Data-Driven Anomaly Detection Techniques for Power Generator Engines

Electricity markets operate and evolve in a world where policy and technology are continuously changing, raising the bar for availability and reliability, which are the most relevant key performance indicators of power plants. In this context, Anomaly Detection plays an important role in power generation applications by indicating significant deviations from normal operating conditions to promote proactive maintenance activities, avoiding unnecessary outages and optimizing the efficiency of generation operations. Power generator engines are used as the main power supply for specific power stations. Typically, this machinery utilizes a centralized automation system which gathers the signals from the multiple sensors installed around the equipment to always control and monitor the behavior of power generators. The multiple temporal sequences of data generated can be analyzed for predicting anomalies. In such a framework, this thesis probes a real-world application, focusing on the analysis of multivariate time series data produced by a single gas engine generating set. At the beginning, this study conducts an accurate analysis of the signals generated by the generating set in order to create a dataset suitable to identify anomalous behaviour on the turbocharger’s subsystem. A preliminary preprocessing is required to change the time series resolution and to perform the time alignment of signals. Subsequently, a series of deep learning-based models is proposed capable of learning complex temporal dependencies in the real process signals and detecting anomalies within the proposed datasets. Detecting anomalies in such a real-world scenario becomes challenging, because the behavior of the engines changes as a result of usage and external factors which are not captured by sensor signals. The amount of load on an engine at a time might change frequently or abruptly and, under such settings, it becomes difficult to predict the time-series values for the near future, rendering inefficient the utilization of standard mathematical prediction models that rely on stationary time-series properties solely. The purpose of this research is to unravel the latent complexity of the real data and improve anomaly detection accuracy through the application of sophisticated deep learning models.