Regime-Adaptive Weighted Ensemble Learning for Computing-Driven Dynamic Load Forecasting in AI Data Centers
MIT Supercloud test shows regime-adaptive ensemble cutting forecast errors below 1%.
A team led by Ziying Wang has developed a regime-adaptive weighted ensemble learning approach to predict the dynamic, bursty power loads of AI data centers. Unlike traditional loads, AI data center workloads are computing-driven with heterogeneous job arrivals, sizes, and durations that are non-stationary. The method combines two machine learning submodels via a weight-learned neural network, dynamically optimizing ensemble weights based on operating regimes. A novel feature engineering strategy incrementally learns from streaming data, enabling calibration to changing conditions.
Tested on the MIT Supercloud dataset, the approach is the first to reduce minute-class forecasting errors for AI data center loads to below 1%, significantly outperforming other combinations. This precision is critical for grid-interactive coordination and demand response, helping prevent instability from the rapid growth of AI infrastructure. The work addresses a key gap in short-term load forecasting for a load type that poses increasing threats to power grid efficiency.
- Achieves <1% minute-class forecasting error on MIT Supercloud dataset, a first for AI data center loads.
- Uses regime-adaptive weighted ensemble learning with a neural network to combine two ML submodels.
- Novel feature engineering incrementally learns from non-stationary data streams for dynamic weight optimization.
Why It Matters
Enables precise load forecasting for AI data centers, improving grid stability and enabling smarter demand response.