Artificial intelligence (AI) is changing nearly every industry imaginable, and the clean energy sector is no exception. With AI now top of mind for many utility leaders, there is value in learning from early adopters’ experiences with a range of technologies, use cases, and governance structures.
The Smart Electric Power Alliance (SEPA) recognizes the need for peer exchange and expert guidance on AI. At February’s RE+ Northeast conference in Boston, SEPA’s Emerging Technology team hosted a panel discussion on the current role and value of one form of AI – machine learning – for grid modernization in the Northeastern US. Thank you to Alexina Jackson with the AES Corporation, Sreedhar Sistu with Schneider Electric, David Stuebe with Camus Energy, and Mark Waclawiak with Avangrid for making the trip to Boston to join this panel. About half of our audience members said they already use AI in their work, but all leaned in close to learn from these experts. It was exciting to hear how utility operators, technology firms, and others are innovating to use AI here and now– as well as what to expect from this fledgling technology as its full capabilities unfold.
We learned that AI and machine learning are a good fit for many of the changes we need to make to optimize the electricity grid. This includes making more refined predictions about EV adoption, providing sharper and earlier visibility into where tree trimming and asset maintenance are needed the most, calculating new dynamic line ratings to maximize existing transmission corridors, and more. Panelists shared a consensus agreement that all AI/ML applications still require humans to think critically about data inputs and model design, and remain in the loop to oversee their use, take action based on the results, and monitor and calibrate performance. With several years of results coming in, panelists also explained how they’re measuring outcomes of AI investment today and why it is poised to help utilities develop innovative approaches for delivering safe, reliable, affordable, and low-carbon electricity to their customers. This blog captures what we learned from the four experts at RE+ Northeast.
(Do you need a refresher on key AI/ML terms before you dive in? See our January 2024 blog.)
Orchestrating the Modern Grid
The Northeastern US faces dynamic, broad-scale changes in the energy ecosystem, including a swell of building electrification and transportation electrification, the quest for more renewable energy generation, and studies about the role that batteries and distributed energy resources play in managing the future grid. Combined, these changes offer tremendous carbon-reduction potential, but create new complexity and uncertainty for developers, utilities, and grid operators. To the extent that these challenges require data-driven solutions, there is a widening opportunity to use advanced analytics to identify new solutions and manage change.
For example, anticipated EV adoption among residential and fleet customers presents new load forecasting challenges. It is increasingly important for distribution operators to be able to track and forecast the location and timing of EV-driven load growth. Coupled with refined grid-edge data from EV chargers and smart meters, machine learning-driven analytics have helped utility staff develop more accurate forecasting models for high-electrification scenarios. As Camus Energy, a grid orchestration platform software provider, and SEPA member, explains in their recent blog,
“Thanks to advances in AI and machine-learning, data from the edges of the grid can be used to generate accurate, meter-level forecasts for a fraction of historical costs. These hours- and days-ahead forecasts can help utilities make informed operational decisions and better manage the challenges of changing load patterns and growing demand.”
For grid operators, machine learning is also proving useful as a segmentation and classification tool to identify new strategies for managing the grid. Building on the machine learning tools in traditional forecasting models, the latest advancements enable utilities to incorporate new data streams– dynamically, and in higher volume. Panelists shared several examples of AI value when paired with the onslaught of digitized data about the grid, environmental conditions, asset health, and other metrics:
Through their Geomesh project, Avangrid is mapping their multi-state service area and forecasting electric system performance during various weather conditions. The project includes collecting and analyzing millions of data points on weather patterns, outage history, population, and vegetation. This comprehensive forecasting affords Avangrid greater visibility into grid performance and allows them to plan upgrades accordingly.
AES uses an emerging technology called dynamic line ratings from LineVision to harness untapped transmission capacity. Machine learning computations are the key to ingesting a wealth of real-time data on air temperature, solar radiation, and wind speed and direction. Models calculate line carrying capacity dynamically, indicating how much energy can flow through a power line at a specific time. The resulting real-time visibility enables AES to maximize existing transmission capacity.
Camus Energy uses machine learning classification algorithms to detect EV charging loads based on complex patterns in its utility customers’ Advanced Metering Infrastructure (AMI) data. This enables utilities to better understand their customers who are adopting EVs, track EV load growth over time, forecast how much EV adoption will occur in the future, and predict grid impacts.
