Photovoltaic cleaning robot operation method and photovoltaic cleaning robot

Abstract

The invention discloses a photovoltaic cleaning robot operation method and a photovoltaic cleaning robot, and the method comprises the steps: collecting the surface pollution image data and environment perception data of a photovoltaic module in real time, and constructing a time-space correlation pollution state sequence; inputting the pollution state sequence into the edge AI prediction model, and outputting a pollution trend prediction value; constructing a dynamic pollution state transition matrix based on the pollution trend predicted value, and constructing a Markov decision process model based on the dynamic pollution state transition matrix; solving the Markov decision process model, and generating a dynamic path planning sequence considering clean energy consumption and power generation income balance; the cleaning robot is controlled to execute the dynamic path planning sequence, and pollution removal effect feedback data is obtained in real time in the operation process; and updating the parameters of the edge AI prediction model and the probability distribution of the state transition matrix on line according to the feedback data. According to the invention, the cleaning efficiency maximization, the energy consumption minimization and the long-term power generation benefit optimization can be realized.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic cleaning robot technology, and in particular to a photovoltaic cleaning robot operation method and a photovoltaic cleaning robot. Background Technology

[0002] With the widespread application of photovoltaic (PV) power generation globally, surface contamination of PV modules has become a significant factor affecting their power generation efficiency and operational lifespan. Contaminants such as dust, bird droppings, and plant debris, when adhering to the PV panel surface over a long period, can lead to attenuation of incident light and exacerbate hot spot effects, thus significantly reducing the overall power generation efficiency of the PV system. According to relevant research, power generation losses due to surface contaminants can reach over 20%, and this problem is particularly pronounced in special environments such as deserts, plateaus, and coastal areas. Therefore, conducting efficient and intelligent PV module cleaning operations has become one of the key directions of current technological development.

[0003] Most existing photovoltaic cleaning technologies rely on timed cleaning or simple visual recognition mechanisms, often lacking dynamic prediction and assessment of pollution diffusion trends, and unable to dynamically adjust cleaning paths according to the spatial distribution and evolution of pollution development. Furthermore, traditional path planning methods are mostly based on static maps or fixed task assignments, making it difficult to fully consider the dynamic balance between "clean energy consumption and power generation gain" during cleaning operations. This leads to unreasonable allocation of cleaning resources, and even inefficiencies such as "over-cleaning" or "mis-cleaning."

[0004] Therefore, there is an urgent need for a photovoltaic cleaning robot operation method that combines pollution images and environmental data to model the pollution state, has edge intelligent prediction capabilities, can dynamically plan the optimal path and achieve adaptive updates, so as to maximize cleaning efficiency, minimize energy consumption and optimize long-term power generation benefits. Summary of the Invention

[0005] The purpose of this section is to outline some aspects of embodiments of the present invention and to briefly describe some preferred embodiments. Simplifications or omissions may be made in this section, as well as in the abstract and title of this application, to avoid obscuring the purpose of these documents; however, such simplifications or omissions should not be construed as limiting the scope of the invention.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a photovoltaic cleaning robot operation method, mainly including the following steps: Real-time acquisition of surface contamination image data and environmental sensing data of photovoltaic modules to construct a spatiotemporally correlated contamination state sequence; The pollution state sequence is input into an edge AI prediction model deployed locally on the robot, and the predicted pollution trend value for a future preset time period is output. A dynamic pollution state transition matrix is ​​constructed based on the pollution trend prediction values, and a Markov decision process model is constructed based on the dynamic pollution state transition matrix. Solving the Markov decision process model yields the optimal clean energy strategy and generates a dynamic path planning sequence that considers the balance between clean energy consumption and power generation revenue. The cleaning robot is controlled to execute the dynamic path planning sequence, and feedback data on the pollution removal effect is obtained in real time during the operation. The parameters of the edge AI prediction model and the probability distribution of the state transition matrix are updated online based on the feedback data.

