Advanced Multi-Scale CNN-BiLSTM for Robust Photovoltaic Fault Detection

Area of Science:

• Renewable Energy Systems
• Artificial Intelligence in Engineering
• Signal Processing and Machine Learning

Background:

• Photovoltaic (PV) systems require reliable fault detection for operational safety and efficiency.
• Existing methods face limitations in noisy, incomplete, and imbalanced industrial data.
• Need for advanced algorithms to address complex fault detection challenges in PV systems.

Purpose of the Study:

• To propose an innovative Advanced CNN-BiLSTM architecture for enhanced PV fault detection.
• To improve accuracy and robustness in complex industrial environments.
• To establish a new paradigm for temporal feature fusion in renewable energy fault detection.

Main Methods:

• Developed an Advanced CNN-BiLSTM framework integrating multi-scale feature extraction (CNNs with kernel sizes 3, 7, 15, 31) and hierarchical attention.
• Employed an adaptive feature fusion network with multi-head attention to combine multi-scale features.
• Utilized a two-layer bidirectional LSTM with temporal attention for final fault classification.

Main Results:

• Achieved 83.25% accuracy under extreme industrial conditions, a 119.48% relative improvement over baseline CNN-BiLSTM (37.93%).
• Multi-scale CNNs contributed 28.0% and adaptive feature fusion 22.0% to performance gains.
• Demonstrated superior robustness against severe noise (σ = 0.20), missing data (15%), and outliers (8%).

Conclusions:

• The proposed Advanced CNN-BiLSTM architecture offers significant improvements in PV fault detection accuracy and reliability.
• The framework's robustness makes it highly suitable for real-world industrial deployment.
• This study sets a new standard for temporal feature fusion in renewable energy fault detection.