Deep reinforcement learning (RL) is a promising tool for control of complex physical systems such as power and energy systems, yet its deployment is often hindered by the lack of explicit stability guarantees. In this talk, I will present a stability-constrained RL framework, where we show that monotonicity in control policies implies Lyapunov stability in power grid control. By parameterizing policies with a novel monotone neural network design, we ensure stability by design, achieving better control performance and rigorous guarantee compared to standard RL methods. In the second part, I will introduce a generative AI approach for analytical Lyapunov function discovery. Using a symbolic transformer-based model trained with RL, our framework can generate interpretable Lyapunov functions for nonlinear systems, including high-dimensional and non-polynomial cases. Together, these efforts highlight new pathways toward efficient and trustworthy physical AI for control of real-world power grid.

Seminar Recording