[0007] As a preferred embodiment of the photovoltaic cleaning robot operation method of the present invention, constructing the spatiotemporally correlated pollution state sequence includes the following steps: Pollution images of photovoltaic modules and corresponding environmental sensing data were collected at multiple time points. By associating the pollution status at different points in time with the spatial coordinates of the corresponding photovoltaic modules using spatial location information, a pollution status time series matrix with temporal continuity and spatial distribution is formed.

[0008] As a preferred embodiment of the photovoltaic cleaning robot operation method of the present invention, the edge AI prediction model adopts a hybrid network architecture that integrates spatiotemporal convolutional neural network and long short-term memory network. The spatiotemporal convolutional neural network is used to extract multi-scale spatial features from pollution images and environmental perception features, while the long short-term memory network is used to model the dynamic evolution of pollution features at each spatial location in the time dimension.

[0009] As a preferred embodiment of the photovoltaic cleaning robot operation method of the present invention, the pollution trend prediction value includes, but is not limited to, changes in pollution level, pollution diffusion range, and pollution probability distribution.

[0010] As a preferred embodiment of the photovoltaic cleaning robot operation method of the present invention, the construction of the dynamic pollution state transition matrix includes: The pollution trend prediction values ​​are discretized to define a pollution level state set. There are multiple pollution levels, each corresponding to a different level of pollution. In each preset prediction time period, based on the probability distribution of pollution level output by the edge AI model, multiple pollution state transition sample pairs are statistically analyzed to form a pollution state transition sample set. By combining historical pollution state evolution data with current prediction results, the transition probability values ​​between pollution states are updated using weighted average or Bayesian estimation methods. Constructing a time-varying pollution state transition matrix The element in the i-th row and j-th column of the matrix represents the state at the current time t. Transition to state The probability of; Disturbance factor calculated based on real-time environmental perception data. For the time-varying pollution state transition matrix Dynamic corrections are made to form an environment-adaptive state transition matrix.

[0011] As a preferred embodiment of the photovoltaic cleaning robot operation method of the present invention, the construction of the Markov decision process model includes the following steps: Define a state space S, which is the set of states after the pollution level is discretized. ; Define action space Action space A includes a variety of optional cleaning behaviors and scheduling strategies; Design reward function Considering the power generation gain and clean energy consumption after the robot performs action A, the specific form is as follows: ,in, Indicates the gain in power generation. Indicates clean energy consumption, and These are weighting coefficients; Using dynamic pollution state transition matrix Construct state transition function .

[0012] As a preferred embodiment of the photovoltaic cleaning robot operation method of the present invention, the method includes: solving the Markov decision process model to obtain the optimal cleaning strategy, and generating a dynamic path planning sequence that considers the balance between clean energy consumption and power generation revenue, including: Constructing the expected long-term return function based on Bellman optimality equation modeling : ; Used to measure the initial state of contamination. Execution strategy The cumulative returns that can be obtained; in: This is a discount factor, representing the importance of future returns; The reward function represents the state. Take action below Instant rewards; Let be the state transition probability, representing the probability in state . Take action below Then transferred to state The probability of; Indicates all possible actions Choose the action that yields the greatest benefit, i.e., find the optimal strategy; This represents all possible next states. Summation is performed to reflect the expected value calculation of the state transition; Indicates the next state The expected long-term return value is used for backtracking calculation of the current state value; The Bellman optimality equation is solved using either value iteration or policy iteration algorithms to obtain the optimal policy. This refers to the optimal combination of cleaning actions to be performed under various pollution conditions. The optimal strategy It is transformed into a dynamic path planning sequence with time series characteristics, which is used to guide the photovoltaic robot to perform the optimal cleaning action.

