Improving the prediction of wind power ramps using texture extraction techniques applied to atmospheric pressure fields

Abstract

Much recent research in wind power forecasting has been focused on predicting large, sudden changes in wind power output, called wind ramps. State-of-the-art wind ramp prediction methods estimate future wind ramps from forecast power time series. We suggest that, by analyzing the weather associated with wind ramps, their forecasting can be improved. In particular, we propose a new method for wind ramp forecasting based on the analysis of forecast atmospheric pressure fields. Feature vectors relating to the pressure gradient are extracted from the pressure fields using an image texture extraction technique, called Gabor filtering. Numerical experiments show that these Gabor feature vectors are well correlated with power generation. They are used as inputs to a new wind power forecasting model. Compared with a basic state-of-the-art wind power forecasting model that does not use Gabor features as input, the proposed model exhibits better performance for power prediction, for two of the three wind farms chosen for this study. However, the ability of the model to forecast actual wind ramps is worse than that of the basic model, as measured by ramp capture rate and forecast accuracy. We also describe a second method to pre- Earth and Atmospheric Sciences, University of Alberta, Edmonton, Canada dict the magnitude of a sudden power change (wind ramp), using several input variables, in addition to Gabor features. Numerical experiments show that this second approach has better performance than the basic model, with respect to ramp capture rate and forecast accuracy. We suggest that it could be used operationally to supplement a current state-of-the-art ramp prediction model. We present an example of using this approach to provide a warning of a potential ramp.

Day‐ahead wind power ramp forecasting using an image‐based similarity search strategy

“…where Mag(k) is the ramp magnitude when the k-th cluster of the wind speed forecast profile is extracted in the time series framework, and Mag(t 0 ) is the ramp magnitude forecast if a ramp occurs at t 0 . Figure 12 gives an example of the proposed day-head wind power ramp forecasting scheme with the ramp magnitude estimation, where the top orange line is the magnitude forecast calculated based on (7), and red error bar represents the measured magnitude of an actual ramp event with an acceptable error range of ±20%, which is one of ramp forecast requirements of the evaluation criteria suggested in [31].…”

“…Li et al. [7] and [8] used additional meteorological variables, atmospheric pressure fields and Foehn wind, respectively, to improve the prediction of wind power ramps. However, they could not be used as a generic solution to wind power ramp forecast due to the high requirement for meteorological expertise and the limitation on specific geographical environments.…”

“…However, these methods are highly dependant on the accuracy of NWP models, and the implemented ramp detection algorithms may even amplify the forecasting errors of the wind power production itself. Li et al [7] and [8] used additional meteorological variables, atmospheric pressure fields and Foehn wind, respectively, to improve the prediction of wind power ramps. However, they could not be used as a generic solution to wind power ramp forecast due to the high requirement for meteorological expertise and the limitation on specific geographical environments.…”

Machine intelligent and deep learning techniques for large training data in short‐term wind speed and ramp event forecasting

“…In ML, the performance on an unseen data is highly dependent on the input feature set, and the dimensionality of input features during training phase governs the computational time and the generalization performance. In Reference 23, the prediction performance of a feature extraction technique based on Gabor filtering and considering atmospheric pressure fields, is found to be better than the state‐of‐the‐art neural network method. Figure 1 illustrates a graphic for ML performance with respect to the number of features.…”

Machine Intelligent Techniques for Ramp Event Prediction in Offshore and Onshore Wind Farms

“…Three wind speed datasets from Spain are considered to evaluate the accuracy of hybrid method for classifying wind power ramp events. In [22], the authors have presented a feature extraction method based on Gabor filtering where the atmospheric pressure fields are taken into consideration. Results are compared with state-of-the-art neural network method and it is found that Gabor method with its change in power as its output gives a better prediction performance.…”