Intelligent cleaning robot method for photovoltaic panel of photovoltaic power station based on deep learning
By using an improved NCDE model and Lie group integral method, combined with computer vision and multi-sensor fusion control, a unified deep learning framework for photovoltaic panel cleaning robots was constructed. This framework solves the problems of low efficiency, path redundancy, and incomplete cleaning in existing photovoltaic panel cleaning robots in complex environments, achieving efficient and intelligent cleaning results.
Patent Information
- Application Number
- CN202511523010.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-10-23
- Publication Date
- 2026-01-20
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
Existing photovoltaic panel cleaning robots cannot make adaptive decisions based on dynamic changes in the degree of pollution, have unreasonable path planning, lack stability in complex environments, have low cleaning efficiency and high energy consumption, and lack a unified deep learning framework for intelligent control.
An improved NCDE model, Lie group integral method, computer vision and multi-sensor fusion control are adopted to construct a unified intelligent model to realize the identification of surface contamination of photovoltaic panels, cleaning task generation, path planning and dynamic execution control, and adaptive cleaning is performed through a deep learning framework.
It achieves high efficiency and intelligence in the photovoltaic panel cleaning process, improves the intelligence level and operational efficiency of cleaning operations, and can optimize paths and reconstruct strategies in complex environments to ensure the stability and accuracy of cleaning.
Smart Images
Figure CN121364718A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the field of intelligent robot control, and in particular to a photovoltaic panel intelligent cleaning robot method based on deep learning for photovoltaic power stations. BACKGROUND
[0002] With the rapid development of global new energy industry, photovoltaic power stations, as an important part of clean energy, their power generation efficiency is directly affected by the degree of cleanliness of the photovoltaic panel surface. During long-term operation, the photovoltaic panel surface is easily covered with dust, bird droppings, sand, snow and other pollutants. These pollutants can significantly reduce the light incidence rate, and thus affect the component output power and overall power generation efficiency. Traditional photovoltaic panel cleaning methods mainly include manual cleaning, regular mechanical cleaning and simple automatic cleaning devices. However, the existing technology generally has low cleaning efficiency, unreasonable path planning, high energy consumption, and insufficient intelligence, which makes it difficult to meet the operation and maintenance needs of large-scale photovoltaic power stations in complex environments.
[0003] In existing photovoltaic panel cleaning technologies, some automatic cleaning robots can complete periodic cleaning by setting fixed paths, but such systems usually rely on pre-set rules or simple visual detection algorithms and cannot make adaptive decisions according to the dynamic changes in pollution levels. For example, a pollution detection model based on traditional convolutional neural networks can only identify pollution areas on static images and lacks the ability to model time series pollution evolution, resulting in a mismatch between cleaning plans and pollution distribution. At the same time, existing path planning algorithms are mostly based on discrete time and static environment assumptions, making it difficult to realize real-time adjustment of the robot's path and cleaning strategy in dynamic environments. In existing deep learning applications, neural networks mostly use gradient update methods with discrete time steps, which cannot effectively handle the motion control and state changes of cleaning robots in continuous time domains. While neural networks based on ordinary differential equations can describe the continuous changes of system states, they have problems such as difficulty in data alignment and discontinuity of control signals when processing multi-source environmental inputs, resulting in insufficient stability and large prediction bias of the model in complex scenarios. Existing research still lacks a unified deep learning framework that combines time series perception, dynamic modeling and strategy optimization to realize adaptive control of the intelligent cleaning process. SUMMARY
[0004] One object of the present application is to provide a photovoltaic panel intelligent cleaning robot method based on deep learning for photovoltaic power stations. The present application makes full use of improved NCDE models, Lie group integral methods, computer vision and multi-sensor fusion control technologies, and introduces a unified intelligent model in the links of photovoltaic panel surface pollution identification, cleaning task generation, path planning and dynamic execution control, to realize the efficiency and intelligence of the photovoltaic panel cleaning process.