[0013] As a preferred embodiment of the photovoltaic cleaning robot operation method of the present invention, the feedback data refers to the data set of the photovoltaic cleaning robot in real time sensing and recording the pollution removal effect during the execution of the cleaning task, including but not limited to pollution image data before cleaning, pollution image data after cleaning, actual pollution level change value, operation action record data and environmental interference factor record. The feedback data is used to verify the prediction accuracy of the edge AI prediction model, to make actual corrections to the pollution state transition matrix, and to support the incremental updates of the subsequent edge AI model and the optimization and adjustment of the Markov policy model.

[0014] The present invention also provides a photovoltaic cleaning robot, applied to the above-mentioned photovoltaic cleaning robot operation method, comprising: The robot body includes a self-driven moving device for driving the robot to move on the surface of the photovoltaic module and a cleaning actuator for removing contaminants from the surface of the photovoltaic module. A control device, which is communicatively connected to the robot body, includes: The image and environmental perception acquisition module is used to acquire pollution image data and environmental perception data on the surface of photovoltaic modules in real time, and to construct a spatiotemporally correlated pollution state sequence for pollution trend prediction. The edge AI prediction module is used to predict the changes in pollution level, diffusion range, and pollution probability distribution within a preset time period based on the pollution state sequence. The path decision and control module is used to construct a Markov decision process model based on pollution trend prediction results and generate the optimal clean path; The feedback perception module is used to collect information on changes in pollution status before and after cleaning, form feedback data, and use it to update the edge AI prediction module and state transition probability online.

[0015] As a preferred embodiment of the photovoltaic cleaning robot of the present invention, the edge AI prediction module includes: A spatiotemporal convolutional unit is used to extract spatial structural features and environmental factor distributions in polluted images; Long Short-Term Memory (LSTM) network units are used to model the dynamic evolution trend of pollution over time series. An online fine-tuning submodule is used to periodically and incrementally update the parameters of the AI ​​prediction model based on feedback data.

[0016] The beneficial effects of this invention are: 1. This invention constructs a spatiotemporal pollution state sequence, which realizes a deep fusion expression of pollution state in time and space dimensions. After inputting it into the edge AI model, it can significantly improve the accuracy of pollution trend prediction and the ability to judge local area diffusion, providing accurate prediction basis for subsequent path planning.

[0017] 2. This invention constructs a dynamic pollution state transition matrix based on pollution trend prediction values, and introduces a Markov decision process model with pollution level as state, clean behavior as action, and power generation gain and energy consumption as reward functions. By solving the Bellman optimal equation, the optimal strategy is obtained, so that the optimal path planning sequence under the trade-off between clean energy consumption and power generation benefit can be output under each pollution state, thereby realizing intelligent and maximized benefit clean operation scheduling.

[0018] 3. This invention uses feedback data obtained during cleaning operations, such as before-and-after image comparisons, pollution level changes, and action and environmental records, to compare the deviation between the actual pollution state and the predicted results. This enables online incremental updates of the edge AI model parameters and probabilistic correction of the state transition matrix based on the actual evolution trajectory of the pollution state. This achieves continuous closed-loop optimization of model prediction and path strategy, improving the system's adaptability and long-term performance in dynamic environments. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the 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. Wherein: Figure 1 This is a flowchart of a photovoltaic cleaning robot operation method according to the present invention.

[0020] Figure 2 This is a partial structural schematic diagram of a photovoltaic cleaning robot according to the present invention.

[0021] Figure 3 This is a schematic diagram illustrating the working principle of a photovoltaic cleaning robot according to the present invention.

[0022] In the diagram: 100, robot body; 101, self-propelled mobile device; 102, cleaning actuator. Detailed Implementation

[0023] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0024] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0025] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0026] Secondly, the present invention is described in detail with reference to the schematic diagrams. When detailing the embodiments of the present invention, for ease of explanation, the cross-sectional views illustrating the device structure may be partially enlarged, not according to the usual scale. Furthermore, the schematic diagrams are merely examples and should not limit the scope of protection of the present invention. In addition, actual fabrication should include three-dimensional spatial dimensions of length, width, and depth.

