ADAPTIVE ENERGY ENTROPY WEIGHTED FEATURE RANKING FRAMEWORK FOR HYBRID POWER QUALITY DISTURBANCE CLASSIFICATION USING MULTI-RESOLUTION DWT–ANN
Keywords:
Power Quality Disturbances; Discrete Wavelet Multiresolution Analysis, Energy–Entropy Feature Ranking; Artificial Neural Networks, Feature Optimization; IEEE Std 1159.Abstract
The increasing penetration of renewable energy systems, nonlinear industrial loads, electric vehicle charging infrastructure, and power electronic converters has significantly intensified the occurrence of complex and hybrid power quality disturbances (PQDs) in modern electrical networks. Accurate classification of such disturbances is essential for reliable smart grid monitoring and protection. Although discrete wavelet transform (DWT) combined with artificial neural networks (ANNs) has demonstrated promising performance, most existing approaches rely on fixed statistical features, heuristic mother wavelet selection, and lack systematic feature ranking mechanisms. Furthermore, hybrid disturbances remain insufficiently analyzed in terms of discriminative feature optimization.
This paper proposes an adaptive Energy–Entropy Weighted Feature Ranking (EEWFR) framework integrated with multi-resolution DWT–MRA for intelligent classification of sixteen IEEE Std. 1159 compliant single and hybrid PQ disturbances. Unlike conventional feature fusion approaches, the proposed method introduces an adaptive weighting strategy combining normalized wavelet energy and Shannon entropy across decomposition levels to rank and optimize feature vectors. The ranked feature subset is then evaluated using Multilayer Perceptron (MLP), Radial Basis Function (RBF), and Probabilistic Neural Network (PNN) classifiers.
Extensive simulations conducted at 10 kHz sampling frequency demonstrate that the proposed adaptive feature ranking reduces dimensional redundancy by 48% while improving hybrid disturbance separability. The optimized DWT–PNN model achieves 98.84% overall classification accuracy under clean conditions and maintains strong stability across 20–50 dB SNR levels. Comparative analysis confirms that the proposed framework outperforms conventional DWT-based fixed feature approaches and offers improved computational efficiency compared to deep learning models.
