Stochastic Dispatch of Battery Storage via Neural Policies for AI-Driven Data Centers

The surge in AI workloads is driving explosive and highly variable electricity demand in data centers.  This talk introduces a two-stage hybrid control framework for Battery Energy Storage Systems (BESS) dispatch under these conditions. We propose a risk-averse deep neural policy trained on real AI workload and ambient traces to optimize economic performance while minimizing outcome variance. A real-time model-based safety adjuster is used to refine policy actions via chance-constrained quadratic optimization and enforce probabilistic bounds on battery state-of-charge and temperature. Simulations with realistic bursty AI computational loads show the policy delivers efficient dispatch results, while the safety layer eliminates all constraint violations and offers a practical, data-driven, and certifiably safe approach for grid-interactive AI infrastructure.