RUL forecasting for wind turbine predictive maintenance based on deep learning.

Area of Science:

• Engineering
• Computer Science
• Renewable Energy

Background:

• Wind farm operation and maintenance (O&M) costs are significant.
• Current predictive maintenance (PdM) methods struggle with the remote nature of wind farms, limiting practical application.
• Accurate Remaining Useful Life (RUL) forecasting is crucial for effective maintenance scheduling.

Purpose of the Study:

• To introduce a novel deep learning (DL) methodology for accurate RUL forecasting in wind turbines.
• To address the limitations of current PdM approaches in remote wind farm settings.
• To enable practical implementation of PdM by providing a reliable advance maintenance window.

Main Methods:

• Developed a multi-parametric attention-based deep learning approach, bypassing traditional feature engineering.
• Proposed two models: ForeNet-2d and ForeNet-3d for RUL prediction.
• Validated the models on forecasting RUL for seven types of wind turbine failures.

Main Results:

• Achieved a 2-week forecast window for RUL.
• Demonstrated high accuracy, with the most precise forecast deviating by only 10 minutes from the actual RUL.
• The least accurate prediction deviated by 1.8 days, with most predictions within a few hours of the actual RUL.

Conclusions:

• The proposed DL methodology significantly enhances the reliability and practicality of PdM for wind farms.
• The accurate RUL forecasts provide a substantial time frame for maintenance, crucial for remote wind turbine access.
• This approach minimizes human error through automated feature extraction, paving the way for efficient wind farm O&M.