It is well known that each year the wind sector has profit losses due to wind turbine failures and operation and maintenance costs. Therefore, operations related to these actions are crucial for wind farm operators and linked companies. One of the key points for failure prediction on wind turbine using SCADA data is to select the optimal or near optimal set of inputs that can feed the failure prediction (prognosis) algorithm. Due to a high number of possible predictors (from tens to hundreds), the optimal set of inputs obtained by exhaustive-search algorithms is not viable in the majority of cases. In order to tackle this issue, show the viability of prognosis and select the best set of variables from more than 200 analogous variables recorded at intervals of 5 or 10 min by the wind farm’s SCADA, in this paper a thorough study of automatic input selection algorithms for wind turbine failure prediction is presented and an exhaustive-search-based quasi-optimal (QO) algorithm, which has been used as a reference, is proposed. In order to evaluate the performance, a k-NN classification algorithm is used. Results showed that the best automatic feature selection method in our case-study is the conditional mutual information (CMI), while the worst one is the mutual information feature selection (MIFS). Furthermore, the effect of the number of neighbours (k) is tested. Experiments demonstrate that k = 1 is the best option if the number of features is higher than 3. The experiments carried out in this work have been extracted from measures taken along an entire year and corresponding to gearbox and transmission systems of Fuhrländer wind turbines.

In this paper, a method to build models to monitor and evaluate the health status of wind turbines using Single-hidden Layer Feedforward Neural networks (SLFN) is presented. The models are trained using the Extreme Learning Machines (ELM) strategy. The data used is obtained from the SCADA systems, easily available in modern wind turbines. The ELM technique requires very low computational costs for the training of the models, and thus allows for the integration of a grid-search approach with parallelized instances to find out the optimal model parameters. These models can be built both individually, considering the turbines separately, or as an aggregate for the whole wind plant. The followed strategy consists in predicting a target variable using the rest of the variables of the system/subsystem, computing the error deviation from the real target variable and finally comparing high error values with a selection of alarm events for that system, therefore validating the performance of the model. The experimental results indicate that this methodology leads to the detection of mismatches in the stages of the system’s failure, thus making it possible to schedule the maintenance operation before a critical failure occurs. The simplicity of the ELM systems and the ease with which the parameters can be adjusted make it a realistic option to be implemented in wind turbine models to work in real time.

Background: Identifying the health status of wind turbines becomes critical to reduce the impact of failures on generation costs (between 25–35%). This is a time-consuming task since a human expert has to explore turbines individually. Methods: To optimize this process, we present a strategy based on Self Organizing Maps, clustering and a further grouping of turbines based on the centroids of their SOM clusters, generating groups of turbines that have similar behavior for subsystem failure. The human expert can diagnose the wind farm health by the analysis of a small each group sample. By introducing post-processing tools like Class panel graphs and Traffic lights panels, the conceptualization of the clusters is enhanced, providing additional information of what kind of real scenarios the clusters point out contributing to a better diagnosis. Results: The proposed approach has been tested in real wind farms with different characteristics (number of wind turbines, manufacturers, power, type of sensors, ...) and compared with classical clustering. Conclusions: Experimental results show that the states healthy, unhealthy and intermediate have been detected. Besides, the operational modes identified for each wind turbine overcome those obtained with classical clustering techniques capturing the intrinsic stationarity of the data.

With the aim of helping researchers to develop intelligent operation and maintenance strategies, in this manuscript, an extensive 3-years Supervisory Control and Data Acquisition database of five Fuhrländer FL2500 2.5 MW wind turbines is presented. The database contains 312 analogous variables recorded at 5-minute intervals, from 78 different sensors. The reported values for each sensor are minimum, maximum, mean, and standard deviation. The database also contains the alarm events, indicating the system and subsystem and a small description. Finally, a set of functions to download specific subsets of the whole database is freely available in Matlab, R, and Python. To demonstrate the usefulness of this database, an illustrative example is given. In this example, different gearbox variables are selected to estimate a target variable to detect whether or not the estimate differs from the actual value provided for the sensor. By using this normality modelling approach, it is possible to detect rotor malfunction when the estimate differs from the actual measured value.

Detecting and determining which systems or subsystems of a wind turbine have more failures is essential to improve their design, which will reduce the costs of generating wind power. Two of the most critical failures, the generator and gearbox, are analyzed and characterized with four metrics. This failure analysis usually begins with the identification of the turbine’s condition, a process normally performed by an expert examining the wind turbine’s service history. This is a time-consuming task, as a human expert has to examine each service entry. To automate this process, a new methodology is presented here, which is based on a set of steps to preprocess and decompose the service history to find relevant words and sentences that discriminate an unhealthy wind turbine period from a healthy one. This is achieved by means of two classifiers fed with the matrix of terms from the decomposed document of the training wind turbines. The classifiers can extract essential words and determine the conditions of new turbines of unknown status using the text from the service history, emulating what a human expert manually does when labelling the training set. Experimental results are promising, with accuracy and F-score above 90% in some cases. Condition monitoring system can be improved and automated using this system, which helps the expert in the tedious task of identifying the relevant words from the turbine service history. In addition, the system can be retrained when new knowledge becomes available and may therefore always be as accurate as a human expert. With this new tool, the expert can focus on identifying which systems or subsystems can be redesigned to increase the efficiency of wind turbines.