War in Ukraine: We Need to Talk About Fossil Fuels

We think there’s a better way. Instead of drastically scaling up power-grid infrastructure, our work at the University of Vermont has focused on how to coordinate demand in real time to match the increasingly variable supply. Our technology takes two ideas that make the Internet fundamentally scalable—packetization and randomization—and uses them to create a system that can coordinate distributed energy. Those two data-communication concepts allow millions of users and billions of devices to connect to the Internet without any centralized scheduling or control. The same basic ideas could work on the electrical grid, too. Using low-bandwidth connectivity and small controllers running simple algorithms, millions of electrical devices could be used to balance the flow of electricity in the local grid. Here’s how.

Electricity demand on the grid comes from billions of electrical loads. These can be grouped into two broad categories: commercial and industrial loads, and residential loads. Of the two, residential loads are far more dispersed. In the United States alone, there are over 120 million households, which collectively account for about 40 percent of annual electricity consumption. But residential customers generally don’t think about optimizing their own electricity loads as they go about their day. For simplicity’s sake, let’s call these residential loads “devices,” which can range from lights and televisions to water heaters and air conditioners.

The latter devices, along with electric-vehicle chargers and pool pumps, are not only large electric loads (that is, greater than a 1-kilowatt rating), but they’re also flexible. Unlike lighting or a TV, which you want to go on the instant you throw the switch, a flexible device can defer consumption and operate whenever—as long as there’s hot water for your shower, your pool is clean, your EV has enough charge, and the indoor temperature is comfortable.

Collectively, there is a lot of flexibility in residential electricity loads that could be used to help balance variable supply. For example, if every household in California and New York had just one device that could consume power flexibly, at any time, the power grid would have the equivalent of around 15 gigawatts of additional capacity, which is more than 10 times the amount currently available from utility-scale battery storage in these states.

Here’s what flexibility means when it comes to operating, say, a residential electric water heater. While heating water, a typical unit draws about 4.5 kilowatts. Over the course of a normal day, the appliance is on about a tenth of the time, using about 10.8 kilowatt-hours. To the homeowner, the daily cost of operating the water heater is less than US $2 (assuming a rate of about 15¢ per kWh). But to the utility, the cost of electricity is highly variable, from a nominal 4¢ per kWh to over $100 per kWh during annual peak periods. Sometimes, the cost is even negative: When there is too much power available from wind or solar plants, grid operators effectively pay utilities to consume the excess.

To reduce demand during peak periods, utilities have long offered demand-response programs that allow them to turn off customers’ water heaters, air conditioners, and other loads on a fixed schedule—say, 4 p.m. to 9 p.m. during the summer, when usage is historically high. If all we want to do is reduce load at such times, that approach works reasonably well.

However, if our objective is to balance the grid in real time, as renewable generation ebbs and flows unpredictably with the wind and sun, then operating devices according to a fixed schedule that’s based on past behavior won’t suffice. We need a more responsive approach, one that goes beyond just reducing peak demand and provides additional benefits that improve grid reliability, such as price responsiveness, renewable smoothing, and frequency regulation.

How can grid operators coordinate many distributed, flexible kilowatt-scale devices, each with its own specific needs and requirements, to deliver an aggregate gigawatt-scale grid resource that is responsive to a highly variable supply? In pondering this question, we found inspiration in another domain: digital communication systems.

Digital systems represent your voice, an email, or a video clip as a sequence of bits. When this data is sent across a channel, it’s broken into packets. Then each packet is independently routed through the network to the intended destination. Once all of the packets have arrived, the data is reconstructed into its original form.

How is this analogous to our problem? Millions of people and billions of devices use the Internet every day. Users have their individual devices, needs, and usage patterns—which we can think of as demand—while the network itself has dynamics associated with its bandwidth—its supply, in other words. Yet, demand and supply on the Internet are matched in real time without any centralized scheduler. Likewise, billions of electrical devices, each with its own dynamics, are connecting to the power grid, whose supply is becoming, as we noted, increasingly variable.

Recognizing this similarity, we developed a technology called packetized energy management (PEM) to coordinate the energy usage of flexible devices. Coauthor Hines has a longstanding interest in power-system reliability and had been researching how transmission-line failures can lead to cascading outages and systemic blackouts. Meanwhile, Frolik, whose background is in communication systems, had been working on algorithms to dynamically coordinate data communications from wireless sensors in a way that used very little energy. Through a chance discussion, we realized our intersecting interests and began working to see how these algorithms might be applied to the problem of EV charging.

Shortly thereafter, Almassalkhi joined our department and recognized that what we were working on had greater potential. In 2015, he wrote a winning proposal to ARPA-E’s NODES program—that’s the U.S. Department of Energy’s Advanced Research Projects Agency–Energy’s Network Optimized Distributed Energy Systems program. The funding allowed us to further develop the PEM approach.

Let’s return to the electric water heater. Under conventional operation, the water heater is controlled by its thermostat. The unit turns on when the water temperature hits a lower limit and operates continuously (at 4.5 kW) for 20 to 30 minutes, until the water temperature reaches an upper limit. The pair of black-and-white graphs at the bottom of “Matching Electricity Demand to Supply” shows the on and off patterns of 10 heaters—black for off and white for on.

Under PEM, each load operates independently and according to simple rules. Instead of heating only when the water temperature reaches its lower limit, a water heater will periodically request to consume a “packet” of energy, where a packet is defined as consuming power for just a short period of time—say, 5 minutes. The coordinator (in our case, a cloud-based platform) approves or denies such packet requests based on a target signal that reflects grid conditions, such as the availability of renewable energy, the price of electricity, and so on. The top graph in “Matching Electricity Demand to Supply” shows how PEM consumption closely follows a target signal based on the supply of renewable energy.

To ensure that devices with a greater need for energy are more likely to have their requests approved, each device adjusts the rate of its requests based on its needs. When the water is less hot, a water heater requests more often. When the water is hotter, it requests less often. The system thus dynamically prioritizes devices in a fully decentralized way, as the probabilities of making packet requests are proportional to the devices’ need for energy. The PEM coordinator can then focus on managing incoming packet requests to actively shape the total load from many packetized devices, without the need to centrally optimize the behavior of each device. From the customer’s perspective, nothing about the water heater has changed, as these requests occur entirely in the…