Presentations | WindEurope Technology Workshop 2024

Follow the event on:

Presentations

Applying wind turbine SCADA data in the early detection of harmful status codes: a novel approach using ANN

Eric Tromeur, Director of Research, Innovation, Service and Expertise, Meteodyn

Abstract

Introduction With the fast development of renewable energy toward the Net-Zeros commitment, especially in wind energy, reducing the maintenance cost of wind turbine becomes more and more important. To anticipate wind turbine failures, several studies proposed installing a condition monitoring system, which is onerous and difficult to be implemented. In this study, an approach using only wind turbine SCADA data is considered using Artificial Neural Network. The objective of the study is to examine the feasibility of using SCADA data to detect the occurrence of harmful status codes before it happens. Approach Data Collection and Preparation The SCADA data of each wind turbine consists of several years of measurements such as temperature, wind speed, power from various sensors and control systems. This data is usually archived as time series,with 10-minute time step. Though, the status codes of the machine bringing information about the operation state are usually archived in the form of an event log. The data preparation consists of formatting, cleansing the SCADA data, converting the status code event log into time series, and synchronizing the 2 datasets. Formulation challenge In the scope of this study, we examine the feasibility of using the SCADA data in the day D-1 to predict the occurrence of harmful status codes in the day D+1. The data model takes the daily SCADA data, each day consisting of 24 input datapoints with 1-hour timestep. The output (label) is then the occurrence (in binary status) of the considered harmful status code. Choice of model: Long Short-Term Memory (LSTM) is chosen as itincorporates additional features to maintain memory of sequential data, allowing them to detect patterns, including anomalies, in predictive maintenance problems. This makes them well-suited for tasks where the order of data points is significant, such as time series analysis. Validation results After several calibrations and enhancements of the prediction model, the validation is made with data from 2016 to 2021 of a wind farm in England with 14 wind turbines, on a harmful status code that has a stop ratio of more than 98%. The training dataset consists of 7842 samples corresponding with 7842 days, the validation dataset consists of 2616 samples. The table below is the validation result: Predicted Negative Predicted Positive Actual Negative 2607 (True Negative) 3 (False Positive) Actual Positive 0 (False Negative) 6 (True Positive) The occurrence of harmful status codes is usually rare. In this case, only 6 times among 2616 samples. All of the 6 are detected by the prediction model. We have 3 cases of false positive (predict that the status code will occur but actually not). The prediction also gives 2607 true negative without any false positive. Conclusions The validation result confirms the possibility of using SCADA data to predict the occurrence of harmful status code, which opens great potential in the field of predictive maintenance. Future work consists of maximizing the number of status codes that could be predicted, with the horizon from 1 week to 1 month (currently 1 day).

Follow the event on:

WindEurope Technology Workshop 2024