Fine-tuning how homes can help the grid as 'virtual power plants'

Can thousands of houses equipped with remote-control thermostats and other devices mimic what big power plants do for utility grids?

EnergyHub, a software and services company that operates virtual power plants (VPPs) for about 70 utilities in 30 states, says yes — and it has a trove of data from three major U.S. utilities that shows how it can be done.

This week, EnergyHub is publishing the results of trial runs it conducted this spring and summer with Arizona Public Service in Arizona, Duke Energy in North Carolina, and National Grid in Massachusetts. Over several afternoons and evenings, EnergyHub tested its capabilities to get thousands of smart thermostats — and for National Grid, rooftop solar-charged batteries as well – to help the grid.

The company says its approach, known as dynamic load-shaping, will allow it to more precisely manage the 2 gigawatts of flexible capacity it now controls from about 1.4 million customer devices, said Paul Hines, EnergyHub’s vice president of power systems. The company is hoping that the successful test cases will help convince utilities and regulators that ​“virtual power plants can become a core part of their infrastructure,” he said.

Getting customers to turn down their air conditioners, water heaters, and other appliances to deal with rising electricity demand is a lot cheaper than building new power plants. Hines cited a U.S. Department of Energy report estimating that 80 to 160 gigawatts of VPP capacity could be unleashed across the country by 2030, enough to meet 10 to 20 percent of U.S. peak grid needs and save utility customers roughly $10 billion in annual costs.

“There’s broad agreement that there’s a need for this stuff,” he said. ​“But when you talk to grid operators, they still have a ton of skepticism about whether virtual power plants can be a valuable part of the core resource mix.”

That’s because traditional demand response programs that pay customers to let utilities turn down their thermostats and other appliances have some well-known features that can make them unreliable, he said.

A typical demand response event – say, one reacting to a heatwave — involves two key steps: The first is ​“precooling,” or ordering thermostats to ramp up air conditioning earlier in the day when the grid isn’t yet stressed, so that homes can ride through the hotter hours ahead. The second is the demand response event itself, during which temperature settings on thermostats are raised to reduce air-conditioning power use when overall grid demand reaches its peak. 

Hines pointed out two significant issues that can crop up as a demand response event unfolds. First, there’s the ​“decaying response” factor. ​“You get a ton of load shed in hour one,” he said. But as people grow less comfortable with rising temperatures and start to override their thermostat settings, the energy savings decrease.

Then there’s the so-called snapback effect — the surge in electricity use when a demand response period ends and a large number of thermostats reset, triggering a bunch of air conditioners to turn on at once. ​“You get this big spike in load at the end,” he said — and that secondary peak can cause grid problems of its own.

These are standard problems for a smart-thermostat-based program, Hines said. ​“The question is, can we turn that into a more reliable, schedulable resource?”

EnergyHub, a subsidiary of Alarm.com, says its technology enables it to do that. In 2022, the company purchased Packetized Energy, the startup that Hines co-founded with two fellow University of Vermont professors in 2016. In simple terms, Packetized Energy’s software analyzes how much energy every device connected to the system needs to do its work — in the case of thermostats, keeping air conditioning and heating on to deliver set temperatures — over long periods.

That data is then fed into mathematical models that analyze how much each individual customer can be expected to reduce their electricity use during a demand response event. Some homes have poor insulation or older air conditioners that don’t perform as well when it’s time to precool homes, and some people may be less comfortable putting up with multiple hours of hotter-than-usual indoor temperatures, for example. Other homes are better insulated and have more efficient cooling systems, or their residents have previously demonstrated a willingness to sweat it out for longer periods.

EnergyHub’s system divides all those customers into different categories. ​“We use some fairly sophisticated machine learning, differential equations, to build a mathematical model” to predict what’s going to happen to the power consumption of each group when their setpoints change, Hines said.

That’s very different from simply broadcasting a signal to every customer’s thermostat to turn up by several degrees and then waiting to see how many customers actually let those commands stand for hours at a time.

This grouping approach also allows utilities to create a three-hour load-reduction event out of programs that allow utilities to ask their customers to turn over their thermostat controls only for two hours, he noted.

That’s what EnergyHub did with Duke Energy, the sprawling utility with operations in six states. In North Carolina, dynamic load-shaping test runs this spring and summer ​“combined a whole bunch of two-hour events, staggered them in a smart way so we got a three-hour constant load-shed event, and minimized snapback, to get it as close to zero in that final hour,” he said. ​“We fed that into an optimizer that essentially explored an enormous space of what-if scenarios, to figure out what is the optimal combination of 15–20 events to get the load shape you want.”

​“The result was a bunch of different schedules from different devices,” he said.

Below is a chart displaying this, with target load reductions in yellow and actual load reductions in black, followed by a second chart that color-codes the different groups that EnergyHub assembled to do the job. (Note that this chart uses positive numbers to indicate how much load reduction is happening and negative numbers to show greater-than-normal electricity consumption during precooling and postevent snapback.)