Presentations

Abstract

Data-driven models using SCADA data have proved useful for detecting some component issues such as upcoming failures, ventilation issues, sensor issues or lubrification problems. These approaches can present owners and operators with multiple issues that they will need to prioritize. The objective of this presentation is to share recent work on estimating the time between a temperature anomaly detection and the exceedance of critical SCADA thresholds leading to turbine shut down or component failure. This time metric allows operators and owners to prioritize, among all turbines requiring attention for components at risk, which one(s) to tackle first.  We have approached the problem by building a testing, training and validation dataset, and then assessing several data-driven methods.  This study is distinguishes itself by its extensive dataset, encompassing actual overtemperature-related downtimes and failures from around 200 turbines of varying sizes and OEMs across multiple different global locations, over two years of data. This diversity not only enhances the robustness of our models and validations, but also addresses various distinct real-life scenarios. Other recent literature seems to focus mostly on small-scale datasets and few failures, for example the interesting study from Vieira, J.L.d.M et al., Remaining Useful Life Estimation Framework for the Main Bearing of Wind Turbines Operating in Real Time. Energies 2024, 17, 1430. https://doi.org/10.3390/en17061430  After building the dataset, we have started by assessing a simple linear regression approach from which we could then quantify the added value of more complex methods.   From there, and in partnership with a PhD student from our academic partner École de Technologie Supérieure in Montréal, Canada, we have assessed the use of AutoRegressive Integrated Moving Average (ARIMA), and Non-linear AutoRegressive with eXogenous inputs Neural Networks (NARXNN), coupled Multi-Layer Perceptron (MLP) methods.  In parallel, our Advanced Analytics and Research department has been evaluating the benefits of Survival analysis to predict the time until temperature thresholds are exceeded, tracking component temperatures alongside covariate signals such as ambient temperature, power generated, accumulated energy production, and wind speed. Cross-validation was used to select the best features. The presentation will focus on:  * Providing an overview of the general context and added value of estimating time to temperature exceedance * Summarizing our work around the validation dataset and proving statistics on the time between a initial anomaly detection and the subsequent downtime/failure * Presenting the different assessment methods, results and validation work * Communicating the main challenges and next steps in our work By sharing this research, we aim to increase the industry's ability to predict turbine component failures. Accurate estimates of time until failure or temperature exceedance are crucial for planning maintenance and avoiding failures, especially as turbines approach the end of their operational life. Our study underlines the potential of using SCADA data and advanced analytics in predictive maintenance, contributing to extending asset lifetimes and optimizing resource allocation in the wind energy sector. The key innovative content of this proposed presentation comes from:  * The extensiveness of our validation dataset  * The variety of the approaches evaluated