Wind Resource Forecasting

Wind power forecasting is a tool that predicts the amount of energy that will be produced from a wind farm over a period of time. The need to schedule energy output from wind generators comes from the requirement to maintain electricity grid reliability in a cost effective manner.

The science behind wind power forecasting continues to advance as more research is dedicated to the cause and as the abilities of computer processors increase. In parallel, power forecasting is evolving as forecast users are faced with new challenges related to additional wind generation on the electricity grid. There is ample evidence that wind power forecasting has and continues to provide a great deal of benefit to the wind energy community. To understand the limitations of wind power forecasting, it is important to first understand the fundamental science of producing a wind power forecast in light of the sensitivities that end users have to forecast accuracy.

Questions about the utility of wind power stem from the reliance on a fuel source that by its nature is variable. Whereas conventional power plants have the ability to provide a constant supply of electricity, wind power output oscillates based on the interaction of atmospheric phenomena with terrestrial features. Wind farms are strategically situated in locations of the world that have optimal potential to provide the most fuel. The wind turbines situated in the farms are positioned to ensure maximum wind energy capture. However, even with an optimized turbine layout and access to available transmission, wind power brings with it requirements that have caused Independent System Operators (ISOs) and other scheduling and balancing agencies to rethink their approach to electricity grid reliability.

Regions that have experienced a modest amount of wind generation installation can see power from wind generators fluctuate by several hundred megawatts hourly or in shorter periods of time in extreme cases. Power swings that take place without advance warning give rise to electricity supply reliability risks and increased costs for the electricity system and consumers. Real-time oversight of power output from a wind farm can provide an ISO with supply-side control including increasing or decreasing output from other sources as needed or decreasing wind generation output by means of curtailing the wind generator. Supply-side control is exacerbated by wind generators because of their inability to increase power output on demand, unless they are operating at a low capacity and enough fuel is available at the time of the request for increased power output. Operating in a manner that is less than optimal has profit consequences to the wind farm operator and potential societal consequences that come with unnecessary additional generation from fossil-fuel sources.

Two Steps

Wind power forecasting is an important tool to help accommodate and promote the inclusion of wind power into the electricity grid. Wind power forecasting is the science of predicting the power output from a wind farm based on two primary components. The first component (we’ll call it Step 1 for the sake of simplifying a complex process) attempts to predict future atmospheric variables (for example, wind speed and direction at 80 meters, air density and so on). Step 1 includes running highly complex numerical weather prediction (NWP) models to determine the variation of atmospheric variables over short time intervals and small spatial scales.

The second component, or Step 2, uses statistical analysis to develop relationships among meteorological variables and also characterize a wind farm’s historical power production based on known historic atmospheric variables occurring at the same time. In Step 2, statistical models are used to determine meteorological variables within the atmospheric models and at the wind farm based on known variables.

During Step 1, NWP models can be computationally demanding and costly to run due to the quantity of computer processors necessary and the speed at which they must solve the physical atmospheric model equations. Examples of equations in the physical atmospheric models include the conservation of mass, momentum and energy that are solved on a three-dimensional data grid spanning hundreds to thousands of miles from the wind farm, both horizontally and vertically. The physical models in Step 1 can also be used to model the variations in wind flow within the wind farm and the resulting energy output by modeling the interactions of the wind with turbines and turbine-induced wake effects.

Accuracy of the output from Step 1 processes is heavily dependent on the quality of input data used during the initialization process. Predicting wind speed and direction from input data gathered from hundreds and potentially thousands of miles away is a complex process subject to the quality and quantity of input data. A small inaccuracy in the input data will result in incorrect values to be carried over in equations over a billion times before the model can predict meteorological variables at the wind farm. The quality and quantity of the input data as well as the assimilation of local and regional data sources is an important element in the Step 1 processes.

