Research
The design and operation of large-scale engineering systems such as power networks and networked control systems require efficient and high-performance schemes for modeling and decision making. I study how machine learning and optimization tools can enable the reliable interactions between data and cyber-physical systems. Our group work on developing both fundamental algorithm tools and practical engineering solutions which can tackle the emerging challenges such as data and model uncertainties, computation tractability, safety and reliability. My research is driven by real-world applications in power and energy systems, transportation, cloud computing, robotics and IoT. Some of my research topics are highlighted as follows.
Learning Generative Models for Scenario Generation
Granular data are essential input for decision making at varying timescales for engineering problems, such as operating sustainable power systems and scheduling stochastic EV charging sessions. We explore how to represent the data and system uncertainties through the lens of generative machine learning. Resulting time-series “scenario” representation can inform system operators to implement reliable and credible decisions.
Related works
Chen, Y., Wang, Y., Kirschen, D., & Zhang, B. (2018). Model-free renewable scenario generation using generative adversarial networks. IEEE Transactions on Power Systems, 33(3), 3265-3275.
Contreras-Ocana, J. E., Chen, Y., Siddiqi, U., & Zhang, B. (2019). Non-wire alternatives: An additional value stream for distributed energy resources. IEEE Transactions on Sustainable Energy, 11(3), 1287-1299.
Optimal Control via Neural Networks
Classical controllers assume perfect knowledge about the underlying systems, while purely data-driven control paradigms such as reinforcement learning may not provide performance guarantees, such as solution optimality. To enable optimal control on complex systems, we envision a novel kind of learning-based controller that is built upon specifically designed input convex neural networks. Such design achieves both computation tractability and algorithm efficiency.
Related works
Chen, Y., Shi, Y., & Zhang, B. (2019). Optimal Control Via Neural Networks: A Convex Approach. In International Conference on Learning Representations.
Chen, Y., Shi, Y., & Zhang, B. (2020). Data-driven optimal voltage regulation using input convex neural networks. Electric Power Systems Research, 189, 106741.
Gu, C., Chen, Y. (2022). Pontryagin Optimal Controller via Neural Networks. arXiv:2212.14566
Integration of Learning and Computation into Clean Energy Systems
Seamless development and integration of computational and learning methods into real engineering systems can bring both economical values and operational benefits. In our research, we have been constructing a set of foundational platforms and computational tools for closing the loop of forecasting, system modeling, streaming measurements, and decision making. Some of our novel platforms span the operation of carbon-aware charging stations, to learning autonomous building energy management. Developed algorithms and computation platforms are open sourced on Github for academic research.
Related works
Cheng, K. W., Bian, Y., Shi, Y., & Chen, Y. (2022). Carbon-Aware EV Charging. IEEE International Conference on Communications, Control, and Computing Technologies for Smart Grids (SmartGridComm).
Zhang, C. Shi, Y., & Chen, Y. (2023) BEAR: Physics-Principled Building Environment for Control and Reinforcement Learning. ACM e-Energy. ACM.
Cheng, K., Chen, Y., and Shi, Y. (2023) GridViz: a Toolkit for Interactive and Multi-Modal Power Grid Data Visualization,” IEEE Power & Energy Society General Meeting (PESGM).