In recent years, the rapid expansion of solar energy storage systems has become a cornerstone of renewable energy integration, addressing the intermittency of solar power generation. As a researcher focused on enhancing the efficiency and reliability of photovoltaic (PV) systems, I have investigated the critical issue of hot spot faults in PV panels, which can severely impact the performance and longevity of solar energy storage setups. Hot spots, caused by localized overheating due to shading or material defects, not only reduce energy conversion efficiency but also pose fire risks, undermining the stability of solar energy storage networks. Traditional inspection methods, such as manual checks with thermal imagers, are time-consuming, hazardous, and inefficient for large-scale solar farms. To address this, I propose a deep learning-based approach that leverages convolutional neural networks (CNNs) for real-time hot spot detection in aerial infrared images of PV panels. This method aims to improve the maintenance and optimization of solar energy storage systems by enabling automated, accurate, and rapid fault identification.
The integration of solar energy storage with PV systems requires robust monitoring to maximize energy yield and prevent failures. Hot spots occur when certain cells in a PV panel operate at higher temperatures than others, often due to partial shading, dust accumulation, or internal defects. This phenomenon can lead to significant energy losses and, in extreme cases, damage the entire panel, affecting the overall solar energy storage capacity. In my work, I analyze the characteristics of hot spots in infrared images, where they appear as bright regions with distinct temperature gradients. However, challenges such as background noise, reflections, and small target sizes complicate detection. To overcome these issues, I developed a two-step deep learning framework: first, using an improved YOLOv4 model to identify and extract PV panels from aerial infrared images, eliminating background interference; second, applying an enhanced DeeplabV3+ model to segment hot spots accurately. This approach not only enhances detection precision but also supports the operational efficiency of solar energy storage infrastructures by enabling proactive maintenance.
In the context of solar energy storage, the reliability of PV panels is paramount. Hot spots can degrade the performance of batteries and inverters in storage systems by causing uneven energy flow and reducing overall efficiency. My research emphasizes the importance of early detection to prevent cascading failures. The dataset for this study was collected using a DJI Matrice 300 drone equipped with a Zenmuse XT2 thermal camera, capturing 2,188 infrared images from a 10MW PV plant in China. After filtering out irrelevant images, 1,557 usable images were enhanced through rotations, contrast adjustments, and resizing to simulate various environmental conditions, resulting in 7,785 images for training. This dataset ensures the model’s adaptability to real-world solar energy storage scenarios, such as different weather patterns and flight altitudes. Data augmentation techniques, including OpenCV-based transformations, were employed to increase diversity and reduce overfitting, crucial for generalizing across diverse solar energy storage applications.
The first part of my framework involves PV panel detection using an optimized YOLOv4 model. YOLOv4 is a one-stage object detection model known for its balance between speed and accuracy, making it suitable for real-time applications in solar energy storage monitoring. I modified YOLOv4 by replacing its backbone feature extraction network, CSPDarknet53, with lightweight MobileNet networks to reduce computational complexity. Specifically, I tested MobileNet-V1, MobileNet-V2, and MobileNet-V3, with MobileNet-V2 yielding the best performance due to its inverted residual structure and depthwise separable convolutions. Further, I replaced standard convolutions in the PANet aggregation network with depthwise separable convolutions, significantly decreasing the model size. The loss function used was Complete Intersection over Union (CIoU), which considers overlap area, center point distance, and aspect ratio, defined as:
$$ \text{Loss}{\text{CIoU}} = 1 – \text{IoU} + \frac{\rho^2(b, b{gt})}{c^2} + \alpha v $$
where ( \rho ) is the Euclidean distance between predicted and ground truth bounding box centers, ( c ) is the diagonal length of the smallest enclosing box, ( \alpha ) is a weight parameter, and ( v ) measures aspect ratio consistency. For training, I applied transfer learning with pre-trained weights, setting the initial learning rate to ( 1 \times 10^{-3} ) for 50 epochs and then ( 1 \times 10^{-4} ) for another 50 epochs. The dataset was split 80:20 for training and validation, with evaluation metrics including Average Precision (AP), Recall, Frames Per Second (FPS), and model size. The results demonstrated that the improved YOLOv4 model with MobileNet-V2 and depthwise separable convolutions achieved an AP of 99.56%, Recall of 98.91%, FPS of 22.1, and a model size of 46.4 MB, making it efficient for deployment in solar energy storage systems.
