Photovoltaic (PV) systems have immense potential to generate clean energy, and their adoption has grown significantly in recent years. A PV fault is a condition of a PV module that is unable to produce optimal power due to environmental factors, such as shading, hot spots, cracks, and other defects. The occurrence of faults in PV systems can present safety risks, shorten system lifespans, and result in waste. Therefore, this paper discusses the importance of accurately classifying faults in PV systems to maintain optimal operating efficiency, thereby increasing the financial return. Previous studies in this area have largely relied on deep learning models, such as transfer learning, with high computational requirements, which are limited by their inability to handle complex image features and unbalanced datasets. The proposed lightweight coupled UdenseNet model shows significant improvements for PV fault classification compared to previous studies, achieving an accuracy of 99.39%, 96.65%, and 95.72% for 2-class, 11-class, and 12-class output, respectively, while also demonstrating greater efficiency in terms of parameter counts, which is particularly important for real-time analysis of large-scale solar farms. Furthermore, geometric transformation and generative adversarial networks (GAN) image augmentation techniques improved the model’s performance on unbalanced datasets.

Due to the accelerated growth of the PV plant industry, multiple PV plants are being constructed in various locations. It is difficult to operate and maintain multiple PV plants in diverse locations. Consequently, a method for monitoring multiple PV plants on a single platform is required to satisfy the current industrial demand for monitoring multiple PV plants on a single platform. This work proposes a method to perform multiple PV plant monitoring using an IoT platform. Next-day power generation prediction and real-time anomaly detection are also proposed to enhance the developed IoT platform. From the results, an IoT platform is realized to monitor multiple PV plants, where the next day’s power generation prediction is made using five types of AI models, and an adaptive threshold isolation forest is utilized to perform sensor anomaly detection in each PV plant. Among five developed AI models for power generation prediction, BiLSTM became the best model with the best MSE, MAPE, MAE, and R2 values of 0.0072, 0.1982, 0.0542, and 0.9664, respectively. Meanwhile, the proposed adaptive threshold isolation forest achieves the best performance when detecting anomalies in the sensor of the PV plant, with the highest precision of 0.9517.

No abstract

No abstract

No abstract