Real-time power quality monitoring and control using FPGA and machine learning
Journal of Engineering Research (Kuwait), 2026 (SCI-Expanded, Scopus)
- Yayın Türü: Makale / Tam Makale
- Basım Tarihi: 2026
- Doi Numarası: 10.1016/j.jer.2026.10.053
- Dergi Adı: Journal of Engineering Research (Kuwait)
- Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Arab World Research Source, Directory of Open Access Journals, Academic Search Ultimate (EBSCO)
- Anahtar Kelimeler: Fault classification, FPGA, Gradient Boosting Decision Tree, Hybrid PV–grid system, Power quality monitoring, PWM inverter control, Support Vector Machine, Vitis HLS
- Kayseri Üniversitesi Adresli: Evet
Özet
Power quality monitoring and smart fault diagnosis are very important for keeping modern hybrid photovoltaic (PV)–grid systems running reliably. Adding renewable energy and power electronic converters to these systems creates important operating challenges. Most existing machine-learning solutions, however, remain software-based and do not combine real-time hardware acceleration with inverter control. This paper presents an FPGA-based framework that combines power quality monitoring, inverter control, and intelligent fault classification for a grid-connected photovoltaic system on a Xilinx Zynq-7000 (xc7z020clg400-1) device. The system is modeled in MATLAB/Simulink, and two complementary hardware modules, a PWM controller for the grid-connected inverter and a Gradient Boosting Decision Tree (GBDT) accelerator for multiclass fault classification, are implemented with Vivado and Vitis HLS and verified separately on the same device. A 172-dimensional feature vector combining statistical, frequency-domain, harmonic, symmetrical-component, and power-quality descriptors was extracted offline from the three-phase voltage and current signals under one normal and eleven fault conditions. The proposed GBDT classifier, trained on only the 60-dimensional statistical feature group, achieved an overall classification accuracy of 98.35%, exceeding the 95.95% obtained by an Enhanced SVM with a radial basis function kernel trained on the complete 172-dimensional feature vector and the 79.58% of a Hybrid SVM based on GBDT leaf encoding. The implemented accelerator occupies 3.81% of the available LUTs, consumes 1.416 W, meets all timing constraints at 100 MHz with an inference interval of approximately 0.20μs and was functionally validated on a physical PYNQ development board in a host-in-the-loop configuration. These results indicate that a compact statistical descriptor set is sufficient for accurate multiclass fault classification, and that the classifier and the PWM controller, implemented separately, together occupy a small fraction of a single low-cost Zynq-7000 device.