Power grids are getting faster. As traditional spinning generators are replaced by inverter-based resources, the system’s inertia falls and frequency can change in hundreds of milliseconds. That is faster than the automatic generation control (AGC) systems operators use today. This paper presents a hybrid controller called AG2C-DT that uses an offline-trained Decision Transformer (a sequence model) together with a two-stage safety stack to manage secondary frequency control without risky online trial-and-error learning.

The researchers trained the Decision Transformer on historical supervisory control and data acquisition (SCADA) records so the model learns to map recent trajectories and desired outcomes to generator dispatch decisions. Because training is done entirely offline on past operational logs, the method avoids online exploration that could endanger the grid. The work is tested in simulation using ANDES, a symbolic-numeric power system simulator, on the Northeast Power Coordinating Council (NPCC) 140-bus model under low-inertia conditions. The paper compares the approach to a linear quadratic regulator (LQR) baseline and examines how it differs from conservative offline reinforcement-learning methods.

At run time the Decision Transformer proposes redispatch actions. Those proposals pass through a two-stage safety stack before being applied. The first stage is a Constraint Verification Unit that runs a fast algebraic screening using power transfer distribution factors (PTDFs) — a way to estimate how changing one generator’s output moves power flows on lines — and operates in under ten milliseconds. The second stage is an aggregate digital twin that checks dynamic stability using simplified swing equations (these equations capture how generator rotors and frequency respond). If a proposed action fails either check, the system falls back to the tuned AGC, so worst-case behavior in simulation is bounded by standard practice.