The rapid growth of datacenter
and AI workloads, coupled with increasing inverter-based resource (IBR)
penetration, is fundamentally reshaping modern power system dynamics. On the
generation side, the displacement of synchronous machines by grid-forming
inverters introduces low-inertia and weak-grid challenges. On the demand side,
large-scale datacenters, especially those dominated by AI training and
fine-tuning workloads, exhibit stochastic and periodic power fluctuations that
can trigger wide-area oscillations. This talk presents a comprehensive
framework to understand, model, and mitigate these emerging stability issues.
It first introduces the usage of AI techniques to identify the uncertainties
from IBRs by accurately and efficiently forecasting their outputs. Design of
loss function, end-to-end approach will be specifically focused. NN-based
carbon tracing method will be also handled. Then, the seminar introduces
stochastic modeling of AI workloads, revealing dominant fluctuation components
near 1 Hz that excite local and inter-area oscillation modes in large
interconnected systems. The analysis also highlights the sensitivity of oscillation
magnitude to datacenter penetration, size, and spatial distribution. To enhance
grid resilience, data-driven oscillation source location based on Dynamic Mode
Decomposition (DMD) and a two-tier localization strategy are proposed. Finally,
a hybrid energy storage system (ESS + supercapacitor) co-located with
datacenters is introduced as an effective mitigation approach for ramping and
oscillatory disturbances. To this end, this talk establishes an integrated
perspective on grid for AI and AI for grid.