[0005] According to the photovoltaic panel intelligent cleaning robot method based on deep learning of the photovoltaic power station, the following steps are included: Collecting original image data of the photovoltaic panel surface and preprocessing; Inputting the preprocessed original image data into a feature encoding module of the improved NCDE model, extracting time sequence feature information of the photovoltaic panel surface, and outputting a pollution feature sequence; Inputting the pollution feature sequence into a control flow construction and time alignment module of the improved NCDE model, generating a unified control signal sequence by using a mapping feature mechanism; Inputting the unified control signal sequence into a dynamics evolution module of the improved NCDE model, introducing a Lie group integral method for state evolution calculation, and outputting a dynamic strategy representation; Inputting the dynamic strategy representation into a readout and strategy coupling module of the improved NCDE model, generating a cleaning task instruction set, and combining array layout information and obstacle information of the photovoltaic panel to generate an optimal cleaning path; Moving the photovoltaic panel intelligent cleaning robot along the optimal cleaning path, adjusting the moving speed, cleaning arm angle and adsorption pressure in real time through an execution mechanism to execute the cleaning task, and generating cleaning path data; According to the cleaning path data, collecting photovoltaic panel surface image data in real time, and inputting the improved NCDE model for cleaning effect detection to generate a compensation cleaning path; Uploading the cleaning path data, the compensation cleaning path and the running state data of the intelligent cleaning robot to the photovoltaic power station operation and maintenance management system to update the parameters of the improved NCDE model.
[0006] Optionally, the original image data includes visible light image data, infrared image data, depth image data and image acquisition metadata, and the preprocessing includes image denoising, brightness correction, distortion correction and edge enhancement.
[0007] Optionally, the output of the pollution feature sequence specifically includes: Arranging the preprocessed original image data in the order of collection time to form a time sequence input sequence; Performing size unification and brightness normalization processing on the time sequence input sequence to generate a normalized image sequence; Inputting the normalized image sequence into the feature encoding module of the improved NCDE model, extracting spatial features through a convolution layer, extracting time sequence correlation features through a time encoding layer, and fusing and mapping the spatial features and the time sequence correlation features to obtain latent feature representations corresponding to time indexes, outputting a latent observation feature vector, and the convolution layer extraction refers to multi-scale convolution operation on local regions of the normalized image sequence, and the time encoding layer extraction refers to time recursive calculation on adjacent frames of the normalized image sequence; The potential observation feature vectors are time-aggregated to generate time-enhanced feature vectors, and the generation process is a nonlinear combination between the potential observation feature vectors to obtain the time evolution relationship of the pollution feature, forming the time-enhanced feature vector; Based on the time-enhanced feature vector, a pollution identification result at a time index is generated, and combined in time sequence to form a pollution feature sequence, and the pollution identification result is obtained by spatial regression, class discrimination and grade estimation on the time-enhanced feature vector, and the pollution area coordinates, pollution type and pollution grade combination result.
[0008] Optionally, the generation of the unified control signal sequence specifically includes: The pollution feature sequence is taken as a time reference to collect environmental and running state data corresponding to the time index, and the pollution feature sequence is synchronously registered in the time dimension to form a time-synchronous data set, and the time-synchronous data set includes illumination intensity, environmental temperature, position data and attitude data; The time-synchronous data set is subjected to time interpolation and data completion operations, and a linear interpolation method is used to continuously process the data segment with missing sampling intervals to generate a time-continuous multi-source input sequence; The time-continuous multi-source input sequence is input into the control flow construction and time alignment module of the improved NCDE model, and based on the feature mapping mechanism of the control flow construction and time alignment module, the pollution area coordinates, pollution type, pollution grade, illumination intensity, environmental temperature, position data and attitude data are fused by feature mapping to generate a continuous-time control signal; The continuous-time control signal is subjected to normalization and smoothing processing to form a unified control signal sequence, and the unified control signal sequence includes pollution state control components, environmental parameter control components, motion state control components and time index identification components.
[0009] Optionally, the output of the dynamic strategy representation specifically includes: The unified control signal sequence is input into the dynamics evolution module of the improved NCDE model to establish a continuous-time nonlinear state evolution equation, and the Lie group integral method is used to solve the equation to obtain a state evolution path, and the state evolution path is obtained by calculating the exponential mapping and group multiplication of the state vector in the Lie group space at each time step, and continuously evolving the states at adjacent time points in the manifold space to obtain the state evolution path from the initial time to the termination time; According to the state evolution path, a feature aggregation operation is performed on the continuous-time state vector to obtain a global dynamic state representation of the photovoltaic panel intelligent cleaning robot; Based on the global dynamic state representation, a dynamic policy representation is obtained by using a nonlinear mapping and feature decoding, the dynamic policy representation including a task weight parameter, a path optimization vector, an execution control parameter, and an environmental feedback factor.
[0010] Optionally, the generation of the optimal cleaning path specifically includes: The dynamic policy representation is input into a readout and policy coupling module of the improved NCDE model, and the pollution area coordinates, the pollution type, and the pollution level are weighted and fused and feature mapped according to the task weight parameter in the dynamic policy representation, so as to construct a pollution task matrix. The pollution task matrix is prioritized and parameter-assigned to generate a cleaning task instruction set, the cleaning task instruction set including the pollution area coordinates, the cleaning priority, and the cleaning mode. Based on the cleaning task instruction set, the cleaning path planning model is established in combination with the photovoltaic panel array layout information and the obstacle information, the establishment being a multi-objective optimization solution based on path continuity constraints, obstacle avoidance constraints, array structure constraints, and energy consumption constraints, so as to form the cleaning path planning model. The cleaning path planning model is solved by using a combination of dynamic programming method and gradient optimization method to generate the optimal cleaning path.
