CAREER: Performance-Guaranteed Learning and Control for Real-world Energy Systems: Stability, Robustness, and Computational Tractability

About This Grant

This five-year CAREER project plan is centered on research aiming to develop performance‑guaranteed learning and control for real‑world energy systems, with the objective of providing stability, computational tractability, and robustness guarantees. We focus on two energy system applications: distribution grid voltage regulation and building system control, motivated by the system operator and end‑user perspectives, respectively. Existing control methods are largely limited to known dynamics and strong assumptions about the models, rendering them insufficient for future energy system control with many unknown active components and complex dynamics. At the same time, more data is becoming available due to the widespread deployment of smart meters and upgraded communication networks. As a result, learning‑based control techniques have attracted surging attention in recent years. Despite the promise, existing learning‑based methods lack reliable performance guarantees, posing significant risks to mission‑critical applications in energy systems. This research plan seeks to develop reliable and computationally efficient learning and control algorithms for real-world energy systems that address three central challenges: (1) Incorporating control‑theoretic tools into reinforcement learning (RL) to obtain stability and steady-state optimality guarantees for distribution grid voltage regulation. Compared to existing control methods, RL with neural network-based controllers has the potential to significantly reduce transient control costs and achieve faster disturbance recovery; (2) Designing operator learning to accelerate control in computationally expensive applications, especially for building control governed by partial differential equations (PDEs). This has the potential to reduce building energy consumption while maintaining safe indoor air quality; and (3) Bridging the gap in deploying learning‑based control algorithms to the real world, specifically under time-varying network topologies in voltage regulation and perturbed sensor inputs in building control. The proposed algorithms will be validated in real-world energy systems, leveraging the NSF-funded DERConnect testbed at the PI's home institution, U.C. San Diego. A comprehensive curriculum, including undergraduate Linear Control System Theory, graduate-level Machine Learning for Physical Applications, and a new research-oriented special topics class on Learning and Control for Energy Systems, is designed to train the next generation of engineers and researchers and provide them with interdisciplinary skills in energy, control, and AI to meet the growing needs of industry, utilities, and academia. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.