OWPI’s Wind Assessment for Oklahoma Wind speed is fundamental to a wind turbine’s productivity. Because the amount of energy available in the Tim Hughes wind is proportional to the cube of the wind speed, small The Environmental Verification changes in wind speed result in relatively large changes in and Analysis Center wind power. Over varied terrain, it is not unusual for the wind University of Oklahoma speed to vary significantly over short distances. 3200 Marshall Ave, Suite 110 Norman, OK 73072-8032 Consequently, knowledge of the variability of the wind over Phone: (405) 447-8412 time and space is critical to developers of wind farms. E-mail: thughes@ou.edu The Pacific Northwest Laboratory in conjunction with the U.S. Department of Energy has established a wind power classification system. The classification provides a practical reference for determining the resource potential based upon an average annual wind speed. Average annual wind speeds are categorized into seven wind power classes. Wind class 1 denotes very light winds (poor wind resource), while larger class numbers indicate stronger winds. One of Oklahoma Wind Power Initiative’s many goals is to develop high-resolution wind power maps for Oklahoma at heights of 10 and 50 meters. Prior to this study, the best resource map available for Oklahoma was developed by the U.S. Department of Energy and Pacific Northwest Labs in 1987. Figure A shows their resource map for Oklahoma, which has resolution of 1/3 deg. Longitude by 1/4 deg. Latitude, or approximately 33 km (20 miles) in each dimension. Since 1993, Oklahoma has become home to one of the premier surface weather networks in the world - the Oklahoma Mesonet. With 114 stations and 5-minute averaging intervals on atmospheric data, this network offers an opportunity to create accurate high-resolution wind Figure A. U.S. Department of Energy, Pacific Northwest power density (WPD) maps to Laboratory estimation of wind resources in Oklahoma, help determine optimal 1987. 1 – poor; 2 – marginal; 3 – fair; 4 – good. placement of both large and small wind turbines. For More Information contact: Selecting Mesonet stations to be used for Wind Assessment Mesonet stations were sited to present the best overall estimate of regional environmental conditions, but not all stations were optimally located to monitor wind resource. Stations with poor exposure to the wind must be excluded from the assessment, because data from these sites would likely bias the wind resource. The following information was used to evaluate the wind exposure (i.e., fetch conditions) for each Mesonet station: 1 Version Date: August 2002 • • • Panorama photos from the Oklahoma Mesonet Web site (okmesonet.ocs.ou.edu), 1-m resolution digital orthophotos (aerial photos), and a 200-m resolution Land Use/Land Cover (LULC) grid derived from the USDA/NRCS MIADS data for Oklahoma. Sites were rated “poor”, “fair”, “good”, and “excellent” based on subjective criteria. For example, stations with short, consistent vegetation cover and no obstructions in the immediate vicinity of the site were rated as having excellent or good fetch conditions. Figure B. 79 Mesonet stations with “good” to “excelStations with tall, inconsistent ent” fetch conditions. vegetation cover or anomalous vegetative cover too close to the site (e.g. a wind break of trees) in prevailing wind directions, were rated as having fair to poor fetch conditions. Of the 114 Mesonet stations, 79 sites were classified as having good or excellent fetch conditions (Fig. B). Analyses Performed Computer models provide an objective method for estimating the effects of terrain on wind flow and for interpolating wind data to locations where data does not exist. An exact mathematical description of the wind flow across the terrain is provided by the Navier-Stokes equations. Because of the great complexity of the equations, they require large amounts of input data and are extremely difficult (often impossible) to solve. Furthermore, it is uncertain whether the large efforts required to run the complex model are rewarded with significantly more accurate wind simulations than those obtained from simpler models. For projects like OWPI, the hardware (e.g., supercomputers) and input data necessary to solve the Navier-Stokes equations is cost prohibitive. Instead, OWPI investigated the use of simple models that can be run on personal computers. Currently, two categories of simple analytical models exist for use in wind resource assessment: mass consistent and Jackson-Hunt. Mass consistent models conserve mass, while Jackson-Hunt models conserve mass and attempt to conserve momentum (approximations made to the Navier-Stokes equations). In general, mass consistent models are used for surveying large areas, while Jackson-Hunt models have been used more extensively for micrositing. Accordingly, the decision was made to use a mass consistent model for the state assessment. Because of the large number of input data-points, a modeling technique employing “neural networks” was also considered a credible tool for wind assessment. Neural nets are empirical models often used for statistical analysis and data modeling. They provide an alternative to conventional analytical techniques for solving nonlinear problems. To our knowledge, neural networks have never been used for the purpose of wind assessment. Nonetheless, the use of neural nets in other fields of study has shown great promise. 