AI demonstration projects show great potential, but are utility companies ready?

Utilities and system operators are exploring new applications of artificial intelligence and machine learning to address reliability threats from load growth, according to multiple organizations and analysts.

According toa 2024 report by the Electric Power Research Institute (EPRI), generative AI models have "burst into the public eye." This surge has spurred massive AI investment commitments for 2025, such asthe $500 billion U.S. Stargate projectandthe €206 billion EU fund. Utilities are beginning to realize new possibilities.

"Even utility executives who were skeptical of AI five years ago are now using cloud computing, drones, and AI in innovation projects," said EPRI's Executive Director of AI and Quantum,Jeremy Renshaw. "Rapid adoption by utilities could make what seems impossible today standard operating practice within a few years."

However, concerns remain that AI/ML algorithms could bypass human decision-making and instead cause the very reliability failures they are meant to prevent.

"But any company that has not yet incorporated its internal knowledge base into a generative AI model that can be queried on demand is not fully leveraging data it has paid to store for years," saidMarc Spieler, Senior Managing Director at NVIDIA. He added that humans will still be involved in decision-making for now, with AI/ML algorithms making decisions smarter by providing more and more relevant data faster.

In practical demonstrations, utilities and software vendors are using AI/ML algorithms to improve tasks ranging from nuclear plant design to electric vehicle charging. But utilities and regulators must confront a dilemma: how to make proprietary data more accessible to new digital intelligence to improve reliability and lower customer costs, while also protecting that data.

Updates to older technologies

Power systems have already deployed AI/ML algorithms in cybersecurity applications, using technologies with cutting-edge learning capabilities to better identify attackers.

Checkpoint Software—a security services provider for global AI chip maker NVIDIA—is working with standards certification bodyUnderwriters Laboratoriesto develop new security standards for consumer devices, said Checkpoint's Global Chief Information Security OfficerPeter Nicoletti. Smart devices "will be required to meet security standards that prevent hacking during software updates," he said.

Another validated advanced computing application for power systems is market price forecasting based on weather, load, and available generation.

Amperonhas been using AI/ML algorithms for weather, demand, and market price forecasting since 2018, said its co-founder and CEO Sean Kelly. But Amperon's short-term modeling now "runs hourly and continuously retrains itself with less energy, smarter and faster, incorporating the advantages of each iteration in ways humans cannot match," he added.

Hitachi Energy'sNostradomus AIforecasting tool has the latest AI/ML capabilities and has "improved price forecast accuracy by 20% over manual forecasts since November," said Jason Durst, General Manager of Asset Management and Enterprise Software Solutions at Hitachi Energy.Jason Durstsaid.

AI/ML-assisted technology has also "become a key pillar of wildfire mitigation strategies," said Rob Brook, Senior Vice President and Managing Director of the Americas at forecasting software providerNeara.Rob Brooksaid. The technology helps utilities identify "wildfire risks in their networks by proactively evaluating more variables than humans can integrate," he added.

Over the past year, AI/ML algorithms have accelerated the deployment of solar construction robots, said Deise Yumi Asami, developer ofAES'sMaximo robots.Deise Yumi Asamisaid. Because Maximo's AI/ML algorithms can autonomously learn the unique characteristics of each solar project before construction begins, the six months previously required to retrain Maximo have been eliminated.

New, more autonomous AI/ML capabilities will provide "scalable improvements in stability, predictability, and reliability," said Nate Melby, Vice President and Chief Information Officer of Midwest generation and transmission cooperativeDairyland Power Cooperative.Nate Melbysaid. Managing system complexity "is exactly where AI can shine," he added.

Utilities are increasingly leveraging new AI/ML capabilities to address the growing complexity brought by variable loads, the surge in distributed energy resources, and other power system challenges.

New demands, new capabilities

Power systems lacking sufficient flexibility "could lead to reduced reliability and safety, and increased operating and capacity costs," summarizedPacific Gas and Electric Company (PG&E)in its2024 R&D Strategy Report. "AI/ML and other new technologies can not only enhance our immediate response capabilities, but also inform long-term planning and policy-making," the report added.

PG&E's total electricity usage will double over the next 5 to 10 years, but peak load growth can be limited to 10% through AI/ML-based optimization of DER on existing infrastructure, said PG&E CEOPatti Poppeat the company's November innovation summit.

Access to AI/ML algorithms is now commercially viable, with capabilities to optimize multiple large scenarios in parallel to support decisions involving millions of variables in power systems, said NVIDIA's Spieler. These algorithms can also write software code, enabling utilities to leverage "petabytes of system data they have stored but not yet used to optimize more operations," he added.

