A Hybrid – Machine Learning and Possibilistic – Methdology for Predicting Produced Power Using Wind Turbine SCADA Data

Abstract

During its operational lifetime, a wind turbine is continuously subjected to a number of aggressive environmental and operational conditions, resulting in degradation of its parts. If left unattended, these degraded components will negatively influence its performance and may lead to failure of the wind turbine. In order to mitigate the risk associated with the failure of components, a wind turbine is regularly inspected and maintained.

Currently, there are two commonly used approaches for making maintenance management (inspection and maintenance) plans. Traditional Approach utilises understanding of failure profile of the components for manually developing maintenance plan for the equipment. Condition-Based Approach utilises data collected by condition monitoring of equipment for developing dynamic maintenance plan. SCADA system offers a low-resolution condition-monitoring data that can be used for fault detection, fault diagnosis, fault quantification and fault prognosis and eventually for maintenance planning.

The monitoring data from SCADA system of a wind turbine is often afflicted with variability and uncertainty. The variability in data is the result of continuously changing environmental conditions and uncertainty arises due to imperfections in the recorded data. The uncertainty may be due to many reasons, including, inherent characteristic of sensing devices, drift in calibration with time, deterioration of sensing devices’ sensitivity and response due to environmental attacks, etc.

For handling variability in monitoring data a number of parametric and non-parametric (statistical) predictive models for different aspects of a wind turbine’s structure and operation have been proposed. Depending upon its type – aleatory or epistemic – an uncertainty can be handled in a number of ways. Since, the dynamic nature of wind turbine operation does not allow collection of multiple values under the same condition; hence, uncertainty is mostly epistemic in nature. Possibilistic Approach, based on Fuzzy Set Theory, is especially suitable for handling epistemic uncertainty that may arise due to imprecision or lack of statistical data.

Aim of the ongoing research is to develop a methodology for detecting sub-optimal operation of a wind turbine by comparing Measured Produced Power against Predicted Produced Power. Unfortunately, variability and uncertainty associated with the recorded data make accurate prediction of produced power challenging.

This paper presents methodologies for predicting produced power using SCADA data while simultaneously accounting for variability and uncertainty. The methodologies utilise either parametric (example, power curve) or machine learning (example, XGBoost) models for handling variability; and Possibilistic Approach for handling uncertainty.