
This project’s purpose was to address the growing need for better short-term solar energy planning by providing clear and reliable predictions. We developed a machine learning model to predict solar coverage of a given area for the next 5 to 10 minutes. The model was a convolutional neural network that took an image of overhead weather and solar radiation at the same time as inputs. Based on the input data, the model learned what features in the image related to varying solar radiation levels and then output a prediction of future solar radiation.
The model studied different types of weather data and adjusted to specific locations, making the predictions useful in various settings. We focused our design on using open-source tools to keep costs low and allow for easy updates or changes. Key project goals included making predictions in less than 10 minutes and reaching at least 70% accuracy.
The project’s results helped solar energy systems work better by improving how energy was used and stored. It also helped keep power grids stable and reliable. This system gave energy managers helpful information that supported smarter decisions about solar power.
Timothy Burman, Anthony Fiandaca, Christopher Guerrero, Judson Ivy, Kim Le
Victor DeBrunner, Ph.D., Stephen Cross, Paul Hynes
Florida Power and Light
Spring