Photovoltaic power station automatic cleaning layout system and method based on digital twinning technology

Abstract

The application provides a photovoltaic station automatic cleaning layout system and method based on digital twin technology. A multi-modal sensor data fusion is performed by a photovoltaic panel environment perception subsystem, and a stain digital twin is generated in real time by an edge computing node. A photovoltaic panel fusion decision subsystem performs cleaning strategy simulation verification in a virtual space, and the influence of different cleaning parameters on stain removal effect is preformed through a digital twin model. The optimal scheme is issued to a physical execution system. Finally, a feedback subsystem feeds back the physical cleaning result to the virtual model, and the decision algorithm is updated through reinforcement learning (DDPG algorithm). A digital twin management and control platform realizes full-process visualization, and an operation and maintenance personnel can monitor the running state of a single robot, view a station cleaning heat map and a power generation gain curve through a three-dimensional interface.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic cleaning technology, and in particular to an automatic cleaning layout system and method for photovoltaic power plants based on digital twin technology, electronic equipment, and computer-readable storage media. Background Technology

[0002] Currently, global photovoltaic (PV) installed capacity has exceeded 1 TW, and the operation and maintenance efficiency of PV power plants directly affects the cost per kilowatt-hour. According to the International Energy Agency (IEA) 2025 Annual Report, surface contamination of PV panels can lead to a 15%-35% decrease in power generation efficiency, and in extreme cases, even more than 50%. Existing cleaning methods are mainly divided into three categories: manual cleaning, traditional automated cleaning equipment, and semi-automated cleaning systems, but all of them have significant technical bottlenecks.

[0003] Manual cleaning relies on a team of 6-8 people, using high-pressure water guns or wiping with cloths. Cleaning a single component takes about 3 minutes, and the daily cleaning volume is less than 500 units.

[0004] Traditional automated cleaning equipment mainly consists of track-mounted robots and unmanned vehicles, which have three major limitations: First, they have poor environmental adaptability, with a failure rate as high as 42% in sites with slopes exceeding 15° or where components are misaligned; second, they rely on single-path planning, which is ineffective in treating stubborn stains such as bird droppings (68% removal rate) and oil stains (53% removal rate); and third, they lack global optimization, making it easy for path conflicts to occur when multiple machines work together, resulting in an equipment utilization rate of only about 65%.

[0005] The core flaw of existing technologies lies in the separation between physical space and information space: (1) the perception dimension is limited to the surface of a single component, without considering macro-environmental factors such as weather (cleaning when wind speed is >8m / s can easily cause component microcracks) and terrain (undulations exceeding 0.5m affect robot passage); (2) the decision-making process lacks virtual simulation verification, and the cleaning path correction rate reaches 38% during actual execution; (3) the equipment status and cleaning effect do not form a closed-loop feedback, resulting in a repeated cleaning rate of 23% for the same stains. Digital twin technology provides a new technical path to solve the above problems by constructing a real-time mapping between the physical world and virtual space. Summary of the Invention

[0006] To address the technical problems existing in the prior art, the present invention provides the following technical solution: On the one hand, an automated cleaning layout system for photovoltaic power plants based on digital twin technology is provided, including: The photovoltaic panel environmental sensing subsystem is used to collect environmental data on the surface of the photovoltaic panel through a multi-source sensor array, and extract information on the location distribution, hardness, and type of stains. A photovoltaic panel integrated decision-making subsystem is connected to the photovoltaic panel environmental perception subsystem. It is used to formulate a global cleaning strategy and generate cleaning instructions based on the stain information and digital twin simulation data through optimization algorithms. The cleaning robot execution system is connected to the photovoltaic panel fusion decision subsystem and is used to receive and execute the cleaning instructions to complete the cleaning operation on the surface of the photovoltaic panel. A feedback subsystem, connected to the cleaning robot execution system, is used to acquire physical cleaning results and feed the cleaning results back to the photovoltaic panel fusion decision subsystem to optimize the cleaning strategy. The digital twin management and control platform is connected to the photovoltaic panel environmental perception subsystem, the photovoltaic panel fusion decision-making subsystem, and the cleaning robot execution system, respectively. It is used to construct a digital twin model of the photovoltaic power station, realize bidirectional mapping between physical space and virtual space, and perform simulation optimization and collaborative management and control of the entire cleaning process.

[0007] Preferably, the photovoltaic panel environmental sensing subsystem includes: The image acquisition unit is used to acquire multi-band images of the photovoltaic panel surface; An ultrasonic testing unit is used to emit ultrasonic pulses to the surface of a photovoltaic panel and receive the echoes, and calculate the hardness information of the stains based on the echo signals. A stain detection unit, connected to the image acquisition unit, is used to identify stains in the image and generate location distribution information; The stain type determination unit is used to fuse image features, ultrasonic features, and environmental features to determine the stain type and output a cleaning mode decision.

[0008] Preferably, the photovoltaic panel integration decision subsystem includes: The data preprocessing unit is used to perform spatiotemporal alignment and feature fusion of physical sensor data and digital twin virtual data to generate a standardized decision matrix; The strategy formulation unit is used to formulate a cleaning strategy that includes cleaning path, parameter configuration and equipment scheduling based on the decision matrix and through the optimization algorithm enhanced by digital twin in a virtual environment. An optimization unit, connected to the feedback subsystem, is used to dynamically adjust the parameters of the cleaning strategy based on the deviation between the physical cleaning results and the virtual prediction results. The resource scheduling unit is used to allocate tasks and resolve path conflicts among multiple cleaning robots based on a digital twin model.

[0009] Preferably, the cleaning robot execution system includes: The instruction receiving unit is used to receive cleaning instructions from the photovoltaic panel fusion decision subsystem; A motion control unit, connected to the instruction receiving unit, is used to control the motion trajectory of the cleaning actuator according to the cleaning instruction based on a predictive control algorithm using a digital twin model. The cleaning execution unit is connected to the motion control unit and is used to perform high-pressure water jet, mechanical brushing, or hot air blowing operations.

[0010] Preferably, the feedback subsystem includes an image acquisition module, a laser contour sensor, and a power generation monitoring module, used to acquire image data, three-dimensional morphology data, and power generation change data of the cleaned surface, so as to calculate the actual cleaning effect.

[0011] Preferably, the digital twin management platform includes: The virtual simulation module is used to simulate the cleaning process in a virtual environment based on the 3D model of the site, the digital twin of the equipment, and the stain evolution model. The collaborative management module is used to enable collaborative scheduling and conflict detection of multiple cleaning robots based on a digital twin model; The model optimization module is used to update the digital twin model and the cleaning decision model based on physical feedback data using a reinforcement learning algorithm.

[0012] Preferably, the stain type determination unit integrates a stain recognition and determination model, which is a multimodal feature fusion model built based on a deep belief network. Its input layer fuses image feature vectors, ultrasonic feature vectors, and digital twin environmental feature vectors, and its output layer outputs a cleaning mode decision vector.

[0013] Preferably, the optimization algorithm used by the strategy formulation unit is an improved genetic algorithm with improved digital twin enhancement, and its fitness evaluation is based on the simulation results in the digital twin virtual environment; the optimization unit uses a digital twin feedback optimization algorithm, which dynamically calibrates the strategy parameters by minimizing the deviation between the virtual predicted clearance rate and the physically measured clearance rate and the cleaning energy consumption.

[0014] Preferably, the cleaning robot execution system adopts a modular hardware architecture, including a physical execution layer and a corresponding virtual twin, and interacts with the digital twin management and control platform in real time through 5G communication and edge computing nodes.

[0015] On the other hand, an automatic cleaning layout method for photovoltaic power plants based on digital twin technology is provided. The method is implemented based on the system described above and includes the following steps: S1. Multi-source heterogeneous data fusion perception: Collect environmental data of photovoltaic power stations through the collaborative collection of physical sensors and digital twin virtual sensors, and extract information on the location, hardness and type of stains; S2. Digital Twin Simulation Optimization Decision: Based on the extracted stain information, virtual simulation optimization is performed in the digital twin platform through optimization algorithms to formulate the global optimal cleaning strategy and generate cleaning instructions; S3. Physical-Virtual Cooperative Execution: The cleaning robot execution system receives the cleaning command, performs the cleaning operation under the guidance of the digital twin, and synchronizes the physical state and virtual state in real time. S4. Feedback-driven continuous optimization: Collect physical cleaning results and compare them with digital twin prediction results. Based on the deviation, update the decision model through optimization algorithms to achieve continuous evolution of the cleaning strategy. S5. Digital Twin Full Lifecycle Management: Real-time monitoring of system status through the digital twin platform, enabling health assessment, fault early warning, and predictive maintenance. On the other hand, an electronic device is provided, comprising: a processor; and a memory storing computer-readable instructions, which, when executed by the processor, implement the method described above.

[0016] On the other hand, a computer-readable storage medium is provided, wherein at least one instruction is stored therein, the at least one instruction being loaded and executed by a processor to implement the above method.

