Continuous power quality monitoring allows grid stakeholders to obtain information about the performance of the network and costumer facilities. Moreover, the analysis of continuous monitoring allows researchers to obtain knowledge on power quality phenomena. Power quality measurements result in a large amount of data. Power quality data can be classified as big data, not only for its volume, but also for the other complexities: velocity, variety, and veracity. Manual analysis of power quality data is possible but time-consuming. Moreover, data reports based on standardized indexes and classical statistical techniques might hide important information of the time-varying behaviour in power quality measurements. Artificial intelligence plays a role in providing automatic tools for proper analytics of big data.  A subset of artificial intelligence called machine learning has enabled computers to learn without explicit programming. Driven by the huge improvements in computer processing, a subset of machine learning based on multiple layers of artificial neural networks has been developed to tackle increasingly complex problems. The so-called deep learning applications teach themselves to perform a specific task by automatically extracting essential features from the raw data. Despite the possibilities of automatic feature extraction, most applications of deep learning to power quality are still the same as expert systems or earliest machine learning tools. Moreover, most of the applications are based on synthetic generated data and supervised techniques. In this context, the main motivation of this thesis is providing a new tool based on unsupervised deep learning to handle analytics of time-varying power quality big data. 

The unsupervised deep learning method proposed in this thesis combines a deep autoencoder with clustering for extracting patterns and anomalies in power quality big data. The deep autoencoder maps the original data to a compressed format that contains the principal features of the data. Automatic results are provided by the deep learning, and inferences can be obtained without requiring prior knowledge of deep learning. The outputs from unsupervised deep learning can serve as a guide for further data analysis, highlighting important time steps within large power quality datasets. By following these indications from the deep learning results, experts gain valuable insights into power quality phenomena, which can be referred to as "learning from deep learning". The interpretation of the deep learning results in this thesis allowed to making proper inferences for patterns and anomalies. For power quality measurements synchronised with 24-h, the results allowed making inferences concerning daily variations, seasonality, and the origins of power quality disturbances. For power quality measurements non-synchronised with 24-h, the results could be interpreted visually through the distribution of the patterns in a physical variable, such as the dynamic operating conditions of an electrical railway power system.

An important contribution of this thesis concerns the physical interpretation of the phenomena is related to the anomalies in harmonics caused by geomagnetically induced currents. An interesting finding by applying the deep anomaly detection to measurements in the Swedish transmission grid is the damping of the anomalies caused by geomagnetically induced currents in the winter due to the heating load. This thesis also demonstrated that the signatures for anomalies in harmonic measurements in a Swedish transmission location are similar to the ones found in a low-latitude transmission location at the South Atlantic Anomaly due to geomagnetically induced currents. Moreover, by cross-checking the anomalies at the South Atlantic Anomaly with protection trips with undetermined causes, this thesis demonstrated that anomaly harmonics due to geomagnetically induced currents can cause protection mal trips.

This thesis demonstrates that unsupervised deep learning can serve as an additional tool for compressing time-varying power quality big data into a more interpretable form. Despite the application of an unsupervised method, power quality experts remain significant in power quality studies. The main conclusion of this thesis is that unsupervised deep learning enhances the understanding of power quality experts and provides a complementary approach for analysing and extracting insights from time-varying power quality big data.