Solar power is unpredictable
Balancing power production and consumption is key when working with solar power. If production exceeds demand, energy storage becomes an issue. If production falls short, consumers face black-outs. As the price of capturing solar energy goes down, predicting solar irradiance has become key to increasing efficiency of power grids and to minimize the use of fossil fuels on spinning reserves.
SunCast uses satellite image data from NOAA, collected by the GOES satellite system and made available every hour for the continental United States. This data is coming to us in NetCDF format and is already enriched with Downward Shortware Radiation information by NOAA.
We built our solution on two fundamental deep learning technologies: convolutional neural networks (CNNs) and Long Short Term Memory (LSTM) neural networks. Combined CNN-LSTM architectures are specifically designed for sequence prediction problems with spatial inputs, like images or videos.
We provide fast APIs to access our predictions for any lat/long in the United States. Our API returns shortwave radiation predictions for 3 hours, along with last 3 hours of actual data. An example GET call and JSON response can be seen here.
Request for San Francisco: