Abstract
Photovoltaic (PV) systems suffer significant performance degradation from shadows, yet current monitoring systems cannot distinguish between static shadows requiring physical intervention and temporary dynamic cloud shadows. This misclassification causes unnecessary maintenance dispatches, costing millions annually. This thesis presents an intelligent discrimination framework using convolutional neural networks (CNN) to classify shadow types based on temporal electrical patterns.A MATLAB/Simulink simulation platform was developed to generate realistic PV responses under various shading conditions. The framework models six static shadow patterns (localized blocks, horizontal strips, uniform coverage, gradients, spots) and stochastic cloud movements, producing 12,000 labeled training samples. This addresses the critical challenge of obtaining labeled data for PV fault detection.
The proposed CNN architecture processes 10-minute sequences of electrical characteristics sampled at 1 Hz. The optimized three-layer architecture (32→64→128 filters) contains only 285K parameters, enabling edge deployment. Comprehensive validation demonstrates 96.8% classification accuracy, outperforming traditional machine learning by 7.3%. The system maintains >94% accuracy under 20 dB signal-to-noise ratio and >93% with 10% missing data.
This research establishes intelligent shadow discrimination as viable for operational PV systems. By accurately distinguishing static faults from dynamic shadows, the framework enables targeted maintenance strategies, reducing operational costs while improving system reliability. The work contributes to sustainable energy development by enhancing PV system efficiency and demonstrates how deep learning can solve practical challenges in renewable energy infrastructure.
| Date of Award | 2026 |
|---|---|
| Original language | English |
| Awarding Institution |
|
| Supervisor | Changying XIANG (Supervisor) |
Cite this
- Standard