What happens when you leverage technologies like IBM's artificial intelligence engine Watson for clean power? The answer is the awesomely named Watt-sun project, a machine learning platform that IBM Research has quietly been building over the last year, and which is now highly accurate at predicting how cloud cover, weather and atmosphere (among many other data points) affect the way solar panel systems operate.
IBM's sky cameras installed at several solar farms to observe cloud cover. Image courtesy of IBM Research, Flickr Creative Commons.
Solar forecasting has been around as long as solar panels have been plugged into the grid. But the forecasting systems historically haven't been all that accurate, given that so many factors can contribute to the amount of sunlight that's able to descend from the sky and onto the solar panel and then get converted into electricity.
To build a better system, Watt-sun created a platform that blends dozens of currently available solar forecasting models created by organizations, government bodies and companies throughout the years. Watt-sun then takes this blended forecasting model and adds in tons of data about environmental and atmospheric conditions, about the solar plants and about the surrounding power grid.
The more information Watt-sun sucks in, the smarter it gets. On a couple test sites IBM Research has installed "sky cameras," fish-eye lens cameras that are pinned onto poles or the rooftops of buildings and continuously stream visual data about the atmosphere and cloud density over the solar site. Increasingly IBM is also adding in satellite data, looking down on the solar system from way up in space.
As Watt-sun adds data and tweaks its platform, it learns over time. After a year in service, and having been tested on over a dozen solar sites, it is now 35 percent more accurate than the next-best solar forecasting model, project manager Hendrik Hamann told me in an interview. In another year, Hamann says it will be 50 percent better; in the third year of the project Watt-sun is expected to be 100 percent better than current models.
"I'm confident we can hit these numbers," Hamman said. "It's becoming an expert system."
Watt-sun has been funded by a grant from the Department of Energy's SunShot program. It's currently an academic project, and IBM Research doesn't have any immediate plans to turn Watt-Sun into a product.
IBM solar forecast for solar farm in Smyrna, TN. Image courtesy of IBM Research, Flickr Creative Commons.
But if they did some day, it'd be a pretty valuable one. Solar panels are plugging into the power grid in the U.S at a dramatic rate. In the first quarter of 2014, there were 1.3 GW of solar panels installed in the U.S., which is 79 percent more than were installed in the first quarter of 2013. That made it the second largest quarter for solar panels installed in the U.S. ever, according to a report from the Solar Energy Industries Association.
As utilities and power companies install more solar power farms -- and also face a swell of customers installing solar panels on rooftops -- they'll need to get better at managing the solar sites and integrating them into the grid. Unlike coal, natural gas or nuclear plants, solar sites turn on and off at various times -- at night, for instance, or on a cloudy day. That makes them a new "variable" asset that utilities have to deal with, and the reality is that many utilities are very worried about having to manage this new type of variable generation.
Better forecasting -- at least 24 hours ahead from Watt-sun -- can help immensely with figuring out how to balance the grid.
While this might be the first time you've heard of Watt-sun, it's actually got an impressive list of partners. Watt-sun has been used across all of California's independent grid system operator's footprint (CAISO) -- that's 2 GW of generation. Vermont local retailer Green Mountain Power is also using it, as are Sacramento utility SMUD and Tucson, Arizona utility TEP. There's another version of the system being developed at Argonne Labs.
Earlier this week the U.S. Environmental Protection Agency proposed that power plants need to reduce their carbon emissions by 30 percent by 2030. That means that many states will be building out more solar panel plants around the country (in addition to closing coal plants), and guess what -- that will just mean better forecasting will be needed. Prediction: look for widespread use of Watt-sun some day. Machine learning and solar data are a perfect fit.