Domain-specific artificial intelligence models have been proliferating in the power sector. Energy companies, third-party vendors, and utilities themselves have been trying to maximize the potential of AI by developing tools trained on sector data to operate like industry experts. And their applications include everything from managing the grid to improving energy forecasting.
But the many efforts, no matter how promising, are mostly uncoordinated. Today, however, the Electric Power Research Institute is launching the Open Power AI Consortium to provide some order to a space that is “still very much the Wild West,” according to an EPRI spokesperson.
The group’s main aim is to “drive the development and deployment of an open AI model tailored for the power sector,” according to the announcement. As a first step, EPRI, alongside consortium members NVIDIA and Articul8, developed a set of domain-specific generative AI models for electric and power systems. The models will be the foundation for the consortium’s future work.
Jeremy Renshaw, EPRI’s executive director of AI and quantum, told Latitude Media that this “early ‘alpha’ model” one of the main focus areas for the consortium, and that it is currently “being evaluated and further trained using reinforcement learning from human feedback to enhance performance around a subset of topics.”
The consortium is global in scope and includes companies such as AWS, Microsoft, Constellation Energy, Duke Energy, PG&E, and NVIDIA. The group will work to develop and maintain the models, as well as the datasets and libraries necessary to train them. It also includes the creation of a “sandbox environment” for AI applications, in collaboration with startups, academia, national labs, utilities, and technology companies.
“We are collecting a range of applications for development,” Renshaw said in an email. “Right now, we have quite a few (over 50), but we will need to pare the list down. We anticipate starting on a list of around five to 10 use cases that span generation to transmission and distribution.”
The intent, Renshaw added, is to create tools “to solve specific use cases of varying size and complexity, which… could include using AI or generative AI technologies to solve individual use cases, up to using a multimodal open-source model trained/fine-tuned on energy industry data.” The consortium will then leverage its members’ resources and expertise to deploy the tools.
EPRI made the announcement at NVIDIA’s GTC event, an annual developer conference that has been dubbed “the Super Bowl of AI.”
This is the second example of EPRI — an organization focused on power system research — embracing the growing role of data centers in the grid in less than six months. Last October, the organization launched DCFlex with the aim of bringing together a large group of utilities, hyperscalers, and other tech companies to coordinate flexibility demonstrations at data centers across the world.
According to Tom Wilson, a principal technical executive in the integrated grid and energy systems division at EPRI, much of the value of this model comes from the collaboration is in bringing to the table players who might be working towards the same end goal, but are not necessarily communicating on a regular basis.
“There are a lot of bilateral conversations going on,” he told Latitude Media in an interview about the DCFlex program earlier this month. “We hope by getting the players together, that we can accelerate the level of conversation.”
