Artificial intelligence solutions for the solar industry have been gathering steam. Today, developers have access to everything from predictive modeling of solar panel production to panel optimization, and from predictive maintenance to energy management.
More recently, however, there’s been an increasing emphasis on using AI to optimize the demand side of the industry. Over the last few years, Solstice, a community solar company owned by Japanese developer Mitsui Group, has been quietly developing machine learning models to target two of the industry’s most important metrics: collection rates and churn (or the measure of participants leaving a program).
Steph Speirs, co-founder and CEO of the company, said that using AI to improve the customer experience has the potential to have a major impact on project revenue.
“The two metrics developers care most about [are] the subscription rate on their project and their collections rate, because that affects the financial viability of a project,” Speirs said. She added that that financial viability also directly informs a customer’s experience with community solar broadly — a canceled project is, of course, a negative outcome for that customer as well.
Expanding community solar’s reach
Solstice initially began experimenting with machine learning when the company was looking for an alternative to FICO credit scores to qualify customers.
“We were racking our brains with how to find a better way, because increasing the customer pool for community solar projects is good for the whole industry,” Speirs said.
Solstice’s polling showed that over half of the developers the company worked with had concerns about bringing low-income households into projects, due to the perception that they wouldn’t pay their bills or would churn at higher rates, Speirs said. The company’s foray into machine learning may ultimately help Solstice use data to alleviate those concerns, she added.
Solstice’s flagship machine learning product is EnergyScore, an alternative credit qualification that uses account level credit profile information and monthly utility payment history data to predict the likelihood of someone failing to make a payment on their solar installation.
The model is trained on data from over 800,000 utility payment performances and 5,000 demographic variables, and according to Solstice, is more accurate than traditional credit metrics.
“With a product like EnergyScore, we can more effectively and equitably qualify individuals, particularly for products like renewable energy,” said Solstice’s chief data scientist Jake Ford.
What Ford found in developing and applying the model is that the perception of risk around low-income customers is greater than the actual risk that those households would default on payments or quit projects. Relative to the 680 FICO cutoff traditionally applied for solar loans, applying EnergyScore decreased the average default rate by 1.9%, and increased the number of low- and medium- income customers approved for projects by 14%, Ford said.
More modeling
More recently, Ford has been working on Solstice’s churn model. It uses internal Solstice data on elements including tenure, allocation and credit data, utility attributes, and customer location to rate the likelihood of a current customer’s churn rate on a scale of zero to 100.
The model is built on a common classification algorithm called Extreme Gradient Booster, which essentially builds a predictive model by enhancing traditional decision trees, including by sequentially adding trees to correct errors made by previous trees.
Solstice initially tested that model over a period of three months last winter, identifying customers that the model rated as most likely to churn off of a project, and developing a series of “interventions” by the customer success team. The hypothesis was that reengagement could help decrease churn rates, Ford said.
Compared to the same ranking of customers over the same three-month period in 2022, Solstice found a “drastic reduction” in churn rates, he added. That first pilot had rates that were four times lower among the highest risk users in 2023.
Ford said that the team is currently replicating that pilot on a much larger scale, but the initial test revealed a few key data points for the broader industry. First, tenure on a project was one of the strongest predictive variables, with longer tenured customers being the least likely to churn off of a project. Billing type, i.e. whether a customer receives one bill from their utility that reflects a community solar credit or receives dual bills, was also important — perhaps unsurprisingly, customers who receive a single bill were more likely to stay on a project.
The weakest indicators of churn rate were primarily demographic — homeownership status has no correlation to churn rate, nor does income, Ford said.
Of course, challenges remain as Ford and his team look to expand their models.
First, there’s the ever-present data challenge: getting the right data from utility companies on a regular basis is harder in some geographies than others, Ford said. There’s also the fact that different markets operate differently, as do consumers in those markets, so scaling isn’t necessarily a linear process.
The initial pilot applied the churn algorithm across Solstice’s entire customer base in each state the company operates in, but things get more complicated as the company tries to zero in on more specific data, such as churn rates by state.
“You have to be flexible, you have to take into account who you trained on,” Ford said.
Moving forward, Solstice’s focus is on applying EnergyScore and the churn model more widely out in the field, and eventually finding partners and opportunities for application outside the company. That will, in part, involve spreading the word about Solstice’s machine learning innovations, and specifically about how they are leveraging AI differently from others in the industry to make a tangible impact on project profitability, Speirs said,
“It helps us increase the amount of education and communication that certain customers are getting,” she added. “I think it makes the customer experience better and makes the projects more durable.”
