Machine Learning Algorithms for Wind Turbine Power Performance Monitoring – Tracking optimal yaw alignment based on SCADA data

Introduction

Stakeholders in wind energy are interested in optimal wind farm operation. Key to an optimal operational wind farm performance is high availability and good quality of power performance in operation.
Wind turbines provide a large amount of data which is regularly used for availability calculation and ad-hoc analyses of specific questions.
An increasing number of wind farm operators as well as consultants are analysing the performance of wind turbines based on SCADA data. Occasionally results show sub-optimal performance. There are multiple reasons why an OEM cannot directly apply parameter changes concluded from third party analysis: contractual arrangements, uncertain technical expertise of consultant or unclear methodologies. Hence, time-consuming validation of these results by the OEM is necessary, which causes delays and may result in yield loss.
To obtain good and reliable results, a deep understanding of turbine technology is required for modelling. Machine learning helps to create more sophisticated models. These algorithms create knowledge based on experience and permanently update this knowledge. In general, the more data available and taken into account, the better the model and results. OEMs have a large amount of data and possess deep knowledge of their turbines ideally positioning them to implement machine learning techniques.
Machine learning aggregates different statistical approaches for supervised and unsupervised learning based on large datasets. This methodology is already applied in wind energy, e.g. for predictive maintenance. Similar to predictive maintenance, we show that machine learning can also be applied to maintain optimal turbine performance.


Approach

A combination of different data sources are used in an ensemble of machine learning algorithms. 10 minute averaged sensor data, wind turbine parameters and operation counters are aggregated and used to calculate key performance indicators for wind turbines. Instead of using a single set of turbine data, a data set of the entire fleet of identical turbine types is taken into account. Fleet characteristics are established and wind turbines are regularly investigated by the algorithm. Turbines showing suspicious operational characteristics are checked for validity in the field and the calculated key performance indicators are added to the training set. This way, machine learning algorithms are trained on an increasing size of data sets and results improve over time.

Main body of abstract

Turbine yaw misalignment is seen as one of the largest sources of power losses during operation. Detection of turbines with yaw errors in the field is not trivial. Typically there is no met mast and thus no reliable reference for wind direction.
In a wind farm with twelve turbines a yaw error was manually triggered on two turbines for a fixed period of time shorter than a month. The presented model was able to successfully detect the turbines with yaw error based on the described approach. The number of training sets as well as the test data is presented and a validation case is demonstrated.
Further case studies show detection of e.g. sensor faults and changes in parameters.


Conclusion

OEMs possesses large amounts of data and a deep understanding of their turbine technology. Both was used to develop statistical models for wind turbine power performance monitoring. Model precision increases with increasing number of training data. The presented methods were able to learn autonomously. This analysis process is highly automatable and can be easily applied to new sites. Corrective actions can be taken immediately – no third party is involved and no additional check of analysis required. Our results indicate that it is possible to detect errors in an early phase to maintain optimal performance.


Learning objectives
Monitoring an operational wind farm with respect to maintain optimal turbine power performance. Turbine yaw misalignments can be detected by machine learning algorithms.