Improved wind turbine monitoring using operational data

With wind energy becoming a major source of energy, there is a pressing need to reduce all associated costs to be competitive in a market that might be fully subsidy-free in the near future. Before thousands of wind turbines were installed all over the world, research in e.g. understanding aerodynamics, developing new materials, designing better gearboxes, improving power electronics etc., helped to cut down wind turbine manufacturing costs. It might be assumed, that this would be sufficient to reduce the costs of wind energy as the resource, the wind itself, is free of costs. However, it has become clear that the operation and maintenance of wind turbines contributes significantly to the overall cost of energy. Harsh environmental conditions and the frequently remote locations of the turbines makes maintenance of wind turbines challenging. Just recently, the industry realised that a move from reactive and scheduled maintenance towards preventative or condition-based maintenance will be crucial to further reduce costs. Knowing the condition of the wind turbine is key for any optimisation of operation and maintenance. There are various possibilities to install advanced sensors and monitoring systems developed in recent years. However, these will inevitably incur new costs that need to be worthwhile and retro-fits to existing turbines might not always be feasible. In contrast, this work focuses on ways to use operational data as recorded by the turbine's Supervisory Control And Data Acquisition (SCADA) system, which is installed in all modern wind turbines for operating purposes -- without additional costs. SCADA data usually contain information about the environmental conditions (e.g. wind speed, ambient temperature), the operation of the turbine (power production, rotational speed, pitch angle) and potentially the system's health status (temperatures, vibration). These measurements are commonly recorded in ten-minutely averages and might be seen as indirect and top-level information about the turbine's condition. Firstly, this thesis discusses the use of operational data to monitor the power performance to assess the overall efficiency of wind turbines and to analyse and optimise maintenance. In a sensitivity study, the financial consequences of imperfect maintenance are evaluated based on case study data and compared with environmental effects such as blade icing. It is shown how decision-making of wind farm operators could be supported with detailed `what-if' scenario analyses. Secondly, model-based monitoring of SCADA temperatures is investigated. This approach tries to identify hidden changes in the load-dependent fluctuations of drivetrain temperatures that can potentially reveal increased degradation and possible imminent failure. A detailed comparison of machine learning regression techniques and model configurations is conducted based on data from four wind farms with varying properties. The results indicate that the detailed setup of the model is very important while the selection of the modelling technique might be less relevant than expected. Ways to establish reliable failure detection are discussed and a condition index is developed based on an ensemble of different models and anomaly measures. However, the findings also highlight that better documentation of maintenance is required to further improve data-driven condition monitoring approaches. In the next part, the capabilities of operational data are explored in a study with data from both the SCADA system and a Condition Monitoring System (CMS) based on drivetrain vibrations. Analyses of signal similarity and data clusters reveal signal relationships and potential for synergistic effects of the different data sources. An application of machine learning techniques demonstrates that the alarms of the commercial CMS can be predicted in certain cases with SCADA data alone. Finally, the benefits of having wind turbines in farms are investigated in the context of condition monitoring. Several approaches are developed to improve failure detection based on operational statistics, CMS vibrations or SCADA temperatures. It is demonstrated that utilising comparisons with neighbouring turbines might be beneficial to get earlier and more reliable warnings of imminent failures. This work has been part of the Advanced Wind Energy Systems Operation and Maintenance Expertise (AWESOME) project, a European consortium with companies, universities and research centres in the wind energy sector from Spain, Italy, Germany, Denmark, Norway and UK. Parts of this work were developed in collaboration with other fellows in the project (as marked and explained in footnotes).