Solar Power Forecasting Initiative

The Solar Power Forecasting Initiative is a large scale, multiple-campus project envisioned and directed by Prof. Carlos F. M. Coimbra of the UC Merced School of Engineering. For more details, click here.

The UC Merced Solar Forecasting Laboratory is developing a network of solar instruments to monitor, map and forecast the solar resource in the state of California with unprecedented resolution and accuracy. The data from a few strategically placed high-grade solar ground stations will be used for the continuing development of a self-learning, stochastic model that UC Merced has created for the accurate forecasting of the solar resource. The model combines remote sensing, ground irradiance measurements, and weather data from hundreds of National Weather Service (NWS) meteorological stations with the objective of providing solar irradiance forecasting capability for the entire state of CA at very high temporal and spatial resolutions.

This project will place California in a strategically important position in terms of solar technology development since the solar resource will be monitored, mapped and forecast in greater detail than any other region in the world, which in turn will enable much higher PV market penetration than is allowed today.  This project aims to overcome the problems associated current grid stability constraints, and the high costs of uncertainty in power ratings given by traditional forecast models for intermittent sources such as solar and wind.

Left: Detail of some of the instruments at the UC Merced Solar Observatory; Center and Right: Solar Maps for Direct Normal Irradiance using remote sensing data acquired and modeled by Prof. Guo’s research group at UC Merced. DNI maps generated for 9:15am and 10:45am, June 28, 2009.

The ultimate goal of this project is to bring down the cost of Solar Power generation by providing a direct and superior service to the power producers, grid regulators and utility companies. The status quo is the following: independent power producers that rely on intermittent sources such as wind and solar are responsible for providing meteorological data from their locations to CAISO, which in turn subcontracts a forecasting service to rate the capacity of each power producers over the next 2 hours in intervals of 10 minutes each. The process is non-optimal because the data provided by the independent producers is inaccurate, sparse and contaminated by several layers of experimental uncertainty. The end result is that, if the power rating is not met by the producer, the producer has to buy the difference between the power generated and the power forecasted at the market value, which is much higher than the cost of the power produced by solar or wind. Because the forecasting is inaccurate, the cost of producing power through intermittent sources like solar and wind is driven upwards, and the higher cost of not knowing the resource well trickles down to the consumer. The end result of the higher costs is that solar and wind power producing farms are less widespread than they should be if the wind and solar resources were well monitored, mapped and forecast in real time. The research activities in our group at UC Merced address this very real and practical problem by providing the developing highly advanced self-learning models for monitoring, modeling and forecasting of these important renewable resources in an effort to bring down the price of uncertainty in the generation process for smart power grid implementation.

Scatter plot of Actual GHI versus Estimated GHI for data taken at the UC Merced Solar Observatory. Left: Training set Verification; Right: Model applied to portion of the data set not used for ANN training. (RMSE of 6% for 24-hour forecast, and less than 3% for the 1-hour forecast).

Our approach in this project is to develop major modeling infrastructure by building on improvements in both hardware and software. Our methodology will enable the accurate monitoring, modeling and forecasting of the solar irradiance resource at much higher levels of spatial and temporal resolutions than is available today. In order to develop the more accurate stochastic models that are needed for detailed short (< 5 minutes) and long term (24-48 hour) forecasting, a much higher density of ground sensors for verification and benchmarking is needed, and this proposal addresses those needs by designing new hardware synergistically with the model development. The resulting system is cost-effective and capable of high-fidelity modeling for integration with remote sensing data analysis.
 

Solar Power Forecasting research is currently funded by the Center for Information Technology Research in the Interest of Society (CITRIS).

Funding is also provided by the National Science Foundation (NSF ASSIST), and the California Energy Commission under a PIER RESCO project.