A data-driven approach to grid-connected PV system reliability assessment: Combining deep learning and hybrid optimization

Faults in PV systems significantly impact system reliability, leading to inefficiencies, performance degradation, and increased maintenance costs. The proposed model aims to address these issues by focusing on precise and timely fault detection, contributing to the overall operational efficiency of PV systems. This paper introduces a novel approach to enhance the accuracy of fault detection in photovoltaic systems. Key contributions of the study include the implementation of advanced data pre-processing techniques, such as mean imputation for handling missing values, Winsorization for outlier detection, and Min-Max scaling for normalization, ensuring data consistency. Furthermore, the statistical features, including mean, variance, median, and correlation, are extracted from the pre-processed data to capture critical patterns associated with faults. A hybrid feature selection method, combining the Zebra Optimization Algorithm (ZOA) and Red Panda Optimization (RPO), is utilized to identify the most relevant features for fault detection. Since operational attributes and variations in photovoltaic system performance can be reflected by statistical measures like mean and variance, therefore these measures have been explored for feature extraction. For feature selection and extraction related to fault detection, the ZOA algorithm has efficiently searched through the feature space for its key characteristics; the ROA approach has further optimized this selection in terms of feature relevance while minimizing redundancy. The outcome of this integrated approach is that it ensures selection of only those features that have a great impact on the model's capability to effectively detect faults. Additionally, a hybrid Convolution Neural Networks (CNN) and Recurrent Neural Networks (RNN) model with an integrated attention mechanism is employed, focusing on crucial features to enhance the accuracy of fault detection and prediction. The proposed model was trained and validated using Power System Intrusion dataset containing fault and operational data from PV systems. Here, 80 % data is used for training and remaining 20 % data is used for testing. The performance results demonstrate that the model achieves a fault detection accuracy of 98.71 %. Additionally, it exhibits a high sensitivity of 0.97, a specificity of 0.98, and an F1-score of 0.97, indicating a significant improvement over existing techniques. The practical implications of this research highlight enhancements in PV system maintenance and reliability, leading to a reduction in downtime and operational costs. The rigorous integration of advanced data-preprocessing techniques, efficient feature selection strategies, and a deep learning model has a great potential in improving the maintenance and reliability of PV systems, leading to lower levels of operational and downtime.