In the field of photovoltaic (PV) power generation, dust accumulation on solar panels is a persistent and critical issue. Dust deposition reduces the light transmittance of the glass cover, leading to a significant drop in power generation efficiency. Moreover, accumulated dust can cause localized heating, hot spot formation, and even corrosion, which further deteriorates the performance and lifespan of solar panels. Traditional monitoring methods rely on manual inspection or indirect electrical parameter measurements, which are labor-intensive, subjective, and unable to provide real-time quantitative assessment. To address these limitations, we propose a novel approach based on convolutional neural learning to directly recognize and analyze the dust status on solar panels from visible-light images. By leveraging the power of deep learning, our method automatically extracts color and texture features related to dust coverage and maps them to the corresponding power efficiency loss rate. This paper presents a comprehensive study, including experimental setup, dataset preparation, model architecture, training process, and field validation, demonstrating that the proposed method achieves high accuracy and low error in identifying 11 grades of dust severity. Our work provides a portable, intelligent, and maintenance-free solution for solar panel condition monitoring, which is essential for optimizing cleaning schedules and improving the overall economic and operational efficiency of PV plants.
1. Introduction
The global shift toward renewable energy has led to a rapid expansion of photovoltaic installations, especially in arid and semi-arid regions with abundant solar resources. However, these regions often suffer from high dust loads due to wind-blown sand, industrial emissions, and agricultural activities. Dust particles settle on solar panels and form a layer that obstructs sunlight. Even a thin layer can reduce the short-circuit current and fill factor, resulting in a power loss of up to 24% after 20 days of exposure, as reported in earlier field studies. In addition, uneven dust distribution can create shading patterns that trigger reverse bias in some cells, leading to hot spots and permanent damage. Therefore, timely and accurate detection of dust accumulation is crucial for PV system operation and maintenance.
Conventional dust monitoring techniques include direct measurement of transmittance, weight-based deposition sensors, and analysis of electrical I-V curves. These methods often require dedicated hardware, frequent calibration, or manual intervention. Image-based approaches have emerged as a promising alternative because they are non-contact, low-cost, and can leverage the rich visual information embedded in the appearance of solar panels. With the advent of deep convolutional neural networks (CNNs), automatic feature extraction and classification have become feasible for complex visual patterns. In our work, we adopt residual networks (ResNet) to overcome the degradation problem in very deep architectures and achieve robust recognition of dust states from natural images captured under field conditions.
The key contributions of this paper are threefold. First, we design and deploy a field experimental system that simultaneously collects high-resolution images of solar panels and synchronous electrical and meteorological data. Second, we establish a labeled dataset with 11 discrete dust severity levels, each characterized by the average power efficiency loss rate calculated after temperature and irradiance normalization. Third, we train and compare multiple CNN architectures, including ResNet-18, ResNet-34, ResNet-50, ResNet-101, VGG-16, and AlexNet, and demonstrate that ResNet-50 achieves the best trade-off between accuracy and computational cost. Our results show that the proposed method can estimate the efficiency loss rate with a root mean square error (RMSE) as low as 0.69% and a precision of 0.81, enabling real-time quantitative dust assessment directly from images.
2. Characteristics of Dust on Solar Panels and Image Features
Dust particles deposited on solar panels originate from soil, industrial emissions, and biological residues. Their composition varies by location, but typical components include SiO₂, Al₂O₃, Fe₂O₃, CaO, and organic matter. The particle size distribution is usually log-normal, with most particles ranging from 10 to 100 µm. The settling process follows Stokes’ law, where the terminal settling velocity w of spherical particles is given by:
$$
w = \frac{2}{9} \frac{(\rho_s – \rho) g r^2}{\mu}
$$
where ρ_s is the particle density, ρ is the fluid density (air), μ is the fluid viscosity, r is the particle radius, and g is the gravitational acceleration. In practice, dust accumulates gradually, forming distinct visual patterns that we categorize into three types: loose dust, spot-like dust, and salt-stained dust. Loose dust appears as a uniform, powdery layer; spot-like dust consists of irregular patches caused by partial cleaning or rain; salt-stained dust results from repeated wetting and evaporation, leaving crystalline residues. These patterns exhibit clear differences in color (from dark gray to light brown) and texture (from smooth to rough), which are easily captured by visible-light cameras. Our analysis focuses on loose dust because it is the most prevalent and can be identified early to prevent progression to more damaging forms.
3. System Design and Methodology
3.1 Field Experimental System
We constructed an experimental monitoring system at a photovoltaic power plant located in a dusty region of Northeast China. The system consists of the following components:
- Solar panel array: each panel is a polycrystalline silicon module (model CSUN260-60P, rated power 260 Wp) mounted at a tilt angle of 43°.
- Environmental monitoring station: measures global irradiance (pyranometer), ambient temperature, humidity, and wind speed.
- Visible-light camera: captures RGB images of the solar panel surface at a fixed position and angle, with a resolution of 1920×1080 pixels.
