Utilities face a complex challenge: meeting growing energy demand while controlling infrastructure costs. To solve this, they need to identify which tools they can use to manage energy consumption without disrupting customers’ daily lives. Many of these tools exist at the grid edge—the connection point between individual customers and the energy system—and include smaller-scale solutions beyond traditional large power plants. 

As utilities shift toward managing energy at the grid edge, they need to understand what distributed energy resources (DERs) are available to meet current energy demands. This includes both devices already installed in households and businesses and those likely to be added soon.

Two types of models provide this insight:

• Detection models identify devices that customers already have
• Propensity models predict what devices customers are likely to install

Utilities can use information from these models to anticipate potential risks to capacity, estimate potential for what is next, make decisions on how to tailor programs to individual needs, identify eligible customers for particular programs, and monitor utility goals such as electrification. Without visibility into existing DERs  or reliable forecasts for what is coming, utilities cannot effectively plan and make decisions for what is needed.  

In the first part of the Uplight series called, “Meet the Models,” I’ll uncover what is behind our detection and propensity models. In the next weeks, my team will post about our disaggregation and bill compare models, measurement, forecasting, and dispatch/optimization. 

How Detection and Propensity Models Work

Detection and propensity models both work by tying known positive cases, for example a known electric vehicle owner or a known smart thermostat purchaser, to other information we have about the user such as demographic data, home profile data or consumption data.  Depending on the availability and complexity of the data, machine learning models or heuristics are employed to expose patterns in the data of known positive cases. These patterns can then be applied to situations where the case is not known. For example, by looking at consumption patterns, a machine learning model can give a probability of electric vehicle ownership. Or, based on demographic and home profile data, models can determine the probability of a future smart thermostat purchase. 

Uplight’s detection models use energy consumption data to identify devices in customers’ homes, such as  electric vehicles, electric hot water heaters or heating fuel types.  The models analyze either hourly Advanced Meter Infrastructure (AMI) data, monthly billing data, or both, depending on the specific application.

Patterns in consumption data are too numerous and complex for a human to evaluate for every household. Instead, Uplight employs AI in the form of machine learning models to identify these patterns and determine which devices are likely consuming energy in each home. These detection models are non-invasive, relying solely on data we already collect, and requiring no additional data collection devices.

Propensity models utilize household data which includes information about both the occupants and physical environment characteristics to predict the likelihood of joining utility programs such as demand response, enrolling in a time of use (TOU) rate or purchasing a smart, controllable device. 

• Life circumstances such as income, marital/parental status and level of education may all influence a household’s ability to purchase certain devices or enroll into various programs. 
• The physical properties of a location such as the presence of central heat and air certainly contribute to the decision of purchasing a smart thermostat or to finding benefit from a TOU program.  

Value of Uplight’s Detection and Propensity Models 

A Deep Dive into Electric Vehicles

Transportation electrification of transportation can dramatically reduce carbon emissions, and utilities are eager to provide that energy. However, poorly timed charging can overload transformers, cause voltage instability, and force utilities to fire up expensive, high-emission peaker plants during peak demand periods. 

Unlike most electrical loads, EV charging is inherently flexible since most drivers don’t charge daily and can typically delay charging by several hours without affecting their daily routines—providing a powerful tool for load management.

Until grid infrastructure can reliably handle widespread EV charging without strain, how do we minimize system burden and emissions while still advancing toward electrification?

The first step is identifying where EVs are, specifically Level 2 chargers where load is high.  Right now, most utilities don’t necessarily know which households have EVs. Many utilities use state databases to understand EV adoption and identify EV drivers. However, few states have comprehensive datasets, and data that is available changes rapidly. In the United States, new electric car registrations totalled 1.4 million in 2023, increasing by more than 40% compared to 2022.  Electrification may be happening faster than the grid is ready for, making knowing where EVs are emerging all the more important.

Enter: Uplight’s EV Detection Model. 

Uplight’s EV detection model uses AMI data to predict whether a household has an electric vehicle. More specifically, it predicts that a household owns an electric vehicle Level 2 charger, and can be used for residential customers that have AMI data. The model uses data from known EV users to identify charging patterns in households that have an EV but have not informed their utility. 

The combination of electric vehicle detection and propensity models can assist utilities in grid planning, highlighting potential areas of grid strain as well as possible resources for capacity. The propensity model can also identify end users who are interested in electrifying their transportation but may need additional encouragement and incentives to do so.  

Additional Detection and Propensity Models

Though electric vehicles are certainly at the forefront of everyone’s mind, electric hot water heaters pose similar risks and opportunities. Encouraging electrification to capitalize on renewable energy is important, but managing that additional strain on the grid is perhaps even more  important. Identifying where electric hot water heaters are can help utilities enroll customers in managed programs or encourage electrification where they are not detected. Similar benefits come from knowing the heating fuel type of the home as well as knowing what devices are currently present to predict where new devices will show up in the future. 

The output of these models can be used to personalize and improve recommended actions for customers. For example, knowing the heating fuel type of a location allows for personalized tips on maintaining the system that result in energy savings. We can more effectively break down usage by type of usage and highlight the biggest consumers. 

DER adoption is growing, and this presents a big opportunity for utilities to leverage these devices to meet growing demand. But in order to control devices at the grid edge, customers need  to enroll, and identifying those likely to enroll is where propensity models become valuable. For example, targeting probable smart thermostat with information about the benefits of demand response programs, or knowing which households would be willing to sign up for TOU and demand response programs helps utilities send directed, personalized content maximizing campaign efforts without overloading individuals with irrelevant content.