Wind Power

How Xcel Saved $22 Million with Weather and Wind Forecasting

We recently covered the launch of IBM’s weather + wind forecasting and modeling system, dubbed HyRef, touting the ability to predict incoming weather patterns and calculate wind turbine performance from 15-minute intervals up to 30 days in advance, with 90+ percent accuracy, and with a stamp of approval from customer Chinese State Grid Corp.

But to be fair, IBM’s HyRef isn’t the first effort at putting together weather forecasting and wind data analysis. One of these efforts, a partnership in the U.S. between the National Center for Atmospheric Research (NCAR) and Xcel Energy, has been fully operational since 2009 and has saved Xcel tens of millions of dollars.

Sue Ellen Haupt, director of weather systems and assessment programs in NCAR’s Research Applications Laboratory in Boulder, Colorado, offered more insights into this work. NCAR’s roots in this technology, dubbed “Variational Doppler Radar Assimilation System” (VDRAS), go back into the 1990s, extending what has been used for “nowcasting” weather at U.S. Army test ranges to the past two summer Olympic Games. (More details were published in an IEEE journal last fall.)

Xcel’s service territory (as of early August) covers 107 wind farms totaling 3,746 turbines and a total capacity of roughly 5.4 GW. A February 2013 presentation at a workshop of the Utility Variable Generation Integration group discloses Xcel’s use of variable generation forecast (NCAR calls it the “Wind Power Forecast System”) for reserve planning, forecasting (real-time, hour-ahead, day-ahead) and ramping, and planning for its power commitments and trading. Calculating a forecasted mean absolute error (MAE), i.e. variation over time in a plant’s performance (installed capacity vs. power production), Xcel determined that from 2009-2012 it saw anywhere from 17-38 percent improvement across its service territories, translating to nearly $22 million in total savings. Of course forecast accuracy depends greatly on location and local conditions, from terrain to atmospheric phenomena, and forecasting accuracy and precision differs greatly further out in time, i.e. a 6-hour forecast will have a lower error than a forecast looking two days out.

To that end, NCAR and Xcel are now enhancing the system to include more custom forecasting for several specific Xcel sites in Colorado, Minnesota, and Texas. Part of that will be to predict potentially damaging icing conditions and how that affects power levels; forecasting energy load; and enhancing ramp forecasting. They also are adding probabilistic predictions on estimates for each time period to increase the confidence level. “Because the atmosphere is inherently chaotic, one can never have an exact prediction,” she explains, so this is expressed in terms of uncertainty, i.e. error bars — anywhere from 20-80 percent, whatever is requested by a customer as a design criteria. Note that this is different than the error of mean prediction which is dependent upon location and operating factors, and which is the 92 percent statistic for the Zhangbei customer that IBM is talking about.

Lead image: Percent shape clouds, via Shutterstock