Virtual Power Plants for Frequency Regulation: A Learning-Based Method With Safety Guarantee

Frequency safety is becoming increasingly critical in modern power systems due to the reduction of rotational inertia and governor damping. With advances in power electronics, the inverter-based resources (IBRs) installed in distribution networks can provide inertia and frequency support if their controllers are properly designed. In this paper, the virtual power plants (VPPs) framework is leveraged to aggregate and coordinate IBRs to provide inertial and primary-frequency responses. In particular, sufficient conditions are derived for the VPP central controller design to guarantee the transient frequency safety, including stability, the rate of change of frequency and Nadir requirements. On this basis, a nonlinear controller parameterization that satisfies these conditions automatically is proposed. The optimal VPPs control policies are trained by recurrent neural network (RNN)-based reinforcement learning and then fairly disaggregated to individual IBRs, ensuring that IBRs share the power injections for frequency response in proportion to their reserve capacities. Moreover, the controllers are equipped with a safety layer with trivial computation burden to satisfy critical constraint during implementation. Case studies demonstrate that the proposed control strategy outperforms linear controller and other cutting-edge learning-based control strategies.