[0017] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: This invention achieves a technological breakthrough through a triple architecture of "physical entity - virtual model - data interaction": First, the photovoltaic panel environmental perception subsystem integrates multimodal sensor data (images, ultrasound, laser), and generates a digital twin of the stain in real time through edge computing nodes (containing four attributes: location, hardness, area, and type); second, the photovoltaic panel integrated decision-making subsystem performs cleaning strategy simulation verification in virtual space, pre-simulating the impact of different cleaning parameters (pressure, path, and intensity) on the stain removal effect through the digital twin model, and disseminating the optimal solution to the physical execution system; finally, the feedback subsystem feeds the physical cleaning results back into the virtual model, updating the decision algorithm through reinforcement learning (DDPG algorithm) (the decision accuracy improves by 4.2% with every 1000 iterations). The digital twin management and control platform achieves full-process visualization, allowing maintenance personnel to monitor the operating status of a single robot (position accuracy ±0.1m), view the site cleaning heat map (updated once per minute), and the power generation gain curve (refreshed in real time) through a 3D interface. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a schematic diagram of the system composition structure of an automatic cleaning layout system for photovoltaic power plants based on digital twin technology, provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of the system composition structure of an image acquisition subsystem provided in an embodiment of the present invention; Figure 3 This is a data preprocessing flowchart provided in an embodiment of the present invention; Figure 4 This is a flowchart of an automatic cleaning layout method for photovoltaic power stations based on digital twin technology, provided by an embodiment of the present invention; Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0020] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0021] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0022] In the embodiments of this invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning. Similarly, the terms "of," "corresponding (relevant)," and "corresponding" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning.

[0023] In this embodiment of the invention, sometimes a subscript such as W1 may be mistakenly written as a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.

[0024] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0025] like Figure 1 The architecture of the photovoltaic power station automatic cleaning layout system based on digital twin technology is shown below: 1. Perception → Decision: The photovoltaic panel environmental perception subsystem transmits the extracted stain feature data (location, hardness, type) to the fusion decision subsystem as input for strategy formulation.

[0026] 2. Decision → Execution: The cleaning instructions (path, parameters, scheduling scheme) generated by the integrated decision-making subsystem are sent to the cleaning robot execution system.

[0027] 3. Execution → Feedback: The execution system uploads the equipment status (location, energy consumption, cleaning parameters) to the feedback subsystem in real time, and the feedback subsystem collects the cleaning results (cleaning rate, power generation).

[0028] 4. Feedback → Decision: The feedback subsystem feeds back the cleaning results (such as cleaning rate deviation and equipment wear and tear) to the decision subsystem to dynamically optimize the cleaning strategy.

[0029] 5. Digital Twin Platform → Subsystems: The platform connects in real time with the perception, decision-making, execution, and feedback subsystems, providing virtual simulation (such as path pre-simulation and conflict detection), data synchronization, and global optimization support.

[0030] The system functions and interactions will be described in detail below.

[0031] This invention provides an automated cleaning layout system for photovoltaic power plants based on digital twin technology, the system comprising: A photovoltaic panel environmental sensing subsystem is used to acquire environmental data, extract the location distribution information and hardness information of stains from the environmental data, and determine the type information of the stains; It should be noted that the photovoltaic panel environmental perception subsystem constructs a physical-virtual fusion perception network through a multimodal sensor array: (1) The physical sensor layer includes a 4K high-definition camera (resolution 3840×2160, frame rate 30fps), an infrared thermal imager (temperature range -20~150℃, accuracy ±1℃), and an ultrasonic hardness tester (measurement range 0-10HA, error <0.5HA); (2) The digital twin virtual sensor layer generates extended perception data through physical sensor data assimilation, such as the prediction of stain accumulation rate based on meteorological station data (wind speed, precipitation). The system uses an improved YOLOV5 algorithm for stain identification. Its network structure adds an attention mechanism (CBAM module) and multi-scale feature fusion (PANet structure) to the original model. The loss function is defined as: L = Lcls + Lbox + Lobj Wherein: Lcls is the class loss (using Focal Loss to address class imbalance), Lbox is the bounding box loss (CIoU Loss), and Lobj is the target confidence loss. By learning the color, shape, and texture features of eight types of stains (dust, bird droppings, oil stains, leaves, insect corpses, scale, rust, and snow), the stain type recognition accuracy is 98.2%, the location accuracy is ±2cm, and the hardness measurement error is <0.5HA.

[0032] A photovoltaic panel integrated decision-making subsystem, connected to the photovoltaic panel environmental perception subsystem, is used to automatically formulate cleaning strategies and generate cleaning instructions using the location distribution information, hardness information, and type information of the stains. It should be noted that the photovoltaic panel integration decision-making subsystem is built on a digital twin platform and achieves globally optimal cleaning path planning through an improved genetic algorithm (DT-GA). This algorithm uses the photovoltaic power station's digital twin model as the decision-making environment and constructs a multi-objective optimization model with stain removal rate (objective function f1), path length (objective function f2), and energy consumption cost (objective function f3). min F(X)=[ω1f1(X),ω2f2(X),ω3f3(X)], Where: X is the path decision variable matrix (including the number of robots, cleaning order, and operation time); ω1-ω3 are weight coefficients (0.5, 0.3, and 0.2 respectively); f1=1-∑(stain residual area / total stain area), f2 is the total robot movement distance (unit: m), and f3 is the cleaning energy consumption (unit: kWh). The algorithm performs 1000 iterations of optimization through digital twin virtual simulation to finally generate the globally optimal cleaning strategy. For example, for dust adhering to the surface of the photovoltaic panel, the system may choose to spray water for cleaning; while for more stubborn stains such as bird droppings or leaves, it will use brushing or a combination of spraying water and brushing. At the same time, the photovoltaic panel integration decision subsystem will also optimize the cleaning path based on the data (cleaning results) fed back by the feedback subsystem to ensure that the cleaning work is both efficient and safe.

[0033] A cleaning robot execution system is connected to the photovoltaic panel integration decision subsystem to receive and execute the cleaning instructions to clean the stains in the photovoltaic power station. It should be noted that the cleaning robot execution system is the core execution component of the photovoltaic power station automated cleaning layout system based on digital twin technology. It is responsible for receiving cleaning instructions from the photovoltaic panel integration decision subsystem and precisely controlling the movement trajectory and cleaning operation of the cleaning equipment according to these instructions. The high precision and flexibility of the cleaning robot execution system ensure the quality and efficiency of the cleaning work. When performing a cleaning task, the cleaning robot execution system first receives cleaning instructions from the photovoltaic panel integration decision subsystem through the instruction receiving subsystem. These instructions contain detailed cleaning strategies, such as the optimal cleaning path generated based on the DT-GA algorithm, cleaning intensity (water pressure 1-10MPa), and cleaning time (5-30s / module). Subsequently, the motion control subsystem uses a model predictive control (MPC) algorithm to calculate the movement trajectory of the cleaning subsystem based on the received cleaning instructions and real-time environmental data (terrain slope, module temperature, etc.) provided by the digital twin model. The control equation is as follows: u*(k)=argminu(k)J(x(k),u(k)), Where: u*(k) is the optimized control input vector, i.e., the control command finally executed by the cleaning robot; u(k) is the control input vector (including the control input vector at the current moment, including parameters such as the robot's motion speed, turning angle, and cleaning water pressure); x(k) is the robot state vector (real-time state information such as current position (3D coordinates), attitude (heading angle), and battery power (SOC); J is the objective function, comprehensively considering trajectory tracking accuracy and energy consumption cost, including trajectory tracking error (weight 0.6) and energy consumption cost (weight 0.4); k: discrete time step, corresponding to the sampling time of the control system; argmin u(k): the operator for solving the control input vector u(k) that minimizes the objective function J. The equation dynamically optimizes the control command through real-time environmental data (such as terrain slope and component temperature) provided by the digital twin model, ensuring that the robot's motion trajectory tracking error is <0.05m, while reducing energy consumption.

[0034] The cleaning subsystem executes the control commands from the motion control subsystem, removing dirt from the surface of the photovoltaic power station through methods such as spraying water, brushing, or blowing air. While performing the cleaning task, the cleaning robot execution system also monitors its own operating status in real time, including parameters such as the speed, position, and cleaning intensity of the cleaning equipment. This data is fed back to the digital twin platform in real time via a 5G edge node (latency <20ms), allowing the photovoltaic panel integration decision-making subsystem to dynamically adjust the cleaning strategy based on the cleaning effect and actual conditions. In this way, a tight closed-loop control system is formed between the cleaning robot execution system and the photovoltaic panel integration decision-making subsystem, ensuring the accuracy and efficiency of the cleaning work. During the cleaning process, if the cleaning robot execution system detects any abnormalities, such as cleaning equipment malfunction or trajectory deviation (exceeding ±0.05m), it will immediately stop the cleaning operation and issue an alarm signal. This not only prevents damage to the cleaning equipment and the photovoltaic power station but also ensures the safety of the operators.

[0035] A feedback subsystem, connected to the cleaning robot execution system, is used to acquire cleaning results and feed the cleaning results back to the photovoltaic panel integration decision subsystem, which optimizes the cleaning strategy based on the cleaning results. It should be noted that the feedback subsystem constructs a multi-dimensional feedback network using a high-definition camera (4K resolution / 30fps), a laser contour sensor (accuracy ±0.1mm), and a power generation monitoring module (sampling frequency 1Hz): It captures real-time images of the surface state after cleaning and compares them pixel-level with the pre-cleaning state in the digital twin model; it generates a 3D point cloud of the component surface through laser scanning and calculates the residual dirt height (a threshold <0.05mm is considered cleanliness compliant); it simultaneously collects 24-hour power generation data before and after cleaning, establishing a quantitative relationship between cleaning effectiveness and power generation gain (power generation efficiency improvement rate after cleaning = (average daily power generation after cleaning - average daily power generation before cleaning) / average daily power generation before cleaning × 100%). The feedback data is transmitted to the digital twin platform via a 5G edge node (latency <20ms), forming a closed loop of "perception-decision-execution-feedback".

