Study Of Pattern-based Wind Speed Prediction

Wind power producers play a more and more important role in electricitymarket. However, the uncertainty of wind is considered as a major barrier againstfurther wind power penetration. The uncertainty increases the level of regulationand reserves required to maintain stability of the grid. Short term wind speedprediction plays an important role in large scale wind power penetration. In thispaper, we focus on the study of wind speed prediction in wind farm.Firstly, we propose a pattern based approach for short-term wind speedprediction based on GPCA. It is well accepted that wind varies in differentpatterns in different weather conditions. Thus, we should use different models todescribe these patterns, while most current works build wind speed prediction ona single model. Based on this observation, we introduce GPCA method todiscover the patterns hidden in data of wind speed. Then we train a model foreach pattern. Finally we use a weighted combination of these models as the finaloutputs. Experiment results show that our approach outperforms some classicaltechniques.Secondly, we introduce Deep Learning theory to find patterns automatically.Deep multi-layer neural networks have a strong ability to represent highly non-linear functions. A deep neural network can extract simple and abstract featureson the top hidden layer. Thus we use this approach to find patterns in wind speedtime series and build a multi-step prediction. The experiment results show thatdeep architectures are very efficient.Thirdly, we propose an approach to search for patterns and importantcharacteristics in frequency domain. We use Wavelet Analysis to decompose rawwind speed series into a set of sub-series. Sub-series with different frequency canbe seen ones with different patterns in frequency domain. A set of predictionmodels are built for each sub-series and the prediction steps are determinedaccording to the self-relevance character. Then a multi-scale ensemble strategy isused give final outputs.A wind speed forecasting system is built for one Ningxia wind farm and thesoftware is under testing. Further study and application can be developed on this system.In this paper, we use three methods to find patterns in wind speed series.GPCA, Deep Learning and Wavelet Analysis are selected as tools in differentviews. According to the experiment results, we can see the effectiveness ofapproaches in this paper as patterns are important physical characteristics ofwind speed.