The California Independent Systems Operator processes hundreds of outages daily — a painstaking task only made more challenging by the sheer complexity and interconnectedness of California’s grid and a rise in extreme weather events and distributed assets.
Managing planned and unplanned interruptions to parts of the grid is a core responsibility of independent system operators. But it’s a continuous, highly manual task distributed among hundreds of employees every single day. Outage requests submitted by transmission owners, generation companies, and utilities looking to take infrastructure out of service — whether for maintenance, upgrades, or unexpected faults — must be assessed for their impact on system reliability, market operations, and customer service.
CAISO operators in departments around the organization (transmission, generation, reliability) read those reports, track down relevant historical data and operational information from multiple sources, input relevant metrics into CAISO’s grid software system, and run calculations to evaluate the impact.
But the organization is moving slowly away from doing everything by hand. For the last several months, CAISO has been piloting a generative AI platform that it hopes will streamline the process and help it meet the compounding demands of outage coordination.
“Genie” was purpose-built for CAISO by energy services provider Open Access Technology International, and is designed to automate the outage validation process. It pulls together information from across the CAISO system and provides system operators with real-time recommendations as they vet and process outage requests each day.
With Genie, California will become the first state to actively manage power outages with AI, explained Abhimanyu Thakur, vice president of platforms, visualizations, and analytics at OATI.
Key to the rollout is the longstanding relationship between CAISO and OATI. CAISO already used a suite of OATI applications, meaning Genie could be more easily integrated into existing operations without disruption.
Genie also runs entirely within OATI data centers: Its large language models (trained on CAISO’s own dictionary) and agents are all hosted internally, where CAISO data already resides, meaning no data leaves the network.
Building a library of AI agents
Most artificial intelligence tools piloted and deployed in the energy sector to date have focused on forecasting, such as predicting load or energy prices. For example, Duke Energy is using AI to run power flow simulations. Google’s moonshot project, Tapestry, is applying AI to the transmission planning process by simulating grid congestion.
What makes Genie unique is that it layers advanced generative AI, like natural language processing, with agentic AI, which can proactively plan and execute multi-step tasks, automate routine analyses, and make recommendations to operators in real time.
The platform is made up of a suite of AI agents, each responsible for a different aspect of outage processing, Thakur explained. “The whole goal here is that by using a combination of multiple agents, you should be able to solve a problem,” he said. “We’ll continue to build different agents as the needs arise, and keep adding them to our library…Eventually we’ll have hundreds of agents available in this infrastructure.”
OATI built Genie in this modular, agent-based way on purpose, he added, to allow precise validation of the results of individual tasks, and help prevent AI hallucinations. Each component is being tested independently, but eventually they will be tested in parallel and then integrated incrementally into the overall Outage Management Application.
For example, there’s an agent focused on keyword extraction. Today, CAISO engineers read through outage reports and requests to identify keywords that come up in specific contexts, like downed lines or transformers, and then run calculations to determine how each item could impact the grid. Genie, meanwhile, can scan requests, as well as any free-form, unstructured notes attached to them, interpret them using CAISO’s own dictionary of operational keywords, and automatically flag potential discrepancies for further review.
Meanwhile, “similarity search” agents sift through historical outage records and compare them to new requests based on equipment, location, and procedural details to anticipate reliability risks. And report-generating agents combine inputs from around the organization to create nightly summaries that can be reviewed by operators in the transmission, reliability, and generation departments.
A simple example of how all of those elements work together, Thakur explained, is an operator using Genie to identify the most relevant documents — out of hundreds — for a particular outage request. Instead of having an employee or several parse through hundreds of PDFs worth of operating procedures, Genie could direct the operator to those that are most likely to be relevant.
“Imagine, for all 200 outages you have to evaluate, the AI system can detect that these three are the most important documents…and within these three documents these are the paragraphs that are most important,” he said. “You’ve suddenly shrunk that overhead. So we’re not really doing predictions here, we’re improving the whole experience of how operations happen inside the control room.”
Timeline and implementation
Building AI tools for the energy sector is a slower, more deliberative process than for many other industries. For example, the pilot started with several months of listening sessions with more than 200 CAISO grid operators, Thakur said. OATI held meetings with over 200 CAISO staff across the organization, tracing workflows and attempting to understand true pain points and their origins.
OATI’s data scientists and engineers set out to tailor Genie’s underlying models, training them on CAISO data including historical outages, internal dictionaries, and operating procedures. Then, they set about creating synthetic test cases to validate and ground the models.
As of July, CAISO’s grid operators are testing each of the four agents individually. “We don’t want to wait to put out this giant AI system and then validate the accuracy, because AI systems can hallucinate,” Thakur explained. The ISO team is in the midst of testing, and then will work on validating the various agents — a critical step to determining whether Genie makes the outage management process more efficient, and more reliable.
The goal is that by December, the pilot will be integrated and available in CAISO’s outage management application, Thakur said. In the meantime, the ISO is also working on defining what success looks like.
“Obviously the goal here is to assist the operators on multiple levels,” Thakur said. But defining exactly how to measure the impact of that assistance is tricky, and requires establishing a baseline, which the ISO is working on as well.
The first metric they’re focusing on is time, he added. Does the use of Genie decrease the aggregate amount of time it takes CAISO operators to review outages?
To create that baseline, CAISO needs to establish how long it takes an operator to review a single outage, and then extrapolate that to all outages managed by each group within the ISO for a given day.
CAISO, for its part, appears to be approaching the pilot with exactly the level of skepticism associated with the slower-moving power sector. A senior advisor at the ISO told MIT Tech Review earlier this month he was cautiously optimistic about where and how broadly Genie could be applied. “I don’t want to overhype it,” he said. “But I see a world where AI agents are able to do a lot more.”
