An Improved Feature Selection Method Based on Random Forest Algorithm for Wind Turbine Condition Monitoring.

Area of Science:

• Engineering
• Data Science
• Renewable Energy

Background:

• Effective wind turbine condition monitoring relies on robust feature selection and dimensionality reduction from SCADA data.
• Existing methods may not optimally identify features critical for accurately assessing turbine operational status.

Purpose of the Study:

• To propose an improved random forest algorithm, Feature Simplification Random Forest (FS_RF), for enhanced feature selection in wind turbine SCADA data.
• To validate the effectiveness of FS_RF in identifying features highly correlated with wind turbine working conditions.

Main Methods:

• Development of the Feature Simplification Random Forest (FS_RF) algorithm, incorporating Euclidean distances to weigh feature importance across samples.
• Application of the random forest algorithm to measure feature importance.
• Validation of selected features using a two-layer gated recurrent unit (GRU) neural network for condition monitoring.

Main Results:

• The FS_RF algorithm successfully identified features closely correlated with wind turbine working conditions.
• The selected features, when used with a GRU network, demonstrated effective wind turbine condition monitoring capabilities.
• Experimental results confirmed the capacity and effectiveness of the proposed FS_RF method.

Conclusions:

• The proposed FS_RF method offers a significant improvement for feature selection in wind turbine condition monitoring.
• This approach enhances the performance and reliability of SCADA data-driven monitoring models.
• FS_RF provides a valuable tool for optimizing wind turbine operational health assessment.