Detection of Wind Turbine Faults Using a Data Mining Approach

Abstract

Wind energy has become one of the leading sources of renewable energy, but faults and unscheduled shutdowns of wind turbines are costly. As the size and number of wind turbines continue to rise, monitoring for faults has become increasingly important for companies to remain competitive. This paper presents a three-phase methodology to detect emerging faults and to possibly reduce the number of unscheduled shutdowns. The first phase monitors the wind turbine online by constructing residual control charts for the turbine’s power curve. Type I errors or type II errors are identified in this phase to avoid false warnings if an unusually high number of observations outside the control limits are present. Given the control charts, the second phase uses autoassociative neural networks (AANNs) to detect anomalous components by comparing the mean square errors (MSEs) of the components. The third phase generates a schedule chart consisting of an anomalous schedule, a preventive schedule, and an ideal schedule. The objective of the schedule chart is to enhance the turbine’s productivity by reducing the number of unexpected shutdowns, with the aim to eventually eliminate them.