- Backsheet temperature sensors: type DS18B20 digital sensors attached to the rear surface of selected panels.
- Electrical parameter meters: measure DC voltage, current, and power from each string via a data acquisition unit.
- Wireless data/image transmission device: transmits all data to a central monitoring computer for storage and processing.
Images are taken daily between 11:30 and 13:30 local time to minimize the influence of varying sun angles. Images with snow cover, equipment shadows, or obvious defects are filtered out. A total of 550 representative images are selected from a three-month period (March to June 2018) and divided into 11 groups according to the calculated efficiency loss rate. Each group contains 50 images, with 40 used for training and 10 for testing. All images are resized to 224×224 pixels in JPEG format to standardize input dimensions.
3.2 Calculation of Power Efficiency Loss Rate
To quantify the dust severity, we compute the power efficiency loss rate η_pl for each image timestamp. The instantaneous DC power output Pm is measured, and the theoretical maximum power Pmax under standard test conditions is adjusted for the actual backsheet temperature Tb and irradiance G using the temperature coefficient β (typically -0.408 %/°C for the panel). The efficiency under current conditions η is:
$$
\eta = \frac{P_m}{A \cdot G} \quad \text{(actual efficiency)}
$$
$$
\eta_{\text{std}} = \frac{P_{\text{max,ref}}}{A \cdot G_{\text{ref}}} \cdot \left[1 – \beta (T_b – T_{\text{ref}})\right] \quad \text{(temperature-corrected reference efficiency)}
$$
where A is panel area, Gref = 800 W/m², Tref = 45°C (nominal operating cell temperature). The efficiency loss rate is then defined as:
$$
\eta_{pl} = 1 – \frac{\eta}{\eta_{\text{std}}}
$$
We average the loss rates of all samples within a group to obtain the representative value for that dust grade. Table 1 summarizes the 11 grades and their corresponding mean loss rates.
| Grade | η_pl (%) | Grade | η_pl (%) |
|---|---|---|---|
| 1 | ≤4.0 | 7 | 11.7 |
| 2 | 5.6 | 8 | 12.9 |
| 3 | 6.9 | 9 | 14.1 |
| 4 | 8.9 | 10 | 15.8 |
| 5 | 9.5 | 11 | ≥16.6 |
| 6 | 10.5 |
4. Convolutional Neural Learning and Residual Networks
Convolutional neural networks (CNNs) have revolutionized image classification tasks by learning hierarchical feature representations directly from raw pixels. A typical CNN consists of convolutional layers, activation functions (e.g., ReLU), pooling layers, and fully connected layers followed by a softmax classifier. Depth is crucial for capturing complex patterns, but simply stacking layers leads to vanishing gradients and degradation. He et al. proposed the residual learning framework, which introduces shortcut connections that allow the network to learn residual mappings F(x) = H(x) – x instead of the original mapping H(x). The building block is defined as:
$$
\mathbf{y} = \mathcal{F}(\mathbf{x}, \{W_i\}) + \mathbf{x}
$$
where x is the input, F(x) is the residual mapping (e.g., two convolutional layers), and the addition is element-wise. If the dimensions increase, a linear projection W_s is used to match the shortcut. The derivative of the output with respect to the input is:
$$
\frac{\partial \mathbf{y}}{\partial \mathbf{x}} = \frac{\partial \mathcal{F}}{\partial \mathbf{x}} + 1
$$
This ensures that the gradient does not vanish even in very deep networks. We experiment with several ResNet variants (ResNet-18, -34, -50, -101) and compare them against VGG-16 and AlexNet. The detailed configuration of each network is given in Table 2.
| Model | Input size | Conv1 | Conv2_x | Conv3_x | Conv4_x | Conv5_x | Fully connected |
|---|---|---|---|---|---|---|---|
| ResNet-18 | 224×224 | 7×7,64 | [3×3,64]×2 | [3×3,128]×2 | [3×3,256]×2 | [3×3,512]×2 | 1 layer |
| ResNet-34 | 224×224 | 7×7,64 | [3×3,64]×3 | [3×3,128]×4 | [3×3,256]×6 | [3×3,512]×3 | 1 layer |
| ResNet-50 | 224×224 | 7×7,64 | [1×1,64; 3×3,64; 1×1,256]×3 | [1×1,128; 3×3,128; 1×1,512]×4 | [1×1,256; 3×3,256; 1×1,1024]×6 | [1×1,512; 3×3,512; 1×1,2048]×3 | 1 layer |
| ResNet-101 | 224×224 | 7×7,64 | [1×1,64; 3×3,64; 1×1,256]×3 | [1×1,128; 3×3,128; 1×1,512]×4 | [1×1,256; 3×3,256; 1×1,1024]×23 | [1×1,512; 3×3,512; 1×1,2048]×3 | 1 layer |
| VGG-16 | 224×224 | [3×3,64]×2 | [3×3,64]×2 | [3×3,128]×3 | [3×3,256]×3 | [3×3,512]×3 | 3 layers |
| AlexNet | 277×277 | 11×11,96 | [5×5,256] | [3×3,384] | [3×3,384] | [3×3,256] | 3 layers |
All models are trained using stochastic gradient descent with momentum 0.9, weight decay 0.0001, and an initial learning rate of 0.01 that is decreased by a factor of 10 every 20 epochs. The loss function is cross-entropy. We augment the training data with random horizontal flips and color jitter to improve generalization. Each model is trained for 60 epochs on a GTX1080Ti GPU.
