Co-op Webinar: How to analyze grid impacts of EVs and Solar in just 10 weeks. Register now
Electrification of transportation, buildings, and industry is transforming energy systems around the world. The potential benefits for society—especially decarbonization and affordability—are enormous. Electrification also offers an important upside for utilities, which can call on electric vehicles (EVs), heat pumps, and other controllable, electrified appliances to help balance their grids.
But utilities also recognize that electrification can create challenges for grid reliability. Even modest load growth in local pockets—such as when several homeowners on the same street purchase EVs—can overload transformers, conductors, and utility equipment at the grid’s edge.
So what can utilities do to identify and manage the challenges of electrification?
One key part of the solution is load forecasting. Forward-thinking utilities have long developed load forecasts to inform both long-term system planning—such as where to upgrade substations—and short-term operations. In both cases, utilities make load predictions based on correlations between weather data and aggregated load, typically forecasting at the feeder-level—which is roughly halfway between the transmission system and the end customer.
This has worked well for decades of slow or non-existent load growth, but the rapid adoption of high-powered devices like EV chargers is quickly highlighting the gaps of feeder-level forecasting – namely that utilities can’t see the hyperlocal impacts of electrification on their equipment. Considering a medium-sized utility manages tens of thousands of transformers, that’s a lot of blind spots.
However, forward-thinking utilities can now turn to meter-level forecasts to identify local electrification impacts and inform their actions. The concept of meter-level forecasting is not new, but it’s long been viewed as too expensive, complex, and slow to be useful. Fortunately, advances in analytical efficiency via machine learning and the low costs of cloud computing have made it possible for utilities to generate meter-level load forecasts for use in real-time operations.
Over the past year, our team at Camus Energy has worked with utility partners to develop a fast, efficient, and accurate meter-level forecasting system. In addition to predicting load, it can forecast generation at sites with solar arrays – even those without telemetry from the solar inverters. We’ve successfully scaled forecasting to more than a million meters, with day-ahead, hourly (soon: 15-minutely) forecasts calculated in just minutes and achieved at a fraction of the cost of prior forecasting efforts. We now have a meter-level forecasting product that we can quickly roll out for new customers and that can scale up or down cost-effectively.
Our forecast system is already empowering operations for our partners. There are three immediate ways it’s proving helpful:
Let’s dig a little deeper into each of these three use cases.
We’ve found that offering to provide net and gross load forecasts for every point on the grid turns heads – and raises questions. How do we achieve that? And how can it be cost efficient?
Our ability to forecast at any point relies on a mix of real-world data streams and a utility’s connectivity model. We start at the edges of the grid, bringing in best available data for meter-level loads and distributed energy resources (DERs). Typically, this includes 15-minute meter data, a list of known DERs (usually solar and battery storage, sometimes EVs) tagged to specific meters, telemetry directly from a limited number of those DERs (often the largest devices), and a GIS-based map of the distribution system. We combine this with best available weather data to create meter-level load forecasts. There’s a little bit of magic happening in that process (see our deep dive blog), but that’s the general idea.
We then roll up our meter-level forecasts to any point on the grid, using the connectivity model to help us provide forecasts for all transformers, as well as other upline locations. For example, if a certain transformer serves five meters, we aggregate the forecasts for those five meters into one forecast for the transformer. These “roll ups” also point out errors in the utility’s connectivity model – such as a piece of equipment that appears to be loaded at 1,000% of its rated capacity for many hours of the day. We work with our utility partners to identify and resolve inaccuracies in grid models until we have an approach they trust.
These aggregated forecasts can enable numerous applications for grid operations teams, including better balancing the grid through orchestration of distributed energy resources. For example, avoiding major equipment upgrades is an increasingly high priority for utilities managing the costs of adapting the grid for electrification. With transformer-level forecasts and active managed charging of downline EVs, we’re able to help utilities defer expensive equipment upgrades or even avoid them entirely. The ability to forecast net and gross loads for specific pieces of equipment is essential to right-sizing a utility’s capital expenditures.
The latest smart meters provide energy consumption data averaged over 15-minute intervals. However, utilities don’t receive the meter data until hours or even days later due to latency in communication and processing systems. That is too delayed to be useful for operations teams. They need to know what’s happening now and what’s happening next, so they can make decisions now for better operational outcomes.
Our forecast system directly addresses this delay. It uses the best available meter data, DER telemetry, and real-time weather data to forecast what’s happening right now. This “nowcast” is updated every hour and provided at 15-minute resolution. Utility operators can clearly see what data is measured and what data is forecasted, better enabling them to manage real-time operations versus looking at a large gap in net or gross load visibility at the grid edges.
A third key application for our utility customers is helping them restore service after outages. When a utility is managing restoration from an outage — whether planned or unplanned — they need to know how much load will be present downline from that location at the time of restoration. Historically, static load profiles have been sufficient to enable safe restoration. But the growth in distributed generation, especially rooftop solar, has meant that there’s often a lot more total (or “gross”) load than what the utility might see when looking at static load shapes.
That becomes a problem during restoration because the solar inverters have safety settings to prevent them pushing energy back onto a de-energized line. As a result, when the utility re-energizes a line, those solar inverters remain turned off — and the actual load can be much higher than the utility expected. The result can mean overloaded lines, voltage issues, and even safety concerns for line workers.
Our meter-level forecasting helps reduce these risks by estimating distributed generation at each meter and incorporating that into forecasts for gross load at every meter. When these forecasts are used in the restoration analysis (e.g., incorporated into the state estimation process), the utility is much better positioned to execute safe, reliable service restoration for their customers.
For any tool to be used in utility operations centers, it needs to be trustworthy, efficient, and accurate. Our meter-level forecast tool uses open-source machine learning algorithms that have been tried and tested. Drawing on the experience of our utility partners, we select inputs for the algorithms and train them on input data specific to that utility. The algorithms learn the relationships among the inputs and inform a machine learning model that forecasts load based on new input data. This type of machine learning, known as gradient-boosted decision trees, has been used in the field for years in mission-critical industries like logistics, healthcare, cybersecurity, and finance.
Much like in other industries, cloud computing is an essential enabler for rapidly—and cost-effectively—crunching large amounts of data for training, testing, and forecasting. With our cloud approach, we can run forecasts for millions of meters in parallel – and immediately ramp those servers back down when not needed. Compared with on-premises approaches, cloud computing is significantly more flexible and cost effective.
To further increase efficiency, we use a cohort-based approach. We divide a utility’s meters into cohorts—groups that share characteristics such as rate class or presence of DERs. We randomly select from each cohort a subset of a few thousand meters and train the algorithms on data from the subset, rather than from the much larger cohort. This makes training much faster and more cost effective.
And, of course, before we launch the system, we need to demonstrate its accuracy. We do so by running our approach on our utility partners’ real-world historical data and comparing its predictions to what historically occurred on their grid. We continue to back-test our models to assess accuracy over time and identify edge-cases the model is less suited to predict.
For a deeper dive into how our forecast works, I recommend reading my colleague David Stuebe’s companion blog post.
Meter-level forecasting is one of the most compelling new utility capabilities, and I believe it will be transformative for utilities as they look to manage the impacts of electrification. When coupled with managed EV charging and DER dispatch, meter-level forecasts provide the foundation for robust orchestration of the distribution grid.
Today, Camus’s DER Orchestration platform is providing our utility partners with the universal foresight to support the shift to a 100% electrified future. If you’re interested in learning more about our meter-level forecasting, reach out to our team directly or subscribe to our updates.
