Techniques for Improving Wind to Power Conversion Gerry Wiener Sue Ellen Haupt Bill Myers Seth Linden Julia Pearson Laura Imbler National Center for Atmospheric Research P.O. Box 3000 Boulder, CO 80307-3000 gerry@ucar.edu haupt@ucar.edu myers@ucar.edu linden@ucar.edu jpearson@ucar.edu imbler@ucar.edu ABSTRACT In forecasting wind farm power output, it is important to obtain an accurate farm power output estimate based on given forecast winds. Generally, the manufacturer's turbine power curves are applied to obtain this estimate. In this paper we will discuss the errors that result from using the manufacturer s power curves at actual wind farms. We will discuss alternative approaches that statistically model the power output based on incorporating air density data with actual wind and power observations at wind farms. We will show how these alternative approaches can reduce the overall conversion error and can thus be superior to using the manufacturer s power curves. INDEX TERMS - wind power, power curve, wind turbine, data mining, wind forecasting, power forecasting 1.0 INTRODUCTION Starting in 2009, the National Center for National Research (NCAR) has been working together with Xcel Energy on the development and implementation of a wind/power forecasting system. This system covers wind farms in the Xcel Energy domain that includes Colorado, New Mexico, Texas and Minnesota. The implemented system makes hourly wind and power forecasts for all the wind farms in the Xcel domain out to 7 days. Short term forecasts out to 3 hours are made every 15 minutes. The system incorporates 4 major pieces: the actual real-time wind speed and power observations from the wind farms; a set of meteorological numerical models including standard National Weather Service models such as the Real-Time Four-Dimensional Data Assimilation Weather Research and Forecast model, the Global Forecast System model, the Rapid Update Cycle model, the Global Environmental Multiscale model and others; a dynamic integrated wind forecasting system that integrates the model forecasts based on underlying skill; and a wind to power conversion module. In implementing this system we discovered that the observed wind to power conversion for the turbines at all the various wind farms can deviate significantly from their industrial power curves. As a result, as part of the implementation, NCAR developed statistical methods for performing the power conversion for all the various turbines in the Xcel domain where observed wind and power data were available. 2.0 MANUFACTURER'S POWER CURVE ERRORS The plot in Fig. 1 illustrates the power curve for a common turbine in the Xcel domain. As can be seen in this figure, there is a unique power for every wind speed and the power 1
cuts out at 25 m/s in order to protect the turbine. Actual observations tell a different story as depicted in Figure 2. Here it can be seen that individual wind speeds actually lead to a wide distribution of powers. The percentiles in the plot in Fig. 2 were determined by gathering turbine wind and power observations for GE 1.5 SLE turbines for a period of approximately one year. The observations were then binned into 0.1 m/s bins and distributions for each bin were formulated and percentiles calculated. The percentiles were then plotted. It is interesting to see that a 10 m/s for the GE 1.5 SLE turbine maps to a 1.2 MW power in Fig. 1. In Fig. 2 the same wind speed maps anywhere from 1.1 MW to close to 1.5 MW, approximately a 25% range in maximum turbine capacity. Another interesting phenomenon to note in Fig. 2 is the shape of the fifth percentile curve which is clearly anomalous. Such power output could be due to a number of factors both non-meteorological and meteorological. In general, power can be curtailed at wind farms owing to market or transmission line conditions and at such times, the generated power will be less than the potential power at a given wind speed. Anomalous power can also be due to meteorological conditions such as snow and ice building up on the turbine blades leading to power loss. In developing a wind to power conversion model that outputs accurate potential power, it is important to filter out anomalous wind/power pairs that are associated with curtailment, turbine malfunctioning, or unusual meteorological conditions such as icing/heavy snow. In the work discussed below, the wind and power input data are pre-filtered by restricting the power of the wind/power pairs to be in the interquartile range between the 25 th and 75 th percentile powers. There is, however, a good rationale for using a larger interpercentile range such as the 25 th to 95 th interpercentile range since curtailments, icing, heavy snow, etc. impact the lower end of the power distribution and it is more rare to see anomalies in the high end of the power distribution (again refer to Fig. 2). 