2 Mass Consistent Model WindMapTM, a software program by Brower & Company, uses a mass-conserving model for predicting and mapping the wind over an area. The software has gained acceptance in the wind industry as it has been used by several other states (e.g., Iowa, Massachusetts, Minnesota, etc.) in the assessment of their wind resources. Data The model can ingest four types of data: surface, elevation, roughness, and upper-air data. For this project, the following data sets were used as inputs into the model: Wind Speed and Direction Mean wind speeds for 16 compass directions and a Weibull parameter were entered for each Mesonet station incorporated into the model.1 A Weibull parameter describes the shape of the frequency distribution of wind speeds. Of the 79 Mesonet stations with good to excellent fetch conditions, wind data for 76 stations were included in the model. Two stations were excluded from the group because both stations were not in existence for the entire period, 1994 to 2000. A third station was excluded because of its close proximity to another Mesonet Figure C. Locations of the 89 Mesonet stations used to station. In addition to the 76 initialize WindMap. stations, 13 Mesonet stations with fetch ratings of fair were included in the model. These stations were located in the eastern portion of the state in areas that lacked stations with good or excellent fetch conditions (Fig. C). WindMap requires a reference station to be chosen from one of the 89 stations. A reference station defines the directional frequencies of the wind. These frequencies are often displayed on a graph called a wind rose. An average wind rose was computed for Mesonet stations with good to excellent fetch conditions. Of these Mesonet stations, ARNE (near Arnett, OK) was determined to have a wind rose most comparable to the average (Fig. D). Hence, ARNE was chosen as the reference station for the model. Topography DEM Data Topography information for the entire state was obtained from the Digital Atlas of Oklahoma produced by the U.S. Geological Survey. The digital atlas provides a 60-m resolution 1 Mean wind speeds, Weibull parameters, and wind roses were all calculated using seven years worth of 5-minute, scalar-average wind speeds (5-minute average of 100 counts taken at 3-second intervals). 3 Digital Elevation Model (DEM) derived from 1:100,000 scale digital topographic maps of Oklahoma. Land Use/Land Cover Data In general terms, "roughness length" represents a height above ground below which friction from obstacles (e.g., vegetation and buildings) effectively stifles air currents. In scientific terms, it is the height above ground at which the neutral wind profile extrapolates to zero. The roughness values were obtained from a landuse/landcover (LULC) grid model put together by the GAP Analysis Program (GAP). GAP’s LULC grid describes specific land use practices and natural Figure D. Wind rose for the Arnett Mesonet Station. vegetation covers for the state at a Percentages represent averages over the period, 1994 resolution of 30 m. Roughness to 2000. values were assigned to these specific land use practices and natural vegetation covers. For example, urban areas were assigned roughness values of 1.0 while agricultural regions were assigned values closer to 0.03. Results from WindMap Figure E illustrates the results from WindMap’s output at 50 meters. Evident in this figure is the good to excellent wind resource in the western 1/3 of the state. A significant amount of Class 5 winds exist near the town of Woodward in the northwest portion of the state. This area has been under study and OWPI believes that it will be one of the first areas developed. Other significant Class 5 areas include Beaver County (in the Panhandle) and the Wichita Mountains/Slick Hills area located in the southwest. The tremendous wind resource in the western half of the state is attributable to its geographic proximity to the Rocky Mountains and Gulf of Mexico, as well as small roughness values due to a lack of trees or other tall vegetations. The Class 3 areas are located in the flattest portions of the state between the arid, high plains of the west and the rugged, woodlands in the east. Interestingly, the transition between Classes 2 and 3 areas follows along the boundary between the “Cross-Timbers” and the tall grass prairies of the southern Great Plains. The Cross Timbers consist of post oak and blackjack oak woodlands that form the