[0011] Optionally, the generation of the cleaning path data specifically includes: The photovoltaic panel intelligent cleaning robot executes the cleaning task along the optimal cleaning path, and drives the execution mechanism to adjust the moving speed, the cleaning arm angle, and the adsorption pressure in real time. In the cleaning process, the position, the posture, and the cleaning pressure of the photovoltaic panel intelligent cleaning robot are monitored in real time by using a multi-sensor fusion method, and the monitored data are fused to form a real-time state vector. According to a difference between the real-time state vector and a target state of the optimal cleaning path, state error information is calculated, and the state error information is input into a feedback control system of the photovoltaic panel intelligent cleaning robot, and is corrected in real time in combination with the dynamic policy representation to form a corrected execution control instruction. Based on the corrected execution control instruction, cleaning path data are generated, the generation being that when the photovoltaic panel intelligent cleaning robot executes the cleaning task, the control system records the motion state and the working parameter of the robot at each time step in real time according to the corrected execution control instruction to form the cleaning path data, the cleaning path data including the path node coordinates, the moving speed, the cleaning arm angle, and the adsorption pressure at the corresponding time.
[0012] Optionally, the generation of the compensation cleaning path specifically includes: Real-time image data of the surface of the photovoltaic panel is collected, and the image data is synchronously registered with corresponding cleaning path data according to time indexing to generate a post-cleaning image sequence; The cleaning path data and the post-cleaning image sequence are input into the improved NCDE model to detect the cleaning effect, residual pollution areas are identified through image feature matching and time sequence difference, and a residual pollution feature value sequence is output; The residual pollution feature value sequence is statistically calculated to obtain an average cleaning effect value, and the average cleaning effect value is obtained by weighted summation and averaging of the ratios of residual pollution of all detection areas; The average cleaning effect value is compared with a preset cleaning threshold value, and when the average cleaning effect value is less than the preset cleaning threshold value, a residual pollution area identification map is generated to form a compensation cleaning path, and the compensation cleaning path is formed by recalculating the path node weight and movement direction constraint according to the area coordinate information in the residual pollution area identification map after the improved NCDE model obtains an updated dynamic strategy representation, and the Lie group integral method is used to generate an optimized path trajectory sequence in the continuous time domain.
[0013] Optionally, the updating process of the parameters of the improved NCDE model is to analyze the uploaded data, calculate the error between the prediction result of the improved NCDE model and the actual execution effect, and adjust the weight coefficient and coupling parameter of the improved NCDE model according to the error.
[0014] The beneficial effects of the present application are: The present application constructs a photovoltaic panel intelligent cleaning robot method based on deep learning for photovoltaic power stations, significantly improves the intelligent degree and operation efficiency of photovoltaic power station cleaning operation, and has outstanding technical effects and application advantages. By introducing the improved NCDE model and Lie group integral method in the system, dynamic, continuous and adaptive control of the photovoltaic panel cleaning task is realized. The method of the present application not only can accurately solve the motion state of the robot in the continuous time domain, but also can optimize the path and reconstruct the strategy according to the changes of real-time environment and pollution condition in the cleaning process, and solves the technical problems of low cleaning efficiency, path redundancy and incomplete cleaning of traditional cleaning robots in complex environment.
[0015] Compared with the prior art, the application realizes a unified deep learning framework from pollution detection to path planning to execution control. By improving the feature encoding module and control flow construction mechanism of the NCDE model, the system can model and predict the dynamic changes of the pollution of the photovoltaic panel in the time dimension, thereby generating a set of cleaning task instructions that match the environmental changes; by combining the dynamics evolution module with the Lie group integral method, the system realizes high-precision description and strategy update of the continuous motion of the robot in the state space, ensuring the physical continuity and energy optimality of the cleaning trajectory; through the readout and strategy coupling module, the system directly converts the dynamic strategy representation into execution control instructions, enabling the cleaning robot to adjust the moving speed, suction pressure and cleaning arm angle in real time, ensuring the stability and accuracy of the cleaning process.
