Back to the programme printer.gif Print




Delegates are invited to meet and discuss with the poster presenters in this topic directly after the session 'Forecasting: Maximizing grid deliverability and leading your business processes to profitability' taking place on Thursday, 13 March 2014 at 11:15-12:45. The meet-the-authors will take place in the poster area.

Andrea Vignaroli WindSim AS, Italy
Co-authors:
Matteo Mana (1) F P
(1) WindSim AS, fossano, Italy (2) Uni Research, Bergen, Norway

Printer friendly version: printer.gif Print

Abstract

Improving wind farm power output forecasting with machine learning and classifications techniques

Introduction

To improve the wind farm performance and obtain an accurate prediction of energy production and loads new mesoscale models were integrated with fine-scale fluid dynamics models by means of machine learning techniques. The obtained forecasts show different performances that depends on several factors, e.g. periods, forecast length, etc. Self organizing map's based classifiers are employed to search for patterns in time series data. Then different supervised machine learning techniques are used selectively for each of the period revealed by classifiers.

Approach

A new procedure for coupling mesoscale and fine-scale fluid dynamics and application of machine learning techniques was invented. The procedure includes three major steps. At first, the self organizing maps type of neural network was used to identify data series with different characteristics and divide them in sub groups (clusters). Each of these sub groups later were used in training of different feed-forward neural networks (FFNN). That ensured accurate downscaling of NWP results down to the reference point (metmast). At the last step, CFD is used to obtain a power forecast from the FFNN output.

Main body of abstract

The test case was based on historical data from Rotello wind farm: the site is located in central Italy, and consisting of 20 turbines all located in complex terrain. Two years of data (2010 - 2012) were used in the experiment: from which the whole year period was used for training and two half years periods were set for validation.
Visualization and analysis of extracted patterns was performed to improve the selection of forecasting techniques.
To test the performance of the forecast the Normalized Mean Average Error (normalized towards the nominal power of the wind farm) between detected and expected output was calculated.
The ability of the model to forecast extreme cases (high production periods and calm periods) was tested and compared to the no-classifiers model's performance. The improvement in forecast is observed.
The benefit of having instant and more accurate forecast with proposed new classifiers techniques vs. cost of observed improvement (shown for example in time consumption for data preparation and training, or in availability of production data for model training and validation) is discussed in addition.

Conclusion

The machine learning classifiers shown to be a powerful tool for pattern recognition in NWP time series.
The forecast accuracy is improved by making the forecasting techniques case dependent: model's architectures varies for different classes of data.
In addition, the links between classified patterns and used forecasting techniques allow better understanding of machine learning approach optimization, beside standard trial-and-error approach, therefore fast selection of best performing method can be achieved.
The proposed methodology enables the inclusion of realistic stratification flow regimes and boundary conditions in CFD simulations of relevance to site and therefore a wind farm optimization in complex terrain.


Learning objectives
Upon completion, a new method for fast and accurate wind power output forecast will be delivered: here supervised machine learning techniques are enforced by classifiers to bridge a gap between mesoscale models and fine-scale computational fluid dynamics approach.