[0036] The digital twin management and control platform, acting as the system's central hub, is connected in real-time via an industrial Ethernet network (1Gbps bandwidth) to the photovoltaic panel environmental perception subsystem, the photovoltaic panel fusion decision-making subsystem, and the cleaning robot execution system. It is used to construct a bidirectional mapping between the physical and virtual spaces of the photovoltaic power plant, enabling visualized monitoring, simulation optimization, and remote management of the entire cleaning process. The platform adopts a layered architecture: a data layer (supporting TB-level time-series data storage), a model layer (containing a 3D model of the power plant, digital twins of equipment, and a stain evolution model), and an application layer (providing functional modules such as cleaning task scheduling, equipment health management, and energy efficiency analysis).

[0037] The core functions of the digital twin management platform include: (1) Virtual simulation: Based on the Unity3D engine, a 1:1 high-precision site model is built, integrating real-time data such as meteorological (wind speed, precipitation, temperature), terrain (slope, altitude), and equipment (robot position, power, cleaning mechanism status), and supports cleaning path pre-play (simulation accuracy error <0.5m); (2) Conflict detection: Collision detection of multi-machine collaborative paths is performed by improving the A* algorithm, and the conflict warning response time is <100ms; (3) Performance optimization: The improved genetic algorithm (DT-GA) is used to schedule the cleaning task globally, and the equipment utilization rate is increased to 89%; (4) Fault diagnosis: Based on the LSTM neural network, the remaining life of the equipment is predicted (prediction accuracy >92%), and potential faults are warned 72 hours in advance. The platform is deployed on an edge server (GPU model NVIDIA A100), supports concurrent control of 100+ robots, and the daily data processing volume reaches 500GB.

[0038] Specifically, the digital twin management platform adopts a "cloud-edge-device" collaborative architecture: edge nodes (deployed in the site control room) are responsible for real-time data acquisition and processing (response time <50ms), the cloud platform (AWS EC2 instance) performs big data analysis and model training, and terminal devices (HMI touchscreen, mobile APP) provide a visual interactive interface. Communication between devices is achieved through the MQTT protocol, and data integration is performed using the OPC UA standard to ensure compatibility with third-party SCADA systems. The platform has a built-in data security module, employing AES-256 encrypted transmission, role-based access control (RBAC), and data anonymization, complying with the ISO 27001 information security standard.

[0039] As an optional embodiment of the present invention, the photovoltaic panel environmental sensing subsystem may optionally include: The image acquisition subsystem is used to acquire image information from the surface of photovoltaic panels and construct the visual input layer for the digital twin model. It includes three sets of industrial cameras (2592×1944 resolution, 15fps frame rate), deployed at the site's highest point (global view), the robot's end (close-up view), and the drone inspection unit (area coverage), respectively. The spatiotemporal consistency of multi-source images is ensured through a time synchronization protocol (PTPv2, synchronization accuracy ±1μs). Each set of industrial cameras includes a visible light camera and an infrared camera.

[0040] like Figure 2As shown, the image acquisition subsystem adopts a "binocular stereo vision + infrared thermal imaging" fusion scheme: a visible light camera (Sony IMX385 sensor) captures the color and texture features of stains, while an infrared camera (640×512 resolution, temperature range -20℃~150℃) enhances the contrast of recognition through the temperature difference between the stain and the substrate (typically 3-8℃). After distortion correction (using Zhang Zhengyou calibration method, reprojection error <0.3 pixels) and image stitching (based on SIFT feature matching, stitching seam error <1 pixel), a 2K resolution panoramic image of the component surface is generated and transmitted in real time to the digital twin platform to construct a virtual surface model.

[0041] The ultrasonic testing subsystem is used to detect ultrasonic feedback information of stains on the surface of photovoltaic panels and calculate the hardness data of the corresponding stains, providing mechanical property input for the digital twin model. It consists of an array of ultrasonic probes (8 channels, center frequency 5MHz, sampling rate 100MS / s), a pulse generator (output voltage ±400V) and a signal processing module, and is integrated into the front actuator of the cleaning robot (30±5mm from the panel surface).

[0042] The ultrasonic probe mounted on the robot operates using the pulse-echo method: it emits ultrasonic pulses with a width of 50 ns (repetition frequency of 1 kHz) towards the surface of the photovoltaic panel, and receives the reflected echoes through a piezoelectric transducer. The difference in acoustic impedance between the dirt and the photovoltaic glass (acoustic impedance 3.7 × 10⁶ kg / (m²·s)) causes characteristic changes in the echo signal in both the time domain (time delay difference Δt) and the frequency domain (power spectrum centroid f_c). For example, bird droppings (acoustic impedance 1.2 × 10⁶ kg / (m²·s)) have a Δt of 2.3 μs and a f_c of 2.1 MHz; dust (acoustic impedance 0.8 × 10⁶ kg / (m²·s)) has a Δt of 1.5 μs and a f_c of 1.3 MHz; and oil stains (acoustic impedance 0.95 × 10⁶ kg / (m²·s)) have a Δt of 1.8 μs and a f_c of 1.7 MHz. The analog signal is digitized by a high-speed ADC (16-bit, 100MS / s sampling rate), and then transmitted to the edge computing unit after FPGA preprocessing (filtering, envelope detection).

[0043] The hardness calculation employs an improved ultrasonic feature fusion algorithm, as shown in the following formula: H = α·(A / A0) + β·E, in: H is the stain hardness index (dimensionless, range 0-10, the larger the value, the higher the hardness). A Shore hardness tester (accuracy ±1HA) was used to calibrate 1000 typical stain samples (dust, bird droppings, oil stains, rust, etc.) in a laboratory environment to establish the mapping relationship between the hardness index and the actual hardness (R2=0.96). A represents the peak echo amplitude (in mV); the envelope is extracted by performing a Hilbert transform on the received signal, and the maximum value of the envelope is taken as the peak amplitude, with a sampling accuracy of ±0.5mV. A0 is the reference amplitude (unit: mV) of the photovoltaic panel cleaning surface; during the system initialization phase, 30 standard cleaning components are ultrasonically scanned, and the average value is taken as the reference value (usually 450±20mV), and automatically calibrated every 7 days; E represents the spectral energy density of the echo signal (unit: dB / Hz); perform a Fast Fourier Transform (FFT, 1024 points) on the echo signal to calculate the energy integral in the 1-5MHz frequency band with a resolution of 0.1dB / Hz; α and β are weighting coefficients (α=0.6, β=0.4), which are obtained by optimization through a genetic algorithm: using the actual hardness of the stain measured in the laboratory as the label, a loss function is constructed, and α and β are iteratively optimized by cross-validation (5-fold) to minimize the root mean square error (RMSE) of the prediction error (final RMSE=0.32).

[0044] Cleaning parameters are dynamically matched based on hardness index: when H≥7 (hard stains, such as bird droppings, cement stains), a high-pressure rotary brushing mode is triggered (brush head speed 300rpm, water pressure 8MPa, brushing time 5s / point); 5≤H<7 (medium-hard stains, such as rust, resin) uses a medium-pressure rinsing + brush combination (water pressure 4MPa, brush head speed 150rpm); H<5 (soft stains, such as dust, fallen leaves) uses low-pressure spray (water pressure 1.5MPa) combined with airflow blowing (wind speed 15m / s). The addition of an ultrasonic detection module improves the stain type identification accuracy to 94.7%, an increase of 12.4 percentage points compared to traditional visual recognition (82.3%), while avoiding the risk of component microcracks caused by mechanical contact (annual failure rate reduced to 0.03%). Detection data is transmitted to the digital twin platform in real time, updating the stain hardness attribute layer in the virtual model and providing mechanical constraints for path planning. The stain detection subsystem is connected to the image acquisition subsystem via an HDMI 2.1 interface (transmission bandwidth 48Gbps). It uses an improved YOLOv5s algorithm (7.5M parameters, inference speed 62fps) to identify stains in the image and generate a location distribution heatmap. It integrates a lidar (scanning frequency 10Hz, angular resolution 0.1°) to acquire the three-dimensional point cloud of the site. Through coordinate transformation (converting image pixel coordinates (u,v) to world coordinates (X,Y,Z)), it calculates the actual area (unit cm²) and center coordinates (accuracy ±2cm) of the stain.

[0045] The improvements in YOLOv5s include: (1) Mosaic-9 data augmentation is used at the input end (random scaling, rotation, and color gamut distortion are added on the basis of traditional 4-image stitching), which improves the recall rate of small target detection by 9.2%; (2) The backbone network introduces the CBAM attention mechanism (channel attention + spatial attention) to dynamically adjust the stain feature weights; (3) The neck network adds the ASFF (adaptive feature fusion) module to solve the multi-scale feature alignment problem; (4) The output end adopts the CIoU loss function, which improves the bounding box regression accuracy by 3.7%. The model is trained on a dataset containing 100,000 photovoltaic stain images (VOC format, 8 types of stains), and the mAP@0.5 reaches 96.3%, and the mAP@0.5:0.95 reaches 82.1%, which meets the real-time detection requirements (single frame processing time <16ms).

[0046] The stain type determination subsystem classifies stains based on a multimodal feature fusion model (CNN+Transformer) and outputs a cleaning mode decision vector. The model input consists of image features (output of the C3 module of YOLOv5s, dimension 512×16×16), ultrasonic features (hardness index H, spectral energy density E), and environmental features (temperature T, humidity RH). After feature concatenation (dimension 512×16×16 + 3 = 131075), the data is input into a Transformer encoder (6 layers, 8 attention heads). After Softmax activation, the output is the probability distribution of 8 types of stains (dust 0.32, bird droppings 0.45, oil stains 0.12, leaves 0.05, rust 0.03, cement 0.01, insects 0.01, others 0.01). The classification result is the one with the highest probability.