[0016] In addition, the application introduces a multi-sensor fusion algorithm and a cleaning effect detection mechanism to build a data-driven closed-loop feedback system. After cleaning, the robot re-collects the surface image of the photovoltaic panel and uses a deep learning model to evaluate the cleaning effect. When the detection result does not reach the set threshold, the system will automatically generate a compensation cleaning path to realize self-correction and secondary optimization of cleaning. Further, the application uploads the cleaning path data, compensation path data and running state information to the photovoltaic power station operation and maintenance management system, and uses the parameter update module to adaptively adjust the weights of the improved NCDE model, so that the model has the ability of continuous learning and performance evolution. BRIEF DESCRIPTION OF DRAWINGS
[0017] The accompanying drawings are included to provide a further understanding of the application, and constitute a part of the specification, together with the embodiments of the application, to explain the application, and do not constitute a limitation on the application. In the drawings:
[0018] Fig. 1 A flowchart of a photovoltaic panel intelligent cleaning robot method based on deep learning for a photovoltaic power station according to the application; Fig. 2 A module structure diagram of an improved NCDE model of a photovoltaic panel intelligent cleaning robot method based on deep learning for a photovoltaic power station according to the application; DETAILED DESCRIPTION
[0019] The application will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams that only schematically illustrate the basic structure of the application, and therefore only show the components related to the application.
[0020] REFERENCE Figs. 1-2 A photovoltaic panel intelligent cleaning robot method based on deep learning for a photovoltaic power station, comprising the following steps: Collecting original image data of the surface of the photovoltaic panel and preprocessing; The pre-processed original image data is input into a feature encoding module of the improved NCDE model to extract time sequence feature information of the surface of the photovoltaic panel and output a pollution feature sequence; The pollution feature sequence is input into a control flow construction and time alignment module of the improved NCDE model to generate a unified control signal sequence by using a mapping feature mechanism; The unified control signal sequence is input into a dynamics evolution module of the improved NCDE model to introduce a Lie group integral method for state evolution calculation and output a dynamic strategy representation; The dynamic strategy representation is input into a readout and strategy coupling module of the improved NCDE model to generate a cleaning task instruction set and generate an optimal cleaning path in combination with array layout information and obstacle information of the photovoltaic panel; The photovoltaic panel intelligent cleaning robot is controlled to move along the optimal cleaning path, and a cleaning task is executed by executing mechanism to adjust moving speed, cleaning arm angle and adsorption pressure in real time to generate cleaning path data; According to the cleaning path data, photovoltaic panel surface image data is collected in real time and input into the improved NCDE model for cleaning effect detection to generate a compensation cleaning path; The cleaning path data, the compensation cleaning path and the running state data of the intelligent cleaning robot are uploaded to a photovoltaic power station operation and maintenance management system to update parameters of the improved NCDE model.
[0021] In the embodiment, the original image data includes visible light image data, infrared image data, depth image data and image acquisition metadata, and the preprocessing includes image denoising, brightness correction, distortion correction and edge enhancement.
[0022] In the embodiment, the output of the pollution feature sequence specifically includes: The pre-processed original image data is arranged in time sequence according to acquisition time to form a time sequence input sequence; The time sequence input sequence is subjected to size unification and brightness normalization processing to generate a normalized image sequence; The normalized image sequence is input into a feature encoding module of the improved NCDE model, spatial features are extracted through a convolution layer, time sequence correlation features are extracted through a time encoding layer, and the spatial features and the time sequence correlation features are fused and mapped to obtain latent feature representations corresponding to time indexes, output latent observation feature vectors, the convolution layer extraction refers to multi-scale convolution operation on local regions of the normalized image sequence, and the time encoding layer extraction refers to time recursive calculation on adjacent frames of the normalized image sequence; The latent observation feature vectors are subjected to time sequence aggregation to generate time sequence enhanced feature vectors, the generation process is nonlinear combination between the latent observation feature vectors to obtain a time evolution relationship of pollution features, and the time sequence enhanced feature vectors are formed; Based on the time sequence enhanced feature vector, a pollution identification result at a time index is generated, and is combined in time sequence to form a pollution feature sequence, wherein the pollution identification result is obtained by spatial regression, category discrimination and grade estimation on the time sequence enhanced feature vector, and the pollution identification result is a combination of pollution region coordinates, pollution type and pollution grade.
