Abstract
The dataset was created in order to be able to train machine learning models in the domain of wind energy, especially for the use cases of predictive maintenance / condition monitoring and anomaly detection. In the wind energy domain, one usually suffers from two main problems: Firstly, the lack of sufficiently many labeled data instances, especially for rarely occurring anomalous states. Secondly, the stochastic nature of the resulting measurements, as changing and turbulent environmental conditions lead to a lot of noise in the recorded sensor signals, which in turn makes the analysis of the data very hard. To this end, a synthetic wind turbine dataset has been created, which addresses these issues. The dataset allows to obtain a more controlled view than measurements in the field would yield, since for different fixed conditions of the wind turbine, simulations have been carried out for a various number of environmental conditions. The goal is to be able to learn data representations for different states of the wind turbine from the simulated data, which would not be possible on real measurements. Moreover, as the simulations can be carried out with a very high number of different sensor outputs, the dataset offers a possibility to uncover relations between different measurements / components of the wind turbine, which are not known yet. The dataset consists strictly of highdimensional time series as output of certain setups of a 5MW wind turbine. They represent various measurements of important characteristics, such as acceleration or bending moments, on the blades, the nacelle, or different parts of the wind turbine. The dataset is created to resemble different locations and degrees of mass imbalance on one of the three wind turbine blades.
Technical Information
The dataset consists of 16 different files, which all correspond to turbulent wind fields, which range between 5 and 20 m/s mean total wind speed (at the reference height 10m). Those files are all packed in a tar.gz file. Upon extracting these files for each wind speed one will obtain hdf5-files consisting of the simulation data together with their meta information corresponding to wind files, which have an angle of attack ranging from -45 ° to 45 ° with respect to the wind turbine plane. For each of those files, there exist simulations for various mass distributions of blade 1 (the mass distribution of blade 2 and blade 3 stay the same as per default). The setup for the mass distribution is as follows: the blade is divided into three zones, the first half, as well as the third and the last quarter. On each of these three zones, the mass is increased up to 4 % in steps of 0.2 %. Each of these individual simulations has been carried out for a runtime of 10 minutes for 47 different output signals, where there are 160 measurements per seconds. The shape of a regular dataset is 96001 rows and 47 columns. One hdf5-file can be read with the h5py-module in python. One file sample.h5 can then be read via hf= h5py.File('sample.h5,'r'). Its meta-information can then be seen with hf.keys(). Among others, these keys contain the simulation data and the channel names. A list of the channel names is included in this repository named 'OutListParameters.xlsx'. They can be read out frmo the hdf5-file via chan = hf.get('channels')
chans = np.asarray(chan['channels']) Similarly, get the data with dat = hf.get('data') . Then the keys dat.keys() give information on the mass distributions available in this dataset. Read the data with mass distribution xxx out with dat_xxx = np.asarray(dat['xxx']) . Note that not all files may contain the same set of wind angles, as some of the speed and wind combinations led to numerical instabilites in the simulator. Also note that the nominal wind speed is at reference heigth 10m, hence the simulated wind speed may differ from the nominal speed slightly.
