Physics-Informed Modeling and Stability-Constrained Intelligent Control of Grid-Forming Battery Storage in Low-Inertia Power Systems
Year: 2026
Type: article
Abstract: The declining synchronous inertia of power systems with high shares of inverter-based resources makes the tuning of grid-forming battery energy storage systems (BESS) both more important and more uncertain, because the inertia that the storage must complement is not measured directly and changes with the unit commitment. This paper proposes a framework that couples physics-informed estimation with stability-constrained learning. First, a physics-informed neural network (PINN) embeds the swing and governor equations into the training loss and identifies the synchronous inertia and frequency stiffness from a single disturbance recorded by phasor measurement units, using the measured BESS power as an input. Second, a neural policy maps these estimates to the virtual inertia and damping of a virtual synchronous machine; it is trained with an augmented Lagrangian method subject to closed-form constraints on the rate of change of frequency (RoCoF), frequency nadir, damping ratio and converter power, tightened by the estimation uncertainty. A safety layer certifies each output against the same constraints. In simulations with inverter-based shares of 50–85 %, the PINN estimated inertia within −10.5 % to +7.3 % and frequency stiffness within 0.3 % for measurement noise up to 20 mHz, whereas the RoCoF-based method was biased by +20 % to +49 %. In closed-loop tests on 300 random operating points of a nonlinear model, the proposed controller produced no violation of the 49.5 Hz nadir and 0.5 Hz/s RoCoF limits and reduced the sustained storage reserve by 61 % relative to fixed gains; a policy trained with a fixed penalty violated the limits in 59.7 % of cases.
Language: English
Subfield: Control and Systems Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W7222807040