Schneider Electric’s EcoStruxure Microgrid Advisor allows operators to optimize among onsite renewables, battery storage, EV chargers, the grid, and facility consumption needs. Using AI and advanced analytics, EcoStruxure Microgrid Advisor automatically connects to distributed energy resources and works alongside the grid to optimize when to store, produce, and consume energy. The software provides information on savings, earnings, and emissions through a cloud interface.
Digital Strategy for a Modern, Resilient Grid
AI/ML tools are one step in the progression to a mature digital-first approach. In that context, Sreedhar (Schneider Electric) explained that the added value of AI “is about making better decisions faster and at a larger scale than what humans can possibly do.” In this context, organizations should consider AI as part of a broader data and analytic strategy. Those who have invested in other digitization efforts may be the best-equipped to take early action on AI. This is because a high volume of digitized utility data enables the coverage and economies of scale useful for territory-wide model development and deployment. And, after all, as Mark (Avangrid) remarked, “Utilities are the ultimate data generator.” What then, can we accomplish by making use of that data?
Many early adopter utilities are focusing their initial AI efforts on resilience and reliability benefits. Generally, entities with transmission or distribution infrastructure perform routine inspections and management around their assets to limit and avoid customer interruptions or adverse events. Traditionally, this management occurs on a cyclical and relatively fixed schedule. AI-based tools help operators transition from a fixed approach to one that more dynamically identifies where risks are higher and action is needed.
Mark (Avangrid) elaborated, “Reliability is core to the day-to-day lives of our customers, and this is why we are treating data science and analytics as core to our business.” In addition to GeoMesh, Avangrid has a project they call HealthAI which uses photos of poles, wires, and grid equipment to identify the different assets and then analyze their health. Doing so allows Avangrid to better identify locations that most urgently require inspections and maintenance, thereby reducing outages, and providing lineworkers with more information in the event of outages.
Alexina (The AES Corporation) also explained that AI can be a resource when filling in data gaps to better characterize the grid. By switching to an ML-based model using historical data, AES can now take a more proactive and risk-based tree-trimming approach informed by vegetation and outage data. AES reports that the new data and new strategy have helped to both reduce costs and improve grid reliability.
Simple Tasks First
For some, the excitement around AI and ML is tempered with uncertainty about where to start. Panelists helpfully clarified strategies to get started on the path to digitization and AI adoption. One point of consensus: it makes sense to move incrementally, deploying aspects of data architecture and AI one step at a time. As Alexina (The AES Corporation) put it:
“Engineers often say we need an orderly transition. We need regulators and customers to see the value in the acts we are taking. We need objective information. Confidence [from early applications] allows us to think about planning the grid in new ways.”
How else can entities build confidence in the safety- and reliability-oriented utility industry? Several panelists elaborated that they are using machine learning to perform familiar, well-defined tasks where performance can be assessed relative to known benchmarks. AES’s tree trimming program is just one of many examples of using machine learning to improve upon a known task. By replicating an existing process in parallel, utility staff can benchmark the new model’s performance against outcomes in the business-as-usual approach. According to Alexina, AES Indiana determined that around 20% of the vegetation transects analyzed in the machine learning model comprised 80% of the risk, and refocusing their vegetation management on these identified areas, “has reduced vegetation-related interruptions by 3.5% in less than two years, a meaningful improvement for customers.” Studying the differences between AI-based and conventional results reveals opportunities to refine AI models for accuracy or improved value. Once the model drives small successes, it builds operator confidence in assigning an AI to work on new processes or to solve more complex tasks.
David (Camus Energy) offered another metric to establish trust: financial outcomes. When his company tests a new advanced ML-based load forecast, they perform a retrospective analysis to understand how well the new model performed relative to actual conditions. A key consideration for their utility clients is cost, so it is helpful to translate statistical precision and reliability to the utility’s costs. Utility operators will gain trust in deploying ML-based forecasts in the future if ML-based models show they can yield costs lower or similar to human-driven decisions.
Designing AI to Complement Human Experts
Although AI may be capable of a certain level of independent problem-solving, our panelists emphasized that AI is ultimately a tool to assist humans. Thus, it is crucial to design AI for the problems we have at hand now and to consult and engage analysts and leaders to learn what would be helpful to them. AI literacy and AI experience are top of mind for utility leaders working to enable their workforce to use AI. David (Camus Energy) underlined the need for training, skill-building, and governance relaying that some “[AI] adoption requires people to use a new tool for a new business function– which will require a lot of training. There is a question of how to build trust in tools that are completing a new business function.”
