Regardless of what fuel source we use to generate it, electricity is not particularly useful if you can’t get it to consumers. Which is why grid reliability is an area that The Fusion Report puts a lot of interest in. For anyone who has been following the electricity sector, the impacts of the rapid increases in electrical demand (coincidentally, much of it from data centers) on grid reliability is a big  concern. One of the hottest topics as of late regarding grid reliability is how artificial intelligence (AI) could potentially help stabilize the same grid that AI datacenters are negatively impacting.

Understanding grid reliability requires understanding the grid. While most people see the grid as one big monolithic structure, in reality there is a great number of smaller grids interconnected by transmission lines and switching gear. Each of these smaller grids are highly instrumented so that power can be balanced and dispatched across these individual grids to where it is needed. This, plus the distributed nature of grids which utilize (at least in part) renewable energy sources, means that rapid changes in either demand or supply can cause local grid instabilities. This is where AI comes in. Some of the most forward-looking work in this area is being performed by the National Renewable Energy Laboratory (NREL). NREL’s work is focused on how generative AI can provide proactive decision support for grid operators such as the California Independent System Operator (CAL-ISO), the Midcontinent ISO (MISO), and similar US Grid operators.

The Potential Impacts of AI on Grid Management

There are a number of ways that AI and machine learning (ML) can help improve grid management:

• Responsiveness to Real-Time Situations: Prior to machine learning models, it took roughly ten (10) minutes for MISO to perform complete grid dispatch planning. With machine learning models this has been reduced to less than 60 seconds, allowing grid operators to respond to rapid changes or unexpected situations far faster than they were able to previously.
• Anticipating Outages Caused by Equipment Failures: Unanticipated changes in electrical demand are not the only cause of grid instabilities; equipment failures can also cause rapid instabilities within local grids that can propagate through entire regions. By training AI and ML models on the grid infrastructure and the characteristics of grid equipment, it becomes possible to proactively predict equipment failures and take the necessary maintenance actions.
• Mitigating the Impact of Weather Events on the Grid: Just as unexpected equipment failures can cause grid issues, adverse weather events can wreak havoc on the grid. By understanding weather conditions and the location of grid infrastructure, mitigation strategies such as the shutdown of vulnerable lines or equipment can be used to prevent weather issues from becoming grid issues.
• Managing the Dispatch of Renewable Resources: Renewable resources such as solar and wind energy now make up much larger parts of our electricity generation capability. With the increasing penetration of wind, solar, and other distributed energy resources, grid reliability can be challenged by the sometimes-intermittent supply characteristics of renewables. AI-powered tools such as graph neural networks can predict and balance this variability, enabling improved optimization of the grid.
• Improving the Utilization of Energy Storage Resources: As we have reported earlier, the utilization of energy storage systems, particularly battery energy storage systems (BESS), can have a strong impact on grid reliability by providing a “sink” for excess energy generated by renewable resources. AI and ML models can help not only to optimize the use of this stored energy, but also to plan where these resources should be deployed in the future, based on today’s power consumption models.

Best Practices for Deploying AI Solutions to Support the Grid

Because of the nature and criticality of the grid, AI and ML models cannot be deployed in the same manner as they would be for business applications. While a cloud-based model is typical for the deployment of most AI and ML solutions, the problem with doing so in the case of the utilities is that a single point of failure could take out the entire management capabilities for the grid.

To minimize the likelihood of this occurring, most experts suggest the following:

• Utilize an Edge-First Approach with Centralized Control: This approach enables local (“edge”) nodes to make decisions, which are then coordinated with those of other nodes through a centralized control point.  This allows quick, independent reaction to local problems, while still maintaining a system-wide picture of the overall utility grid situation.
• Share AI Learning Through a Federated Learning Approach: Rather than having all learning occur at the “central core”, federated learning spreads the AI learning across a number of major nodes. Each of these nodes learns from the endpoints in its local network, and then sends the learned parameters to the central server, which updates the global model.
• Ensure a Strong System for Data Management: By nature, a system like this would have multiple stakeholders, including multiple government agencies and utilities. Changes made by one party can affect the entire network, which requires all parties to agree on a change management protocol that gives everyone visibility. Obviously, a strong cyber-security approach would also be critical to minimizing system vulnerability.

Conclusion: AI Holds Lots of Promise for the Grid

Maintaining grid reliability, stability, and resilience on today’s complex grid is no simple feat, but the benefits of utilizing AI for grid management promise the potential to “make things work”. By planning ahead and taking the unique nature of the utility grid into account, AI solutions can be deployed that can maximize grid resilience without making the grid vulnerable to single points of failure.