In this project, our analysts worked with a wind farm operator to investigate the likely reasons for observed operational reductions in turbine availability and production output during operational wind windows (wind speeds >3ms-1 and <25ms-1) and to explore the potential for developing an AI-based predictive maintenance programme to help:
– reduce maintenance-related downtime disruption,
– minimise maintenance costs,
– providing enhanced lead times to enable timely ordering of parts.
A rich set of operational wind turbine performance data was made available from the operator’s SCADA system. This included error information such as timestamps, duration and severity.
Process
The data allowed the team to investigate the changes in the frequency of the errors over time using the 10-minute window dataset and to evaluate the most common types of errors. This was then compared to 5 years of meteorological data at the same temporal resolution to look for any correlations with factors such as wind speed, wind direction, temperature, and rainfall.
Similarly, power curves for each turbine were analysed for any sign of output loss correlated with the error information. All the data was brought together and visualised through a series of plots and charts, two of which can be seen in the examples below.
Solution
Our analysis demonstrated that the frequency of errors that lead to downtime of the wind turbines had increased from 2015 to 2019, both in total count as well as for all levels of severity. While these errors affect the availability of the turbines, they did not affect the power production to a large extent, with the exception of one turbine. While weather variables were seen to act on all turbines equally, the fact that all turbines present different patterns for the occurrence of errors showed that the errors were independent for each turbine.
Our data scientists also investigated the feasibility of developing a system of two complimentary predictive models to help optimise maintenance scheduling and minimise maintenance costs. One model would focus on the prediction of component-specific failures (given suitable error data to train the model on) whilst the other would predict indiscriminate failures in the turbines based on recent operating performance, faults, and maintenance programmes.
By the end of the project, the team built a successful proof-of-concept which was able to predict significant failures for selected lead times. We also identified how best to improve the outputs by fine-tuning parameters, using different lead times and timeframes for training data, and developing specific models for each turbine or component system.
Impact
Our analysis demonstrated that the frequency of errors that lead to downtime of the wind turbines had increased from 2015 to 2019, both in total count as well as for all levels of severity. While these errors affect the availability of the turbines, they did not affect the power production to a large extent, with the exception of one turbine. While weather variables were seen to act on all turbines equally, the fact that all turbines present different patterns for the occurrence of errors showed that the errors were independent for each turbine.
Our data scientists also investigated the feasibility of developing a system of two complimentary predictive models to help optimise maintenance scheduling and minimise maintenance costs. One model would focus on the prediction of component-specific failures (given suitable error data to train the model on) whilst the other would predict indiscriminate failures in the turbines based on recent operating performance, faults, and maintenance programmes.
By the end of the project, the team built a we built a successful proof-of-concept that was able to predict significant failures for selected lead times. We also identified how best to improve the outputs by fine-tuning parameters, using different lead times and timeframes for training data, and developing specific models for each turbine or component system.
Tell us about your data challenge