Scientific Machine Learning for Power System Simulation

Speeding up power system simulation using machine learning to address the complexity associating with modeling a growing penetration of renewable energy sources.

Trained surrogate models can replace computationally expensive physics-based models to accelerate computer simulations

Background and Motivation

With the growing penetration of renewable energy sources in power systems that are connected to the grid through fast-acting inverters, the dynamic behavior of these systems is rapidly evolving. One of the challenges with simulating large deployments of inverters is the increased number of equations and small numerical timesteps required. To address the computational intensity, this project explores the use of scientific machine learning to accelerate power systems simulations. This research attempts to speed-up these computer simulations by 1) reducing the number of equations we need to solve to model the dominant system dynamics and 2) using machine learning to solve these equations faster.

Objective

This project uses machine learning as a means to more cheaply, quickly, and accurately predict the dynamic response of the power grid to a disturbance in the system. In addition to producing reports and journal publications, all code will be developed in the Julia programming language and made available as open-source libraries.

Team

Duncan Callaway

Ciaran Roberts

Collaborators

Bri-Mathias Hodge (NREL/Boulder)
Chris Rackauckas (MIT)
José Daniel Lara (UC Berkeley)
Rodrigo Henriquez-Auba (UC Berkeley)
Matt Bossart (CU Boulder)
Marena Trujillo (CU Boulder)
Avik Pal (MIT)
Ranjan Anantharaman (MIT)
Nicholas Klugman (MIT)

Recent Publications

Roberts, Ciaran, José Daniel Lara, Rodrigo Henriquez-Auba, Matthew Bossart, Ranjan Anantharaman, Chris Rackauckas, Bri-Mathias Hodge, and Duncan S. Callaway."Continuous-time echo state networks for predicting power system dynamics."Electric Power Systems Research 212 (2022) 108562. DOI

Interested in learning more about how machine learning approaches, e.g. continuous-time echo state networks, can speed-up computer simulations of power systems? Check out these Julia code repositories:

https://github.com/Energy-MAC/CTESN_PSCC - A Julia library showing how continuous-time echo state networks can accelerate prediction of power system dynamics following an unexpected disturbance.
https://github.com/HodgeLab/PowerSimulationNODE - A Julia library for training surrogate models based on Deep Equilibrium layers and Neural ODEs that can be seamlessly integrated into a power system simulation.
https://github.com/NREL-SIIP/PowerSimulationsDynamics.jl - A flexible modeling framework for power system modeling and simulation of power systems dynamics.

Sponsor

This project was funded by the Department of Energy Office of Electricity.