Two useful steps in the trust-building process are encouraging energy professionals to become familiarized with machine predictions and providing them with continuous educational opportunities on AI use. AI is not a “set it and forget it” tool. The longer human operators can stay “in the loop,” the more we can learn about its best uses and how to improve its value. This includes dedicating ongoing resources not just to problem identification, data cleaning, model development, query, and decision making– but also allowing space for reflection and recalibration. As people use AI more and more, they can learn from AI models and continue pointing this new tool in the correct direction to improve AI model accuracy.
AI’s potential to amplify embedded bias is one of the driving concerns around AI that warrants “staying in the loop” after a model is trained and deployed. The concerns are that both the data used to train AI/ML models, and the systems in which they are deployed, may make implicit biases worse, and result in negative outcomes for some groups. While our panel focused on machine learning– and not large language models– panelists agreed that this potential danger is worthy of examination as entities consider what an AI-enabled future might look like.
This said, all noted risk assessment and mitigation opportunities. Staff responsible for developing AI can, for example, measure whether AI generates biased results and then take the time to make corrections to ensure a more equitable outcome. Furthermore, AI can be useful in allowing us to see flaws in our data, and at times it can reveal biases we may not have noticed.
Scaling AI: Pacing Ourselves for the Long Haul
As the benefits of AI are just coming into focus now, we will continue to track the net benefit of AI for supporting lower-carbon grid modernization pathways. Today there are ample innovation opportunities for determining how we can use AI as a powerful tool in our decarbonization strategy. Two points help us cap the discussion around ML for grid modernization.
First, industry attention is needed to steer AI towards net benefit. Against the backdrop of today’s digitized world, AI is expected to drive growth in data center energy consumption. The impact on carbon emissions depends on how and when these new computations are powered. As Schneider Electric explains in a recent blog:
“The final verdict on whether AI is good for the planet depends on every single use case that businesses and organizations decide to implement. Our approach is that the energy saved must outweigh the energy needed to power an AI model. With a dedicated decision framework to make these kinds of choices, we ensure this technology is responsible, ethical, and serves a meaningful purpose.”
Second, organizations piloting AI today will soon need to confront questions about the long-term scope, value, and outcomes of these new technologies. Schneider Electric’s experts weighed in again here. Sreedhar relayed several learnings gained from their work on scaling AI.
First, it is important to work closely with stakeholders to identify problems together.
Second, AI requires investment and a team to build and maintain adequate staff capacity; as adoption scales, clear governance and processes help maintain alignment.
Last, Sreedhar also urges those setting out on AI adoption to make space for experimentation – or, put another way, building in the time and budget to learn within the innovation process. While working methodically and thoughtfully is important, taking a few calculated risks can lead to bolder innovation. Budget time and manage expectations assuming that not all AI pilots will become full projects and that not all projects will thrive.
Conclusion
Today, the industry is in a rapid growth stage and poised to continue learning what is possible with AI. The panelists all agreed that while AI and ML are in their infancy today, the technology’s untapped potential may have benefits we haven’t even begun to consider. As Mark (Avangrid) put it, “We are in the midst of a huge transition– and very much at the beginning of what is possible with the electric grid.” Onward!
Stay in Touch
If you are looking for more AI and ML insights catered to the energy transition, check out our other resources:
All SEPA members: please join us for a rotating series of AI deep dives at your member-only SEPA Working Groups this year. We are bringing you peer examples and lessons from the edge of AI, including how AI enables transportation electrification, microgrids, and more. Ask your Working Group Lead for more information.
Join executives and decarbonization leaders at the new Energy Evolution Summit, April 29-May 1, 2024 in Coronado, California for exclusive and candid conversations, including a roundtable on “Navigating Data Peaks: Digital Transformation and Carbon Reduction.” Register to attend here.
SEPA is planning a member peer-learning project to support you in assessing AI, including machine learning and generative AI applications as part of the digitally-enabled energy transition. If you have a success story to share, a question to ask, or an interest in participating in or sponsoring this research, email us at [email protected].