Statistical models used in Step 2 are employed both within the atmospheric models and models that determine energy output from a wind farm by characterizing the relationship between meteorological variables and the farm’s actual energy output. Appropriately tuned statistical models will directly impact the forecast’s usefulness as they can be tailored to achieve specific performance metrics such as the lowest mean absolute forecast error or minimization of the mean squared error. The desired performance metric will guide all processes in Step 2 often at the expense of other performance metrics.

An example of a performance metric that is often used is minimizing the mean absolute error (MAE). A forecast that has a guiding Step 2 principle of minimizing MAE will tend to predict power output from a wind farm that avoids spikes in actual power output from that farm. In many instances, MAE minimization can misrepresent the amplitude and/or the duration of variations in energy output. When averaged over a finite period of time (that is, weekly or monthly) a state-of-the-art forecast can significantly minimize the MAE. The example in Figure 1 (above) depicts a forecast tuned to minimize MAE, but which at several points significantly misses the amplitude and timing of the actual energy output.

Another often used performance metric that is the target of Step 2 processes is the minimization of the root mean square error (RMSE). Again, there are benefits and limitations to a forecast optimized to minimize RMSE. A forecast tuned to RMSE will tend to place importance on the large increases or decreases in energy output and less on the smaller variations. RMSE amplifies the importance of the large energy variations but will often miss the smaller variations that might have been captured if the forecast was tuned to minimize MAE.

Step 2 processes focus the forecast performance on end user sensitivities to forecast accuracy. The ability of an end user to define these sensitivities will affect the forecast’s overall utility. User sensitivity to forecast accuracy is almost always tied to a cost function, be it a profit motive or a system reliability motive. Of the many users who utilize wind power forecasts, there are two general groups that receive the forecast: the commercial regime and the security regime.

Commercial Regime

The commercial regime includes wind farm owners and operators. Their objective is to minimize cost and maximize profitability. Quantifying sensitivities to forecast error for this category of user is often straightforward. The cumulative effects of small errors are important over time as they can directly relate to the owner’s profit function. The best forecasting approach for these organizations is to suitably optimize a forecast that directly relates to their cost/profit function. Depending on the market in which these wind farms participate, performance metrics are often easily identifiable.

Over the past decade, the majority of forecasting for the commercial regime has fallen into two separate forecasting category types: hour-ahead and day-ahead forecasts. Both methods predict expected energy output over specific time intervals such as a 1-hour period or 15-minute periods. Where they differ is the focus of the forecasting technique. Hour-ahead forecasts predict the output from a farm starting in the next full hour period and might span the next 6 to 8 hours.

The modeling techniques used in the hour-ahead forecasts place a high degree of importance on the data that is measured directly at the wind farm, including wind speed and direction and actual power generated over the past interval period. Day-ahead forecasts predict the energy output over the same intervals, but focus on longer periods of time, often picking up where the hour-ahead forecast stops and continuing for several days. Day-ahead forecasts rely more on NWP models than the direct plant feedback as the atmospheric variables that are important to a wind farm two days from now might be found 800 to 1,000 miles away from where the wind farm is situated.

Security Regime

The security regime includes balancing authorities, ISO and scheduling entities whose objective is to maintain the electricity grid within operating specifications. The security regime is less motivated by profit than by the need to manage tens of hundreds of generators and to schedule reserve capacity as effectively as possible. Where it is important to forecast the timing and amplitude of changes in wind generation output in the commercial regime, the importance for the security regime is based on anticipated electricity demand. The focus of wind power forecasting for the security regime is often to predict the minimum amount of wind energy output during peak load hours. In this way, the security regime operates on a different set of values than the commercial regime. The security regime as a whole is a neutral player in the electricity market and while it does administer the electricity markets, it does so with the primary objective of system reliability rather than profit maximization.