| Model | AP (%) | Recall (%) | FPS | Model Size (MB) |
|---|---|---|---|---|
| YOLOv4 | 99.66 | 98.57 | 13.7 | 244 |
| YOLOv4-V2 (Improved) | 99.56 | 98.91 | 22.1 | 46.4 |
The second step focuses on hot spot segmentation using an enhanced DeeplabV3+ model. Semantic segmentation is essential for pixel-level identification of hot spots, allowing precise analysis of their shape and size, which is critical for diagnosing root causes in solar energy storage systems. I replaced the original Xception backbone with MobileNet-V2 to reduce model parameters and improve inference speed. Additionally, I modified the loss function from cross-entropy to Dice loss, which is more effective for imbalanced datasets where hot spots represent small regions. The Dice loss is defined as:
$$ \text{Dice Loss} = 1 – \frac{2 \sum_{i=1}^{N} p_i g_i}{\sum_{i=1}^{N} p_i + \sum_{i=1}^{N} g_i} $$
where ( p_i ) is the predicted probability and ( g_i ) is the ground truth label for pixel ( i ). I also adjusted the downsampling factor from 16x to 8x to preserve finer details of small hot spots. The training involved transfer learning with a learning rate of ( 1 \times 10^{-3} ) for 35 epochs and ( 1 \times 10^{-4} ) for another 35 epochs, using 1,596 images with hot spots (20% of the PV panel dataset). Evaluation metrics included Mean Pixel Accuracy (MPA) and Mean Intersection over Union (mIoU). The improved DeeplabV3+ model achieved an MPA of 95.99%, mIoU of 85.58%, FPS of 24.5, and a model size of 22.3 MB, demonstrating its suitability for real-time solar energy storage monitoring.
| Model | MPA (%) | mIoU (%) | FPS | Model Size (MB) |
|---|---|---|---|---|
| DeeplabV3+ | 73.26 | 71.34 | 13.7 | 209 |
| DeeplabV3+ with MobileNet-V2 and Dice Loss | 95.99 | 85.58 | 24.5 | 22.3 |
The integration of these models addresses key challenges in solar energy storage management. By first isolating PV panels from complex backgrounds, the YOLOv4-based detector minimizes false positives caused by environmental noise. Then, the DeeplabV3+ segmenter accurately delineates hot spots, even in cases of reflections or small sizes. This two-step process ensures that maintenance teams can quickly identify and address faults, thereby enhancing the reliability and efficiency of solar energy storage systems. For instance, in a solar farm, rapid detection of hot spots can prevent energy losses and extend the lifespan of batteries and other storage components. The lightweight nature of the models allows for deployment on drones or edge devices, facilitating continuous monitoring without significant computational resources.
In experimental analyses, I compared various models to validate the effectiveness of my approach. For PV panel detection, the improved YOLOv4 model outperformed alternatives like Faster-RCNN and YOLOv5s in terms of speed and accuracy, crucial for real-time applications in solar energy storage. Similarly, for hot spot segmentation, the modified DeeplabV3+ model showed superior performance over PSPNet and other variants, with higher MPA and mIoU values. The use of Dice loss significantly improved the handling of class imbalance, as hot spots occupy a small fraction of the image. This is particularly important for solar energy storage systems, where early detection of minor faults can prevent major disruptions. The training process benefited from transfer learning, reducing the required epochs and improving convergence, which is advantageous for adapting to new solar energy storage environments with limited data.
Looking ahead, the proposed framework can be further optimized for broader solar energy storage applications. Future work could involve collecting data at different altitudes and under varying weather conditions to enhance model robustness. Additionally, integrating this detection system with predictive maintenance algorithms could enable proactive management of solar energy storage assets, reducing downtime and costs. The continuous evolution of deep learning models, such as newer versions of YOLO or transformer-based architectures, may offer opportunities for even greater accuracy and efficiency. By advancing these technologies, we can ensure that solar energy storage systems operate at peak performance, supporting the global transition to sustainable energy.
In conclusion, my research demonstrates the viability of deep learning for hot spot detection in aerial infrared images of PV panels, directly contributing to the optimization of solar energy storage. The improved YOLOv4 and DeeplabV3+ models provide a lightweight, accurate, and fast solution for real-time monitoring, addressing critical challenges in maintenance and efficiency. As solar energy storage continues to grow, such automated detection systems will play a pivotal role in maximizing energy yield and ensuring system reliability. Through ongoing innovation, we can further enhance the integration of AI in renewable energy, paving the way for smarter and more resilient solar energy storage infrastructures.