Wind Turbine Performance Monitoring using Artificial Neural Networks With a Multi-Dimensional Data Filtering Approach

Abstract

The wind power sector has grown rapidly and has become a substantial part of the global sustainable energy production. Performance and condition monitoring systems are gaining ground, but most faults are still detected during planned maintenance. This can lead to long time periods of underperformance, which translates to lost revenues. In this thesis, Artificial Neural Networks (ANN) are used to model the normal behaviour of a wind turbine, which could be used for real-time monitoring of operations. A number of other studies that use ANN’s to predict wind power output were found during the literature study; but this thesis presents a new direction where the standard deviation of the wind speed is used as an input to the model, as well as a multi-dimensional filtering method, meant to exclude outliers in the training set with higher accuracy than conventional filtering techniques. The study follows the method of (Schlechtingen, et al., 2013a), who made a comparative study of different data-mining approaches, to be able to compare the model results. The proposed model shows an improvement in prediction performance of between 16 % and 22 %, depending on performance parameter. The results from the multi-dimensional filtering shows that unhealthy data situated inside what is conventionally thought of as normal operating range can be excluded with the proposed method. It is concluded that the model is well suited for performance monitoring, but its applicability to fault prediction could ultimately not be concluded due to a lack of suitable faults during the period. Finally, it is concluded that if the proposed model had been used for performance monitoring in the turbine that was the main subject in this study, earlier maintenance could have resulted in an additional electricity generation of up to 270 MWh during the three years of data used.