[0047] The classification model training employs a transfer learning strategy: pre-trained weights are derived from the ImageNet dataset, fine-tuned on the photovoltaic stain dataset (learning rate 0.001, batch size 32, 500 epochs), and early stopping (patience=20) is used to prevent overfitting. The final classification accuracy reaches 97.8%, and the F1-score reaches 0.96. The type determination results are encapsulated in JSON format (including stain ID, type, confidence score, location coordinates, and hardness index) and published to the digital twin platform's message queue via the MQTT protocol (QoS=1) for use by the fusion decision subsystem. The system supports online learning, automatically triggering incremental model updates every 500 new samples (using a federated learning framework to protect data privacy), ensuring continuous optimization of classification performance.

[0048] As an optional embodiment of the present invention, the stain type determination subsystem may optionally implement cleaning mode decision-making through a stain identification and judgment model driven by a digital twin. This model is integrated into the model layer of the digital twin management and control platform and continuously optimizes parameters through bidirectional feedback of physical sensor data and virtual simulation data.

[0049] The stain recognition and judgment model adopts an improved deep belief network (DBN) architecture, which includes an input layer (multimodal feature fusion), a hidden layer (3-layer restricted Boltzmann machine RBM), and an output layer (cleaning mode decision). Its structure is linked in real time with the virtual stain database of the digital twin platform, supporting online incremental learning.

[0050] The input layer employs a multi-source heterogeneous feature fusion mechanism, expressed as: X = [Ximg; Xult; Xtwin], Where X represents the fused feature vector (dimension 1×517), which consists of three parts: image feature vector Ximg (output features of YOLOv5s, 512-dimensional), ultrasonic feature vector Xult (hardness index H, spectral energy E, 2-dimensional), and digital twin environment feature vector Xtwin (temperature T, humidity RH, light intensity L, 3-dimensional); the total number of input features (n=512+2+3=517) is processed by feature concatenation and standardization (Z-score, μ=0, σ=1) and then input into the hidden layer.

[0051] The hidden layer adopts a 3-layer RBM stacked structure. Each layer is pre-trained using the contrastive divergence (CD-k) algorithm and then fine-tuned through backpropagation. The expression for the first layer is: , , The formula is the core probability calculation formula of the Restricted Boltzmann Machine (RBM) in the stain recognition and judgment model. It is used to realize the bidirectional probability mapping between the input layer and the hidden layer. The definitions of each term are as follows: Conditional probability, representing the probability that the j-th neuron in the hidden layer will be activated (with a value of 1) given the visible layer state vector v; Conditional probability, representing the probability that the i-th neuron in the visible layer will be activated (with a value of 1) given the hidden layer state vector h; sigmoid(⋅): An improved sigmoid activation function, expressed as σ(z;θ)=1 / (1+e−θz), where θ=1.2. This was calibrated through digital twin simulation experiments to maximize the gradient of the activation function within the stain feature interval [0.3,0.7], thus improving discriminative power. The activation output (0-1) of the j-th neuron in the hidden layer represents the intensity of the neuron's response to the input features; The visible layer's i-th input feature value comes from multimodal fusion features (image features, ultrasound features, and digital twin environment features). The connection weights between the i-th neuron in the visible layer and the j-th neuron in the hidden layer are initialized through pre-training using a digital twin virtual sample library (range [-0.5, 0.5], precision ±0.001). Their values ​​are positively correlated with the importance of stain features (e.g., the texture feature weight of bird droppings is 2.3 times higher than that of dust). : The bias of the j-th neuron in the hidden layer, with an initial value of 0.1, is dynamically adjusted based on the cleaning error fed back by the digital twin platform (updated once every 100 cleaning tasks). : The bias of the i-th neuron in the visible layer, used to adjust the baseline response level of the input features (the initial value is not specified in the document, but is usually set according to the distribution of the training data).

[0052] The two formulas are core components of the hidden layers of a Deep Belief Network (DBN). Through pre-training with the Contrastive Divergence (CD-k) algorithm, they enable unsupervised learning of multimodal stain features (images, ultrasound, environmental data). Specifically: The previous formula is used to infer the hidden layer state from the input features (visible layer) and extract the abstract features of the stain (such as hardness, higher-order correlation of texture). The latter equation is used to reconstruct the visible layer input from the hidden layer state. By minimizing the reconstruction error, the network parameters are optimized to ensure the robustness of feature extraction.

[0053] Combining digital twin technology, weight and bias The model can be pre-trained and initialized using a virtual sample library (100,000+ virtual stain scenarios) and dynamically adjusted based on physical cleaning feedback, thereby improving the model's accuracy in recognizing complex stain types to 98.2%.

[0054] The output layer employs a Softmax regression combined with a digital twin simulation verification mechanism to output a cleaning mode decision vector (dimension 1×8, corresponding to 8 cleaning modes), the expression of which is: , , , in: : The linear combination result of the kth neuron in the output layer, corresponding to the original score of the kth cleaning mode; The connection weights from the j-th neuron in the hidden layer to the k-th neuron in the output layer are optimized through the reinforcement learning module of the digital twin platform (reward function R = cleaning success rate − energy consumption penalty, updated once every 1000 iterations). : The bias of the kth neuron in the output layer, with an initial value of 0.05, which is positively correlated with the probability of the stain type (e.g., bird droppings correspond to c_2=0.3). The linear weighted sum of hidden layer features and weights reflects the contribution of input features to the k-th cleaning mode. : The probability that the input feature X belongs to the k-th cleaning pattern (normalized to [0,1]); Softmax(⋅): The output layer activation function, which converts linear scores... Convert to a probability distribution, ensuring that the sum of the probabilities of all patterns is 1.

[0055] Temperature parameters ( =0.8), controlling the sharpness of the probability distribution ( The smaller the value, the more concentrated the probability distribution; digital twin simulation shows... (Decision robustness is best when the value is 0.8). The specific calculation formula for the Softmax function is achieved through probability transformation via exponential normalization. : Optimal cleaning mode decision value, i.e., the index of the cleaning mode with the highest probability (e.g., k=2 corresponds to "high-pressure rotary brushing mode"). : Take the pattern index corresponding to the maximum probability to realize the mapping from probability distribution to specific decision.

[0056] These three formulas constitute the decision output module of the stain recognition and judgment model, and its core logic is as follows: Through formula The abstract features from the hidden layer are converted into raw scores for each cleaning mode; Through formula The scores are normalized into a probability distribution to quantify the degree of fit of different modes; Through formula Select the optimal mode to generate the final cleaning decision.

[0057] Combining digital twin technology, weight and bias The model can be pre-trained using virtual simulation data (100,000+ virtual stain scenarios) and dynamically optimized based on physical cleaning feedback, resulting in a model error rate of only 0.8% for matching cleaning patterns of 8 types of stains, and a secondary cleaning rate reduced from the traditional 23% to 3.7%.

[0058] The model training and optimization process is deeply integrated with digital twin technology: (1) Data augmentation: 100,000+ virtual stain samples (covering different lighting, temperature, and mixed stain scenarios) are generated through the digital twin platform and combined with 50,000 real samples collected physically to form a training set (virtual samples account for 67%), solving the problem of scarcity of real samples; (2) Simulation verification: The cleaning mode output by the model is first pre-performed in the digital twin environment, and the cleaning effect (such as the amount of stain residue and the risk of component damage) is evaluated by a virtual robot. If it does not meet the standard, the model parameters are adjusted (such as increasing the weight W of the corresponding stain type); (3) Online update: After the physical cleaning is completed, the feedback subsystem transmits the actual effect (cleaning rate, power generation gain) to the platform and optimizes the weight V through gradient descent (learning rate η=0.001, iteration 50 times / day). Therefore, the model classification accuracy reaches 98.2% (11.5% higher than pure physical sample training), the decision time is reduced to 8ms (meeting the real-time requirements), the cleaning mode matching error rate is only 0.8%, and the secondary cleaning rate of stubborn stains is reduced from the traditional 23% to 3.7%. As the core intelligent module of the digital twin, this model realizes closed-loop optimization of "virtual training - physical execution - data feedback", providing high-confidence decision support for the formulation of cleaning strategies.

[0059] The stain recognition and judgment model is based on a deep belief network (DBN) enhanced by digital twins. Its training process integrates physical sensor data and virtual simulation data: virtual samples are created in batches (2000 sets per day) through the virtual stain generator of the digital twin management platform (supporting customizable stain morphology, lighting, and material parameters), forming a hybrid training set with physically collected samples (100,000 sets / year). The model is deployed on edge computing nodes (NVIDIA Jetson AGX Orin), using TensorRT to accelerate inference (frame rate ≥ 50fps), and version control and online updates (update cycle ≤ 24 hours) are achieved through the model management module of the digital twin platform.

[0060] As an optional embodiment of the present invention, the photovoltaic panel integration decision subsystem is optionally deeply integrated with the digital twin management and control platform, including: The data preprocessing subsystem is used to perform spatiotemporal alignment and feature enhancement on physical sensor data (stain type, hardness, location density) and digital twin virtual data (simulated cleaning effect, equipment wear prediction), and output a standardized decision matrix (500×8 dimensions, rows represent cleaning tasks, and columns represent decision features).

[0061] like Figure 3 As shown, the data preprocessing process includes: (1) Data cleaning: using the 3σ criterion to remove outliers (error > ±3σ) from the ultrasonic sensor, and smoothing the lidar position data through Kalman filtering (Q=0.01, R=0.1); (2) Spatiotemporal registration: aligning image frames (30fps) and ultrasonic samples (1kHz) based on timestamps (accuracy 1ms), and unifying physical and virtual coordinates through a coordinate transformation matrix (provided by the digital twin platform as the site reference coordinate system); (3) Feature fusion: using principal component analysis (PCA) to reduce the high-dimensional image features (512 dimensions) to 64 dimensions, and concatenating them with hardness index H, location density D, and environmental parameters (T, RH) to form a 72-dimensional feature vector; (4) Virtual enhancement: randomly extracting 10% of the virtual feature vectors from the digital twin virtual sample library, and using the SMOTE algorithm to solve the imbalance of physical sample categories (bird droppings samples account for only 8%). The preprocessed data is pushed to the strategy formulation subsystem through a message queue (Kafka, throughput > 1000 messages / second).