[0027] Example 1 Reference Figure 1 The first embodiment of the present invention provides a method for operating a photovoltaic cleaning robot, comprising the following steps: S1: Real-time acquisition of surface contamination image data and environmental perception data of photovoltaic modules to construct a spatiotemporally correlated contamination state sequence.

[0028] It should be noted that surface contamination image data includes visible light image data, infrared thermal imaging data, or multispectral image data, reflecting the color, area, thermal distribution, and adhesion area of ​​pollutants, and is collected through visible light cameras. Environmental sensing data includes, but is not limited to, light intensity, wind speed, temperature, humidity, PM concentration, GPS location information, and collection timestamps, used to construct pollution influencing factors, and is collected through edge sensing nodes deployed at photovoltaic sites. The pollution state sequence refers to the synchronous association of pollution images and environmental sensing data according to time and space dimensions at a specific time and spatial location, forming a multidimensional sequence of data that can be input into edge AI models, serving as the basis for pollution trend prediction.

[0029] Specifically, constructing the spatiotemporally correlated pollution state sequence includes the following steps: Pollution images of photovoltaic modules and corresponding environmental sensing data were collected at multiple time points. By associating the pollution status at different points in time with the spatial coordinates of the corresponding photovoltaic modules using spatial location information, a pollution status time series matrix with temporal continuity and spatial distribution is formed.

[0030] It should be noted that during the specific construction process, the photovoltaic site was divided into multiple spatial regions according to a pre-set grid. The system collects the aforementioned pollution images and environmental sensing data at designated sampling time points (e.g., every 15 minutes) within each grid region. For each set of collected data, the pollution images and environmental sensing data from the same time point and region are paired and labeled to form a pollution status observation pair. ,in, Indicates time Spatial location Pollution images obtained from the site; This represents the set of environmental sensing parameters at the same spatiotemporal location. Then, the series of pollution state observation pairs are arranged according to the time dimension and grouped according to the spatial grid to construct a pollution state time series matrix with a three-dimensional structure.

[0031] This step integrates collected pollution image data with environmental perception data to construct a pollution state time series matrix that has temporal continuity, spatial distribution, and diverse pollution characteristics. This matrix serves as the input data source for the edge AI model, thereby improving the spatial accuracy and temporal foresight of pollution trend prediction and providing highly reliable predictive support for subsequent intelligent clean path planning.

[0032] S2: Input the pollution state sequence into the edge AI prediction model deployed locally on the robot, and output the pollution trend prediction value for a future preset period.

[0033] It should be noted that the edge AI prediction model adopts a hybrid network architecture that integrates a spatiotemporal convolutional neural network and a long short-term memory network. The spatiotemporal convolutional neural network is used to extract multi-scale spatial features from polluted images and environmental perception features, while the long short-term memory network is used to model the dynamic evolution of pollution features at various spatial locations over time. Specifically, the pollution state time series matrix is ​​first input into the spatiotemporal convolutional neural network for spatial feature extraction. The extracted spatial feature tensor is then flattened and used as a time series input into the long short-term memory network for pollution trend prediction. Finally, a Softmax or Sigmoid prediction output layer is connected to the output of the long short-term memory network to generate the pollution trend prediction value.

[0034] It should be noted that the pollution trend prediction values ​​include changes in pollution levels, the extent of pollution diffusion, and the probability distribution of pollution. Specifically, changes in pollution levels refer to the dynamic trend of the pollution level on the surface of photovoltaic modules predicted by the AI ​​prediction model over one or more preset time periods in the future. This trend can be quantified through the pollution level change value. To express, for example This reflects the degree and direction of pollution intensification or mitigation. Pollution diffusion range refers to the model's prediction of the extent of pollution spread within the photovoltaic array in a spatial dimension. It is mainly used to determine whether pollution has a spreading trend and to assist in multi-regional operation path scheduling. Pollution probability distribution refers to the predicted probability of a specific area or component's pollution level falling into various possible values ​​at a future time, forming a probability distribution vector of pollution states.

