Conference programme

Back to the programme printer.gif Print




Delegates are invited to meet and discuss with the poster presenters during the poster presentation sessions between 10:30-11:30 and 16:00-17:00 on Thursday, 19 November 2015.

Lead Session Chair:
Stephan Barth, ForWind - Center for Wind Energy Research, Germany
Alla Sapronova Uni Research, Norway
Co-authors:
Alla Sapronova (1) F Catherine Meissner (2) Matteo Mana (2)
(1) Uni Research, Bergen, Norway (2) WindSim, Tonsberg, Norway

Share this poster on:

Printer friendly version: printer.gif Print

Presenter's biography

Biographies are supplied directly by presenters at EWEA 2015 and are published here unedited

Dr. Sapronova is a senior researcher at Uni Research Computing and external censor at Natural Science Department, University of Bergen, Norway. Her scientific interests are focused in machine learning and data mining techniques. Since 2010 she is actively involved into several national R&D projects that require various predictive models development and application for the wind energy sector.
She received her PhD in Physics and Mathematics in 2004 and worked as Postdoctor at Uni Research, University of Bergen in the field of computational modeling and artificial intelligence since 2007.


Poster

Poster Download poster (9.87 MB)

Abstract

Machine learning for short time ahead wind power forecast

Introduction

With wind power generation growing, new complications are introduced to wind energy producers: in order to work effectively changes in the wind power have to be anticipated. This is only possible with accurate wind predictions.
Wind speed and wind direction can be forecasted with quite accuracy for onshore wind parks located in complex terrain with models based on a coupling between artificial neural networks and mesoscale weather predictions and computational fluid dynamics. Yet some weather regimes remain unresolved and forecasts of such events fail. The accuracy of the forecast improved significantly when categorization information is added as an input to the artificial neural network. The improved model was able to resolve extreme events and converged faster with significantly smaller number of hidden neurons. The new model performed equally well on test data sets from both onshore and offshore wind parks.

Approach

In this work, short-term prediction of wind speed and energy output of 20 Vestas V90 wind turbines park was performed. For that Numerical Weather Prediction (NWP) mesoscale models at a coarse resolution of some kilometers and observation at wind park location are coupled with ANN to forecast wind speed and direction, the most important components for energy yield prediction. In order to forecast the wind flow near the ground in the complex terrain, where roughness and complexity affect the flow at microscale the ANN is employed to predict nonlinear multivariable functions in the environment where explicit physics-based models either have limited application or are not available.

Main body of abstract

The mesoscale-microscale coupling model uses NWP forecast timeseries for temperature, pressure, relative humidity, wind speed, and wind direction to issue a site-specific forecast. ANN is trained against historical observations of wind speed and direction and corrects the NWP forecasts of mean hourly wind speed and direction. On the test data sets the model predicted the wind speed in a very satisfactory manner with mean square error (MSE) 1.8 m/s. Even though such accuracy found to be superior to that based on polynomial fittings as well as auto-regressive moving average (ARMA) models, the detailed statistic on model performance shows that for some weather regimes the MSE is significantly higher.
The ANN-based model is improved by using data categorization approach. According to the intrinsic margin, physical nature of the problem, etc, training data sets grouped into several discrete categories. The discrete categories allows identical category values to be treated in the same manner. One logical approach is to categorize numeric data, a wind speed in this case, similar to typical human concepts and then try to generalize. The information obtained from categorization of a single variable is supplied to the ANN input in addition to the NWP variables.
The model with categorization shows MSE =1.1 m/s for wind speed prediction on the test data sets and resolves all weather regimes at the same level of accuracy.

Conclusion

The ANN-based coupled model for mesoscale-microscale coupling can be improved significantly by adding categorization information. It is observed that a model with added categorization information has nearly twice as low RMSPE (root mean square percentage error) than a regular model (5.4% vs 9.8%) with the number of hidden neurons lowered twice in the categorization model.


Learning objectives
ANN based model's generalization arises from the model's ability to find similarity in the training data that usually consists of continuous numeric data. Since numbers are rarely exactly the same from one example to the next, the model can fail in selecting the margins for identical properties. In this case, the generalization can be improved by classification. If the training data is limited or incomplete a categorization model can be used.
The results show that the choice of methods for categorization is irrelevant to the generalization improvement. Methods, quite different by nature, give similar and reasonable results and all lead to improved generalization.