Structured Control and Learning for Sustainable Power Systems

Abstract

With decarbonization efforts in renewable integration and electrification, the electric grid needs to adapt and serve a larger system that is becoming more distributed, having less inertia, and facing more uncertainties. These changes have reduced the safety margins of the grid and significantly increased the costs of risk management. Machine learning tools can potentially unlock design freedoms found in the increased controllability from inverter-interfaced resources (e.g., solar, wind, and electric vehicles), and reshape the landscape of power systems for more efficient operations. However, such algorithms typically do not provide guarantees about safety-critical constraints, making them difficult to implement in practice. The dissertation proposes to bridge the gap between learning and safety-critical constraints through structured neural networks guided by control theory and the physics of power systems. Using Lyapunov theory, we extract stabilizing controller structures for transient stability problems, and parameterize the structures by neural networks. On this basis, we design Neural-PI controllers to further achieve provable guarantees on optimal resource allocation and frequency restoration at the steady state. In addition, we propose a modular approach for transient stability analysis with lossy transmission lines. This provides a simple yet effective approach to optimize control efforts with guaranteed stability regions. The structured approach for learning-based control provides end-to-end guarantees that are independent of the learning process, which in turn provides large flexibility for learning algorithm design. To relieve the burden of centralized coordination in voltage control, we propose a decentralized safe learning approach to train local neural network controllers at each node in a model-free setting. To overcome key barriers on the requirement of a large number of system interactions to learn a good control policy, we develop a sample-efficient trajectory generation algorithm that adapts to the distributional shift of trajectories resulting from updated control policies and also extends to partially observed systems.