Utilities can upload their internal knowledge bases—including research papers, rate cases, and analyses of wildfire and safety issues—into generative AI models and query them, Spieler said. Query responses can then explain system anomalies based on performance and maintenance history, or provide the data and precedents needed to write ordinary rate cases and other regulatory filings, he added.

Utility demonstration projects are validating new AI/ML capabilities.

From distributed energy to nuclear plants

Several demonstrations focus on how AI/ML algorithms can optimize distribution system resources.

Utilidata's's Karman software platformand NVIDIA GPU chips have been embedded inAclara smart metersand will soon be applied to other distribution system hardware, said Utilidata's Vice President of ProductsYingchen Zhang. Karman reads high-resolution distribution system raw data 32,000 times per second and identifies individual customer electricity usage in real time, he added.

In a practical demonstration, Karman read and responded to fine-grained real-time data, finding that utilities can quickly stabilize voltage fluctuations caused by EV charging,University of Michigan-Utilidata researchnoted.

Within a year of implementing software from data disaggregation specialistBidgely,Avista Utilitiesreduced service orders dispatched due to high-bill complaints by 27%, reported Andrew Barrington, Manager of Products and Services at Avista Corp.Andrew Barringtonreported. Bidgely's software analysis identified customer electricity behaviors causing bill spikes rather than sending personnel to check meters, he added.

A Bidgely disaggregation analysis evaluatedEV charging for 10,000 Ameren Missouri customers, said its Vice President of DeliveryCaroline Cochranin aStanford-EPRI conference report. The analysis identified 73 customers who could avoid or defer high-cost infrastructure expenditures otherwise needed to manage EV charging load through better management, she added.

A similar Bidgely disaggregation analysis in 2023 of100,000 NV Energy EV charger ownersidentified "hot spots that might need infrastructure investment first," thereby limiting larger distribution system capital investments, aSmart Electric Power Alliancereport titled"AI for Transportation Electrification Insights Brief"released in January said.

AI/ML algorithms are also improving efficiency and reducing costs and safety risks at nuclear plants.

PG&E is usingAtomic Canyon generative AI softwaretrained toNuclear Regulatory Commissionstandards at itsDiablo Canyon nuclear plant，, said Patrick White, Research Director of theNuclear Innovation Alliance.Innovative AI/ML plant design, operations, and predictive preventive maintenance are limiting costs and improving plant safety, he added.

However, utilities and vendors recognize that much work remains to more fully leverage accelerating AI/ML capabilities.

The work ahead for utilities

Effectively harnessing the benefits of AI/ML algorithms begins with recognizing their potential and acquiring and using the right hardware and software, utilities and third parties say.

Avista's successful adoption of third-party AI/ML "started with a mindset," Barrington said. The key questions are "how to enhance customer interactions, how to integrate customer data with system operations, and how to enhance system visibility and support proactive strategies," he added.

AI/ML algorithms are now extracting real-time data and making actionable recommendations, said Utilidata's Zhang. But "utilities cannot leverage these recommendations because they lack the corresponding technology and communication ecosystem," he added.

Utilities need communication technology, advanced metering and edge computing infrastructure, and data processing and storage technology, said EPRI's Renshaw. At the distribution system level, utilities should also have software that can be safely updated as customers adopt new technologies, added Utilidata's Zhang.

Balancing the protection of security and customer privacy with providing data to train AI/ML algorithms remains a significant challenge.

Protecting utility data requires "robust cybersecurity practices," said Melby of Dairyland Power. But utilities need to access and manage data in ways that "AI platforms can leverage," he added.

Recently, "utilities have begun conducting penetration tests to prove that their data is as secure in our systems as in their own," said Cochran of Bidgely. They have also "established AI committees to conduct additional thorough reviews of their data users," she added.

"Utilities are conservative about data privacy for good reason, but AI/ML power system applications currently do not pose a threat," said Zhang of Utilidata. Federated learning or foundation models are methods to provide data for algorithm training while protecting privacy, he added.

Federated learning allows utilities to protect proprietary data by building synthetic models of their specific challenge data, which can be shared in secure locations for further training, said Zhang.

But some believe federated learning may be too limited to handle the complexity of power systems. Foundation models would use orders of magnitude more anonymized data and be pre-trained with as much power system information as possible, said Renshaw et al. of EPRI.

Utilities might be able to create foundation models to enable shared learning and protect their data, said the Senior Director of Grid Research, Innovation, and Development at PG&EQuinn Nakayamasaid.

"The bottom line is—collect more high-quality data, use, store, and protect it correctly, and feed it into models that are trained and updated for the right tasks," Renshaw concluded.