Officials from CAISO, NYISO and French grid operator RTE joined MISO on the final day of its Market Symposium on Friday to discuss the challenges of developing data analytics to support system operators’ decision-making.

Elliot Mainzer, who became CAISO’s CEO last month after 18 years at the Bonneville Power Administration, said one of his first actions in his new job was creating a new chief operating officer position.

“We’re integrating our operations and transmission infrastructure and market policy and technical groups under one executive so that we can maximize alignment among those groups and make sure that the technical platform evolves as efficiently across the organization as necessary.”

When BPA started adding wind generation more than a decade ago, it had no tools to address ramping issues and curtailments. The difference between “accommodating” renewables and “integrating” them in the system was developing those tools, he said in a conversation with Todd Ramey, MISO’s chief digital officer.

“If you want to really integrate them as efficiently as possible, you have to take that time to do the design work,” Mainzer said.

BPA introduced intra-hour scheduling, held a competition to find the best wind forecasting provider and aligned its tariff and pricing mechanisms to encourage operators to use the new tools.

He said he learned the need to involve control center operators in the design of the systems from the beginning.

“Something that was very important was, first of all, making sure that the systems were integrated so that folks weren’t running around having to make 12 decisions at the same time. … We just don’t have a lot of room for a lot of friction in the system anymore as we’re trying to meet our resource adequacy requirements.”

David Edelson, NYISO’s manager of operations performance and analysis, said operators need to make “second-to-second decisions.”

“There’s really little time to interpret data; therefore, that data needs to be presented to control room operators very clearly, in ways that suit their preferences so that they can make quick decisions — generally binary decisions,” he said during a panel discussion moderated by Keri Glitch, MISO’s chief information security officer.

Avoiding False Positives

“They can’t be presented with unnecessarily large volumes of data — large numbers of false alerts — because that’s going to lead to mistrust of the data, as well as hesitations in their response,” Edelson said.

False positives was also the subject of remarks by Mykel Kochenderfer, a Stanford University associate professor who develops applications for aerospace and automated vehicles. “Many of the challenges are exactly the same” as in the power sector, he said, recalling his work on an aircraft collision avoidance system.

“In this situation, you have imperfect sensor information, so you don’t know exactly the current state of the world. And you also have imperfect information about how the world will evolve: You don’t know the future trajectories of the other aircraft. And you have competing objectives. On one side, you want it to be extremely safe, and on the other side, you want to be efficient. You don’t want to be alerting the pilot constantly to avoid collisions when there’s not a significant threat present.”

The system took about a decade to develop, “and much of that time was just establishing trust that the system will behave correctly in operation,” he said.

For that reason, Kochenderfer said, not all artificial intelligence is suited for mission-critical systems. “A lot of artificial intelligence is just using statistics and optimization together, but it has also come to mean … the use of neural networks.

“Neural networks are incredibly powerful. We’ve had major breakthrough in terms of computer vision applications and natural language processing applications. But in those domains, failure is tolerable. If Alexa doesn’t recognize your question correctly, people won’t be losing power; airplanes won’t be crashing.”

Edelson said the power sector will have to overcome its conservatism to get the most out of advanced analytics.

“We operate the system conservatively, justifiably so, because of its importance. … We apply margins large enough to accommodate fairly infrequent events. [Getting] system operators to rely on more advanced data analytics that allow for the system to be operated leaner will require organizational, cultural changes. That’s going to be a challenge.”

Changes will be required as the grid moves away from the traditional dispatchable thermal resources to much more variable generation, Mainzer said. As “we start running into real resource adequacy challenges … using every megawatt of available supply in the system — both bulk [system] resources and the behind-the-meter and distributed energy resources — is going to become increasingly important,” he said.

‘Trash in/Trash Out’

Anthony Papavasiliou, associate professor in the Department of Mathematical Engineering at the Catholic University of Louvain in Belgium, talked about the “trash in/trash out” challenge in estimating the need for system reserves.

“One of the reasons why … stochastic unit commitment is difficult is that you need to create reasonable inputs, the scenarios: Which resource should be outaged? Which forecast error should we consider? Building that input so that you get a meaningful answer from the optimization itself can be as difficult an exercise as actually solving the optimization problem that gives you the answer.”

Kochenderfer said early AI applications sometimes failed because they did not properly account for uncertainty.

“Another potential pitfall is using overly complicated methods. … We should definitely strive to test out the simplest possible ones first and then only use more complicated methods if we can justify that complexity in terms of performance on well defined metrics.”

Some complexity can’t be avoided, however, said Antoine Marot, AI team lead for RTE, the French transmission system operator.

“There’s been a lot of research for the last 10 years about how do we endure more uncertainties in the system. How do we go beyond [the] N-1 deterministic role [to] considering more probabilities?” he said. “Since we have a lot more risk and uncertainties to assess …. the thing we’d like to do for sure is speed up the computation of the simulations.”