AES and Tapestry, Google’s moonshot for the electric grid, have unveiled a broad vision for architecting the AI-centric digital grid.
The two companies, which are partnering to virtualize AES’ planning and operations — and increasingly implementing machine learning and AI — just published a 20-year framework for building a semi-autonomous grid system.
Their goal: coordinate the power industry to establish a set of standards and protocols for orchestrating digital assets on the grid that will “support complex transactions that lead to efficient markets assisted by AI.”
AES’ own investments into machine learning and artificial intelligence may provide a view into how it could play out.
“We’ve been on the digital journey for quite a while and AI has risen to the top of that list,” explained Alexina Jackson, the vice president of strategic development for AES, in an interview.
There’s no shortage of visions for a digital electric grid. For the last 20 years, the power industry has been saturated with reports on the technical standards, business models, and reliability imperatives for building an intelligent grid modeled after the internet.
The smart grid first envisioned in the early 2000s hasn’t materialized. But the technology environment has changed dramatically since then. Advanced meters now cover 70% of residential customers; there are more than 5 million solar systems; public electric vehicle chargers are catching up to gas stations; and 272 gigawatts of distributed resources and flexible capacity are likely to hit the U.S. grid by 2027. These are all digital assets generating vast amounts of data.
Meanwhile, artificial intelligence is rapidly advancing — and presenting real value in dozens of applications across the grid.
The AES-Tapestry report covers familiar territory around the need for a digital transformation, outlining the cost and efficiency benefits of building a more flexible grid. However, the exploration of autonomous AI agents adds a new twist.
“At the infrastructure level, novel schemes for routing and buffering of electrons ensure a highly flexible and responsive grid,” the research found. “Interactions are facilitated by applications that support AI-based autonomous agents, and which in turn rest on continuous availability, exchange and brokerage of information.”
As AES worked with Tapestry beginning in 2021 to virtualize Midwest distribution grids and leverage more sophisticated machine learning for decision making, the teams started discussing a broader framework for the industry.
“When we first started working with Tapestry, we were really thinking about the specific impact on our utility,” Jackson told Latitude Media. “What could you do if you were able to better visualize not just the assets, but also the power flow? I think inevitably you get into conversations about how Google thinks about the internet, how it thinks about indexing data on the internet, and where are the opportunities within the grid for a similar approach.”
Jackson calls the digital grid architecture framework a “North Star concept” for her team. But how would it work in practice? To understand how a grid assisted by autonomous AI agents could emerge, it’s helpful to look at how AES itself views the technology.
AES, which operates multiple utilities and a vast portfolio of renewable energy and storage projects, has leaned heavily into machine learning for vegetation management, asset management, and expanding grid capacity through dynamic line rating (or DLR).
Tapestry, a project launched by Google’s moonshot factory that describes itself as a “Google Drive for energy,” has been working with AES to virtualize infrastructure (i.e. create digital twins) for the last few years.
Tapestry’s software allows utilities and network operators to organize and collaborate across the grid. The company has multiple planning and operational tools that incorporate multimodal machine learning for equipment identification, forecasting, power flow analysis, and data correction.
Tapestry has only a few announced customers, but those customers tend to go big. AES is one of them.
Building trust in AI
Jackson described how AES has moved machine learning “up the ladder of trust” inside the organization.
“It’s very helpful for the user not to expect AI to exist out of the box,” she said. “If you want good assisted intelligence from a machine, you’re going to invest the time to teach it. And once it learns, man, it can come up with great answers really quickly. And then the human can be in the loop to test it and ensure that the result makes sense and that between various results.”
She gave an example of how the machines learn and potentially evolve into those semi-autonomous agents.
During AES’ DLR deployment, one of the company’s poles was struck by lightning. When it was put back in service, the DLR information went through an algorithm containing a historical data set. The machine identified an unequal retention of the line, which might not have been noticed otherwise. “In a traditional analog system, you wouldn’t have identified it,” said Jackson. “It would’ve stressed the system, maybe reducing performance or the life of the asset.”
The O&M team at AES was intrigued. “It created trust in the tool,” said Jackson. And as that trust builds, Jackson sees teams relying on other machine-led insights for network upgrades — which is exactly what Chile’s grid operator is doing with Tapestry.
“The machine may actually be able to look at an area of the grid and say, ‘You know what? You’ve got a series of network upgrades identified across 10 lines,” she said. “‘If you put some DLR here, a little bit of storage there, maybe some power flow control over there, this combined result is a fifth of the cost and gets you 80% of the solution. So then you can just reconductor this single element over here and get to 100%.’”
Those potential insights, Jackson said, “is where we start seeing true efficiencies come online. And I think that’s the second step.”
The final step, then, is when humans have enough confidence in the machines to outsource certain tasks to AI in real-time operations.
In this phase, the team would “actually have confidence that we have operational tools and assistance from machines that we don’t have to use planning to solve all of the one-in-one-millionth probability of occurrence events.”
Jackson believes the grid will eventually be planned and operated this way, with lots more ML and AI getting layered in over the next decade. And the digital grid architecture report mirrors this vision. It’s a chance to start “talking about these grand concepts and the path toward that future” through 2040, said Jackson.
In the meantime, though, the hurdles are less about the technology itself and more about the structure of the electricity business: outdated regulation, conservative utility culture, lack of data sharing, and an absence of standards. One of the main intentions of the report is to encourage development of open, federated standards and protocols to enable distributed grid agents like the ones Jackson described.
“If we can create a framework that people can take from the conceptual — which this paper is — to the concrete and have engineers say, ‘yes, we do see this as possible,’ then we’ll have created a foundation that allows us to take the digital disruptors of today and make them the enablers of a modern grid,” she said.
