Advertisement

Machine 'learners' compute cloud cover to balance power supplies

Machine 'learners' compute cloud cover to balance power supplies
Solar panels on the roof are connected to IBM's solar forecasting system. (Jon Simon / IBM)

Hendrik Hamann is into cloud computing — as in real clouds, those puffy things in the sky.

Working at IBM alongside some of the computer giant's most advanced systems, Hamann and his team seek a breakthrough in cloud-cover forecasting. They're aiming to help ease the introduction of solar electricity into the nation's major power grids, as solar-generated power is increasingly being loaded onto the grid, propelled by government mandates and solar-technology price declines.

Advertisement

There's a big problem with solar power that the IBM team is trying to solve: You can't pump out much electricity on a cloudy day.

But the demand for electricity doesn't decline when a solar plant can't produce energy. Another source of power has to take its place. So utilities keep coal and natural-gas power plants humming, ready to pitch in when solar can't do its job. That burns more fuel and costs more money.

Utilities would like to know in advance just how much sun will shine on a solar plant at any given time. Then they could plan their power loads more efficiently, saving money and fuel while cutting back on gas emissions that contribute to global warming.

Several forecasting companies already sell services to the renewable energy industry. But the technology will need to advance rapidly, Hamann said.

"As more solar power comes on line, the more accurate you have to be," because the impact of cloudy days will broaden, he said.

That's especially true in California, where the three biggest utilities are under state mandate to make renewable sources 33% of their energy supply mix by 2020, up from the current 20%. Gov. Jerry Brown has proposed increasing that amount to 50% by 2030.

"There has to be a radical transition from the 100-year-old grid technology to the grid of the future," said energy consultant Justin Sharp. "It will involve much deeper analytics to provide real-time optimization of supply and demand."

Analytics is just a part of what IBM's advanced equipment does. Hamann works in a field called machine learning, a subset of artificial intelligence. Called "learners," the computers learn, on their own, how to predict cloud-cover patterns. IBM claims that the machine learners' predictions already are 30% more accurate than National Weather Service forecasts. That's impressive, but not enough to make a big difference in the real world, where other companies claim to be just as accurate with their own technologies. But Hamann is shooting for more than 50%, when the technology could start to be a game-changer.

IBM is not the only company that senses opportunity.

AutoGrid Systems in Redwood Shores, Calif., sells software to forecast energy demand, minutes or seconds in advance, by analyzing consumption patterns. It creates a "living model" of the power grid based on physics and human behavior. The team runs various scenarios through its algorithms, asking questions such as, What if the temperature spikes? What if power rates go up?

"If the penetration of solar or wind gets very high, then you need something to compensate for the fact that power output is changing very quickly," said product manager Basile Bouquet. "The key concept for utilities in the future is to be able to manage flexibility."

Clean Power Research, based in Napa, Calif., provides forecasts using satellite data. Other companies in the field include General Electric and Finland-based Vaisala — both are working in machine intelligence.

But the most advanced approach comes from IBM, whose research is part of the U.S. Department of Energy's SunShot Initiative.

In simple terms: IBM scientists feed a machine learner enormous volumes of observational weather and forecast data. The machine learner looks for statistical correlations between the observations and the forecasts, designing a prediction program based on the methods it found to work best. The machines can keep learning as observations become more detailed and predictions become more accurate.

Advertisement

Hamann feels good about where the project is going.

"I'm certainly very confident that we'll commercialize this technology," he said.

Twitter: @dainabethcita

Advertisement
Advertisement