In 2017 alone, close to 50,000 households installed Tesla solar panels. BUT, around 40,000 people went through a partial assessment but decided that solar panels do not make sense. As a result, the sunken cost of conducting expensive assessments is a huge burden for companies like Tesla. So, can we estimate who will be eligible for solar panels based on basic data collected in the very early stage of the process?
What it does
The machine learning models we tested compare datasets with and without attributes that are difficult to collect. The aim of the project is to assess the necessity of conducting expensive onsite assessments, and to understand which basic (easy to collect) attributes contribute to decisions to install solar panels. We also estimate avoided costs of only using basic attributes to target customers.
How we built it
We took into account census, Zillow, solar insolation, as well as installed and cancelled solar installation projects data from Tesla. Using these data, we trained and tested machine learning models such as LASSO, Random Forest, and AdaBoost; identified significant attributes for solar panel installation; and computed their accuracy, false positive/negative rates, and finally avoided costs of $399/potential customer!
Accomplishments that we're proud of
Tesla could have saved $32 million in 2017 from avoided onsite assessments!
What's next for Sunthing Random
We aim to improve the granularity of the assessment based on more identified attributes, such as proportion of neighbors with solar homes, demographics, and socioeconomic factors. Moreover, self-assessment tools for potential customers are easily built to assess technical viability, and cost & financing options.