western boundary of deciduous forests in Texas, Oklahoma, and southeastern Kansas. Further east into the Cross Timbers, the woodlands become denser and the terrain becomes more rugged. Consequently, the wind resources become highly variable (i.e., higher 4 Figure E Wind Resource at 50 meters (164 ft) AGL WindMap[TM] Computer Model Creation Date: 03/22/02 5 terrain has Class 2 and even 3 winds while lower terrain has very little wind resource). WindMap was able to emphasize the terrain’s affects on the wind resource for some of the more predominant terrain features such as the Ouachita Mountains (southeast Oklahoma) and the Arbuckle Mountains (south central Oklahoma). Unfortunately, these features are barely evident in the figure provided. [Note: Mountains in Oklahoma typically have terrain relief of less than 1500 feet.] Empirical Model Neural networks represent a relatively new method for using computers to solve problems. Specifically, a neural network or “neural net” is a linked assembly of processors or processing elements whose interconnections are similar to those between neurons in a brain. By a process of adaptation, the computer is able to learn from a set of training patterns. Thus, neural networks are often viewed as a type of artificial intelligence (AI). Data The neural network (NN) model was "trained" using 50 of the 76 Mesonet stations with good and Figure F. NN model developed using two sets of Mesonet excellent fetch conditions. stations – a Training group and a Control group. The remaining 26 stations were set aside as the control group (Fig. F). The model incorporated the following information about the Mesonet stations: • calculated wind power density, • elevation, • terrain exposure (or relative elevation) • and roughness length (vegetative influence). Calculated Wind Power Density Using 7 years of data from Mesonet stations, wind power density (WPD) values were calculated for the 10-meter level according to the equation below: WPD = n 1 * ∑ ( ρ i * vi 3 ) 2 * n i =1 where ρ is air density and v is wind velocity (scalar-averaged wind speed) for a particular station. The above equation was applied to all valid five-minute data (n) for the time period. For each station, n was approximately 735,000. Air density was explicitly calculated using temperature and pressure data from the station. 6 Terrain Exposure “Terrain Exposure” (also often referred to as "relative elevation") is defined as the distance a point sits above or below the average elevation of a surrounding area. The surrounding area can be defined in numerous ways. Some models have calculated terrain exposure relative to a circular area with some diameter, typically 10 to 20 km, depending on the scale of the region being modeled. For this model, terrain exposure values were calculated relative to north and south "pie-wedge" areas. This method was based on the assumption that surface terrain and vegetation characteristics in the north and south directions have the greatest impact on the wind resource. To test this assumption, an average wind-energy rose diagram was developed with data collected from sites with excellent and good fetch ratings (Fig. G). The figure displays the mean percent time and mean percent energy of the wind in 16 compass directions over the seven-year period. From the wind rose, it was determined that the wind direction was from the NW to NE (inclusive) and SE to SW (inclusive) sectors 77% of the time, and more importantly 89% of the wind energy was from these north and south sectors. Furthermore, when one considers that winds from east and west sectors tend to be light (not strong enough to turn large turbines), the percentage of realizable wind power from the north and south sectors becomes even more significant. Consequently the development and use of the "wedge-method" appears justified. To determine the terrain exposure inputs to the model, the Figure G. Average wind-energy rose for average elevations were calculated Mesonet stations with 'good' or 'excellent' fetch using the north and south “pieconditions. wedges” with 10-km (6.2-mi) radials. The north wedge subtends the northeast to northwest (Cartesian coordinates 34° to 146°), and the south wedge subtends the southwest to southeast (Cartesian coordinates 214° - 326°). The degree readings correspond to Cartesian degree coordinates rather than compass degrees, since ArcView (the GIS software used) requires the former as inputs. ArcView "spatial analysis" tools were used to calculate the average elevations in these pie-wedges. North and south terrain exposure values were then determined for each Mesonet station used in the model by subtracting the average elevation from the actual elevation at a site. A positive number represents a site that sits above an adjacent wedge area on average; a negative number represents a site that sits below an adjacent wedge area on average. 