[0062] The strategy formulation subsystem, connected to the data preprocessing subsystem and the model layer of the digital twin management and control platform, formulates a globally optimal cleaning strategy based on an improved genetic algorithm (GA). It outputs a decision package containing a three-dimensional cleaning path (X, Y, Z coordinate sequence, accuracy ±0.05m), a cleaning parameter matrix (force P: 1-10MPa, time t: 1-10s, brush head rotation speed v: 50-300rpm), and an equipment scheduling scheme. The cleaning path planning employs a digital twin spatial simulation mechanism: In the virtual site model, the photovoltaic array is divided into 0.5m × 0.5m grid cells. The dirt density Di,j (unit: cells / m²) of each cell is calculated using Gaussian kernel density estimation (bandwidth h = 0.3m), generating a heatmap (color range: blue (D<5) - red (D>50)). Based on this, the digital twin-enhanced genetic algorithm (DT-GA) encodes the path as chromosomes (length N = number of grid cells), with the fitness function defined as: F = α·(1 / Tpath) + β·(1 / Econs) + γ·Crate, where Tpath is the path time (s), Econs is the energy consumption (kWh), Crate is the predicted cleanliness rate (%), and the weights are α = 0.4, β = 0.3, and γ = 0.3 (calibrated through digital twin simulation experiments). The algorithm iterates 50 times (population size 50, crossover probability 0.7, mutation probability 0.05), pre-simulating the collision risk of each path in a virtual environment (a distance <0.2m from an obstacle within the site is considered a collision), and finally outputs the optimal path with no collisions and high coverage (>99%). Cleaning parameter decisions are implemented through fuzzy logic reasoning: input stain type (8 categories), hardness index H (0-10), output force P=2+H×0.8 (MPa), time t=1+H×0.9 (s), rotation speed v=50+H×25 (rpm). The rule base is generated by training on over 100,000 virtual cleaning experimental data from the digital twin platform.

[0063] The optimization subsystem, together with the strategy formulation subsystem and the feedback subsystem, forms a closed loop. It employs a digital twin feedback optimization algorithm (DT-FOA) to dynamically adjust strategy parameters based on physical cleaning effect feedback (cleaning rate Creal, equipment wear ΔE), and outputs optimized cleaning commands (update frequency 1Hz). This algorithm integrates multi-objective particle swarm optimization (MOPSO) with a virtual-real deviation compensation mechanism to achieve dynamic parameter calibration.

[0064] The optimization process includes three stages: (1) Deviation analysis: Calculate the absolute error ε = |Csim-Creal| between the digital twin predicted cleaning rate Csim and the physical measured Creal. Optimization is triggered when ε > 5%; (2) Parameter adjustment: Use MOPSO to optimize the cleaning parameters (P, t, v), with the objective function min(ε, Econs), constraints P ≤ 10MPa (avoid component damage), t ≤ 15s (efficiency constraint), particle dimension 3, population size 30, and 20 iterations; (3) Virtual verification: Input the optimized parameters into the digital twin model for simulation. If ε < 3% in 3 consecutive simulations, it is sent to the execution system; otherwise, it is readjusted. Through this mechanism, the cleaning parameters dynamically respond to changes in stain characteristics (e.g., when the hardness of oil stains decreases by 20% after rain, P is automatically reduced to the original 75%), so that the actual cleaning rate is stabilized above 95%.

[0065] The resource scheduling subsystem, based on the digital twin multi-machine collaborative mechanism, improves the A* algorithm to achieve task allocation and path conflict resolution for N cleaning robots (N≤50), increasing equipment utilization to 92% (compared to 65% for traditional scheduling).

[0066] The scheduling process is as follows: (1) Task division: The digital twin platform divides the site into N sub-regions equally according to the number of robots N. K-means clustering (the distance metric is Manhattan distance) ensures that the total amount of dirt in each region is less than 10%; (2) Conflict detection: Based on the improved A* algorithm, the robot position is tracked in real time in the virtual space (update frequency 10Hz). When the distance between two robots is less than 1.5m, conflict resolution is triggered (priority mechanism: the one with higher power is given priority, and if the power is the same, the area with higher dirt density is given priority); (3) Dynamic adjustment: If a robot fails (the remaining lifespan is predicted to be less than 2 hours by the digital twin model), its task is automatically assigned to an idle device and the path is replanned (response time < 5s). The scheduling command is issued through the 5G industrial module (latency < 20ms) to ensure that the synchronization error of the movement trajectory between the physical robot and the digital twin is less than 0.1m.

[0067] As an optional embodiment of the present invention, the strategy formulation subsystem may optionally formulate a cleaning strategy through a digital twin-enhanced genetic algorithm (DT-GA), which uses virtual simulation results as a key basis for fitness evaluation, thereby significantly improving optimization efficiency.

[0068] The DT-GA algorithm process includes: (1) Chromosome encoding: real number encoding is used. Each chromosome contains a path sequence (grid cell index) and cleaning parameters (P, t, v). The length L = number of grids + 3; (2) Population initialization: 50 chromosomes are randomly generated, of which 30% are from the digital twin's historical best path library; (3) Fitness evaluation: the cleaning process corresponding to each chromosome is simulated in parallel in the digital twin platform (time < 10s / chromosome), and Tpath, Econs, Csim are output, and the F value is calculated; (4) Genetic operations: selection (roulette wheel selection, elite retention rate 10%), crossover (arithmetic crossover, probability 0.7), mutation (Gaussian mutation, mean 0, variance 0.1, probability 0.05); (5) Convergence judgment: when the maximum F value change rate is < 1% for 5 consecutive generations, the optimal chromosome is output. Through digital twin simulation, the number of algorithm iterations was reduced from 200 generations in pure physical experiments to 50 generations, the optimization time was shortened by 75%, and the actual execution success rate of the optimal path was increased to 98% (compared to only 82% in traditional GA).

[0069] As an optional embodiment of the present invention, the optimization subsystem may optionally perform dynamic calibration of the cleaning strategy through the Digital Twin Feedback Optimization Algorithm (DT-FOA), the core of which is to establish an error compensation mechanism between virtual prediction and physical measurement.

[0070] The objective function expression of the DT-FOA algorithm (used for dynamically calibrating the cleaning strategy to minimize the deviation between virtual predictions and physical measurements, as well as energy consumption costs) is as follows: , Wherein, objective function Each item is defined as follows: J: The overall optimization objective should be the value minimized by the algorithm, reflecting the overall performance of the cleaning strategy; , Weighting coefficients are used to balance the optimization priorities of error and energy consumption (usually calibrated through digital twin simulation experiments, for example, they can be set as follows). =0.7, =0.3 (to prioritize cleaning accuracy). The relative error between virtual prediction and physical measurement is used to quantify the prediction accuracy of digital twin models. Energy consumption cost of the cleaning process (unit: kWh), reflecting the economics of the strategy; Error term Each item is defined as follows: The stain removal rate (%) predicted by the digital twin model was obtained through virtual simulation; The stain removal rate (%), measured by physical testing, is calculated by the feedback subsystem using data collected by a high-definition camera and a laser contour sensor. The absolute deviation between the predicted clearance rate and the actual clearance rate, divided by... The relative error is then obtained, avoiding the influence of dimensions; The objective function is the core of the DT-FOA algorithm, which achieves two key optimization objectives by minimizing J: Reduce the discrepancy between virtual and real data: by This ensures the predictive accuracy of the digital twin model, when Parameter adjustment is triggered when it exceeds 5%; Controlling energy costs: through To avoid energy waste caused by over-cleaning, for example, limit water pressure or scrubbing time when the removal rate of stubborn stains is achieved (e.g., >95%). By combining digital twin technology, the algorithm can dynamically adjust the weights. , (such as improving under extreme weather conditions) Prioritizing cleaning effectiveness, the actual removal rate is kept stable at over 95%, while energy consumption is reduced by 35%.

[0071] As an optional embodiment of the present invention, the cleaning robot execution system may optionally adopt a modular hardware architecture driven by digital twins, including a physical execution layer and a virtual twin, and realize real-time data interaction through 5G+edge computing (upload frequency 100Hz, control command issuance frequency 50Hz).

[0072] The instruction receiving subsystem adopts an industrial-grade communication module (model: Huawei ME909s-821 5G module), supports MQTT / OPC UA dual protocols, and is used to receive cleaning instructions (JSON format, including path point coordinates, cleaning parameters, and action sequence) issued by the digital twin management and control platform. The receiving latency is <20ms and the packet loss rate is <0.1%.

[0073] The command receiving subsystem has a built-in edge computing unit (NVIDIA Jetson Nano 2GB) that parses and verifies received commands (CRC32 checksum ensures command integrity), converts the cleaning path into the robot's local coordinate system (using the coordinate transformation matrix T = site reference matrix provided by the digital twin platform × robot initial pose matrix), and caches 3 commands in case of communication interruption (supports offline operation for ≥5 minutes). Simultaneously, it uploads the command receiving status (received / in execution / completed / abnormal) to the digital twin platform in real time, forming a closed-loop command tracking system.

[0074] The motion control subsystem, connected to the command receiving subsystem via an EtherCAT bus (1ms cycle), employs a composite algorithm of model predictive control (MPC) and fuzzy PID based on digital twins. It controls the motion trajectory of the cleaning subsystem according to the cleaning commands and the virtual simulation trajectory (positioning accuracy ±0.02m, trajectory tracking error <0.05m). MPC handles path planning layer optimization, while fuzzy PID handles dynamic compensation at the execution layer. The two interact through the state observer of the digital twin platform.

