Project Background

Cascading failures pose significant threats to power system reliability and security with the likelihood of widespread blackouts. Consequently, the cascades would result in millions of people left without power. Achieving resilient power system necessitates an optimal restoration process as well as prevention, timely detection, and mitigation of the cascades. However, such tasks are extremely challenging because 1) decision making processes are required for mitigation and restoration, 2) the failures are rare and probabilistic, and 3) the system is governed by physical laws with a massive number of cyber-physical components (e.g., grid substations, transmission lines, control systems). This project aims to develop a deep learning framework that integrates dynamic decision-making process for mitigating cascading failures and restoring the system back to operation.

Scientific Opportunities

The project aims to develop a scalable deep learning framework for training optimal control of cascading failures and restoration in power systems, where the stochastic optimal control model minimizes load lost by exercising a control policy (e.g., load shedding, line switching, restoration) for mitigating the cascade based on probabilistic observations of the system states and recovering the system to the original state. Solving such a complex control problem is extremely challenging mainly because of the curse of dimensionality, and also because of the physical laws governed by partial-differential equations and the discrete (i.e., discontinuous) changes in the topology of the network as components are either removed from or recovered back to operation. Moreover, initial contingency event and cascading failures are probabilistic and rare events and thus can be infinitely many.

Research Goal

The goal of this project is to develop new variants of graph convolutional networks that directly encode physical power system and predict the power system resilience (i.e., total load lost between initial contingency and full recovery) for a given initial contingency event. The proposed algorithms will be developed as a software package that can run on Argonne’s high performance computing systems.

Research Plan

The project will consider single time-period alternating current optimal power flow (ACOPF) and multiple time-period ACOPF under line contingency. We plan to incorporate machine learning approaches to (i) the acceleration and (ii) the approximation of ACOPF optimization solutions, which can be achieved by novel integrations of optimization and machine learning.

Deliverables and Impacts

The deliverables of this project include multiple open-source software tools for solving optimal power flow by (partly) integrating machine learning techniques and the corresponding publications.

Team and Contact

Argonne National Laboratory

Team Members

• Dr. Kibaek Kim (PI)
• Dr. Mihai Anitescu (co-PI)
• Dr. Prasanna Balaprakash (co-PI)
• Dr. Youngdae Kim
• Mr. Jisung Hwang
• Mr. Sihan Zeng