3.0 STATISTICALLY MODELED POWER CURVES The GE 1.5 SLE observed wind/power plot in Fig. 2 does not account for other meteorological variables such as air density, and it is known that power production increases linearly with air density. Air density decreases at higher elevations such as those in Colorado. It also decreases as temperature increases. Thus in statistically modeling power conversion it makes sense to incorporate air density and/or meteorological variables associated with air density. There are four different power curve statistical models presented in this paper: 1. Model based on wind 2. Model based on wind and temperature 3. Model based on wind, temperature, air pressure, dew point and a derived air density 4. Model based on based on past wind, past power and current wind The model based on wind uses a training set consisting of 15 minute averaged wind and power observations gathered for all turbines having the same turbine type at a particular wind farm. The model based on wind and temperature is similar to the previous model but augments it by adding the average temperature obtained from the Real-Time Mesoscale Analysis (RTMA) model. The model based on wind, temperature, air pressure, dew point and derived air density is similar to the previous model but augments it by adding other meteorological variables from the RTMA model. Finally, the model based on past wind, past power and current wind uses previous observations of wind and power and the current observation of wind in order to estimate the current observed power. Different data mining techniques were explored in order to construct the above models. The Cubist regression tree model developed by Ross Quinlan at Rulequest (www.rulequest.com) was subsequently chosen owing to its simplicity of use and good performance. 4.0 MODEL AND POWER CURVE PERFORMANCE In order to evaluate the performance of the models discussed in the previous section, approximately 1.75 years worth of data were gathered from a wind farm in Minnesota consisting of approximately 100 GE 1.5 SLE turbines. The wind and power observations were averaged over one minute time intervals and then were filtered using an interquartile range filter. The one minute filtered wind and power observations were then averaged over 15 minute time intervals and the data mining models described above were then applied. The first two thirds of the data set were used for training of the regression tree model and the last one third was used for testing. The RTMA data used for the additional meteorological variables are generated hourly so these data were matched with the nearest observed wind and power data from the wind farm. The first two thirds of the resulting data set were then used for training and the last one third was used for testing. The mean absolute error (MAE) results are as follows rounded to the nearest kilowatt: 1. Model based on wind: 2
a. Training set MAE - 28 kilowatts b. Test set MAE - 22 kilowatts 2. Model based on wind and temperature: a. Training set MAE - 22 kilowatts b. Test set MAE - 16 kilowatts 3. Model based on wind, temperature, air pressure, dew point and a derived air density: a. Training set MAE - 15 kilowatts b. Test set MAE - 12 kilowatts 4. Model based on past wind, past power and current wind: a. Training set MAE - 10 kilowatts b. Test set MAE - 10 kilowatts 5. GE 1.5 SLE industrial power curve: a. Larger test set MAE - 48 kilowatts b. Smaller test set MAE associated with the hourly RTMA data - 43 kilowatts Cole Boulevard, Golden, Colorado 80401-3393. Subcontract Number: AFW-0-99427-01 Note that the model based on past wind, past power and current wind had the lowest overall training and test set errors. The addition of other meteorological variables improved results over statistically modeling wind to power. Utilizing more meteorological variables with observed turbine wind speeds in the statistical modeling resulted in improved MAE over simply using temperature and turbine wind speeds to model power. 5.0 SUMMARY The results presented in this paper illustrate that statistical models can outperform the standard industrial power curve when applied to wind and power observations that have been quality controlled to remove anomalies. In practice such statistical models can be used to reduce overall error and produce a better power forecast. 6.0 REFERENCES William P. Mahoney, Keith Parks, Gerry Wiener, Yubao Liu, Bill Myers, Juanzhen Sun, Luca Delle Monache, Thomas Hopson, David Johnson and Sue Ellen Haupt. A Wind Power Forecasting System to Optimize Grid Integration. Submitted to IEEE Transactions on Sustainable Energy. TSTE-00160-2011 William Myers, Gerry Wiener, Seth Linden, and Sue Ellen Haupt; A Consensus Forecasting Approach for Improved Turbine Hub Height Wind Speed Predictions; American Wind Energy Association (AWEA) 2011 Keith Parks, Yih-huei Wan, Gerry Wiener, and Yubao Liu. Wind Energy Forecasting A Collaboration of the National Center for Atmospheric Research (NCAR) and Xcel Energy. Prepared for: National Renewable Energy Laboratory, 1617 3
Fig. 1: An ideal power curve for the GE 1.5 SLE turbine 4
Fig. 2: A power curve formulated using actual observations from multiple wind turbines at a single wind farm 5