[0035] In some specific implementations, the steps for obtaining pollution level changes include: First, defining a pollution level scoring system, that is, dividing the pollution state into N levels (such as L1-L5) based on features such as image grayscale, color difference, and thermal distribution, with each level corresponding to a specific pollution degree range. For example: S1: Clean, pollution value 0–0.2; S2: Lightly polluted, pollution value 0.2–0.4; S3: Moderately polluted, pollution value 0.4–0.6; S4: Heavily polluted, pollution value 0.6–0.8; S5: Severely polluted, pollution value 0.8–1.0; The pollution value can be obtained by training an AI model to perform multi-feature fusion regression analysis on polluted images and perceived features; Second, the edge AI model takes the pollution state sequence of each grid cell as input and outputs the pollution level prediction value at the corresponding time step, such as pollution score 0.73 and predicted level L3; Third, calculate the pollution level change value. For example, if the current pollution level is St and the predicted value is St+△, then the pollution level change value = St+△-St. If the value is positive, the pollution has worsened; if the value is negative, the pollution has improved.

[0036] In some specific implementations, the steps for obtaining the pollution spread range include: calculating the maximum difference value between the pollution level of each spatial grid and its neighboring grids based on the predicted pollution level values ​​of each spatial grid at the predicted time; when the difference value is greater than a preset threshold, the grid is determined to be a pollution spread point; all pollution spread point grids constitute the pollution spread range, which helps to determine the spatial spread trend of pollution and guides the intelligent scheduling of cleaning paths.

[0037] In some specific implementations, the steps for obtaining the pollution probability distribution include: First: The final output layer of the edge AI model is a Softmax structure, corresponding to the set of pollution level states. Second: Using historical pollution state sequences as training data, each sample corresponds to a pollution level label. The model is trained using the cross-entropy loss function, enabling the model to learn to output the probability of occurrence for each level. Third: For each spatial location at each prediction time, a set of pollution state probability distribution vectors is output, such as: (corresponding to S1-S5), representing the probability that the location will belong to each pollution level at a future time.

[0038] In summary, step S2 inputs the spatiotemporally correlated pollution state sequence into an edge AI prediction model that integrates a spatiotemporal convolutional neural network and a long short-term memory network. This enables the efficient extraction of the spatial distribution and temporal evolution of pollution features at the edge, thereby accurately predicting the changes in pollution levels, pollution diffusion trends, and their probability distributions in various regions within a preset time period. This not only significantly improves the spatiotemporal resolution and reliability of pollution trend prediction but also provides timely and accurate data support for subsequent clean path optimization and scheduling.

[0039] S3: Construct a dynamic pollution state transition matrix based on the pollution trend prediction values, and construct a Markov decision process model based on the dynamic pollution state transition matrix.

[0040] Specifically, the construction of the dynamic pollution state transition matrix includes: S3-11: Discretize the pollution trend prediction values ​​to define a pollution level state set. There are multiple pollution levels, each corresponding to a different level of pollution. S3-12: In each preset prediction time period, based on the pollution level probability distribution output by the edge AI model, multiple pollution state transition sample pairs are statistically analyzed to form a pollution state transition sample set; where the pollution level probability distribution refers to the set of occurrence probabilities of each pollution level output by the edge AI model within the prediction time period. S3-13: Combining historical pollution state evolution data with current prediction results, update the transition probability values ​​between pollution states using weighted average or Bayesian estimation methods; S3-14: Constructing the time-varying pollution state transition matrix The element in the i-th row and j-th column of the matrix represents the state at the current time t. Transition to state The probability of; S3-15: Calculate the disturbance factor based on real-time environmental perception data. For the time-varying pollution state transition matrix Dynamic corrections are made to form an environment-adaptive state transition matrix.

