(2it) Scalable Decision-Making for Decarbonized Energy Systems

I am interested in developing theory, algorithms, and software for large-scale optimization problems arising from the design, planning, and operation of decarbonized energy systems.

Motivation

The global society is facing an unprecedented challenge of decarbonizing the energy infrastructure to mitigate climate change. Decarbonizing our energy system will require all sectors of our society to transform their energy usage patterns. Specifically, residential, commercial, industrial, and transportation sectors will need to (i) electrify a substantial part of their fossil fuel-based energy consumptions; (ii) incorporate renewable generation and energy storage/conversion facilities; (iii) deploy energy-efficient and low-carbon technologies; and (iv) participate in demand response programs. A successful transformation will require continued advancement in not only the enabling technologies such as materials and catalysts, but also the system-level integration techniques based on design, planning, and operations.

The challenge in the operation of the energy systems under the decarbonization scenario arises from (i) the intermittent nature of the energy generations/prices/demands and the increasingly frequent disruptions caused by extreme weather events; and (ii) the interconnected, multi-physical, and multi-scale nature of the energy systems. At the same time, however, the massively increased degree of controllability, brought by spatially distributed energy storage and controllable loads will dramatically increase the flexibility of the overall system, and in turn, provide the opportunity to enhance its reliability and efficiency. As such, harnessing the increased flexibility while systematically dealing with the uncertainties and complexities is the grand challenge of the system-level studies of decarbonized energy systems. The technical challenge arises from the intractability of the resulting optimal operation problems, which have to account for a large number of future uncertainty scenarios as well as interconnected components. As such, developing scalable algorithms and software tools for large-scale optimization is a fundamental challenge in the system-level studies for decarbonized energy systems.

Optimization and Learning for Sequential Decision-Making Under Uncertainty

As the operation of the decarbonized energy system is subject to severe uncertainties, the decision-making problems must rigorously take the future uncertainties into account. There are different paradigms for sequential decision-making under uncertainty; e.g., stochastic control, multi-stage stochastic programming (MSSP), and reinforcement learning (RL). Due to the non-stationary and exogenous nature of the uncertainties in energy systems, naively applying stochastic control and RL techniques may lead to suboptimal performance. Furthermore, due to the inherent intractability of MSSP, applying it to real-world, real-time decision-making problems is practically impossible. This motivates the study of near-optimal decision methods for sequential decision-making under uncertainties. Our recent study suggests that by using predictive decision policies with truncated prediction horizons, a near-optimal performance guarantee can be achieved with dramatically reduced computational/sample complexity. This result suggests that either with optimization-based planning or policy optimization via reinforcement learning, one can construct efficient decision policies with significantly reduced computation and data needs. In the future, my research group will investigate theories and algorithms for sequential decision-making under uncertainties. Further, I am interested in applying these techniques to energy management and long-term planning problems for critical infrastructures, such as utility-scale energy storage, commercial building complexes, and electrified chemical manufacturing facilities.

Graph-Based Distributed Decision-Making for Networked Systems

An efficient operation of the future energy system will require the systematic coordination of its resources, such as energy storage systems, controllable loads (e.g., electrified chemical plants participating in demand response programs), and dispatchable generators (e.g., conventional power plants). Due to their spatially distributed and independent natures, distributed decision-making is preferred over a centralized approach. However, its system-wide performance may have limitations because each agent only has access to a limited amount of information. Our recent study suggests that graph-based distributed decision-making allows for effectively balancing the system-wide performance and practical advantages of decentralization; specifically, distributed control with limited-range communication based on graph distance allows for the distributed control to achieve near-optimal performance. This result can serve as a guiding principle for controller synthesis and policy parameterization in multi-agent RL. In the future, I am interested in applying graph-based modeling, optimization, and RL methods to solve optimization/control problems for large-scale interconnected systems. Ultimately, I am interested in applying these methods to challenging problems that arise in power and gas transmission/distribution systems.

