Predictive Maintenance of Wind Generators based on AI Techniques

Abstract

As global warming is slowly becoming a dangerous reality, governments and private institutions are introducing policies to minimize it. Those policies have led to the development and deployment of Renewable Energy Sources (RESs), which introduces new challenges, among which the minimization of downtime and Levelised Cost of Energy (LCOE) by optimizing maintenance strategy where early detection of incipient faults is of significant intent. Hence, this is the focus of this thesis.

While there are several maintenance approaches, predictive maintenance can utilize SCADA readings from large scale power plants to detect early signs of failures, which can be characterized by abnormal patterns in the measurements. There exists several approaches to detect these patterns such as model-based or hybrid techniques, but these require the detailed knowledge of the analyzed system. As SCADA system collects large amounts of data, machine learning techniques can be used to detect the underlying failure patterns and notify customers of the abnormal behaviour.

In this work, a novel framework based on machine learning techniques for fault prediction of wind farm generators is developed for an actual customer. The proposed fault prognosis methodology addresses data limitation such as class imbalance and missing data, performs statistical tests on time series to test for its stationarity, selects the features with the most predictive power, and applies machine learning models to predict a fault with 1 hour horizon. The proposed techniques are tested and validated using historical data for a wind farm in Summerside, Prince Edward Island (PEI), Canada, and models are evaluated based on appropriate evaluation metrics. The results demonstrate the ability of the proposed methodology to predict wind generator failures, and the viability of the proposed methodology for optimizing preventive maintenance strategies.