Deep Learning Powered Wind Farm Data Analytics


Student thesis: Doctoral Thesis

View graph of relations


Related Research Unit(s)


Awarding Institution
Award date31 Jan 2023


The proportion of wind energy in the worldwide energy portfolio expands continuously due to the mission of the carbon neutrality in 2050. Meanwhile, the rapid growth of installed wind turbines (WTs) also boosts the demand for more efficient and effective WT operations and maintenance (O&M) technologies which can reduce the wind farm O&M cost. WTs as emerging energy systems are fully instrumented. Supervisory control and data acquisition (SCADA) systems have been deployed in most of commercial wind farms to continuously collect data from sensors of WTs, which can be considered as the internet-of-things platform. By analyzing the massive volume of data collected from wind farm SCADA systems, more advanced O&M technologies can be developed, which could significantly promote the efficiency of wind farms.

Today, the deep learning technology has shown its great power in function fitting and have already been successfully applied to many computer science tasks. In this dissertation, focusing on developing more advanced O&M technologies via analyzing the wind farm data based on deep learning methods, four issues from two significant topics, the WT condition monitoring and the wind speed prediction (WSP), are investigated. First, as commercial WTs are typically working at the harsh environment, the aging of WTs can increasingly induce various subsystem faults or failures. Studying advanced methods for monitoring the WT blade breakage, which is one of the most severe failures, is significant and valuable. Second, the gearbox is another critical component of WTs, which adopts different failure features from the blade. Thus, effective methods for accurate gearbox condition monitoring are highly interesting and beneficial to wind industry. Third, due to the stochastic nature of wind speed, the wind speed acts as a critical role in determining the wind power generation as well as in unit commitment planning, power system dynamics estimation, and maintenance scheduling considering wind energy systems. It is important to propose a more accurate deterministic wind speed prediction (DWSP) method to address challenges. Finally, predicting the wind speed uncertainty also plays a critical role in both the wind industry and the power grid. Studying probabilistic wind speed prediction (PWSP) is valuable for the reliability and efficiency of the power grid.

To address the aforementioned issues, deep learning models based on wind farm SCADA data are presented in this dissertation. In the WT blade breakage detection problem, a conditional convolutional autoencoder (CCAE) based method is proposed for detecting impending WT blade breakages in advance. To develop a more effective deep learning based architecture for WT gearbox failure detection, a joint variational autoencoder (JVAE) based deep framework is presented. In DWSP problem, a novel deep attention convolutional recurrent network with k-Shape and enhanced memory (DACRN-KM) is developed for offering a more accurate DWSP. To achieve more reliable PWSP, a deep architecture improved mixture density network via wasserstein distance-based adversarial learning (WA-IMDN) is proposed.

The effectiveness of the presented deep learning models in this dissertation has been verified based on real-world wind farm datasets.