Grid edge

AI-backed energy trading has a data problem

Optimizing returns on renewables is a key part of accelerating the energy transition, but AI might not be a silver bullet for energy markets.

Listen to the episode on:
An energy trader observes a row of computers, with a cityscape and lines of code in the background.

AI-generated image credit: Gold Flamingo

An energy trader observes a row of computers, with a cityscape and lines of code in the background.

AI-generated image credit: Gold Flamingo

Facing thin margins, developers and owners of wind, solar, and battery projects are looking for any financial advantage. They may find it in artificial intelligence, where automation could help boost revenue by improving forecasting accuracy and, ultimately, maximizing market bids.

According to the International Energy Agency, increasing the trading price for renewables is one of the most effective ways to boost the profitability of a project. A 5% increase in the price of onshore wind, for example, is projected to result in a 1.2% increase in equity returns.

Boosting power prices, though, is particularly challenging for renewables. There are major obstacles complicating short- and long-term price forecasting, including resource variability, a shifting policy landscape, and data granularity. 

To maximize returns amid that uncertainty, companies across the energy sector are paying more attention to AI’s potential role in trading. However, the technology isn’t necessarily a simple — or quick — fix for forecasting woes.

AI models, which typically assume that the future will somewhat mirror the past data that they’re trained on, aren’t as easy to apply to the constantly changing electricity grid as they may be elsewhere. Plus, there are data availability and data quality issues — problems that also limit AI’s applications elsewhere in the power sector.

Listen to the episode on:

Renewable energy prediction models face different constraints than the large language models that have grabbed attention over the last year, said Puneeth Kalavase, vice president of data science and engineering at Fluence, a large-scale storage developer launched by Siemens and AES.

“If you feed ChatGPT the entire corpus of the internet, you’re effectively going to have a better language model, right?” Kalavase explained. But for price forecasting, “that’s not necessarily true.” 

It might seem like feeding two years of grid data to a forecasting model would result in more accurate outputs than feeding it a single year, he added, but that’s just not the case. The core difference between large language models (like ChatGPT) and algorithms that predict energy prices is the rate at which training data evolves. 

“Every six months, the grid is fundamentally looking different, and you can see it in the data,” Kalavase said. “It’s a fascinating challenge for training some of these AI models.”

The changing topography of the grid is one reason AI alone isn’t sufficient for making consistently reliable price predictions, said Lauren Kuntz, CEO and cofounder of Gaiascope, an AI forecasting startup.

AI models need to be resilient in a system that’s “changing under your feet,” Kuntz said. And because historical data isn’t necessarily an accurate representation of the current (or future) grid, building in that resilience requires deep industry expertise, she added.

Every six months, the grid is fundamentally looking different.
Puneeth Kalavase, director of data science at Fluence

But it isn’t just that historical data is unreliable. There’s also a limit to the type and amount of data that’s available. Plus, utility data, both past and present, tends to be messy.

“One of the biggest challenges is just clean data in this space,” Kuntz said. “A huge chunk of our time entering a new market just goes into ingesting data and making sure it's in a format that we can work with.”

There isn’t a consistent format for data transmission across system operators, nor is there consistency in which data is available in which markets, Kalavase added.

“Helping standardize and democratize some of the electricity market data will help people build better models,” Kalavase said, pointing to the Federal Energy Regulatory Commission as a possible source of data regulation. 

"If FERC just mandated that all price data had to look a certain way and was available in a central place,” he said, “it would be pretty easy.”

Lack of data standardization is, in part, a policy issue. 

Of course, policy changes — as well as the addition of new system operator products — are a potential force to improve energy markets and grid dynamics. But they also serve as yet another constraint on price prediction models.

“You have to be super careful with AI,” Kuntz said. “If it was trained on historical data and then your market fundamentally changes because the policy changes, you better be able to incorporate that in some way.”

Pricing models therefore have to be able to adapt as quickly as today’s energy landscape. That need for agility limits the types of models being built, Kalavase said, “because we want things to be easy to test and easy to iterate on.”

But that doesn’t necessarily mean the energy sector is moving quickly to embrace AI-powered predictions. The shifting landscape of renewables and the high stakes of a functioning grid mean the “speed over caution” approach to AI favored by many in tech just isn’t an option.

Kuntz said she recommends that everyone in the grid space ask themselves a simple question when they consider incorporating AI: “Why do I trust that it’ll keep working when the grid is changing so rapidly?”

Both Puneeth Kalavase and Lauren Kuntz will be presenting at Latitude Media’s Transition-AI: New York conference on October 19th. See the agenda here.

No items found.
No items found.
No items found.
No items found.
big data
federal policy
energy markets