How to Measure and Prevent Generator Shaft Fatigue Caused by AI Data Center Loads

Why AI Data Center Loads Are Different

Data Centers are used to high power demands and load variations across the day, but the introduction of AI loads has created new challenges. AI loads, specifically during training, impose high and synchronized power demand and create periodic oscillation in the load. Oscillation during the training phase is the result of GPUs conducting identical operations in lockstep. The training phase creates three distinct profiles which correspond to the three phases of a mini-batch training sequence. 

The Hidden Risk: Generator Shaft Fatigue and Mechanical Damage

This rapid periodic variation of load is at relatively low frequencies and can propagate through the grid up the generators. For both multi and single mass systems, forced oscillation causes torsional vibration which increases the stress amplitude experienced at the critical point of the power system. The larger mechanical stress amplitude experienced accelerates fatigue and shortens the lifespan of the element.

Why Conventional EMT and PSCAD Studies Are Not Enough

Quantifying the damage to the lifespan requires characterization of oscillation, material properties, and a life-fatigue model to connect all key factors. Although conventional EMT and PSCAD studies can model the oscillations of torque, they do not offer the tools needed to characterize the oscillation, connect shaft information to a fatigue model, and assess the lifespan of the system under different loading conditions. 

EdgeTunePower Fatigue Life Degradation Assessment Tool

At ETP, we have developed a PSCAD native tool that enables life-fatigue modeling with periodic steady-state torque data. Simply connect the tool to critical torque, select the generator type and mechanical properties from the dropdown menu, input expected usage and run the study. The tool will then output the expected lifespan of under the new conditions, quantifying the degradation caused. 

How we Quantify Fatigue Damage

The tool characterizes the stress profile given and then uses the stress-life (S-N) method to calculate the degradation of the lifespan. The S-N method (an empirical method commonly used for fatigue-life modeling) enables inclusion of key factors such as shaft material, critical dimensions, and temperature. The expected AI training usage is then combined with the baseline usage to assess the cumulative fatigue damage. To accelerate studies, the tool only needs a small representative sample of the torque data and does not require long studies.