[0041] It should be noted that the historical pollution state evolution data comes from pollution image data and corresponding environmental sensing data recorded during the long-term operation of photovoltaic power plants, as well as the historical pollution state sequence constructed by the system. This sequence, after manual or model annotation, records the actual trajectory of pollution level evolution over time. The current prediction result comes from the pollution level probability distribution sequence output by the edge AI prediction model predicting the current pollution state sequence. By fusing the two, a weighted strategy is used to enhance the stability and real-time performance of the pollution state transition probability estimation, thereby constructing a pollution state transition matrix with time-sensitive sensing capabilities. .

[0042] In some specific implementations, the step of statistically analyzing multiple pollution state transition sample pairs to form a pollution state transition sample set is as follows: First, for each grid cell, the pollution states within consecutive time periods are paired up. For example, the predicted state of a certain grid cell at time t is... The predicted state at time t+1 is Then construct a set of pollution state transition sample pairs Repeat the above operation for all spatial locations to generate multiple pollution state transition sample pairs; then, summarize all pollution state transition sample pairs generated in the above steps to form a pollution state transition sample set for the current prediction time period. This set serves as the basic data input for subsequent calculation of state transition probabilities and is used to construct the pollution state transition matrix.

[0043] In some specific implementations, the step of updating the transition probability values ​​between pollution states is as follows: Statistically count the frequency of state pairs, that is, based on the aforementioned pollution state transition sample set, statistically count the frequency of each state transition pair. The frequency of occurrence of each state pair is used to obtain the historical transition frequency. Then, using this historical transition frequency as a long-term statistical basis, the probability of transition for each state is weighted and fused by combining it with the pollution level probability distribution sequence output by the edge AI prediction model; where the pollution level probability distribution sequence refers to the model's predicted probability output for the state set at each spatial location and prediction time; for any state pair... The updated state transition probability value is calculated using the following formula: ; in: This indicates that in the sample set of pollution state transitions, state pairs The frequency of occurrence, and the historical transition probability obtained after normalization. Specifically, the statistical state. Transfer to The number of sample pairs, and The ratio of all transfers to the total is... ; This indicates that based on the probability distribution sequence of pollution level output by the edge AI prediction model at time t, it is inferred that the state at time t is... The position transitions to state at time t+1. The predicted probability.

[0044] In some specific implementations, a time-varying pollution state transition matrix is ​​constructed. The steps are as follows: Set the pollution level status As row and column indices for constructing the state transition matrix; for each state in the set This summarizes the states that the data may transition to in the current forecast period. The probability values ​​constitute the state transition probability vector. Arrange the transition probability vectors of all states row-wise to generate the pollution state transition matrix at the current time. It is an n×n two-dimensional probability matrix that reflects the evolution trend of time-varying pollution states.

[0045] In some specific implementations, the disturbance factor The calculation method is as follows: based on the degree of deviation between the current environmental perception data and the historical statistical average, the disturbance factor is dynamically calculated. The calculation formula is: ; in: This indicates the number of times the data is acquired at the current time t. Environmental parameter values, such as wind speed, temperature, PM concentration, etc.; This represents the mean of the parameter over a historical statistical window period; The disturbance sensitivity coefficient of the corresponding environmental parameter reflects the weight of the parameter's influence on the transition of pollution state. The number of environmental parameters involved in the calculation; This represents the perturbation factor, and its numerical range is generally limited to [0,1] or processed by normalization; if a parameter is invalid or missing, its corresponding perturbation term can be set to 0.

[0046] In some specific implementations, the time-varying pollution state transition matrix The steps for dynamic correction are as follows: using the pollution state transition matrix at the current moment... Based on, combined with disturbance factors The probability values ​​of the affected state pairs in the matrix are weighted and adjusted, for example: when a certain state pair... When affected by drastic changes in wind speed, its transfer probability can be adjusted to To correct the direction and magnitude, each corrected row of specific probability vectors is then normalized to ensure the sum of state transition probabilities is 1, conforming to the definition of a Markov process. Finally, all corrected transition probability vectors are combined to form the final dynamic pollution state transition matrix. This is used to subsequently construct Markov decision process models.

