A Physics-Informed Neural Network Modeling Approach for Energy Storage-based Fast Frequency Support in Microgrids

Energy storage systems (ESSs) can provide fast-frequency support to keep frequency variation in lower-inertia microgrids within allowable bounds. Model-based control strategies like model predictive control (MPC), are efficient ways to enable fast-frequency support in ESSs. However, the efficacy of these control strategies depends upon the accuracy of the underlying model. Previous research has commonly utilized simplified equivalent generator models of power systems. However, such simplified models become insufficient as the system operating conditions and dynamics of the systems change over time resulting in poor performance of frequency support mechanisms. This paper proposes a physics-informed neural network (PINN)-based modeling approach to model the frequency dynamics in microgrids. Rather than relying solely on data-driven black box modeling approaches or simplified equivalent generator models, we propose to integrate both into the training process. Specifically, we utilized a single generator equivalent frequency dynamics model as a template for developing the PINN-based multi-machine equivalent model by supplementing it with the measurement data from the system. The measurement data is acquired by briefly perturbing the system using a square wave signal on an ESS's power dispatch, minimizing data requirements without affecting ESS functionality. The proposed approach exhibits a greater degree of goodness of fit in comparison to training that relies solely on system physics.