5. Field Experimental Results and Analysis
5.1 Training and Test Performance
The training accuracy curves for the six models are shown in Figure 1 (not cited in text). After 45 epochs, ResNet-34 and ResNet-50 achieve accuracies above 85%, while ResNet-18 and ResNet-101 lag behind. VGG-16 and AlexNet plateau near 30%, indicating their inability to learn meaningful features from the limited dataset. The training loss values for ResNet variants converge below 2.15, whereas VGG and AlexNet losses diverge. Table 3 summarizes the quantitative evaluation metrics on both training and test sets: precision, recall, F-score, RMSE, and mean absolute percentage error (MAPE).
| Model | Set | Precision | Recall | F-Score | RMSE (%) | MAPE (%) |
|---|---|---|---|---|---|---|
| ResNet-18 | Train | 0.780 | 1.000 | 0.880 | 2.39 | 0.148 |
| Test | 0.750 | 1.000 | 0.850 | 2.90 | 1.970 | |
| ResNet-34 | Train | 0.850 | 1.000 | 0.920 | 1.41 | 0.057 |
| Test | 0.810 | 1.000 | 0.900 | 1.62 | 0.650 | |
| ResNet-50 | Train | 0.890 | 1.000 | 0.940 | 0.83 | 0.030 |
| Test | 0.810 | 1.000 | 0.900 | 0.69 | 0.330 | |
| ResNet-101 | Train | 0.900 | 1.000 | 0.950 | 1.42 | 0.061 |
| Test | 0.720 | 1.000 | 0.840 | 0.95 | 0.680 | |
| VGG-16 | Train | ~0.30 | ~0.30 | ~0.30 | high | high |
| AlexNet | Train | ~0.32 | ~0.32 | ~0.32 | high | high |
From Table 3, ResNet-50 achieves the best overall performance on the test set: precision of 0.81, F-score of 0.90, RMSE of only 0.69%, and MAPE of 0.33%. ResNet-101 exhibits higher training precision (0.90) but lower test precision (0.72), suggesting overfitting due to its larger capacity given the limited dataset. ResNet-34 also performs reasonably but with higher RMSE (1.62%). VGG-16 and AlexNet fail to converge, confirming their inadequacy for this task.
5.2 Comparison of Recognition Results with Experimental Data
We compare the model-predicted dust grades against the actual ground truth labels for the test samples. The scatter plot (not shown here) indicates that ResNet-50 predictions cluster tightly around the diagonal, with most misclassifications occurring between adjacent grades. For instance, a sample with actual loss rate 9.5% might be predicted as grade 4 (8.9%) or grade 6 (10.5%). This is acceptable because the impact on cleaning decision is minimal. The RMSE of 0.69% means that on average the predicted efficiency loss rate deviates by less than 0.7 percentage points from the true value, which is sufficient for practical monitoring.
5.3 Discussion
The superior performance of ResNet-50 over deeper ResNet-101 can be attributed to the limited size of our dataset (550 images). Deeper networks require more data to avoid overfitting. Data augmentation helps but cannot fully compensate. For future improvement, we plan to expand the dataset by collecting images under varying illumination, tilt angles, and dust compositions from different geographic regions. Furthermore, we will explore transfer learning by pretraining on large-scale datasets like ImageNet and then fine-tuning on our solar panel dust images. This could further boost accuracy and robustness.
Another important aspect is the influence of dust composition on image features. We analyzed dust samples from two different regions (Northeast and Northwest China) and found that the main components (SiO₂, Al₂O₃, Fe₂O₃, CaO) and particle size distributions were very similar. This suggests that our model, trained on one region, may generalize to other regions with comparable dust types. Validation across multiple sites is under way.
6. Conclusion
We have developed an effective method for recognizing and analyzing dust accumulation on solar panels using convolutional neural learning. By constructing a field experiment system and preparing a labeled dataset with 11 grades of dust severity, we demonstrated that the ResNet-50 model achieves high recognition accuracy (precision 0.81) and low estimation error (RMSE 0.69%) for the power efficiency loss rate. Compared to traditional methods that require electrical parameter measurements or manual inspection, our image-based approach is direct, real-time, and maintenance-free. The method can be readily integrated into portable inspection devices or fixed monitoring cameras, enabling intelligent operation and maintenance of large-scale solar panel fields. Future work will focus on dataset expansion, cross-site validation, and model optimization to further improve generalization and practicality.