[0023] In the embodiment, the generation of the unified control signal sequence specifically includes: With the pollution feature sequence as a time reference, environmental and operating state data corresponding to the time index are collected, and are synchronized and registered with the pollution feature sequence in the time dimension to form a time sequence synchronization dataset, wherein the time sequence synchronization dataset includes illumination intensity, environmental temperature, position data and attitude data; The time sequence synchronization dataset is subjected to time interpolation and data completion operations, a linear interpolation method is used to continuously process the data segment with missing sampling intervals, and a time-continuous multi-source input sequence is generated; The time-continuous multi-source input sequence is input into the control flow construction and time alignment module of the improved NCDE model, and based on the feature mapping mechanism of the control flow construction and time alignment module, the pollution region coordinates, pollution type, pollution grade, illumination intensity, environmental temperature, position data and attitude data are fused by feature mapping to generate a continuous-time control signal; The continuous-time control signal is subjected to normalization and smoothing processing to form a unified control signal sequence, wherein the unified control signal sequence includes a pollution state control component, an environmental parameter control component, a motion state control component and a time index identification component.
[0024] In the embodiment, the output of the dynamic strategy representation specifically includes: The unified control signal sequence is input into the dynamics evolution module of the improved NCDE model, a continuous-time nonlinear state evolution equation is established, and the Lie group integral method is used for solving to obtain a state evolution path, wherein the state evolution path is obtained by calculating the exponential mapping and group multiplication operation of the state vector in the Lie group space at each time step, and continuously evolving the states at adjacent time points in the manifold space to obtain the state evolution path from the initial time to the terminal time; According to the state evolution path, a feature aggregation operation is performed on the continuous-time state vector to obtain a global dynamic state representation of the photovoltaic panel intelligent cleaning robot: ; wherein, is the global dynamic state representation, is an activation function, is a time decay coefficient, is a time to time is integrated continuously in time, is an exponential decay function, is an aggregated weight matrix, is a nonlinear mapping function, is a feature mapping weight matrix, is a state vector, is a feature mapping bias term, is a time differential term, is a bias term; Based on the global dynamic state representation, a dynamic policy representation is obtained by using a nonlinear mapping and feature decoding, the dynamic policy representation including a task weight parameter, a path optimization vector, an execution control parameter, and an environmental feedback factor.
[0025] In the embodiment, the generation of the optimal cleaning path specifically includes: The dynamic policy representation is input into a readout and policy coupling module of the improved NCDE model, and the pollution area coordinates, pollution types, and pollution levels are weighted and fused and feature mapped according to the task weight parameter in the dynamic policy representation, to construct a pollution task matrix; The pollution task matrix is prioritized and parameter-assigned to generate a cleaning task instruction set, the cleaning task instruction set including pollution area coordinates, cleaning priorities, and cleaning modes; Based on the cleaning task instruction set, combined with photovoltaic panel array layout information and obstacle information, a cleaning path planning model is established, the establishment being a multi-objective optimization solution based on path continuity constraints, obstacle avoidance constraints, array structure constraints, and energy consumption constraints, to form the cleaning path planning model; The cleaning path planning model is solved by using a dynamic programming method combined with a gradient optimization method, to generate the optimal cleaning path.
[0026] In the embodiment, the generation of the cleaning path data specifically includes: The photovoltaic panel intelligent cleaning robot executes a cleaning task along the optimal cleaning path, to drive an execution mechanism to adjust a moving speed, a cleaning arm angle, and a suction pressure in real time; In the cleaning process, a multi-sensor fusion method is used to monitor a position, a pose, and a cleaning pressure of the photovoltaic panel intelligent cleaning robot in real time, and the monitored data is fused to form a real-time state vector; According to a difference between the real-time state vector and a target state of the optimal cleaning path, state error information is calculated, and the state error information is input into a feedback control system of the photovoltaic panel intelligent cleaning robot, combined with a dynamic policy representation to be corrected in real time, to form a corrected execution control instruction; Based on the modified execution control instruction, the cleaning path data is generated, and the generation process refers to that when the photovoltaic panel intelligent cleaning robot executes the cleaning task, the control system records the motion state and operation parameters of the robot at each time step in real time according to the modified execution control instruction, and forms the cleaning path data, wherein the cleaning path data includes path node coordinates, moving speed at the corresponding time, cleaning arm angle and adsorption pressure.
[0027] In the embodiment, the generation of the compensation cleaning path specifically includes: Real-time image data of the photovoltaic panel surface is collected, and the image data is synchronously registered with the corresponding cleaning path data according to the time index to generate a post-cleaning image sequence; The cleaning path data and the post-cleaning image sequence are input into the improved NCDE model to detect the cleaning effect, residual pollution areas are identified through image feature matching and time sequence difference, and a residual pollution feature value sequence is output; The residual pollution feature value sequence is statistically calculated to obtain an average cleaning effect value, and the average cleaning effect value is obtained by weighting and summing the ratios of residual pollution of all detection areas and taking an average value; The average cleaning effect value is compared with a preset cleaning threshold value, and when the average cleaning effect value is less than the preset cleaning threshold value, a residual pollution area identification map is generated to form a compensation cleaning path, and the compensation cleaning path is formed by recalculating path node weights and movement direction constraints according to the area coordinate information in the residual pollution area identification map after the dynamic strategy representation of the improved NCDE model is updated, and using Lie group integral method to generate an optimized path trajectory sequence in a continuous time domain.