[0047] In the above, by fusing the pollution level probability distribution output by the edge AI prediction model with historical pollution state evolution data, and combining the perception and adjustment of environmental changes by perturbation factors, dynamic modeling of the transition law of pollution level state at each time point is achieved, resulting in a state transition matrix. It possesses both timeliness and environmental adaptability, thereby significantly improving the prediction accuracy and robustness of subsequent Markov decision process models in path planning and strategy optimization.

[0048] Specifically, the construction of a Markov decision process model includes the following steps: S3-21: Define the state space S, which is the set of states after the pollution level is discretized. ; S3-22: Define the action space Action space A includes a variety of optional cleaning behaviors and scheduling strategies, such as zone cleaning, path switching, skipping the current zone, or waiting for the next cycle. S3-23: Designing the Reward Function Consider the power generation gain and clean energy consumption after the robot performs action A; S3-24: Utilizing the dynamic pollution state transition matrix Construct state transition function ; At this point, the contaminated state space S, action space A, and reward function are defined. and state transition function All conditions have been fully defined and meet the requirements for constructing the Markov decision process model, thus completing the construction of the Markov decision process model.

[0049] In some specific implementations, the reward function The expression for is: ,in, This represents the increase in power generation revenue obtained by performing action A in state S; This represents the energy consumption corresponding to action A. and As a weighting factor, This is a weighting factor for power generation gain, reflecting the importance of power generation benefits in total revenue; This is a clean energy consumption weighting factor, reflecting the suppressive effect of energy costs on total revenue.

[0050] In the above, the increase in power generation revenue The method for obtaining the data is as follows: Based on real-time power generation data of photovoltaic modules before and after cleaning, a comparative analysis is performed on the changes in power generation efficiency per unit time under the same illumination and temperature conditions, and the power generation difference ΔP before and after performing cleaning action A is calculated; ΔP is then integrated or estimated over a unit time scale to obtain the cumulative power generation gain brought about by action A from pollution level state S, i.e. .

[0051] In the above, the energy consumption corresponding to action A The method for obtaining the energy consumption is as follows: Based on the robot's operations involved in action A (such as travel distance, cleaning time, cleaning intensity level, etc.), and combined with the robot's power consumption model, calculate the energy consumption per unit time; if action A includes compound operations (such as path switching + area cleaning), calculate and sum the energy consumption of each sub-action to obtain the overall action energy consumption. .

[0052] It should be noted that in the default operating mode, the system can use preset fixed weight values, such as... =1.0、 =0.5, at which point the default objective is to maximize power generation revenue while also considering the impact of clean energy costs. Additionally, the system can incorporate adaptive adjustment strategies based on fuzzy rules, interval weight mapping tables, or machine learning to achieve [further adjustments]. , Automatic optimization and adjustment. Furthermore, to ensure computational stability and optimization convergence, and The range of values ​​for can be constrained to: , Alternatively, more detailed sub-intervals can be set based on simulation and on-site debugging experience.

[0053] In some specific implementations, a dynamic pollution state transition matrix is ​​used. Construct state transition function The process includes the following steps: Based on the changing trend of the pollution state under action A, the dynamic pollution state transition matrix is ​​used. Extracting state pairs Determine the transition probabilities and establish the state transition function. Its expression is: ; This formula indicates that the system is currently in a certain state of contamination. In this case, after performing an operation A (such as cleaning, waiting, etc.), the system transitions to another state. What is the probability of it?

[0054] In the above formula, This indicates the probability that the pollution state will change on its own without any action being taken. For example, if today is "moderate pollution", what is the probability that it will automatically become "light pollution" tomorrow? This value comes from the dynamic pollution state transition matrix constructed earlier or is calculated by combining the historical pollution state change data recorded by the system and the trend prediction results of the edge AI model. This represents the additional impact on the probability of state transition after considering an operation A, such as cleaning or skipping it. For example, the state might have remained the same, but cleaning increases the probability of the contamination becoming lighter. This value can be set based on past experience or rules, such as cleaning usually reducing contamination, or by statistically calculating the probability of change after cleaning based on historical performance data of cleaning equipment. It is a weighting coefficient used to adjust the proportion of "natural changes" and "action effects" in the final result. It can generally be set based on experience. If a certain action has a great impact (such as heavy cleaning). A smaller value indicates greater trust in the changes brought about by the action; if the action has almost no impact, then... Taking a larger value indicates a greater reliance on natural variations.

