Correlation based feature importance analysis for improving machine learning stability predictions in hybrid PV...

Area of Science:

• Electrical Engineering
• Computer Science
• Data Science

Background:

• Accurate grid voltage and stability prediction are crucial for modern power systems, especially with increasing renewable energy integration.
• Existing machine learning models require rigorous evaluation for predicting performance in grid-connected hybrid photovoltaic (PV) systems.

Purpose of the Study:

• To rigorously evaluate five machine learning models (Random Forest, Extra Trees, Support Vector Regression, Cat Boost, and Gradient Boosting) for their predictive performance in grid-connected hybrid PV systems.
• To identify the most accurate and robust model for both grid voltage and stability prediction.

Main Methods:

• A multimetric framework including R², MAE, RMSE, and MAPE was used for evaluation.
• Advanced visual diagnostics such as error distributions and temporal trend analysis were employed.
• A controlled MATLAB/Simulink dataset was generated to capture nonlinear hybrid PV operating regimes.
• Correlation-weighted feature engineering was applied to enhance model interpretability.

Main Results:

• Gradient Boosting (GB) emerged as the top-performing model, demonstrating superior accuracy and robustness.
• For grid voltage prediction, GB achieved R² = 0.9785 and the lowest MAPE = 0.25%.
• For stability score forecasting, GB achieved R² = 0.9300 and the lowest MAE = 0.75, outperforming all other models.

Conclusions:

• Gradient Boosting is a highly accurate and robust solution for smart grid forecasting, offering actionable insights for real-time monitoring and control.
• GB's balanced performance across static and dynamic conditions makes it suitable for resilient grid management in renewable-rich environments.
• The study provides a unified benchmarking of ML models, identifying GB as the most reliable predictor for voltage and stability forecasting.