Forecasting

eWind® - Proven, Accurate, Valuable

eWind provides highly reliable forecasts of wind speed, wind direction, and plant output to plant operators, power marketers, utilities, and Independent System Operators (ISOs). Developed by AWS Truewind, one of the wind industry’s leading consulting firms, eWind has proven its value to clients such as Southern California Edison, the California ISO, FPL Energy, Shell Energy, enXco, and International Energie. Thanks to its technical superiority, high value, and ease of use, eWind has become the US forecasting market leader. eWind is also available in Europe, Asia, and South America.

Value

As wind power becomes widespread, more utility companies and ISOs are demanding accurate forecasts. They know that being unable to predict the output of wind projects a day ahead, or even an hour ahead, imposes real costs on utility system operations, and increasingly they are passing that cost along to the wind plant operators through imbalance charges and other fees and penalties.

eWind provides a solution to these challenges. Depending on the client and power market, eWind can greatly reduce imbalance charges, minimize incremental reserve costs, facilitate plant dispatch scheduling, inform spot-market trading, increase capacity payments, and optimize plant maintenance.

The monetary benefits of next-day forecasts for plant scheduling alone typically range from 6 to 15 times the cost of the service. This increased value translates into greater market acceptance and higher profits for wind plant developers and lower integration costs for utilities and ISOs.

The eWind Service

eWind employs sophisticated physics-based atmospheric numerical models and adaptive statistical techniques that have been customized to produce the most accurate possible wind forecasts. The system predicts the wind speed and direction at a wind project site, and then converts those predictions into plant output. Data collected from the site provide continual feedback to maintain optimal accuracy at all times. AWS Truewind, unlike most other forecast service providers, maintains a global forecasting capability inhouse. That means that your wind project can receive the same forecast feed whether it is based in India, China, or Brazil.

For one-hour-ahead forecasts, the error is typically 15 to 25% lower than that of persistence forecasts. For next-day forecasts, eWind typically produces a 40 to 60% improvement over benchmark climatological and persistence forecasts.

Don’t take our word for it: eWind’s performance has been put to the test. In objective evaluations in both Europe and the United States, eWind has decisively outperformed its leading competitors in accuracy.

Delivery

eWind forecasts are delivered automatically to the user via the Web, FTP, fax, or e-mail, and can be updated on a customized schedule to meet a client’s needs. Forecasts can be provided from several minutes to several days in advance. The eWind web interface presents all the information the user needs in a clear and simple format. With a click of the mouse, the customer can access tabular or graphical forecasts as well as examine the recent record of forecast accuracy. AWS Truewind can also provide regular written reports of forecast performance.

Technology

The eWind system is composed of three basic components: (1) a set of high resolution, three-dimensional meteorological models; (2) adaptive statistical models; and (3) a forecast delivery system.

Atmospheric Models Used in eWind

MASS – Mesoscale Atmospheric Simulation System. A proprietary model developed by TrueWind partner MESO, Inc., and used for both research and commercial forecasting services. The current version, MASS 6, is a non-hydrostatic model similar to MM5.

WRF – Weather Research and Forecasting model. The latest release in the MM5 series of models developed by the forecasting community.

OMEGA – This model features a unique, unstructured grid with varying resolution, which allows the highest resolution to be centered on features of interest such as mountains and coastlines, or even around moving weather systems. Developed by TrueWind partner MESO, Inc., in collaboration with Science Applications International Corp.

The meteorological models (denoted by MASS in Figure 1) are conceptually similar to those used for general large scale weather forecasting by government forecast centers such as those operated by the US National Weather Service. Such models solve equations that represent the basic physical principles of conservation of mass, momentum and energy, as well as the physics of phase changes of water and turbulent processes. They produce forecasts of wind speed and direction, as well as temperature, cloud cover, and other weather parameters, for up to several days ahead.

In contrast to the models used in routine weather forecasting, however, AWS Truewind’s models operate at a high resolution in order to simulate the effects of topography and land cover in the vicinity of a wind project. The project may be located in a windy pass, for example, that is not resolved at all by the large scale government-operated national weather models, but is resolved by the eWind models. Also, unlike most other forecast service providers, AWS Truewind has a variety of models at its disposal. These can be combined to produce "ensemble forecasts," which are more accurate than forecasts generated by any single model.

The forecasts from the atmospheric models are then fed into adaptive statistical models. Their job is to continually fine-tune the forecasts using data from the plant site to provide the best possible accuracy at all times. They also handle the task of translating wind and weather forecasts into plant output. The advanced statistical models deployed in eWind "learn" with experience, so that forecast accuracy actually improves over time.

The final stage is the forecast delivery system. AWS Truewind understands that customers have a wide range of needs and capabilities, and our forecast delivery is tailored accordingly. Some customers want to receive a forecast only once or twice a day ("next-day forecasts"); others require forecasts to be delivered as often as every 10 minutes ("next-hour"). Some prefer to view forecasts on a web page, others to download the data from servers directly via FTP. Our forecast delivery system also handles the acquisition of data from the plant site – for example, it can download the most recent hour’s plant output and meteorological information directly from the customer’s SCADA server.

Statistical Models Used in eWind

Multi-variate linear regression. This type of statistical model automatically selects the best predictive variables from the atmospheric model and calculates regression coefficients that transform those variables into a site forecast with the least error and bias. In addition to the wind forecast, the input variables may include temperature, turbulence, cloud cover, and other parameters.

Neural network. Sometimes referred to as an artificial intelligence model, a neural network also determines a set of coefficients that transforms model variables to a site forecast. The difference is that the functional relationships can be highly nonlinear. The neural network used in eWind employs Markov Chain Monte Carlo training, which trains an ensemble of networks by taking many samples from the distribution of the parameters and searching for the optimal solution.

Accuracy

Two common measures of accuracy are mean absolute error (MAE) and skill score. MAE is expressed as a percentage of the plant’s rated capacity. Skill score is the percent reduction in forecast accuracy compared to a baseline such as persistence (for next-hour forecasts) or climatology (for next-day). For nexthour forecasts, the MAE typically ranges from about 4% to 6%, which represents a 15 to 25% improvement in skill score over persistence. For next-day forecasts (12 hours and beyond), the MAE typically ranges from 14% to 22%, a 40 to 60% improvement in skill score over benchmark climatological models.

Independent tests of eWind have been directed by the Electric Power Research Institute (EPRI) under funding from the California Energy Commission (CEC). In these tests, a year of next-day forecasts was produced for two California wind projects, one in Altamont Pass, the other in San Gorgonio Pass. At Altamont, the monthly MAE for next-day forecasts ranged from a low of 2.4% to a high of 22.9%, with an annual average of 14.1%. The corresponding range at San Gorgonio was 10.4% to 22.7%, with an annual average of 16.6%.

For comparison, the Risø Prediktor model, which is widely used in Europe, achieved a mean annual MAE of 14.2% and 22.3% at the Altamont and San Gorgonio plants, respectively, for the same period. While Risø’s Altamont error was comparable to eWind’s, its San Gorgonio error was 34% larger. This illustrates the value of the high resolution mesoscale model used in AWS Truewind’s forecasting approach.