Digital Twin Protocol for Adaptive Power Distribution in Simulated Rack Level Direct Current Systems under Graphics Processor Load Surges

Rapid changes in graphics processing unit (GPU) demand motivate coordinated control of power supply units (PSUs) and uninterruptible power supplies (UPSs). This protocol describes a digital twin framework for a simulated 54 V direct current (DC) rack, combining a two-layer gated recurrent unit (GRU), model predictive control (MPC), and recursive least squares (RLS). The supervisory interval is 0.25 s, with a 5 s prediction horizon and explicit storage-energy accounting. An independently integrated plant and an analytical prediction model support four synthetic workload scenarios. Five paired load realizations per scenario compare static allocation, threshold-triggered control, conventional MPC, forecast-informed fixed-model MPC, and adaptive MPC. Additional tests impose forecast bias, PSU parameter shifts, sensor noise, and delayed commands; weight and forgetting-factor sweeps examine tuning sensitivity. The results identify the contribution of parameter adaptation and the limits of short-horizon forecasting under abrupt workload transitions. Voltage regulation, PSU peaks and ramps, storage depletion, and equivalent cycling are reported alongside measured forecasting and optimization latency. The framework provides a reproducible numerical basis for supervisory power coordination. Its scope is an equivalent rack model, with switching dynamics, manufacturer-specific protection, and physical-rack deployment treated as separate engineering considerations.