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 'How does the wind blow behind wind turbines and in wind farms?' taking place on Tuesday, 11 March 2014 at 16:30-18:00. The meet-the-authors will take place in the poster area.

Cristóbal Gallego Universidad Politécnica de Madrid, Spain
Cristóbal Gallego (1) F P García-Bustamante Elena (2) Cuerva Alvaro (1) Navarro Jorge (3)
(1) Universidad Politécnica de Madrid, Madrid, Spain (2) Universidad de Murcia, Murcia, Spain (3) CIEMAT, Madrid, Spain

Printer friendly version: printer.gif Print

Presenter's biography

Biographies are supplied directly by presenters at EWEA 2014 and are published here unedited

Cristobal Gallego is assistant professor at the Universidad Politecnica de Madrid, where he teaches wind turbine technology.
He studied aerospace engineering at the same university.
His research is focused on wind power forecasting, with special attention to extreme events.
He recently presented his PhD thesis, entitled "Statistical models for short-term wind power ramp forecasting".


On the relation between wind power ramps and the global/synoptic scales


Nowadays, one of the challenges to achieve large-scale integration of wind energy is to improve the forecast of abrupt power output variations (ramp events) [1,2]. Predicting ramps is difficult since these events can be motivated by meteorological processes occurring at different time/spatial scales. Ramps are also caused by features such as rotor yaw misalignment, wind turbine shut-down and rotor wake interactions. We introduce an innovative methodology to relate different stages of the wind-to-power conversion process to ramp events. In particular, results concerning the contribution of the global/synoptic scales to ramp occurrence at the wind farm level are presented.


Ramps events represent a particular case of wind power dynamics. Thus, ramps can be motivated by processes occurring at any stage of the wind-to-power conversion process (see the following figure).

Monitoring these stages generates massive datasets that can be employed to explore ramp underlying causes. For example, outputs generated with Numerical Weather Prediction (NWP) models carry information on meteorological processes occurring at different time/spatial scales: Global Circulation Models (GCMs) capture meteorological phenomena in the global/synoptic scales while Limited Area Models (LAMs) are able to capture mesoscale and microscale processes. In addition, observational data gathered at weather stations characterise local weather conditions, while data gathered by the SCADA system embody information on the wind farm state.

Because these datasets are likely to contain large amounts of irrelevant information for ramp explanation, appropriate techniques are required to efficiently identify relevant features in big data. In this regard, an innovative methodology to systematically relate massive datasets to wind power ramp events is here introduced. The methodology is inspired in feature-selection literature [3]. It consists of a two-staged approach: first, Principal Component Analysis (PCA) is employed to identify relevant features of the raw dataset. Second, the notion of Mutual Information (MI) is employed to assess the dependency between these features and ramp occurrence. Ramp events are characterised through the ramp function, introduced in [4].

We apply this methodology to relate reanalysis data generated with a GCM and ramps observed at a multi-megawatt wind farm located in the north of Spain. The considered period ranges from 1st November 2007 to 16th September 2008. By doing this, we focus on the relation between the global/synoptic scales and ramp occurrence at the wind farm level.

Main body of abstract

The meteorological dataset employed in this work originates from the CFSR reanalysis performed by NCEP [5]. The dataset consists of 4557 hourly time series, resulting from considering a horizontal spatial domain of 21x31 nodes and seven variables per node. The resolution of the spatial domain is 0.5º, and it roughly covers the Iberian Peninsula (see the following figure).

The seven variables involve the following six geophysical fields: horizontal wind field at 10 metres (zonal and meridional components, U10, V10), horizontal wind field at 850 hPa (U850,
V850), horizontal wind field at 500 hPa (U500, V500), mean sea level pressure (MSLP), geopotential height at 850 hPa (Z850) and geopotential height at 500 hPa (Z500).

The hourly wind power time series originates from a wind farm located in the north of Spain (red square in the figure). The wind farm consists of 35 wind turbines and the rated power is 24.5 MW. A preprocessing of the data was performed in order to identify and remove power data related to abnormal generation circumstances, such as TSO stop order, breakdown of the wind turbines or periods related to maintenance labours. In order to characterise ramp events, the ramp function introduced in [4] was employed. The ramp function provides a continuous index of the ramp intensity at each time step. This function depends on a single parameter that was set according to typical ramp durations (1-4 hours) [4].

The methodology is as follows: First, a subset of the raw meteorological dataset is defined. This is done by selecting one of the six aforementioned geophysical fields together with a spatial subdomain. Eighteen spatial subdomains have been considered, ranging from D1 (the single node closest to the wind farm) to D18 (the broadest domain, covering the Iberian peninsula). The figure highlights the particular case of D4 and D18.

