Regime-Adaptive Weighted Ensemble Learning for Computing-Driven Dynamic Load Forecasting in AI Data Centers

Overview

The work is an arXiv preprint by Ziying Wang, Ying Zhang, Lei Wang and Yuzhang Lin (arXiv:2604.27207, 2026; Oklahoma State University / New York University). Its subject is short-term, minute-level load forecasting for AI data centres whose load is driven by bursty, non-stationary job arrivals.

The method combines two submodels, XGBoost and a 1D-CNN, through a small neural network. That network learns time-varying ensemble weights from recent submodel performance and load-increment features, so the forecast adapts across idle, ramp-up, high and ramp-down regimes. The data are cluster-level series built from the MIT Supercloud GPU logs: job-level logs at 0.1 s resolution were resampled to 1-min node-level series, aggregated to cluster level, and idle slots were filled with a 25 W GPU baseline.

The data were split chronologically: 60% for submodel training, 20% for ensemble training and the remainder for testing. Forecast errors (NRMSE/NMAE) are SVR 3.04/2.30%, LSTM 2.66/2.11%, XGBoost 2.46/1.86%, 1D-CNN 2.45/1.85% and the proposed ensemble 0.83/0.37%. Relative to the strongest baseline (1D-CNN), the method improves NRMSE by 66.1% and NMAE by 80.2%. In ramp regimes both submodels lag, and in the high regime both underestimate; the adaptive weights track the ground truth more closely across all four regimes. Limit: the evaluation rests on a single dataset (MIT Supercloud).

Key points

1. MIT Supercloud job-level GPU logs at 0.1 s resolution were resampled to 1-min node series and aggregated to cluster level; idle slots filled with a 25 W GPU baseline.
2. Data split chronologically: 60% for submodel training, 20% for ensemble training, remainder for testing.
3. Forecast errors (NRMSE/NMAE): SVR 3.04/2.30%, LSTM 2.66/2.11%, XGBoost 2.46/1.86%, 1D-CNN 2.45/1.85%, proposed ensemble 0.83/0.37%.
4. Relative to the strongest baseline (1D-CNN) the method improves NRMSE by 66.1% and NMAE by 80.2%.
5. In ramp regimes both submodels lag and in the high regime both underestimate; adaptive weights track the ground truth more closely across all four regimes.

Using it in design

• For operations and grid-interaction features, use regime-aware (idle/ramp/high) minute-resolution IT load forecasting with job-scheduler signals (active jobs) as inputs, rather than only historical kW.
• Report forecast error by regime, since ramps are where single models fail and where utility ramp limits bite.
• Use such forecasts to schedule storage or demand-response actions ahead of job starts.
• Treat the results as findings from a single dataset; validate with your own workload data.

Who should read it

Engineers and researchers working on grid interaction, storage and energy forecasting for AI data centres.

Publication details

Authors

Ziying Wang; Ying Zhang; Lei Wang; Yuzhang Lin

Publisher

arXiv preprint (Oklahoma State University / New York University)

Year

2026

Document number

arXiv:2604.27207

Document language

English

Source type

Scientific

Access

Free to read

DOI

10.48550/arXiv.2604.27207

Licence

arXiv distribution permission — reuse conditions remain with the rights holder

Citation