Using Census Data to Predict Solar Panel Deployment

New renewable energy sources require improvements to the current electric grid. The recent surge in the number of intermittent energy generation facilities, such as solar and wind farms, has resulted in a need for improved monitoring and control methods for the electric grid due to increased supply-side uncertainty. Mitigating the uncertainty in the amount of electricity privately produced would greatly increase power generation efficiency, resulting in less waste of fossil-fuel generated electricity. One major component of supply-side uncertainty comes from residential solar panel installations. Today, installing solar panels on residential homes is easy and affordable, and will only become easier and more affordable as time progresses. As a result, it is difficult to know how many solar panels exist and supply power to the grid. If energy companies had more insight into this piece of the supply-side puzzle, they could better model an area’s energy production and balance power plant production accordingly, resulting in lower energy costs and less environmental impact. For this project, we implemented and optimized an artificial neural network (NN) and a support vector regression (SVR) algorithm to predict the number of solar installations in a given tract from census data. The input to the model consists of geographical and demographical characteristics, such as land area, average household income, climate data, number of residents of age 30-39, etc. The model takes these census data and then outputs the number of solar systems in a given tract. The model is trained using supervised learning on a labeled dataset. We also used the models and principal-component-analysis (PCA) to determine which features have the most influence on modeling the number of solar systems (i.e. which features are most strongly correlated with solar deployment density).