A robust multi-location evaluation of a machine learning framework for wind power forecasting.

Area of Science:

• Renewable Energy
• Computational Science

Background:

• Accurate wind power prediction is crucial for consistent energy generation from wind farms.
• Machine learning (ML) offers advanced solutions for improving wind power forecasting reliability and efficiency.

Purpose of the Study:

• To analyze and compare the performance of various ML algorithms for wind power forecasting.
• To identify the most effective ML models for diverse geographical wind farm datasets.

Main Methods:

• Data preprocessing included outlier elimination using Z-score and IQR methods.
• Three ML algorithms (XGBoost, Random Forest Regressor (RFR), and Support Vector Regression (SVR)) with different kernels (RBF, polynomial, linear) were trained and evaluated.
• Performance was assessed using R2 and Mean Absolute Error (MAE) metrics.

Main Results:

• XGBoost achieved R2 values of 0.99 across all locations with MAE ranging from 11.10 to 15.94.
• Linear-kernel SVR also attained R2 values of 0.99 with low MAE on all datasets.
• RFR showed strong performance but had a significantly lower R2 (0.83) and higher MAE (600.81) at one site.

Conclusions:

• XGBoost and linear-kernel SVR are highly effective for wind power forecasting on diverse datasets.
• These models provide high accuracy and low error rates, essential for optimizing wind farm energy production.