Adaptive PolyKAN-based autoencoder for fault detection and classification in wind and solar power systems

This paper presents an advanced fault diagnosis framework for renewable energy systems by leveraging a novel Adaptive Polynomial Kolmogorov Arnold Network (Adaptive PolyKAN). The proposed method is evaluated on two distinct applications: a wind energy conversion system and a grid-connected photovoltaic (PV) system, each characterized by complex, nonlinear fault patterns. A comprehensive comparison is conducted against a range of classical and neural classifiers, including Random Forest (RF), Support Vector Machine (SVM), and others. Experimental results demonstrate that Adaptive PolyKAN consistently achieves superior classification accuracy, reaching 99.96 % for wind data and 95.61 % for PV data, outperforming conventional methods across all performance metrics. To improve computational efficiency, an autoencoder-based dimensionality reduction strategy is incorporated, resulting in a reduction of execution time by over 88 % and memory usage by 40 %, while preserving high diagnostic accuracy, maintaining 99.96 % on the wind data and increasing to 96.47 % on the PV data. The results confirm the robustness, adaptability, and efficiency of the proposed framework, highlighting its potential for intelligent fault diagnosis in complex renewable energy systems.