[0028] In the embodiment, the updating process of the parameters of the improved NCDE model is to analyze the uploaded data, calculate the error between the prediction result of the improved NCDE model and the actual execution effect, and adjust the weight coefficient and coupling parameter of the improved NCDE model according to the error.
[0029] Example 1 Taking a certain photovoltaic power station as an example, the installed capacity of the power station is about 100 MW, the number of photovoltaic panels is about 260,000, and the area of a single panel is 2 square meters. The local climate is dry and windy, and the photovoltaic panel surface is seriously covered with dust all year round, which reduces the photoelectric conversion efficiency by about 10% to 15%. The traditional manual cleaning method is low in efficiency and high in labor intensity, and each person can clean at most 200 photovoltaic panels per day, and the cleaning is uneven and high in energy consumption. Therefore, the photovoltaic panel intelligent cleaning robot method based on deep learning proposed by the present application is verified on site in the power station.
[0030] In the experiment, the system first acquires the surface image data of the photovoltaic array through a high-definition camera, and uses the image preprocessing module to suppress noise and balance brightness. Subsequently, the preprocessed image sequence is input into the feature encoding module of the improved NCDE model, the spatial pollution features are extracted through the convolution layer, and the time sequence change features are extracted by the time encoding layer to form the pollution feature sequence.
[0031] The system further inputs the pollution feature sequence into the control flow construction and time alignment module, fuses the ambient light intensity, temperature and robot pose data through the mapping feature mechanism, and generates a unified control signal sequence. The dynamics evolution module solves the continuous time nonlinear state equation using Lie group integral method, outputs dynamic strategy representation, and is used to guide the generation and path planning of the cleaning task. The readout and strategy coupling module generates the cleaning task instruction set and the optimal cleaning path according to the photovoltaic panel array layout and obstacle information.
[0032] In the cleaning stage, the intelligent cleaning robot moves automatically along the planned path, and the actuator dynamically adjusts the moving speed, adsorption pressure and cleaning arm angle according to real-time feedback. The multi-sensor system carried by the robot monitors the path deviation, adsorption load and attitude angle in real time, and the control module corrects according to the feedback data to ensure the continuity of the path and the stability of the pressure. In the continuous task execution, the average cleaning area of a single robot is about 75㎡ per hour, which is nearly three times higher than that of traditional manual cleaning.
[0033] After cleaning, the system reacquires the surface image of the photovoltaic panel, inputs the cleaning path data and the image sequence after cleaning into the NCDE model for cleaning effect detection. If the cleaning effect does not reach the threshold of 95%, the system automatically calculates the residual pollution distribution and generates a compensation cleaning path using the dynamics evolution module. The compensation operation usually only involves the residual pollution area, avoiding repeated cleaning and further saving energy consumption and operation time.
[0034] Through continuous operation for one month, the experimental results show that this method has significant improvement in cleaning efficiency, energy consumption control, path precision and cleaning quality, etc.
[0035] Table 1 Performance comparison table of intelligent cleaning robot for photovoltaic power station
[0036] As can be seen from table 1, the intelligent cleaning method based on the improved NCDE model is superior to the traditional manual and conventional cleaning robot in various key indicators. First, in terms of cleaning efficiency, the method generates a dynamic strategy and optimizes the path through a deep learning model, so that the average cleaning efficiency of a single robot reaches 78 square meters per hour, which is 2.8 times that of manual work. Secondly, the introduction of the Lie group integral method in dynamics solving makes the path continuity stronger and the control signal smoother, and the path deviation is only 2.1 cm, which is significantly better than other methods.
[0037] In terms of cleaning quality, thanks to the time series modeling of pollution characteristics and the feedback compensation mechanism, the system can accurately identify the residual pollution area and perform secondary cleaning, and the final cleaning qualified rate reaches 97.6%. At the same time, due to the use of dynamic strategy control and multi-sensor fusion, the energy consumption is reduced by an average of 31%, and the repeated cleaning rate is only 2%, indicating that the system can efficiently complete the task and has higher resource utilization.
[0038] Long-term operation data shows that in the actual use of 30 consecutive days, the overall power generation efficiency of the photovoltaic power station is improved by about 11.8%, and the equipment maintenance cycle is prolonged by about 1.5 times, indicating that the method has strong stability and economic benefits in complex environments. The present application not only effectively solves the problems of low efficiency, high energy consumption and inaccurate path planning of traditional cleaning methods, but also provides a popular technical path for the intelligent operation and maintenance of photovoltaic power stations.
