Machine Learning for Reliable Grid Planning in the U.S. Electric Grid

What is the amount of electricity that will be required by people for the next day? Now answering that query is more like looking ahead to one’s morning commute—slightly predictable, but definitely not ironclad.

Machine Learning for Reliable Grid Planning in the U.S. Electric Grid" />

In order to control the intrinsic uncertainty in estimating power requirements and prevent undue surprises, electric grid operators tend to depend on computer models that help in estimating everything from traffic patterns to power demand.

Such difficulty of considering both the unknown and uncertainty to supply electricity under different kinds of conditions involves a sequence of extraordinarily intricate math problems. Now, with the help of artificial intelligence (AI), scientists at the U.S. Department of Energy’s (DOE) Argonne National Laboratory are devising innovative methods to obtain a better understanding from a large amount of data on the electric grid, with the aim to guarantee greater efficiency, resiliency, and reliability. The work integrates Argonne’s years of grid know-how with its sophisticated computing experts and facilities.

A Better Grasp of Uncertainties

It is well known that grid operators have invariably handled challenges and a certain amount of uncertainty from factors ranging from equipment failures to adverse weather conditions. Now, inconsistent supplies of renewable energy—with some of the energy flowing from consumers that have rooftop solar panels equipped with intelligent meters—are raising the number of variables that need to be considered by grid operators.

Researchers at Argonne National Laboratory are working to improve models that utilize machine learning, a type of AI, to replicate the electric system as well as the severity of numerous challenges.

In an area that has 1,000 electric power assets, like transformers and generators, an outage of only three assets can create almost a billion scenarios of possible failure. Therefore, which of those possibilities will require the utmost attention?

However, it is rather time-intensive to solve such a complicated model. With resources like the Argonne Leadership Computing Facility (ALCF)—a DOE Office of Science User Facility—scientists can replicate many different situations in parallel, moving the process along more rapidly.

The idea is to generate a large number of scenarios and train the machine learning model to tell us the answer. Instead of solving a number of difficult optimization models over several hours or days, we train the model ahead of time and then get the answer right away.

Kibaek Kim, Assistant Computational Mathematician, Argonne National Laboratory

The scientists train the machine by inputting a set of data that contains the solutions, as if the machine was examining earlier ​“exams” prior to attempting new ones. This is known as supervised learning. Unsupervised learning is another method in which a computer is fed with raw data and is allowed to filter out patterns without being aware of any “answers.”

In another research, a kind of model known as a graph convolutional neural network was used by Kim and his team in order to propose optimal controls that would inhibit transmission lines from overloading in case there was an issue with any of the lines. The team discovered that this specific model, which employed machine learning to rapidly reach a solution, created relatively fewer errors when compared to more traditional ones.

The study was carried out using Argonne’s Laboratory Computing Resource Center (LCRC) and its Joint Laboratory for System Evaluation. Kim’s work also involves collaboration across Argonne National Laboratory and the LCRC.

Kim’s group is exploring ways to make these models even sturdier, providing grid operators more robust guidance that can inform more dependable planning and operations for contingent events such as equipment malfunctions, storms, and major fluctuations in the generation of renewable energy.

Planning Ahead

At Argonne National Laboratory, other work involves applying AI to accelerate the day-to-day calculations that go into the planning of the regional electric system. Security constrained unit commitment, or SCUC, is one such calculation that helps grid operators to set a schedule for both hourly and daily generation of power.

In power systems, this SCUC problem is solved multiple times a day. Since this problem is solved repeatedly, we can accumulate a lot of data and discover patterns that could be used to solve the next round.

Feng Qiu, Principal Computational Scientist, Argonne National Laboratory

Instead of substituting present analytics with machine learning, the strategy is to boost the existing ones using machine learning to provide ​“hints” learned from earlier solutions, said Qiu.

Leveraging LCRC’s Bebop cluster, a research team headed by Alinson Santos Xavier, a postdoctoral appointee at Argonne National Laboratory, created AI that is capable of solving SCUC as much as 12 times faster, on average, when compared to traditional techniques. Tests were conducted at Midcontinent Independent System Operator (MISO), where an early version of the method was effectively utilized. MISO supervises electricity delivery across one Canadian province and 15 states.

“All this can lead to a more efficient market and more cost-effective electricity production,” Qiu added. ​“For long-term planning, it could help grid operators consider more scenarios and make better expansion plans.”

Programming for a Smarter Grid

Up-to-date grids increasingly integrate sensors that are capable of tracking conditions all through the system, and these also provide new opportunities for improved data processing. For instance, devices located at substations and on transmission lines act as sentinels that alert grid operators about equipment issues whenever they take place.

The research team at National Laboratory has assessed a year’s worth of sensor data provided by ComEd—a utility that serves almost four million customers in the U.S. Midwest. The team utilized unsupervised learning this time, inputting the data to the machine and instructing it to search for anomalies in the sensor outputs.

It’s not always known to the operator when things do not work as they should. Our approach decides whether the current conditions of the system are expected based on past behavior, or whether something is new and different. This information can be used to alert the operator that they may have something they don’t expect on the grid.

Mihai Anitescu, Senior Computational Mathematician, Argonne National Laboratory

Anitescu also worked on the project.

A classification work like this can even be used in weather forecasting for renewable energy, rectifying for underestimates of wind resources that are close to bodies of water, for instance, and integrating numerical models with physical measurements to enhance precision.

According to Anitescu, plenty of AI work involves pure data— analyzing a picture, or detecting speech patterns, for instance: ​“There aren’t many physical rules,” he added.

For huge real-world systems such as the electric grid or the weather, that is not the case. ​“You really have to reconcile data, even if there’s a lot of it, with the physical information,” he stated. ​“This is very much in its infancy, and it’s really where supercomputers are necessary.”

Argonne National Laboratory’s AI work for the grid has been funded by DOE’s Office of Science, the Advanced Grid Research and Development Division in DOE’s Office of Electricity and Argonne’s Laboratory Directed Research and Development Swift program for short-term projects.