Electric utilities are seeking innovative ways to manage more complex grids with growing adoption of rooftop solar, electric vehicles, battery storage and other local energy resources. At Camus, we envision AI as an essential part of the solution, but not a silver bullet.

AI will not be independently managing the grid any time soon. In particular, today’s AI technologies are not suitable to replace trained operators for critical operational functions—like real-time switching of distribution feeders to avoid or restore outages. Nevertheless, we believe AI will be transformative for grid operators.

Why does AI hold so much potential?

AI-driven algorithms can discern patterns in—and learn from—vast amounts of grid data, enabling them to instantly find effective solutions to complicated operational scenarios and planning decisions.

While there is no shortage of conjecture about how AI will transform the future of grid operations, we’d like to take a more pragmatic angle. This blog is the first in a series of articles delving into how Camus is using AI to help utilities operate their grids more effectively today (and in the very near future). The series will focus on capabilities available now or within the next 18 months, with an aim to demystify how AI works and to illustrate its real-world grid benefits.

Our expectation for AI is not that it will replace utility systems like Advanced Distribution Management Systems (ADMS). Rather, AI will be used to improve and augment existing systems–serving as an intelligent copilot for utility operators and planners. The idea is to make these subsystems more efficient, help them find optimal solutions, and surface actionable insights to grid operators and planners–who ultimately make the decisions. 

Many of our current AI applications are not particularly sexy or earth-shaking. But they are making life much easier for utility staff and enabling enhanced reliability and affordability. Our success to date has demonstrated that utilities don’t need to adopt an all-or-nothing approach to AI. Instead, they can take a crawl-walk-run approach, in which AI supports their needs in increasingly strategic ways.

This first post details our AI-driven approach to enhance situational awareness throughout the distribution grid by speeding up the use of model-based power flow analyses. 

Power flow models are a slow path to grid awareness

When it comes to grid visibility, distribution utilities have a big blind spot. They have monitoring devices at substations and smart meters tracking customer usage. But they have very little awareness of conditions at everything in between: the conductors, transformers, switches, and reclosers. This blind spot presents a big challenge for utilities. With rapid load growth from electrification and accelerating adoption of controllable, distributed energy resources at homes and businesses, utilities need middle-of-the-grid visibility more than ever to operate the grid reliably and prevent equipment damage.

Some utilities use physics-driven power flow models to estimate conditions in parts of the grid that lack direct monitoring. The hope is that by understanding local grid conditions, control room staff and operational systems are better equipped to reliably balance energy supply and demand in various scenarios. 

A power flow model is essentially a mathematical representation of the arrangement of assets in a distribution grid accounting for their physical, electromagnetic, and operational characteristics and constraints. By solving the set of electrical equations that comprise a power flow model, a utility can determine the power and voltage at different points on the grid. The models include solver software that attempt to solve the model’s equations by taking an initial guess and refining the solution iteratively until a desired accuracy is achieved. This is a computationally intensive process and executing these models to generate meter-level forecasts for the next 48 hours across large distribution grids may take hours or not even be possible.

‍

Consider the following metaphor to help illustrate this complex iteration process. The grid operator is attempting to find the lowest elevation, which occurs at the lowest point across many valleys. The power flow solver software starts by throwing a dart randomly in one of the valleys. It then moves in one direction–and sees if the results were higher or lower.

Usually the model eventually makes it to the lowest point, but it could take a long time if the dart initially lands far away. Depending on the initial dart location, the solver may end up exploring the wrong valley, never finding the viable solution and requiring the dart throwing process to start over. 

Reliance on physics-based power flow models frustrates utilities in a number of ways. Grid operators may have only seconds to respond when cloud cover suddenly reduces output of a large number of distributed solar arrays, leading to a load spike. Running a new grid-wide power flow model simply isn’t feasible in fast-changing operational scenarios. Even for planners, who work on longer timescales,  the reliance on running slow power flow solvers for solar and wind interconnection studies has resulted in significant deployment delays.

Accelerating how utilities engage with power flow models can help utilities make better decisions faster. AI can cut computational time from days to minutes or minutes to seconds.

Accelerating power flow analyses with Machine Learning

To speed up the calculations of power flow models, Camus is training machine learning (ML) models which can nowcast and forecast load, every 15 minutes, with high accuracy. As a result,  initial guesses provided to the solver are much closer to reality which in turn give the power flow solver a head-start in solving its physics calculations. This framework significantly decreases the computational effort and at the same time, provides visibility at the midpoints of the distribution grid with high fidelity. It’s essentially a hybrid approach in which AI is helping the power flow solver do its job more quickly and effectively.

We’ve used these models and techniques with historical grid data from our partner utilities and found that they can accurately inform power flow analyses in scenarios such as managed electric vehicle charging and solar plant curtailment. Based on our testing, we believe the models can support continuous power flow analyses on every feeder operated by a large distribution utility–working quickly and cost effectively enough to be used around-the-clock.

What about utilities with no existing power flow models?

For the many utilities that don’t have existing power flow models, developing approximate models from existing data may be much easier than utilities expect. We’ve worked with utility partners to construct power flow models by standardizing and integrating data from readily-available sources. 

We start by defining how all the components in a utility’s distribution grid—transformers, conductors, fuses, solar power plants, electric vehicle chargers, and much more—are connected. This “connectivity model” also contains information on the various attributes and constraints of these assets, such as transformer ratings, regulator control settings, and rooftop solar capacities.

The information in the connectivity model is gathered from geographic information systems (GIS), customer information systems (CIS), ADMS, and interconnection documents for large-scale solar and battery plants. These models are never 100% accurate, potentially resulting in significant errors in power flow simulations. So we use AI-driven statistical tools, which learn from the data to correct errors and fill gaps.

We construct the power flow model by using the connectivity information as inputs in the model’s electrical equations. Here’s what solving the power flow model might look like: Every five minutes, we input energy usage and production data from customer meters as well as meter-level forecasts into the model. This enables us to calculate power and voltage at any point on the grid—now and over the next 48 hours, and in a variety of different switching scenarios. 

This solution eliminates what was previously a big blind spot for the utility, providing situational awareness—now and in the near future—all across the grid between the meter and substation. Camus’ platform can work with such a power flow model to orchestrate devices based on this awareness.

Subscribe for more examples of AI as a copilot

Future installments of this series will dive into other ways in which AI can be used to support grid operators, including interpolating grid conditions from metering plus connectivity models, anticipating coincident system peaks, and computing dynamic operating envelopes.

To stay on top of the latest and greatest insights, subscribe to our newsletter.