Claims
1.A deep learning-based photovoltaic panel intelligent cleaning robot method for photovoltaic power stations, characterized in that, The method comprises the following steps: Collecting original image data of the surface of the photovoltaic panel and preprocessing; Inputting the preprocessed original image data into the feature encoding module of the improved NCDE model, extracting the time sequence feature information of the surface of the photovoltaic panel, and outputting the pollution feature sequence; Inputting the pollution feature sequence into the control flow construction and time alignment module of the improved NCDE model, generating a unified control signal sequence by using a mapping feature mechanism; Inputting the unified control signal sequence into the dynamics evolution module of the improved NCDE model, introducing a Lie group integral method for state evolution calculation, and outputting a dynamic strategy representation; Inputting the dynamic strategy representation into the readout and strategy coupling module of the improved NCDE model, generating a cleaning task instruction set, and combining the array layout information and obstacle information of the photovoltaic panel to generate an optimal cleaning path; Moving the intelligent cleaning robot for the photovoltaic panel along the optimal cleaning path, adjusting the moving speed, cleaning arm angle and adsorption pressure in real time through the execution mechanism to execute the cleaning task, and generating cleaning path data; According to the cleaning path data, real-time collection of the image data of the surface of the photovoltaic panel, and inputting the improved NCDE model for cleaning effect detection to generate a compensation cleaning path; Uploading the cleaning path data, the compensation cleaning path and the running state data of the intelligent cleaning robot to the photovoltaic power station operation and maintenance management system to update the parameters of the improved NCDE model. 2.The photovoltaic panel intelligent cleaning robot method based on deep learning for photovoltaic power station according to claim 1, characterized in that, The original image data includes visible light image data, infrared image data, depth image data and image acquisition metadata, and the preprocessing includes image denoising, brightness correction, distortion correction and edge enhancement. 3.The photovoltaic panel intelligent cleaning robot method based on deep learning of claim 1, wherein, The output of the pollution feature sequence specifically includes: Arranging the preprocessed original image data in time sequence to form a time sequence input sequence; Performing size unification and brightness normalization on the time sequence input sequence to generate a normalized image sequence; Inputting the normalized image sequence into the feature encoding module of the improved NCDE model, extracting spatial features through a convolution layer, extracting time sequence correlation features through a time encoding layer, and fusing and mapping the spatial features and the time sequence correlation features to obtain a latent feature representation corresponding to a time index, outputting a latent observation feature vector, wherein the convolution layer extraction refers to multi-scale convolution operation on the local region of the normalized image sequence, and the time encoding layer extraction refers to time recursive calculation on adjacent frames of the normalized image sequence; Performing time sequence aggregation on the latent observation feature vector to generate a time sequence enhanced feature vector, wherein the generation process is a nonlinear combination between the latent observation feature vectors to obtain the time evolution relationship of the pollution features, forming the time sequence enhanced feature vector; Based on the time sequence enhanced feature vector, a pollution identification result at the time index is generated, and combined in time sequence to form a pollution feature sequence, wherein the pollution identification result is obtained by spatial regression, class discrimination and grade estimation on the time sequence enhanced feature vector, and the pollution area coordinates, pollution type and pollution grade combination result. 4.The photovoltaic panel intelligent cleaning robot method based on deep learning for photovoltaic power station according to claim 1, characterized in that, The generation of the unified control signal sequence specifically includes: With the pollution feature sequence as a time reference, environmental and operating state data corresponding to the time index are collected, and are synchronized with the pollution feature sequence in the time dimension to form a time series synchronization dataset, the time series synchronization dataset including illumination intensity, environmental temperature, position data, and attitude data; Time interpolation and data completion operations are performed on the time series synchronization dataset, and a linear interpolation method is used to continuously process data segments with missing sampling intervals to generate a time-continuous multi-source input sequence; The time-continuous multi-source input sequence is input into a control flow construction and time alignment module of the improved NCDE model, and based on a feature mapping mechanism of the control flow construction and time alignment module, pollution region coordinates, pollution types, pollution levels, illumination intensity, environmental temperature, position data, and attitude data are fused by feature mapping to generate a continuous-time control signal; The continuous-time control signal is normalized and smoothed to form a unified control signal sequence, the unified control signal sequence including a pollution state control component, an environmental parameter control component, a motion state control component, and a time index identification component. 