Members of the security regime often employ a centralized forecasting service which covers all wind farms within its territory. The sensitivities of the balancing authority often create the need for forecast performance metrics that are only useful to the balancing authority as it acts as a reliability agent. An energy output forecast at specific time intervals is often just the start of centralized forecasting needs. The forecasting requirements of balancing authorities often include bands of expected wind output over specific time intervals, often coined as a probabilistic forecast. These forecasts often take the shape of 80 percent and 20 percent probability of exceedence values where the forecast doesn’t provide a specific energy output level but instead predicts confidence levels of minimum amounts of energy that will be produced. This helps balancing authorities schedule wind generation in its mix without prescribing a specific power output as in the case of traditional generators. Figure 2 (below) offers an example of a probabilistic forecast that at hour one shows an 80 percent chance of at least 1,250 MW being generated and only a 20 percent chance of more than 1,875 MW being produced during the same period.

The various ISO and balancing authorities in the security regime are faced with unique local and regional constraints that in turn place unique requirements on the wind power forecast. An example of this is the Electric Reliability Council of Texas (ERCOT), which earlier in 2009 reported it has over 8,000 MW of installed wind generation capacity in its territory. The majority of existing wind farms in ERCOT are clustered in a few pockets of high wind areas. During the windy season, Texas can experience enough wind to generate 8,000 MW of wind energy on its system. Such significant amounts of wind energy require available transmission, which ERCOT is currently designing, and increased reserve generation that could supplement any sudden decrease in wind power. Both of these requirements are quite expensive.

Likewise the Bonneville Power Administration (BPA) in the Pacific Northwest reports it has over 2,000 MW of installed wind generation capacity, much of it in and around the Columbia River Gorge. BPA relies on 31 hydroelectric dams among other generation sources including wind. Often, hydro power is seen as a complementary generation source to wind power and can be responsive to the load demands. The BPA at times has a supply-side problem in that it is hampered in its ability to increase or decrease its hydro generation to accommodate the variability in wind generation. BPA’s dilemma is to be responsive to changes in wind generation by moving hydro generation up or down while simultaneously responding in ways that are responsible to environmental factors, mainly to protect migrating salmon. Simply increasing or decreasing hydro output to accommodate wind is not possible at times due to the environmental constraints BPA faces.

New unique forecasting tools are under development to satisfy the needs of the security regime and the balancing authorities like ERCOT and BPA. Their sensitivity to forecast error includes the need to act in a manner that is responsible to utility customers, environmental concerns and to keep an open and fair market for all power generation sources, wind farms included. To be sensitive to all parties, they must balance load with energy supply from all producers and maintain enough reserve generation to offset any sudden decrease in wind generation. To meet this unique challenge, forecasting tools like ramp event forecasts are being developed that rethink the approach used in Step 1 and Step 2 processes.

The impetus behind ramp event forecasting development is the need for new performance metrics that can meet the security regime’s obligation to minimize the amount of unnecessary online reserve generation. Ramp event forecasts are a mixture of various forecasting techniques and, in theory, can predict the amplitude and duration of a significant change in wind energy output and associate that change within a distinct set of probability distributions. Moving forward with this type of forecasting solution, the security regime is finding new ways to include variable renewable energy on the electricity grid while minimizing the costs associated with reserve generation.

Balancing authorities are scattered throughout North America and all have unique sensitivities in respect to the inclusion of wind energy including transmission constraints, reserve constraints and even antiquated scheduling procedures. These various locations across North America also bring with them unique challenges when it comes to forecast performance which can and does vary as a function of location, season, weather patterns and the size and diversity of the wind farms in the balancing authority’s area. As more is understood about the needs of the forecast users, forecast providers are taking strides to improve the state of the art and create solutions that address their unique requirements.

Given the strong support for integrating more renewable sources into the generation mix from policy makers additional sensitivities to forecast error are sure to crop up. As the industry continues to grow and mature, it will be essential for all parties to work collectively to clearly identify the challenges of integrating renewables so that specialized forecasting techniques and solutions can be developed.


Kenneth Pennock is the forecasting business manager at AWS Truewind, LLC where he manages strategic planning and implementation of the firm’s renewable energy forecasting services and wind integration studies. Mr. Pennock holds an MA in Geography from the University at Albany and an MBA from the College of Saint Rose.

No posts to display