, chat with them and export them. Continue with Google Create a free account Log in

Export BibTeX · RIS (Zotero, Mendeley, EndNote) · Obsidian (Markdown) · Notion (Markdown)

Advanced tools Premium

For researchers who follow a field closely: citation alerts, journal-club slides, version comparison and Zotero export. Discover Premium

Zotero

Citation alerts Citation alerts guide

Who works on it? The people and institutions who publish on its topic

Weekly letter Presented in "Computing beyond AI this week" of 9 November 2026. Computing beyond AI this week: What noise reveals, and what guarantees conceal

Journal club Premium+ Slides to present it Propose for the journal club

Versions What changed between versions

Talk about it in public Premium+ Press kit: release, questions, hype check Press kit guide

Papers around this one

See them on a map

Ranked by how close their title and abstract are to this paper’s. Select some to read them together.

Closest papers

Close Another approach to the problem
Probabilistic Load Forecasting via Point Forecast Feature Integration
Qicheng Chang, Yishen Wang, Xiao Lu et al. · 2019 · Machine Learning · cited 10 times · AI summary available

Probabilistic Load Forecasting via Point Forecast Feature Integration predicts short-term power-system load probabilistically using point forecasts as an input feature.

Abstract
Short-term load forecasting is a critical element of power systems energy management systems. In recent years, probabilistic load forecasting (PLF) has gained increased attention for its ability to provide uncertainty information that helps to improve the reliability and economics of system operation performances. This paper proposes a two-stage probabilistic load forecasting framework by integrating point forecast as a key probabilistic forecasting feature into PLF. In the first stage, all related features are utilized to train a point forecast model and also obtain the feature importance. In the second stage the forecasting model is trained, taking into consideration point forecast feature…
Close Broader context
Benchmarking CPU, GPU, and ML Power-Flow Solvers from 14 to 70,000 Buses
Benjamin Donnot, Sami Tazi, Eva Boguslawski et al. · 2026 · Distributed, Parallel, and Cluster Computing · AI summary available

Benchmarking CPU, GPU, and ML Power-Flow Solvers compares computational methods for power-flow calculations rather than forecasting AI-cluster demand.

In short
Optimized parallel CPU and batched GPU physics solvers match or exceed tested GNN power-flow surrogates in throughput while delivering substantially higher accuracy.
Close Broader context
DxPU: Large Scale Disaggregated GPU Pools in the Datacenter
Bowen He, Xiao Zheng, Yuan Chen et al. · 2023 · Distributed, Parallel, and Cluster Computing · cited 18 times · AI summary available

DxPU studies pooling and allocating datacenter GPUs, addressing GPU infrastructure rather than probabilistic power forecasting.

Abstract
The rapid adoption of AI and convenience offered by cloud services have resulted in the growing demands for GPUs in the cloud. Generally, GPUs are physically attached to host servers as PCIe devices. However, the fixed assembly combination of host servers and GPUs is extremely inefficient in resource utilization, upgrade, and maintenance. Due to these issues, the GPU disaggregation technique has been proposed to decouple GPUs from host servers. It aggregates GPUs into a pool, and allocates GPU node(s) according to user demands. However, existing GPU disaggregation systems have flaws in software-hardware compatibility, disaggregation scope, and capacity. In this paper, we present a new implem…
Close Broader context
Matching Workloads to Blackwell: Power-Resolved Benchmarking of NVIDIA B200 and B300 across HPC and AI Workloads
Mohsen Zand, Trung Dac Nguyen, H. Birali Runesha · 2026 · Distributed, Parallel, and Cluster Computing · AI summary available

Matching Workloads to Blackwell measures GPU power and performance across AI and HPC workloads, rather than forecasting cluster load for reserve decisions.

In short
Across AI and HPC benchmarks, B300 does not uniformly outperform B200: advantages depend on workload, while kernel granularity, double-precision capability and software maturity shape substantial regressions.
Related Broader context
Electricity Demand Forecasting with Hybrid Statistical and Machine Learning Algorithms: Case Study of Ukraine
Tatiana Gonzalez Grandon, Johannes Schwenzer, Thomas Steens et al. · 2023 · Machine Learning · preprint, not peer reviewed · cited 6 times · AI summary available

Electricity Demand Forecasting with Hybrid Statistical and Machine Learning Algorithms forecasts national Ukrainian electricity demand from hourly consumption data.