[0075] The motion control subsystem hardware includes: (1) a main controller (Siemens S7-1214C DC / DC / DC PLC); (2) a servo drive system (Delta ASD-A2-0421-L servo driver + 400W servo motor, encoder resolution 17-bit); and (3) a multi-sensor fusion unit (IMU: BNO055 9-axis sensor, sampling rate 100Hz; laser rangefinder: Sick TIM561, measurement range 0.05-10m, accuracy ±15mm). The control algorithm offline tunes the PID parameters (P=1.2, I=0.8, D=0.3) through the digital twin platform, and compensates the control quantity in real time according to the position deviation between the physical robot and the digital twin (estimated by Kalman filtering) so that the dynamic response time is ≤0.1s.

[0076] The cleaning subsystem is connected to the motion control subsystem via a PROFINET bus and integrates a multi-modal cleaning actuator to execute the control commands of the motion control subsystem. It supports independent or combined operation of three modes: high-pressure water jet (pressure 1-10MPa continuously adjustable), mechanical brushing (brush head speed 50-300rpm), and hot air blowing (temperature 50-80℃, wind speed 10-20m / s).

[0077] The core component parameters of the cleaning subsystem are as follows: (1) High-pressure water pump (Italian AR RK15.25N, maximum flow rate 15L / min, pressure regulation accuracy ±0.2MPa); (2) Brushing mechanism (nylon brush head diameter 150mm, bristle hardness 65 Shore A, equipped with torque sensor (accuracy ±0.1N·m) to monitor contact pressure in real time); (3) Hot air module (heating power 2kW, nozzle outlet diameter 50mm). During the execution, the cleaning effect is simulated in real time through the digital twin platform (based on CFD flow field simulation to predict the water jet impact force distribution). When the virtual cleaning rate is <95%, the water pressure / speed parameters are automatically increased (adjustment step 0.5MPa / 20rpm). Key status parameters (such as water pump pressure and brush head temperature) are uploaded to the motion control subsystem through the Modbus RTU protocol, and then synchronized to the digital twin through the 5G module to achieve physical-virtual state consistency (error <3%).

[0078] like Figure 4 As shown, an automated cleaning layout method for photovoltaic power plants based on digital twin technology is described. This method, implemented using the aforementioned system, achieves intelligent cleaning throughout the entire process through a two-way mapping between physical and virtual spaces. The method includes the following steps: S1. Multi-source heterogeneous data fusion perception: The photovoltaic power station environmental data is acquired through the collaboration of physical sensors and digital twin virtual sensors, and the location distribution information (accuracy ±2cm), hardness information (error <0.5HA) and type information (classification accuracy 97.8%) of stains are extracted.

[0079] Implementation process: (1) Physical perception: The image acquisition subsystem (4K camera + infrared thermal imager) acquires images of the entire site every 30 minutes, and the ultrasonic detection subsystem (5MHz probe) scans the hardness of the stains simultaneously; (2) Virtual enhancement: The digital twin platform generates a virtual stain distribution prediction map based on historical data (the evolution pattern of stains in the past 3 months) (prediction accuracy of 89%); (3) Data fusion: The federated Kalman filter algorithm is used to fuse physical and virtual data to generate a spatiotemporally aligned stain feature matrix (dimension N×5, where N is the number of stains, including location (X,Y), hardness H, type T, and confidence C). The digital twin virtual sensor makes up for the blind spots of the physical sensor coverage (such as occluded components), and the federated learning framework is used to achieve collaborative training between edge nodes and cloud models to ensure data privacy and security. Therefore, the data acquisition time of a single site is shortened to 15 minutes (traditional manual detection takes 2 hours), and the feature extraction accuracy reaches 98.3%, which is 6.2 percentage points higher than pure physical perception.

[0080] S2. Digital Twin Simulation Optimization Decision: Based on the stain features extracted in S1, a cleaning strategy is formulated in the digital twin platform by improving the genetic algorithm (DT-GA), generating an optimal cleaning instruction package that includes path planning, parameter configuration, and equipment scheduling.

[0081] Implementation process: (1) Virtual modeling: Load the three-dimensional model of the site (accuracy 0.1m), the robot digital twin (including kinematic / dynamic model) and the stain digital twin (assign physical properties: density, adhesion) in the digital twin platform; (2) Strategy simulation: The DT-GA algorithm is iteratively optimized for 50 generations in the virtual environment, and each generation pre-simulates the collision risk (minimum distance to obstacles ≥0.3m), energy consumption (≤0.5kWh / thousand square meters) and cleaning rate (≥95%) of 50 paths; (3) Instruction generation: Encode the optimal strategy into machine-executable instructions (containing 200-500 path points, each point containing coordinates, cleaning parameters, and dwell time). By replacing physical trial and error with virtual simulation, the cost of strategy formulation is reduced (traditional physical testing requires 300L of water / time, while virtual simulation only requires 0.5kWh of electricity). Therefore, the time for strategy formulation is reduced from the traditional 2 hours to 10 minutes, the total length of the cleaning path is reduced by 35%, and the predicted cleaning rate reaches 98.2%.

[0082] S3. Physical-Virtual Collaborative Execution: The cleaning robot execution system receives cleaning instructions and accurately executes cleaning operations under the guidance of a digital twin, uploading physical status data in real time for dynamic calibration with the virtual twin.

[0083] Implementation process: (1) Trajectory synchronization: The physical position of the robot (by encoder + IMU fusion positioning, accuracy ±0.03m) is compared with the position of the digital twin (by simulation calculation) in real time. When the deviation is >0.1m, the PID parameters of the motion control subsystem are adaptively adjusted. (2) Parameter closed loop: During the execution of the cleaning subsystem, the actual water pressure (feedback accuracy ±0.1MPa), brush head speed (±5rpm) are compared with the digital twin command value, and the deviation is dynamically compensated by the fuzzy PID controller. (3) Anomaly handling: When the digital twin model predicts the remaining lifespan <2 hours (based on vibration sensor data, prediction accuracy 92%), it automatically switches to the backup robot. High-precision trajectory tracking and adaptive control are achieved based on the virtual-real interaction of the digital twin. Therefore, the robot trajectory tracking error is ≤0.05m, the cleaning parameter execution accuracy reaches 98%, and the task completion rate is 100% (the failure rate of the traditional non-twin system is 8%).

[0084] S4. Feedback-driven continuous optimization: The feedback subsystem collects physical cleaning results (cleaning rate, component power generation, equipment loss), compares them with digital twin prediction results, and updates model parameters through reinforcement learning to achieve continuous evolution of the cleaning strategy.

[0085] Implementation process: (1) Result acquisition: The feedback subsystem captures images of the components after cleaning using a high-definition camera (4K resolution, 30fps), an infrared thermal imager (640×512 resolution) detects local temperature anomalies (to determine whether the components are damaged due to excessive brushing), and a power generation sensor (accuracy ±0.5%) records the changes in power generation 24 hours before and after cleaning; (2) Virtual-real comparison: The digital twin platform calculates the deviation ε between the virtual cleaning rate Csim and the physically measured cleaning rate Creal, ε = |Csim-Creal|. When ε > Model optimization is triggered at 5%; (3) Reinforcement learning update: Deep deterministic policy gradient (DDPG) algorithm is adopted, state space S=(stain type, hardness, environmental parameters), action space A=(cleaning intensity P, time t, rotation speed v), reward function R=0.6×Creal+0.3×(1-Econs / Emax)-0.1×ΔE (ΔE is equipment loss), and the policy network parameters are updated through experience replay pool (capacity 10000) (learning rate η=0.001, iteration 50 times / day). A virtual feedback environment is constructed based on digital twin, and the self-evolution of cleaning strategy is realized through reinforcement learning. Therefore, after 100 days of iterative optimization, the system cleaning rate is stable at 96.5%±1.2%, and the average mean time between failures (MTBF) of the equipment is extended to 1800 hours (1200 hours for traditional system).

[0086] S5. Digital Twin Full Lifecycle Management: Real-time collection of physical subsystem status data (sensor data in the perception layer, algorithm operation logs in the decision layer, and device parameters in the execution layer) through the digital twin management and control platform, construction of system health assessment model, and realization of remote monitoring, fault early warning and predictive maintenance.

[0087] Implementation process: (1) Data acquisition: Each subsystem uploads status data via the edge gateway (Huawei AR502H) using the MQTT protocol (perception layer: image features, sensor values, sampling rate 10Hz; decision layer: algorithm iteration count, fitness value, sampling rate 1Hz; execution layer: motor current, water temperature, battery SOC, sampling rate 100Hz); (2) Virtual mapping: The digital twin platform constructs a system-level virtual model, synchronizes the physical equipment status (position, attitude, energy consumption) in real time, and automatically calibrates when the deviation exceeds the threshold (>5%); (3) Health assessment: LSTM neural network is used to predict the remaining life (RUL) of key components. The input features include vibration spectrum (1024 dimensions) and temperature curve (24-hour sequence), and the output is the predicted RUL value (error <10%); (4) Maintenance decision: When the predicted RUL is <72 hours, a maintenance work order is automatically generated (including AR guidance on spare parts model and replacement steps). The full-element virtual mapping and life cycle management of the physical system are realized through digital twin. As a result, the system operation and maintenance response time has been shortened from the traditional 4 hours to 30 minutes, the annual maintenance cost has been reduced by 42%, and the fault prediction accuracy rate has reached 93%.

