In the context of global efforts toward carbon neutrality, the development of clean and efficient power generation technologies is crucial for the future of the energy sector. Solar photovoltaic power generation, as a renewable energy technology, offers significant advantages in safety, efficiency, economy, and environmental sustainability. However, solar panels are often installed in outdoor environments where they are exposed to pollutants such as dust and bird droppings. These contaminants reduce light transmittance, decrease power generation efficiency, and can lead to hot spots that damage the panel’s structure over time. Traditional cleaning methods, such as manual washing, are inefficient and hazardous, especially in complex terrains like wetlands and hilly areas where solar farms are commonly located. To address these challenges, we have developed a solar panel cleaning robot that utilizes vision-photoelectric sensors for adaptive path planning and cleaning strategies. This robot is designed to operate autonomously on inclined surfaces, ensuring comprehensive dust removal and targeted cleaning of stubborn pollutants like bird droppings.
The robot’s hardware components are mounted on a chassis and include side brushes, a roller brush, misting nozzles, tracked wheels, photoelectric sensors, a vacuum generator, a control box, a vision sensor, and a water tank. During operation, the vision sensor captures surface characteristics of the solar panel and transmits this information to the control system, which adjusts the motor’s rotation direction accordingly. The tracked wheels, driven by rear motors, enable the robot to move forward, backward, brake, and rotate. The vacuum generator provides negative pressure to adhere the robot to the glass surface, allowing it to work on solar panels at various inclinations. As the robot moves forward, water from the tank is sprayed onto the panel via misting nozzles, while the side and roller brushes clean dust and bird droppings. Photoelectric sensors around the robot detect panel edges and send low-level signals to the control system, facilitating decision-making for direction changes based on pre-mapped cleaning paths.
To ensure stable operation on inclined solar panels, we analyzed the adhesion conditions and required suction force. The driving force for the tracked wheels generates friction between the tracks and the panel surface, which must not exceed the adhesion force to prevent slipping. The adhesion force $F_s$ is given by:
$$F_s = \mu F_n$$
where $\mu$ is the adhesion coefficient between the track and the panel, and $F_n$ is the normal force. For no-slip conditions, the tangential force $F_x$ due to the driving torque must satisfy:
$$\frac{M – M_f}{r} = F_x \leq F_s$$
Here, $M$ is the driving torque, $M_f$ is the rolling resistance torque, and $r$ is the radius of the tracked wheel. The adhesion rate $C$ is defined as:
$$C = \frac{F_x}{F_n} \leq \mu$$
Considering the robot as a rigid body on an inclined panel at angle $\theta$, the normal force $F_N$ includes the gravitational component and the suction force $F_r$ from the vacuum generator:
$$F_N = mg \cos \theta + F_r$$
where $m$ is the total mass of the robot (including water). To prevent slipping, the friction force $F_S$ must counter the gravitational component along the incline:
$$F_S = mg \sin \theta < f_s F_N$$
with $f_s$ as the static friction coefficient. For typical solar panel installations in southern regions, $\theta = 20^\circ$. On dry surfaces, $\mu = 0.52$, but on wet surfaces, $\mu$ drops to 0.25. Without suction ($F_r = 0$), the robot cannot operate on wet panels. Thus, a minimum suction force $F_r > 33.62\, \text{N}$ is required. The suction force is calculated using:
$$F_r = \frac{\pi D^2 p}{4n}$$
where $D = 50\, \text{mm}$ is the suction cup diameter, $p$ is the vacuum pressure, and $n = 4$ is the safety factor. Solving for $p$ yields a required vacuum pressure of $68.49\, \text{kPa}$, which is met by standard vacuum generators capable of up to $88\, \text{kPa}$. Two vacuum generators are used to mitigate pressure loss when crossing gaps between panels.
Path planning is critical for autonomous operation. The robot uses a fusion of vision and photoelectric sensors to adjust its position and orientation upon deployment. The vision sensor captures images of the solar panel’s aluminum frame, which is processed using deep learning-based object detection algorithms. After noise reduction with Gaussian filtering and grayscale conversion, edge detection is performed using the Roberts operator. Thresholding and area filtering isolate the frame edges, and Hough line detection fits straight lines to determine the nearest border. The robot then rotates to align itself with this border and moves toward it. Photoelectric sensors at the front and sides detect edges, providing feedback for precise alignment. For example, when both front-left and front-right sensors detect an edge, the robot stops and adjusts its orientation until it is parallel to the border. This process ensures the robot starts cleaning from a designated initial position, such as the bottom-left corner of the panel.
The cleaning strategy involves two phases: comprehensive dust cleaning and localized bird dropping removal. For dust, the robot follows a zigzag (S-shaped) path to cover the entire panel surface. This path is pre-programmed based on panel dimensions, with adjustments for gaps between panels. The robot moves laterally across the panel, turning at edges detected by photoelectric sensors to avoid falls. For bird droppings, the vision sensor captures images after dust cleaning. Image processing techniques, including Gaussian filtering, grayscale conversion, and Roberts edge detection, identify bird droppings and panel features like busbars and grid lines. The centroid of the bird dropping is calculated, and its distance to the nearest feature line is determined using pixel-to-actual distance conversion. Given pixel distances $l_M$ and $l_N$ from the centroid to the left and right borders, and pixel lengths $l_P$ and $l_Q$ of the feature lines, the actual distances $l_1$ and $l_2$ are:
$$l_1 = \frac{l_M l_4}{l_Q}, \quad l_2 = \frac{l_N l_3}{l_P}$$
where $l_3$ and $l_4$ are the actual lengths of the feature lines. This allows the robot to navigate to the pollutant’s location for repeated cleaning.
We conducted experiments to evaluate the robot’s performance. A testbed was built using tempered glass panels (1 m × 1 m) with aluminum frames, tilted at $20^\circ$. Dust and bird droppings were simulated using sifted soil. The robot was deployed arbitrarily, and the time to reach the initial position and complete full cleaning was recorded. Bird dropping localization error was calculated by comparing detected and actual positions. Results from five trials showed an average initial positioning time of 49.7 seconds and full cleaning time of 72.0 seconds, yielding a cleaning efficiency of $50\, \text{m}^2/\text{h}$. The average localization error for bird droppings was $1.38\, \text{mm}$. To assess cleaning quality, images were processed to count white pixels (representing dust) before and after cleaning. The dust removal rate was calculated as:
$$\text{Dust Removal Rate} = \frac{\text{White Pixels}_{\text{before}} – \text{White Pixels}_{\text{after}}}{\text{White Pixels}_{\text{before}} – \text{White Pixels}_{\text{clean}}} \times 100\%$$
Based on average values—26,122 white pixels for a clean panel, 35,677 before cleaning, and 26,967 after cleaning—the dust removal rate was 91.16%.
| Trial | Initial Positioning Time (s) | Full Cleaning Time (s) | Localization Error (mm) |
|---|---|---|---|
| 1 | 49.3 | 72.6 | 1.5 |
| 2 | 47.6 | 70.3 | 1.2 |
| 3 | 53.1 | 68.8 | 1.4 |
| 4 | 51.8 | 74.4 | 1.3 |
| 5 | 46.7 | 73.9 | 1.5 |
In summary, our solar panel cleaning robot effectively addresses cleaning challenges in complex environments through vision-photoelectric sensor fusion. The theoretical analysis ensures stable adhesion on inclined surfaces, while adaptive path planning and targeted cleaning strategies enhance efficiency and reliability. Experimental results confirm high cleaning performance, contributing to improved solar energy utilization. Future work will focus on optimizing sensor algorithms and expanding the robot’s applicability to various photovoltaic system configurations.