Famous supercomputer Watson provides inspiration for solar power predictions

By Jared SagoffSeptember 12, 2013

ARGONNE, Ill. – There’s a saying in Chicago: If you don’t like the weather, wait five minutes. However, not knowing the weather for even five minutes could get very expensive for solar plant operators who rely on up-to-the-minute predictions of how much sunlight their panels will receive.

The goal of the field known as solar power forecasting is to estimate precisely how much energy a solar plant will provide to the grid within the next day. Scientists around the world, including research meteorologist Edwin Campos of the U.S. Department of Energy’s (DOE) Argonne National Laboratory, develop and refine technologies to more accurately predict short-term changes in cloud cover in the area around and directly above the planet.

“Climate and energy merge at the electric grid,” Campos said. “If we want to better integrate the growing amount of renewable energy we’re generating into the power grid, we have to get better at forecasting.”

When a solar plant experiences a generation shortfall, other plants – typically fossil-fuel plants that run on coal and natural gas – need to increase generation in order to make up the difference, or else the grid could experience sporadic outages or even a blackout. These energy sources are relatively less environmentally friendly than solar.

Additionally, grid operators may need to purchase electricity within a short time frame or continually carry additional back-up power in the absence of accurate and timely solar generation forecasts, which can then increase the cost of electricity. These increases are usually passed on to the consumer.

Underestimating the amount of electricity that a solar plant might generate in the following period is almost as much of a problem as overestimating it. Because grid operators need to instantaneously balance electricity production with demand, their operation plan relies as much as possible on accurate power forecasting. Plant managers and grid operators may curtail extra unanticipated electricity produced by a solar plant on a day when it is sunnier than expected, resulting in wasted energy. “The goal is to get our forecast estimates as close as possible to reality,” Campos said.

To improve the accuracy of solar power forecasting, Campos and his colleagues have partnered with IBM to build a forecasting technology based on IBM’s Watson supercomputer, made famous by its 2011 victory over human champions on the television quiz show Jeopardy!. Campos hopes that the information he gains by integrating big data processing, machine learning and cloud modeling into a Watson-like platform will help grid managers and power plant operators develop more efficient strategies for allocating their resources to manage the unevenness of solar generation.

“Even five minutes of clouds when we thought there’d be sun can equate to tens of thousands of dollars of extra costs from having to buy other forms of power from across the grid,” Campos said. “The more accurately we can forecast clouds, the more money we can ultimately save homeowners on their electric bills.”

One of the big challenges that supercomputers like Watson can help Campos address involves what scientists refer to as multi-scalability – the ability to look at interactions that occur within small, medium and large-scale systems.

“We want to be able to teach a machine to think and integrate different atmospheric aspects at scales from a truck size to a continental size, which can enable a forecasting system to become much more dynamic and flexible,” Campos said. “But to do this, we need the ability to generate and handle tremendous quantities of data.” 

According to Campos, the kind of “nowcasting” required by energy planners has proved particularly suitable for multi-scale forecasting, largely because of the complexities of having to account for climatological and meteorological patterns over very different areas and periods. “If there’s even just a wandering cloud or two that we didn’t plan for, it can unfortunately make a big difference,” he said.

The partnership between Argonne and IBM has grown to include other partners from industry (Northrop Grumman and 3TIER) and academia (Northeastern University and the University of Arizona) as well as the National Renewable Energy Laboratory and the National Oceanic and Atmospheric Administration. The team also includes electric grid operators (Tucson Electric Power, Green Mountain Power, ISO New England, and California ISO) and members from the solar power industry (Petra Solar, Juwi Solar, and Prime Solutions).

This research consortium, sponsored by the DOE SunShot initiative, will last for three years and involves a core team of Argonne’s atmospheric scientists, decision modelers and statisticians. Campos pointed to the interaction with industry as a particularly important part of the project, as he claimed that new ideas for the incipient “smart grid” would have to come from a collaboration of academia, industry and government.

“The kinds of interactions we’re researching ultimately impact millions of people,” Campos said. “It’s important for us to figure out how to incorporate renewables onto the grid as best as we possibly can.”

Argonne National Laboratory seeks solutions to pressing national problems in science and technology. The nation's first national laboratory, Argonne conducts leading-edge basic and applied scientific research in virtually every scientific discipline. Argonne researchers work closely with researchers from hundreds of companies, universities, and federal, state and municipal agencies to help them solve their specific problems, advance America's scientific leadership and prepare the nation for a better future. With employees from more than 60 nations, Argonne is managed by UChicago Argonne, LLC for the U.S. Department of Energy's Office of Science.