Optimized Sensor Data Preprocessing Using Parameter-Transfer Learning for Wind Turbine Power Curve Modeling.

Area of Science:

• Renewable Energy Engineering
• Data Science
• Machine Learning

Background:

• Wind turbine power curve modeling is crucial for operational efficiency and maintenance.
• Data anomalies in SCADA (Supervisory Control and Data Acquisition) systems hinder accurate power curve predictions.
• Existing methods struggle with data quality issues and extensive model tuning.

Purpose of the Study:

• To develop a parameter-transfer learning strategy for robust wind turbine power curve modeling.
• To jointly optimize anomaly detection and regression models for improved accuracy.
• To reduce the computational burden of model tuning and enhance applicability to new turbines.

Main Methods:

• A framework combining anomaly detection (iForest, LOF, DBSCAN) and WTPC (Wind Turbine Power Curve) regressors (MLP, RF, GP).
• A multi-metric objective function for joint optimization of preprocessing and modeling.
• Transfer learning with randomized search for source domain hyperparameter exploration and Bayesian optimization for target domain refinement.

Main Results:

• Achieved a 90% reduction in optimization iterations compared to traditional methods.
• Consistently improved target domain performance across various turbine models and locations.
• Demonstrated no performance loss when source and target turbines differed in site or rated power, with larger gains for similar pairs.

Conclusions:

• The proposed parameter-transfer learning pipeline is practical, model-agnostic, and accelerates preprocessing and modeling.
• It effectively preserves or improves model fit, offering significant benefits for newly installed turbines with limited data.
• The strategy enhances the reliability of wind turbine performance monitoring and forecasting.