How advanced forecasting is making it easier to integrate solar onto the grid
Utilities and grid operators back research to make solar more predictable.
Rapid variations in solar output caused by shifting clouds will no longer keep utilities and system operators from moving to higher grid penetrations of solar power.
The independent system operators (ISOs) for New York, California, and New England and utilities from Hawaii to Vermont are already working with public-private research efforts that show the potential to make day-ahead and 5 minute solar forecasting 25% to 35% more accurate.
That kind of forecasting, researchers say, will make solar a better deal for ratepayers and give utilities and grid operators new ways to improve reliability.
One research group’s $4.1 million grant from the Department of Energy’s SunShot High Solar Penetration initiative, with $2.1 million from private sector partners, will fund refinements to the widely used Weather Research and Forecasting (WRF) model, according to Dr. Sue Ellen Haupt, head of the University Corporation for Atmospheric Research (UCAR)-led project.
“We have been forecasting for wind for some time but only recently has solar penetration been high enough for utilities to call for better solar forecasting,” Haupt explained. “We are able to go into the guts of the WRF model and improve the specific algorithm that calculates irradiance.”
The UCAR researchers, Haupt said, are working to improve the range and detail of variables in NOAA’s WRF calculations for global horizontal irradiance (GHI), direct normal irradiance (DNI), and diffuse irradiance (DIS).
The project is directed at forecasting utility scale solar power but the irradiance calculations can be used for distributed solar power, she added.
“We asked what factors in today’s WRF need to be improved to get irradiance right,” Haupt explained. They are focusing on algorithms that will make WRF modelling of clouds, aerosols, convection, cloud physics, and a few other technical factors more accurate.
Even on clear days, convection — the upward motion caused by surface heating — causes less dense warm air to rise, cool, and condense. Water vapor particles then condense into clouds that block the sun in the afternoon and reduce the amount of solar energy generated electricity produced.
The UCAR group can fine tune the models it derives from understanding these dynamics for any location’s specific characteristics, Haupt said, “because the physics of the atmosphere works the same anywhere.”
A 25% better prediction
An “aggressive ballpark estimate” by the UCAR team suggested they might improve the prediction of the GHI by 25% (in terms of root mean square error), Haupt said.
Such forecasting would allow grid operators more certainty about what the weather means for today’s peak solar output or tomorrow’s overall solar production.
Utilities are looking at renewables as an economic solution, Haupt said. But the metric for how much UCAR’s improved solar forecasting will save can only come after the group’s models have produced a year of data, sometime in early 2016. ”Utilities will have to calculate their energy production costs both with and without forecasts and estimate what the difference is,” Haupt said.
Xcel Energy, she added, calculated that UCAR’s 37.1% improved wind forecasting saved the utility’s customers $36.8 million between 2009 and 2013 in all their service territories and reduced CO2 emissions by 238,000 tons.
“When Xcel thought its wind capacity was affecting system operations, they decided to fund wind forecasting advances,” Haupt said. “They have reached that point with solar.”
UCAR’s A Public-Private-Academic Partnership to Advance Solar Power Forecasting includes Long Island Power and Light, Public Service of Colorado, Sacramento Municipal Utility District, Hawaiian Electric System, Xcel Energy, Southern California Edison, New York Power Authority, and the New York and California ISOs.
IBM’s supermodel and a 30% better forecast
The IBM-led Watt-Sun: A Multi-Scale, Multi-Model, Machine-Learning Solar Forecasting Technology is working with two utilities, Green Mountain Power and Tucson Electric Power, and two grid operators, ISO-New England and the California ISO, explained IBM group lead Dr. Hendrik Hamann.
Both SunShot projects are advancing forecasting, he said. Various WRF models have strengths and weaknesses and UCAR’s algorithms will improve them. “But there is always a trade-off between complexity and computational efficiency because the physics are so complicated you cannot model these things to a very high resolution,” Hamann said.
Instead of refining individual WRF models, the IBM team uses them as “expert systems” in conjunction with big data fed into machine learning technologies. They identify which models work best under specific circumstances and give that model “a high weight” in forecasts for those or similar circumstances.
The result is a forecast from “a multi-model, machine learned blend, a supermodel learned from all existing models,” Hamann said. “We are building a platform with which to better use existing weather prediction models.”
Over several months at fourteen U.S. locations, Hamann said, “the supermodel has proven it improves existing forecasts by 30% to 35%, to the root mean square error.”
IBM’s utility partners and grid operators are already evaluating the forecasts and providing real world data feedback, Hamann said.
ISO New England is using forecasts to better understand how solar modulates load. The California ISO is using IBM’s day-ahead PV power forecasts across its statewide region. Green Mountain Power is using data on 1,200 Vermont PV installations to understand the impacts of forecasting on distributed solar.
“Utilities are very concerned with ramps,” Hamann said. “Forecasting critical ramp events will test the supermodel’s ability. That is an important part of the feedback we are trying to get from our utility and ISO partners.”