[0088] Figure 5 This is a schematic diagram of the hardware architecture of the digital twin cleaning device provided in an embodiment of the present invention. The device 410, as the core node of physical-virtual interaction, integrates an edge computing unit and a digital twin model operating environment, including: a first processor 2001 (digital twin main controller), an edge computing module, a multi-protocol communication interface, and a highly reliable power supply system.

[0089] Optionally, the cleaning equipment 410 hardware configuration includes: (1) Main controller: NVIDIA Jetson AGX Orin (8-core ARM Cortex-A78AE, 200TOPS AI computing power); (2) Memory: 256GB NVMe SSD (for storing digital twin models and historical data) + 32GB LPDDR5 (running memory); (3) Transceiver: Huawei ME909s-821 5G module (supports SA / NSA dual mode) + Siemens X204-2 Ethernet module (supports PROFINET / Modbus TCP); (4) Sensor interface: 8 analog inputs (4-20mA), 16 digital I / Os.

[0090] The first processor 2001 is connected to the memory 2002 (read / write speed 3.5GB / s) and transceiver 2003 (5G peak rate 2.5Gbps) via PCIe 4.0 bus, and communicates with the execution layer devices via industrial real-time bus (EtherCAT, period 1ms) to ensure the synchronization accuracy (<10ms) between the digital twin model and the physical system.

[0091] The following is combined with Figure 5 The functional implementation of the core hardware modules is explained below: The first processor 2001 serves as the core of the digital twin engine, running three major functional modules: (1) Model calculation module: constructing a virtual scene of a photovoltaic power station based on the Unity 3D engine (1 million+ polygons, 30fps rendering frame rate); (2) Data fusion module: using a CUDA-accelerated Kalman filter algorithm (processing latency <5ms) to fuse multi-source sensor data; (3) Optimization decision module: deploying DT-GA and DDPG algorithms (inference time <20ms / time). Its computing power allocation is as follows: digital twin simulation (40%), AI inference (30%), data processing (20%), and system management (10%).

[0092] Optionally, the first processor 2001 deploys digital twin applications via Docker containerization: (1) virtual scene service (based on ROS 2 Humble); (2) model training service (TensorFlow 2.10); (3) data platform service (InfluxDB time series database, write rate 100,000 points / second). Kubernetes is used to achieve elastic scaling of services, supporting collaborative control of 1-50 robots.

[0093] In a specific implementation, as one example, the cleaning equipment 410 is configured with a dual redundant processor architecture: the main processor (Jetson AGX Orin) and the backup processor (Jetson Xavier NX) check each other through heartbeat signals (100ms interval). When the main processor fails (three consecutive heartbeats are lost), the backup processor takes over all functions within 500ms to ensure system availability (MTBF>50,000 hours).

[0094] The memory 2002 adopts a hierarchical storage architecture: (1) Cache: 32GB LPDDR5 (bandwidth 102GB / s) stores real-time simulation data; (2) Local storage: 256GB NVMe SSD (random read / write IOPS 300,000) stores digital twin models (approximately 50GB / site) and historical cleaned data (retaining the most recent 3 months, approximately 200GB); (3) Cloud backup: Key data is uploaded to Alibaba Cloud OSS at regular intervals (2 AM daily) via the 5G module (encrypted transmission, AES-256 encryption). The first processor 2001 controls the memory read / write via the NVMe protocol to achieve fast loading of the digital twin model (startup time <30 seconds).

[0095] Optionally, the storage unit 2002 uses an industrial-grade wide-temperature SSD (operating temperature -40℃ to 85℃), supporting TRIM commands and SMART health monitoring. Data redundancy is achieved through RAID 1, automatically switching to the backup disk in case of single disk failure (switching time <1 second). It is connected to the first processor 2001 via a PCIe 4.0 x4 interface, with measured sequential read and write speeds reaching 3200MB / s and 2800MB / s respectively, meeting the real-time rendering requirements of the digital twin model (containing over 100,000 triangular meshes). The system also features 8GB of DDR4 spare memory, automatically activated in case of main memory failure to ensure continuous operation of core algorithms.

[0096] The transceiver 2003, as the nerve center of the digital twin system, supports 5G+industrial Ethernet dual-mode communication to realize real-time data interaction between physical devices and the digital twin platform. Specifically, it includes: (1) 5G communication: using Huawei ME909s-821 module (supporting 3GPP R15 standard, peak rate 2.5Gbps downlink / 1.2Gbps uplink), responsible for transmitting high-definition images (4K / 30fps, bandwidth requirement 15Mbps) and digital twin model update packets (about 50MB / time); (2) Industrial Ethernet: connecting to the execution layer device through Siemens X204-2 module (PROFINET IO-Link protocol, cycle 1ms), transmitting control commands (≤256 bytes / line) and real-time sensor data (sampling rate 100Hz).

[0097] Optionally, transceiver 2003 is integrated with the first processor 2001 via a PCIe 3.0 interface. VLAN isolation (Virtual LAN ID = 100-105) is used at the data link layer to ensure security, and the MQTT-SN protocol (Lightweight Message Queuing) is deployed at the application layer to achieve low-power communication. Spatial-temporal alignment of physical and virtual data is achieved through hardware timestamps (1μs precision), ensuring the synchronization accuracy of the digital twin model (<10ms).

[0098] It should be noted that, Figure 5The hardware architecture shown supports the collaborative control of 1-50 cleaning robots, and load balancing (CPU utilization ≤70%) is achieved through the distributed computing module of the digital twin platform. The expansion interfaces include: (1) 4 USB 3.0 ports (for connecting local debugging devices); (2) 2 RS485 ports (Modbus RTU protocol, for connecting environmental sensors); (3) 1 HDMI 2.1 port (for outputting real-time rendering of digital twin scenes).

[0099] The digital twin cleaning equipment has the following advantages: (1) Communication delay: physical-virtual data transmission end-to-end delay <20ms (traditional system 50ms); (2) Reliability: communication availability is 99.99% through dual-channel redundancy design; (3) Computing power support: a single device can run 3 digital twin models (photovoltaic station + 2 robots) at the same time, and the model update cycle is ≤5 minutes; (4) Energy consumption optimization: dynamic power management is adopted (CPU frequency is reduced to 1.2GHz when idle), and the average power consumption is ≤35W (traditional industrial control computer 80W).

[0100] It should be understood that the first processor 2001 adopts a heterogeneous computing architecture: (1) CPU: 8-core ARM Cortex-A78AE (2.2GHz, supports lockstep cores, meets functional safety ASIL-D level); (2) GPU: NVIDIA Ampere architecture 2048 CUDA core (1.3GHz, 200TOPS AI computing power); (3) Dedicated accelerator: 2 NVDLA deep learning accelerators (support INT8 / FP16 mixed precision computing). The hard real-time performance of digital twin simulation is guaranteed by the NVIDIA Jetson Linux 35.1 real-time operating system (PREEMPT_RT patch, kernel latency<100μs).

[0101] The memory adopts a three-level storage architecture: (1) L1 / L2 cache: 12MB (CPU private cache, access latency 1ns); (2) LPDDR5 memory: 32GB (bandwidth 102GB / s, storing real-time simulation data of digital twins); (3) NVMe SSD: 256GB (3DTLC NAND, random read / write IOPS 300,000, storing model files and historical data). It supports hot memory backup (automatically switching to the backup memory area when an ECC error is detected) to ensure the continuity of digital twin model operation.

[0102] The digital twin system software stack adopts a layered architecture: (1) Infrastructure layer: Kubernetes container orchestration (managing 5 core service containers); (2) Data layer: InfluxDB time series database (storing 100,000+ sensor data points / second) + Redis cache (accelerating access to model parameters); (3) Model layer: Unity 3D engine (virtual scene rendering) + TensorFlow 2.10 (AI model training); (4) Application layer: WebGL visualization interface (supporting remote access to digital twin scenes). The system supports OTA upgrades (upgrade time <10 minutes via 5G differential upgrade package) to ensure continuous optimization of algorithms and models.

[0103] The core technical features of the digital twin system include: (1) virtual and real two-way mapping: the physical equipment status (location, energy consumption, fault code) and the digital twin are synchronized in real time (synchronization frequency 100Hz, deviation <0.1%); (2) full-element modeling: including photovoltaic modules (100,000+ virtual units), cleaning robots (50 digital twins), and environmental factors (wind field, light, temperature field) multi-physics coupling model; (3) closed-loop optimization: continuously correct the virtual model parameters through physical feedback data (50 iterations per day), so that the prediction accuracy is improved from the initial 85% to 98%.

[0104] The key innovations of this invention are: (1) Digital twin-driven cleaning strategy optimization: by replacing physical trial and error with virtual simulation, the cost of strategy formulation is reduced by 75%; (2) Multimodal data fusion perception: physical sensors and virtual sensors work together to achieve 100% stain detection coverage; (3) Distributed robot collaborative control: based on the conflict resolution algorithm of digital twin, the equipment utilization rate is increased to 92%; (4) Full life cycle predictive maintenance: the LSTM remaining life prediction model reduces downtime due to failure by 60%.

[0105] System deployment and implementation process: (1) Site modeling: laser scanning (point cloud accuracy 5mm) + BIM modeling (Revit format) to build a digital twin base (7 days / site); (2) Equipment access: access 10 types of sensors (camera, ultrasound, IMU, etc.) through OPC UA protocol, with data acquisition delay <50ms; (3) Model training: use transfer learning (pre-trained model + 10,000 sets of site data), shortening the model convergence time to 48 hours; (4) Trial operation: digital twin simulation verification (1,000 virtual cleaning times) + physical small-scale test (3 robots), continuously optimizing parameters.

