SVM based imbalanced correction method for Power Systems Transient stability evaluation.

Area of Science:

• Electrical Engineering
• Power Systems Analysis
• Machine Learning Applications

Background:

• Power systems require high safety and reliability, leading to rare unstable samples in real-world data.
• Existing models trained on imbalanced data exhibit bias, often neglecting data quality imbalances.
• Standard evaluation methods prioritize data quantity over quality, compromising model performance.

Purpose of the Study:

• To propose an imbalanced correction method addressing both quantity and quality imbalances in power system data.
• To enhance the reliability and accuracy of power system fault detection models.
• To develop a robust classifier resilient to data quality variations.

Main Methods:

• Utilizing Support Vector Machine (SVM) to calculate classification hyperplanes and normalized Euclidean distances for fault severity assessment.
• Grouping training samples into multilevel sets based on calculated fault severity.
• Pretraining Stacked Sparse Auto-Encoder (SSAE) to quantify inter-class imbalances within multilevel sets.
• Generating a cost-sensitive correction matrix based on multilevel set imbalance information.
• Modifying the SSAE loss function with the correction matrix to create the final classifier.

Main Results:

• The proposed method effectively quantifies and corrects for imbalances in both data quantity and quality.
• Simulation results on the IEEE 39-bus system and a realistic Eastern China power system demonstrate superior performance.
• The developed classifier shows high accuracy and reliability in identifying power system faults.

Conclusions:

• The SVM-based imbalanced correction method significantly improves the performance of power system classifiers.
• Addressing data quality imbalance is crucial for reliable power system operation.
• The integration of SSAE with a cost-sensitive matrix offers a robust solution for imbalanced learning in power systems.