7 Results from NN Model For the initial wind resource map, the NN model’s 50-m WPD estimates appear conservative. Based on data collected from a single 40-m tower in northwest Oklahoma (near Buffalo), the wind resource for that area should be closer to 500 W m-2 than the model predicted 386 W m-2. The model also appears to underestimate the part of the state with more vegetation (roughly, east of I-35). The underestimation of WPD at a height of 50 meters most likely is the result of the extrapolation technique (power law profile) used on the 10-m model output. The exponent, m, used in the power law is dependent on both the surface roughness and stability. Consequently, m varies across the state, but how and by what amount? Unfortunately, Oklahoma lacks a network that measures winds at heights above 100 feet, thus the mapping of m would be purely speculation. Alternatively, a correction factor was applied to the entire map in order to reduce underestimation. The 10-m WPD grid was multiplied by a correction factor of 1.23. The correction factor was based on discrepancies between predicted versus actual wind power densities at 10 meters (i.e., the slope of the trend is increased to equal one). The result was a map with significantly more Class 3, 4, and 5 winds (Fig. H). The neural network model predicts a good to excellent wind resource for the ridges and hills in Western Oklahoma at the 50-meter level. The model output emphasizes the topography across the state for two reasons. First, the neural network model has a relatively fine resolution of 60 meters. The fine resolution allows the model to recognize terrain features that might not be evident with larger resolutions (e.g., the current national resource atlas has a resolution of roughly 20 miles across Oklahoma). Second, topography is emphasized because many of the model inputs are directly or indirectly related to elevation. Model Comparison Based on DOE/PNL’s assessment (Fig. A), both the NN and WindMap models successfully represented the large-scale variation of wind energy across the state. That is, wind power increases from east to west for all assessments. As expected, the transition from east to west follows the general rise in elevation across the state. However, the model output from the NN and WindMap provides much finer resolution than PNL’s map. The increased resolution in an assessment map is important to the wind power industry for which small changes in wind power translate to millions of dollars. In spite of this general agreement, large discrepancies are evident between DOE/PNL’s map and OWPI’s maps. First, DOE/PNL’s map classifies the Ouachita Mountains in southeast Oklahoma as having Class 4 winds, while the map produced from WindMap categorizes the majority of the region as Class 1 or 2. Unfortunately, verification data does not exist for these mountains; hence, two different conclusions are drawn. A side-by-side comparison of OWPI’s models shows distinct contrasts between the two models. The NN model shows greater detail in local variations of WPD than WindMap. WindMap produced broad areas of winds with similar WPDs, while the NN model’s output closely adheres to changes in topography, hence more local variation in WPDs. Some of the detail can be attributed to the better resolution of the NN model (i.e., 60 vs. 372 m). WPD estimates from both models appear to be conservative for certain areas (e.g., ridges). To determine the accuracy of the model's WPD estimates and to improve the output, more wind data is needed from heights around 50 meters. Recently, OWPI help put into 8 Figure H Wind Resource at 50 meters (164 ft) AGL Neural Network Computer Model Creation Date: 03/22/02 9 operation two more tall towers. Future plans are to instrument more towers. The locations of these towers will likely be in areas with the greatest potential for development of wind energy, but attempts will be made to have these towers spread evenly across the state. Until further validation data from tall towers are obtained, the current results are OWPI’s best estimates of the wind resource across Oklahoma. The NN model’s output may be viewed in conjunction with WindMap’s output to provide a good idea of the energy potential for most locations in Oklahoma. More specifically, the 50-m output from WindMap would be used to provide a first estimate of the wind resource, because these results were derived from physical equations and did not require the application of a correction factor. The 10 or 50-m WPD map from the NN model could be used to adjust the estimate for a local area. For example, the NN maps clearly show significant hills and ridges for the west half of Oklahoma. If the land in question is located on one of these features, WindMap’s estimate may be viewed as conservative. For more information, contact: Tim Hughes, Project Director The Environmental Verification and Analysis Center The University of Oklahoma 3200 Marshall Ave, Suite 110 Norman, OK 73072-8032 Phone: (405) 447-8412 E-mail: thughes@ou.edu 10