By applying PCA to a specific subset, the modes (referred to as Empirical Orthogonal Funcions, EOFs) together with the Principal Components (PCs) are obtained. EOFs are the main features of the data as they represent the way in which data combine to jointly exhibit maximum variance. Eventually, EOFs can be associated to specific meteorological processes [6]. PCs show how the data are organised into modes across time.

Next, the notion of MI [7] is employed to assess the dependency between the PCs and the ramp function. An important point of MI is that no assumptions about the model of dependency are required. Thus, MI captures both linear and non-linear dependencies.

Results are shown in the following figure:
This figure shows the MI obtained for the first four PCs (each of the four plots), the six aforementioned geophysical fields (related to each stack of bars) and the eighteen spatial domains (specified in the legend). The red horizontal dashed-lines represent the bias in the MI estimation. "a.u." stands for arbitrary units.

It can be seen that wind power ramp events experienced at the considered wind farm were mostly associated with geophysical fields related to low levels of the atmosphere (10 metres and 850 hPa), as no evidences of dependency were found for those geophysical fields at 500 hPa. We have performed an analysis on the underlying meteorological processes related to the highest MI marks. It was found that, for the case of UV10 (first stack of bars), high MI marks for PC1 and PC2 are related to regional weather regimes provoked by the channelling effect of the Ebro valley (namely, Cierzo and Bochorno) [8,9]. On the other hand, for the case of MSLP (fourth stack of bars), the observed high MI marks for PC3 are related to zonal pressure gradients, indicative of the passage of large scale weather systems.


In order to improve ramp forecasting, the role of the different stages involved in the wind-to-power conversion process needs to be analysed. To this end, we have proposed an innovative methodology oriented to systematically relate ramp events to massive datasets.

We focused on the relation between wind power ramp events and the global/synoptic scales by considering renalaysis data generated with a GCM. The proposed methodology allowed us to identify spatial domains, geophysical fields and meteorological features relevant in explaining ramp occurrence at the wind farm level. In particular, results suggested that, for the case study considered, ramp events were mainly associated to regional weather regimes conditioned to orographic features and to zonal pressure gradients.

It is worth noting that the meteorological processes considered in this work were limited to those captured by the GCM. For this reason, in order to obtain a comprehensive understanding on the ramp underlying causes, the introduced methodology could be applied to other datasets. For instance, considering NWP outputs with higher spatial/temporal resolutions (i.e. generated by LAMs) would allow researchers to analyse the contribution of subgrid processes non captured by GCMs to ramp occurrence. The influence of local weather conditions could be investigated by considering observational data gathered at weather stations, while the role of other effects such as rotor yaw misalignment, wind turbine shut-down and rotor wake interactions could be explored by considering data from the SCADA system and CFD simulations.

Finally, we point out that the results obtained with the proposed methodology could eventually allow for the elaboration of meaningful explanatory variables to feed ramp forecasting models. This is so because the methodology extracts low-dimensional information from big data on the basis of the estimated dependency with ramp events.

Learning objectives
- An innovative methodology to systematically relate big data with wind power ramp events was introduced.

- The methodology was employed to investigate the relation between the global/synoptic scales and ramp occurrence at the wind farm level.

- Regional weather regimes and zonal pressure gradients were identified as the main causes of ramp events at the considered case study.

[1] Ferreira, C et al (2010) A survey on wind power ramp forecasting. Technical report. Argonne National Laboratory.

[2] Potter, C, et al (2009) Potential benefits of a dedicated probabilistic rapid ramp event forecast tool. IEEE/PES Power Systems Conf. and Exposition, Seattle, Washington. pp 1-5.

[3] Blum, A and Langley, P (1997) Selection of relevant features and examples in machine learning. Artificial Intelligence, 97(1-2):245-271.

[4] Gallego, C et al (2013) A wavelet-based approach for large wind power ramp characterisation. Wind Energy, 16(2):257-278.

[5] Saha, S et al (2010) The NCEP Climate Forecast System Reanalysis. Bulletin of the American Meteorological Society. 91:1015-1057.

[6] Wilks, D (2006) Statistical methods in the atmospheric sciences. Elsevier Academic Press publications (Burlington, USA).

[7] Shannon, C (1948) A mathematical theory of communication. Bell system technical journal 27:379-423, 623-656.

[8] Jimenez, P-A. et al (2009) Climatology of wind patterns in the northeast of the Iberian Peninsula. International Journal of Climatology 29(4): 501-525.

[9] García-Bustamante, E et al (2012) North Atlantic atmospheric circulation and surface wind in the Northeast of the Iberian peninsula: uncertainty and long term downscaled variability. Climate Dynamics, 38(1-2): 141-160.

[10] García-Bustamante, E et al (2013) Relationship between wind power production and North Atlantic atmospheric circulation over the northeastern Iberian Peninsula. Climate Dynamics, 40(3-4): 935-949.