5.The photovoltaic panel intelligent cleaning robot method based on deep learning for photovoltaic power station according to claim 1, characterized in that, The output of the dynamic policy representation specifically includes: The unified control signal sequence is input into a dynamics evolution module of the improved NCDE model to establish a continuous-time nonlinear state evolution equation, and the Lie group integral method is used to solve the equation to obtain a state evolution path, the state evolution path being obtained by calculating the exponential mapping and group multiplication of the state vector in the Lie group space at each time step, and continuously evolving the states of adjacent time points in the manifold space to obtain the state evolution path from the initial time to the termination time; According to the state evolution path, a feature aggregation operation is performed on the continuous-time state vector to obtain a global dynamic state representation of the photovoltaic panel intelligent cleaning robot; Based on the global dynamic state representation, a dynamic policy representation is obtained by using nonlinear mapping and feature decoding, the dynamic policy representation including a task weight parameter, a path optimization vector, an execution control parameter, and an environmental feedback factor. 6.The photovoltaic panel intelligent cleaning robot method based on deep learning for photovoltaic power station according to claim 1, characterized in that, The generation of the optimal cleaning path specifically includes: The dynamic policy representation is input into a readout and policy coupling module of the improved NCDE model, and according to the task weight parameter in the dynamic policy representation, the pollution region coordinates, the pollution types, and the pollution levels are fused and mapped by weighting to construct a pollution task matrix; The pollution task matrix is prioritized and parameterized to generate a cleaning task instruction set, the cleaning task instruction set including pollution region coordinates, cleaning priorities, and cleaning methods; Based on the cleaning task instruction set, combined with photovoltaic panel array layout information and obstacle information, a cleaning path planning model is established, the establishment being a multi-objective optimization solution based on path continuity constraints, obstacle avoidance constraints, array structure constraints, and energy consumption constraints to form the cleaning path planning model; The cleaning path planning model is solved by using a combination of dynamic programming and gradient optimization methods to generate an optimal cleaning path. 7.The photovoltaic panel intelligent cleaning robot method based on deep learning of claim 1, wherein, The generation of the cleaning path data specifically includes: The photovoltaic panel intelligent cleaning robot performs a cleaning task along an optimal cleaning path, and drives an execution mechanism to adjust a moving speed, a cleaning arm angle and a suction pressure in real time; During the cleaning process, a multi-sensor fusion method is used to monitor a position, a posture and a cleaning pressure of the photovoltaic panel intelligent cleaning robot in real time, and data obtained by monitoring is fused to form a real-time state vector; According to a difference between the real-time state vector and a target state of the optimal cleaning path, state error information is calculated, and the state error information is input into a feedback control system of the photovoltaic panel intelligent cleaning robot, and a dynamic strategy representation is corrected in real time to form a corrected execution control instruction; Based on the corrected execution control instruction, cleaning path data is generated, and the generation process refers to that, when the photovoltaic panel intelligent cleaning robot performs a cleaning task, a control system records a motion state and a working parameter of the robot at each time step in real time according to the corrected execution control instruction to form the cleaning path data, and the cleaning path data includes path node coordinates, a moving speed, a cleaning arm angle and a suction pressure at a corresponding moment. 8.The photovoltaic panel intelligent cleaning robot method based on deep learning for photovoltaic power station according to claim 1, characterized in that, The generation of the compensation cleaning path specifically includes: Real-time image data of a photovoltaic panel surface is collected, and the image data is synchronously registered with corresponding cleaning path data according to a time index to generate a post-cleaning image sequence; The cleaning path data and the post-cleaning image sequence are input into an improved NCDE model to detect a cleaning effect, residual pollution areas are identified through image feature matching and time sequence difference, and a pollution residual feature value sequence is output; Statistical calculation is performed on the pollution residual feature value sequence to obtain an average cleaning effect value, and the average cleaning effect value is obtained by weighted summation and averaging of ratios of residual pollution in all detected areas; The average cleaning effect value is compared with a preset cleaning threshold value, and when the average cleaning effect value is less than the preset cleaning threshold value, a residual pollution area identification map is generated to form a compensation cleaning path, and the compensation cleaning path is formed by recalculating path node weights and moving direction constraints according to area coordinate information in the residual pollution area identification map after the dynamic strategy representation of the improved NCDE model is updated, and using Lie group integral method to generate an optimized path trajectory sequence in a continuous time domain. 9.The photovoltaic panel intelligent cleaning robot method based on deep learning for photovoltaic power station according to claim 1, characterized in that, The updating process of parameters of the improved NCDE model is to analyze uploaded data, calculate an error between a prediction result of the improved NCDE model and an actual execution effect, and adjust weight coefficients and coupling parameters of the improved NCDE model according to the error.