An attack-resistant and stable deep learning approach for voltage-reactive power support in grids with inverter-based resources

This paper presents an attack-resistant deep reinforcement learning-based model for voltage-reactive power stability support in transmission grids with inverter-based resources (IBRs). The growing demand for grid decarbonization and changes in the resource mix have led to an increase in IBR adoption, mostly replacing conventional generators. IBR’s proliferation introduces more uncertainties to the operation of power grids, making it more challenging to maintain frequency or voltage stability and necessitating more robust grid stability decision support systems. Model-free artificial neural network-based solutions have been proposed to provide operators with situational awareness and decision-making capabilities during regular operations and contingencies. These solutions, however, may become inaccurate if their input system state is attacked, for instance, by falsifying input data or by noisy measurements. Therefore, a comprehensive support system that detects and responds to perturbed system states is essential for power system operators to rely on these solutions. This paper proposes an attack-resistant solution for assisting operators during Q-V scheduling circumstances using generative adversarial neural network (GAN) architecture combined with deep reinforcement learning (DRL). Voltage stability indices are incorporated into the proposed DRL framework to ensure the voltage stability of the grid. Using the modified IEEE 30-Bus transmission system, the proposed architecture is evaluated for different operational scenarios, and its ability to protect against falsified system states while facilitating voltage stability is demonstrated.