Scalable Numerical Software for Optimization and Learning: GPU and Julia

The design, planning, and operations for future energy systems will require efficient and robust numerical optimization software. While the state-of-the-art nonlinear optimization, automatic differentiation, and linear algebra software are capable of solving the optimization problems at a moderately large scale, they are still limited in dealing with the problems that embed scenario trees, multi-scale spatio-temporal phenomena, and neural networks. The recent advances in the computing power of graphics processing units (GPUs) provide new opportunities for developing more efficient numerical optimization software. Furthermore, Julia, a fast-growing programming language for scientific computing, enables the development of flexible and high-performance numerical software at a rapid pace. Remarkably, its core features (abstract type definition, multiple dispatch, and just-in-time compilation) facilitate flexibly exploiting the problem structures and efficient new hardware devices. By leveraging these, we recently reported, for the first time, a competitive performance of a nonlinear optimization solver that runs on GPU. The optimization solvers running on GPU open up a lot of possibilities and allow for solving the problems that have been impossible to solve with the existing tools (e.g., MSSPs and neural-network-embedded optimization). In the future, my research group will aim to develop efficient modeling/solution tools for large-scale optimization problems.

Teaching Interests

I am interested in teaching control and design to chemical engineering undergraduate students. I believe control and design are essential in the chemical engineering undergraduate curriculum because in these courses the students learn how the chemical engineering principles are integrated into complex engineered systems and top-level decision-making. I am interested in enhancing these courses in the following ways.

Besides the conventional chemical engineering applications, which are heavily based on the refining and petrochemical industry, I will additionally introduce examples of energy systems, such as battery storage control and market participation strategies, smart building temperature control and energy management, and operation of microgrids with energy storages and renewable generations. Learning about these examples will provide the students with the opportunity to learn how chemical engineering core principles are playing an enabling role in the decarbonization of energy systems.

In the long term, I am interested in incorporating optimization into the undergraduate control and design curriculum. While optimization has always been the underlying principle for control and design, the concept of optimization has not been emphasized enough in the standard curriculum, perhaps because it has been considered to be an advanced concept. However, I believe the basic theory of optimization and skills to use the optimization software can be effectively taught to undergraduate students because (i) the basic concept of optimization is highly intuitive; (ii) its basic theory can be taught based on undergraduate level calculus and linear algebra; and (iii) modeling tools for optimization have become very easy to use and freely available (e.g., JuMP and Pyomo). Learning about optimization concepts will allow the students to approach control and design problems from a unifying perspective, and proficiency in using open-source optimization software is likely to be highly useful for any profession. Accordingly, I am interested in developing course materials for control and design that (i) formally introduce the basic theory of optimization; (ii) provide the tutorial for using open-source optimization tools; and (iii) explain the control and design principles based on the language of optimization.

Selected Honors and Awards

- 2020 AIChE Annual Meeting CAST Directors’ Student Presentation Award
- 2021 IFAC ADCHEM Young Author Award
- 2021 IFAC NMPC Young Author Award

Selected Publications

[1] S. Shin, M. Anitescu, and V. M. Zavala. Exponential decay of sensitivity in graph-structured nonlinear programs. SIAM Journal on Optimization, 2022.
[2] S. Shin and V. M. Zavala. Diffusing-horizon model predictive control. IEEE Transactions on Automatic Control, 2022.
[3] S. Shin, V. M. Zavala, and M. Anitescu. Decentralized schemes with overlap for solving graph-structured optimization problems. IEEE Transactions on Control of Network Systems, 2020.
[4] M. Anitescu, K. Kim, Y. Kim, A. Maldonado, F. Pacaud, V. Rao, M. Schanen, S. Shin, and A. Subramanian. Targeting Exascale with Julia on GPUs for multiperiod optimization with scenario constraints. SIAG/OPT Views and News, 2021.
[5] S. Shin, Y. Lin, G. Qu, A. Wierman, M. Anitescu. Near-Optimal Distributed Linear-Quadratic Regulator for Networked Systems. Under Review. arXiv 2204.05551.
[6] S. Shin, S. Na, M. Anitescu. Near-Optimal Performance of Stochastic Predictive Control. In Preparation.