Towards digital twins: wind farm operation analysis and optimization using model-supported data analytics

For the past decades wind turbine technology has significantly evolved and wind energy became a mainstream energy resource. However, further optimization and cost reduction of the wind energy systems is required in order to meet the global demand for sustainable energy. Expenses related to Operation and Maintenance (O&M) activities of modern wind farms is one of the main cost drivers.

Therefore, more advanced wind farm monitoring techniques are necessary to optimize O&M strategies, and to move from corrective towards predictive maintenance. In this context, so-called digital twin technologies are being developed which aim to combine various data sources from the systems to monitor their behaviour throughout the whole lifetime and improve their operation based on the gathered information.

The main purpose of the thesis is to develop analytical methods for supporting digital twin applications that can be used to analyze and optimize the operation of wind farms. For this purpose statistical methods and machine learning approaches are used to combine mathematical models of the physical system with operational data of an offshore wind farm.

The thesis aims at developing analytical methods for improving wind farm O&M knowledge and practices by (1) improving surrogate models for load estimations, (2) suggesting a method for finding connections between loads and failure of turbines in wind farms, and (3) investigating how physics constraints can be included in data-driven monitoring models and analysing the implications on the model performance and anomaly detection.

Firstly, the thesis analyses surrogate models which can be used to replace computationally expensive aeroelastic simulations and to compute simplified load estimations of turbines. A benchmark is conducted to compare the use of artificial neural networks (ANNs) against commonly used methods which have proven to be successful in previous research.

It is found that ANNs outperform the other methods in terms of computational time, model accuracy and improved robustness of model accuracy when using less simulation samples for the model training. Furthermore, the uncertainty propagation through the ANN based surrogate is analysed and a sensitivity study is carried out to quantify the influence of the selected inputs on the variance of the load estimations.

Secondly, a methodology for comparing wake-induced fatigue loads with turbine failures is suggested. This can be used to map the variations of fatigue loading, performance and estimated lifetime within a wind farm including different operational states of the turbines (i.e. normal operation, start-up, shutdown).

For this a surrogate model is calibrated on a large simulation database considering multiple wake conditions with up to four upstream turbines as wake sources. The site-specific estimations can be made by using the wind field distributions obtained from the wind farms’ SCADA data. This methodology can help to better define a relationship between loads and reliability of turbines in a wind farm.

It is presented and discussed on an example offshore wind farm with recorded main bearing failures. Finally, the benefits of including physical information into data-driven monitoring models of turbines in a wind farm are investigated. Previous research has shown that machine learning methods which are based on SCADA data of operating turbines bear large potential for monitoring their performance and detecting abnormal behaviour that can indicate a fault or failure.

This thesis introduces a transfer learning method where a normal behaviour model is pre-trained on aeroelastic simulations and recalibrated on SCADA data. It is demonstrated that when only one month of SCADA data is available, adding aeroelastic simulations to the training process can improve the model performance and precision of the anomaly detection.

The developed methods and findings of this work can potentially be used by wind farm operators to leverage the use of operational data and simulated data for getting useful insights and moving towards more advanced O&M strategies.