NRF CRP 35th Call
Grid Stability and Control for AI Data Centre Integration
Hyperscale AI data centres should no longer be treated as passive loads. Because thousands of processors can vary power in a highly synchronised way, they can act as active disturbance sources and create new stability risks.
This proposal asks three intriguing questions:
-
How to understand this risk
-
How to control it
-
How to translate the results into grid standards and policy for Singapore​​​​
Why this matters
-
AI workloads are becoming grid-relevant​​
Highly synchronised GPU activity can inject quasi-periodic oscillations into the 0.1 to 3 hertz electromechanical band. Server power supplies and associated converter interfaces can behave like negative incremental impedance and interact with inverter-based resources, extending the problem into the 5 to 50 hertz sub-synchronous range.
Training job transitions can create very large power ramps in a short time. Shared protection thresholds can trigger sequential disconnections across data centres and other grid-connected assets.
This means that risk depends on synchronisation, not only MW size. A smaller but highly synchronised facility may create greater dynamic risk than a larger but staggered one. That is the scientific gap this programme addresses.
​
-
Singapore is an especially exposed island grid​​
Key Finding: 1 GW zone exceeds IEC 2.0 Hz/s threshold standard at ~175 MW; larger zones remain below up to 500 MW
​
Singapore is an especially relevant case for three reasons.
-
It is an islanded grid, so it has no synchronous neighbours to absorb fast disturbances.
-
Decarbonisation is changing system dynamics through more solar, DERs, and imports, which can reduce inertia and alter stability margins.
-
Data-centre demand is already large and still expanding, while current roadmaps do not explicitly evaluate load-driven oscillation risk.
So Singapore is not just an example application. It is a highly policy-relevant stress test for this emerging problem.
The preliminary results reinforce this point: the same facility size can have very different ROCOF impacts depending on where it connects. This supports our argument that risk should be assessed with a location- and synchronisation-aware metric, not by MW size alone.​​​​​
Research objectives
Three Falsifiable Hypotheses: H1 (Workload governs risk), H2 (Multi-mechanism suppression), H3 (Spatial disjointness of benefits/burdens)​
Programme structure / roadmap
The programme is organised into three connected sub-projects.
SP1 builds the scientific foundation. It develops the stochastic forced-oscillation theory, SRI and FOIM, SSCI analysis, step-load thresholds, and cascade boundaries.
SP2 converts that understanding into engineering solutions. It develops the unified grid-forming BESS controller, grid-aware scheduling, and validates the framework using a digital twin and hardware-in-the-loop testing.
SP3 translates the validated results into practice: grid-connection guidance, standards contributions, and the SG-AEII equity framework.