Abstract
This article presents a novel hybrid approach using statistics and machine learning to forecast the national demand of electricity. As investment and operation of future energy systems require long-term electricity demand forecasts with hourly resolution, our mathematical model fills a gap in energy forecasting. The proposed methodology was constructed using hourly data from Ukraine's electricity consumption ranging from 2013 to 2020. To this end, we analysed the underlying structure of the hourly, daily and yearly time series of electricity consumption. The long-term yearly trend is evaluated using macroeconomic regression analysis. The mid-term model integrates temperature and calendar reg…
Related Broader context
Not All GPUs Are Created Equal: Characterizing Variability in Large-Scale, Accelerator-Rich Systems
Prasoon Sinha, Akhil Guliani, Rutwik Jain et al. · 2022 · Distributed, Parallel, and Cluster Computing · cited 47 times · Not summarized yet

Not All GPUs Are Created Equal characterizes variability in accelerator-rich systems, while this paper forecasts aggregate AI-cluster power and grid reserves.

Abstract
Scientists are increasingly exploring and utilizing the massive parallelism of general-purpose accelerators such as GPUs for scientific breakthroughs. As a result, datacenters, hyperscalers, national computing centers, and supercomputers have procured hardware to support this evolving application paradigm. These systems contain hundreds to tens of thousands of accelerators, enabling peta- and exa-scale levels of compute for scientific workloads. Recent work demonstrated that power management (PM) can impact application performance in CPU-based HPC systems, even when machines have the same architecture and SKU (stock keeping unit). This variation occurs due to manufacturing variability and th…
Related Another approach to the problem
TimeGPT in Load Forecasting: A Large Time Series Model Perspective
Wenlong Liao, Fernando Porte-Agel, Jiannong Fang et al. · 2024 · Machine Learning · cited 92 times · Not summarized yet

TimeGPT in Load Forecasting considers a pretrained time-series transformer for load forecasting, whereas this paper evaluates probabilistic forecasts and reserve decisions for AI-cluster power.

Abstract
Machine learning models have made significant progress in load forecasting, but their forecast accuracy is limited in cases where historical load data is scarce. Inspired by the outstanding performance of large language models (LLMs) in computer vision and natural language processing, this paper aims to discuss the potential of large time series models in load forecasting with scarce historical data. Specifically, the large time series model is constructed as a time series generative pre-trained transformer (TimeGPT), which is trained on massive and diverse time series datasets consisting of 100 billion data points (e.g., finance, transportation, banking, web traffic, weather, energy, health…
Related Broader context
A hybrid-model approach for reducing the performance gap in building energy forecasting
Xia Chen, Tong Guo, Martin Kriegel et al. · 2022 · Advanced Engineering Informatics 52 (2022): 101627 · cited 39 times · AI summary available

A hybrid-model approach for reducing the performance gap in building energy forecasting combines physics-based and machine-learning models to predict building energy use.

Abstract
The performance gap between predicted and actual energy consumption in the building domain remains an unsolved problem in practice. The gap exists differently in both current mainstream methods: the first-principles model and the machine learning (ML) model. Inspired by the concept of time-series decomposition to identify different uncertainties, we proposed a hybrid-model approach by combining both methods to minimize this gap: 1. Use the first-principles method as an encoding tool to convert the building static features and predictable patterns in time-series simulation results; 2. The ML method combines the results as extra inputs with historical records simultaneously, trains the model t…

Go further

How are these papers found?

We compare the title and abstract of this paper with those of every paper summarized on this site, and with the papers Semantic Scholar recommends, those it cites and those citing it (for a much-cited paper, its most recent citations and, from OpenAlex, the most cited ones). Among the prior work and the follow-ups, the papers that use each other in depth, a method it uses and the much-cited follow-ups come a little higher. The levels (very close, close, related) come from the similarity of the texts, not from a judgment of quality; * marks a paper without an abstract, compared by its title. The sentence under a paper is written by an AI from the two abstracts: check it in the papers. A paper marked not relevant by three readers leaves the list.

Updated on 9 November 2026.

Disclaimer: The AI-based summarization tool and virtual assistant provided on this website may not always provide accurate and complete summaries or responses. We encourage you to carefully review and evaluate the generated content to ensure its quality and relevance to your needs.

Grid-Oriented Probabilistic Load Forecasting for AI Data Centers

The essentials

Question

Method

Findings

Limitations

Why it matters

Read further

Definitions

Why AI Data-Center Forecasts Need to Look Beyond Averages

An Open Benchmark Based on GPU Power

Forecast Accuracy Depends on the Metric

Changing Workloads Can Break Forecast Calibration

Realistic Scenarios Must Preserve Demand Over Time

What Grid Operators Can Take Away

Ask questions about this paper to our AI assistant

Tools for this paper

Advanced tools Premium

Papers around this one

Closest papers

Go further

Cite this paper (BibTeX)