[0106] Compared with existing technologies, the core advantages of this invention are: (1) Innovation in technical architecture: the first "physical-virtual-data" three-element fusion cleaning system, breaking through the intelligent bottleneck of traditional automated cleaning; (2) Innovation in algorithm model: the DT-GA and DDPG fusion optimization algorithm is proposed to realize the self-evolution of cleaning strategy; (3) Innovation in engineering application: the modular digital twin hardware platform is developed to support rapid deployment (3 days / site) and flexible expansion (supporting 1-50 robots).

[0107] Future iteration directions: (1) Edge-cloud collaboration: migrate some digital twin simulation tasks to the cloud (Alibaba Cloud ECS g7 instance) to achieve collaborative optimization of larger-scale sites (GW level); (2) Digital twin federation: realize the sharing of model parameters of multiple sites through federated learning technology (data does not leave the site, privacy is protected); (3) Digital thread technology: connect the data link of the entire process of design-simulation-operation and maintenance to realize full life cycle digital twin management.

[0108] This invention relates to an automated cleaning layout system and method for photovoltaic power plants based on digital twin technology. By constructing a two-way mapping and real-time interaction between physical and virtual spaces, it achieves intelligent and precise operation throughout the entire photovoltaic cleaning process. Compared to traditional cleaning methods, this solution has the following significant technical advantages: 1. A breakthrough in both cleaning efficiency and effectiveness: Through multimodal perception and intelligent decision-making driven by digital twins, the removal rate of stubborn stains is improved compared with traditional manual and traditional automated equipment. The daily cleaning area of ​​a single robot can reach 12,000㎡ (compared to 8,000㎡ for traditional automated equipment), and the energy consumption per unit area is reduced by 35% (0.08kWh / ㎡).

[0109] 2. Deep Environmental Perception and Global Optimization Decision-Making: By integrating data from physical sensors (high-definition cameras, infrared thermal imagers, and ultrasonic hardness testers) and digital twin virtual sensors, the system achieves full-element perception of stain type (8 categories), hardness (0-10HA, error <0.5HA), and location (±2cm). Through an improved genetic algorithm (DT-GA), the system generates the globally optimal cleaning path, reducing the total path length by 35% and the collision risk to below 0.1%.

[0110] 3. Closed-loop control of physical-virtual collaboration: The digital twin and the physical robot are synchronized in real time (synchronization frequency 100Hz, position deviation <0.05m). The cleaning parameters (water pressure 1-10MPa, rotation speed 50-300rpm) are dynamically adjusted through model predictive control (MPC) and fuzzy PID algorithm to ensure the accurate execution of the cleaning process, with trajectory tracking error ≤0.05m.

[0111] 4. Distributed robot collaborative scheduling: Based on the multi-machine conflict resolution algorithm (improved A* algorithm) based on digital twins, the task dynamic allocation of 50 robots is realized, the equipment utilization rate is increased to 92% (65% of traditional scheduling), and the fault response time is shortened from 4 hours to 30 minutes.

[0112] 5. Predictive maintenance throughout the entire lifecycle: LSTM neural networks predict the remaining life (RUL) of critical components with an accuracy of 93%, extending the mean time between failures (MTBF) to 1800 hours (compared to 1200 hours for traditional systems) and reducing annual maintenance costs by 42%.

[0113] 6. Continuous evolution driven by digital twins: Through reinforcement learning (DDPG algorithm) and virtual-real feedback deviation compensation mechanism, the system iteratively optimizes the cleaning strategy every day. After 100 days of operation, the cleaning rate stability has been improved to 96.5%±1.2%, and the model prediction accuracy has been improved from the initial 85% to 98%.

[0114] This invention breaks through the intelligent bottleneck of traditional cleaning methods by establishing a full-process digital closed loop of "perception-decision-execution-feedback-optimization". It provides a highly efficient, accurate and low-consumption automated cleaning solution for GW-level photovoltaic power plants, with significant economic value and ecological benefits.

Claims

1. An automated cleaning layout system for photovoltaic power plants based on digital twin technology, characterized in that, include: The photovoltaic panel environmental sensing subsystem is used to collect environmental data on the surface of the photovoltaic panel through a multi-source sensor array, and extract information on the location distribution, hardness, and type of stains. A photovoltaic panel integrated decision-making subsystem is connected to the photovoltaic panel environmental perception subsystem. It is used to formulate a global cleaning strategy and generate cleaning instructions based on the stain information and digital twin simulation data through optimization algorithms. The cleaning robot execution system is connected to the photovoltaic panel fusion decision subsystem and is used to receive and execute the cleaning instructions to complete the cleaning operation on the surface of the photovoltaic panel. A feedback subsystem, connected to the cleaning robot execution system, is used to acquire physical cleaning results and feed the cleaning results back to the photovoltaic panel fusion decision subsystem to optimize the cleaning strategy. The digital twin management and control platform is connected to the photovoltaic panel environmental perception subsystem, the photovoltaic panel fusion decision-making subsystem, and the cleaning robot execution system, respectively. It is used to construct a digital twin model of the photovoltaic power station, realize bidirectional mapping between physical space and virtual space, and perform simulation optimization and collaborative management and control of the entire cleaning process.

2. The system according to claim 1, characterized in that, The photovoltaic panel environmental sensing subsystem includes: The image acquisition unit is used to acquire multi-band images of the photovoltaic panel surface; An ultrasonic testing unit is used to emit ultrasonic pulses to the surface of a photovoltaic panel and receive the echoes, and calculate the hardness information of the stains based on the echo signals. A stain detection unit, connected to the image acquisition unit, is used to identify stains in the image and generate location distribution information; The stain type determination unit is used to fuse image features, ultrasonic features, and environmental features to determine the stain type and output a cleaning mode decision.

3. The system according to claim 1 or 2, characterized in that, The photovoltaic panel integration decision-making subsystem includes: The data preprocessing unit is used to perform spatiotemporal alignment and feature fusion of physical sensor data and digital twin virtual data to generate a standardized decision matrix; The strategy formulation unit is used to formulate a cleaning strategy that includes cleaning path, parameter configuration and equipment scheduling based on the decision matrix and through the optimization algorithm enhanced by digital twin in a virtual environment. An optimization unit, connected to the feedback subsystem, is used to dynamically adjust the parameters of the cleaning strategy based on the deviation between the physical cleaning results and the virtual prediction results. The resource scheduling unit is used to allocate tasks and resolve path conflicts among multiple cleaning robots based on a digital twin model.

4. The system according to claim 1, characterized in that, The cleaning robot execution system includes: The instruction receiving unit is used to receive cleaning instructions from the photovoltaic panel fusion decision subsystem; A motion control unit, connected to the instruction receiving unit, is used to control the motion trajectory of the cleaning actuator according to the cleaning instruction based on a digital twin model predictive control algorithm. The cleaning execution unit is connected to the motion control unit and is used to perform high-pressure water jet, mechanical brushing, or hot air blowing operations.

5. The system according to claim 1, characterized in that, The feedback subsystem includes an image acquisition module, a laser contour sensor, and a power generation monitoring module, used to acquire image data, three-dimensional morphology data, and power generation change data of the cleaned surface in order to calculate the actual cleaning effect.

6. The system according to claim 1, characterized in that, The digital twin management platform includes: The virtual simulation module is used to simulate the cleaning process in a virtual environment based on the 3D model of the site, the digital twin of the equipment, and the stain evolution model. The collaborative management module is used to enable collaborative scheduling and conflict detection of multiple cleaning robots based on a digital twin model; The model optimization module is used to update the digital twin model and the cleaning decision model based on physical feedback data using a reinforcement learning algorithm.

7. The system according to claim 2, characterized in that, The stain type judgment unit integrates a stain recognition judgment model, which is a multimodal feature fusion model based on a deep belief network. Its input layer fuses image feature vectors, ultrasonic feature vectors, and digital twin environmental feature vectors, and its output layer outputs a cleaning mode decision vector.

8. The system according to claim 3, characterized in that, The strategy formulation unit employs a digital twin-enhanced optimization algorithm, which is an improved genetic algorithm. Its fitness evaluation is based on simulation results in the digital twin virtual environment. The optimization unit adopts a digital twin feedback optimization algorithm, which dynamically calibrates strategy parameters by minimizing the deviation between the virtual predicted clearance rate and the physically measured clearance rate, as well as the cleaning energy consumption.

9. The system according to claim 4, characterized in that, The cleaning robot execution system adopts a modular hardware architecture, including a physical execution layer and a corresponding virtual twin, and interacts with the digital twin management and control platform in real time through 5G communication and edge computing nodes.

10. A method for the layout of an automated cleaning system for photovoltaic power plants based on digital twin technology, characterized in that, The method is implemented based on the system of any one of claims 1 to 9, and includes the following steps: S1. Multi-source heterogeneous data fusion perception: Collect environmental data of photovoltaic power stations through the collaborative collection of physical sensors and digital twin virtual sensors, and extract information on the location distribution, hardness and type of stains; S2. Digital Twin Simulation Optimization Decision: Based on the extracted stain information, virtual simulation optimization is performed in the digital twin platform through optimization algorithms to formulate the global optimal cleaning strategy and generate cleaning instructions; S3. Physical-Virtual Cooperative Execution: The cleaning robot execution system receives the cleaning command, performs the cleaning operation under the guidance of the digital twin, and synchronizes the physical state and virtual state in real time. S4. Feedback-driven continuous optimization: Collect physical cleaning results and compare them with digital twin prediction results. Based on the deviation, update the decision model through optimization algorithms to achieve continuous evolution of the cleaning strategy. S5. Digital Twin Full Lifecycle Management: Real-time monitoring of system status through the digital twin platform, enabling health assessment, fault warning, and predictive maintenance.

Images