[0055] Example: Pollution level state set S ; in: This indicates a pollution level of "clean," meaning the impact on power generation is minimal. This indicates "mild" pollution, which may lead to a decrease in some power generation efficiency. This indicates "moderate" pollution, which significantly reduces power generation efficiency. This indicates "severe" pollution, requiring priority cleaning and treatment.

[0056] The action set is defined as: , in: This indicates that cleaning operations should be performed immediately; This indicates that cleaning will be temporarily suspended, pending the next round of scheduling. This indicates that the current area will be skipped, and other areas will be processed first.

[0057] Then, for each state-action pair, calculate its corresponding reward function: ; Suppose: the current pollution state of the photovoltaic module is... In a (severely polluted) situation, if the action is performed... (Immediate cleaning), the edge AI model predicts recoverable power generation efficiency, and the system estimates energy consumption and gain as follows: Power generation gain Assuming that after cleaning, the current photovoltaic modules can restore approximately 5 kWh of power generation capacity per day; Clean energy consumption This includes the energy consumption for moving to the target area and the operational energy consumption after the cleaning system is turned on, totaling approximately 1.5 kWh; Weighting coefficient , (This reflects that the system prioritizes power generation revenue over energy consumption costs).

[0058] Substitute into the reward function:

[0059] =2.4 indicates that a cleaning action has been performed. In the current state The expected net benefit is 2.4 kWh, which means that after taking into account the returns on power generation and energy consumption costs, the overall clean benefit of the system is positive.

[0060] Additionally, a certain component is currently classified as "heavily contaminated". and perform actions (Immediate cleaning) serves as an example to illustrate how to obtain the numerical value of the state transition function: Based on historical evolution patterns and edge AI model predictions, the probability of natural transfer is: =0.25; =0.75; After considering the impact of cleaning operations, the adjustment probability is set as follows: =0.5 (Cleaning may directly improve the condition to light contamination). =0.4; =0.1; Assuming fusion coefficient =0.4 indicates greater confidence in the effectiveness of "action intervention", then: For example, calculating the state Transfer to The probability is: ; The probabilities of other state transitions can be calculated similarly to obtain the execution action. The subsequent overall transfer distribution: =0.30; =0.34; =0.36; In this way, the transition probability of each state-action pair can be calculated, which can be used to support the solution of Markov decision process models and policy generation.

[0061] In summary, in step S3 above, by integrating historical pollution evolution data with the real-time pollution level probability distribution output by the edge AI model, and dynamically correcting the transition probability using environmental perception disturbance factors, the constructed pollution state transition matrix can reflect the current pollution state evolution trend and environmental impact in real time. This provides a more accurate, timely, and environmentally adaptive basis for the subsequent Markov decision model, enhancing the scientific and intelligent level of the entire photovoltaic power station's pollution control and scheduling strategy. By clearly defining the pollution level, optional cleaning actions, and their corresponding benefits and costs, and dynamically generating a state transition function based on the pollution trend predicted by edge AI, the optimal cleaning strategy can be intelligently selected for different pollution states. This not only improves the power generation efficiency of photovoltaic modules but also significantly reduces clean energy consumption and scheduling costs, enhancing the intelligence, refinement, and economy of overall operation and management.

[0062] S4: Solve the Markov decision process model to obtain the optimal clean strategy and generate a dynamic path planning sequence that considers the balance between clean energy consumption and power generation revenue. Specifically, the optimal clean energy strategy is obtained by solving the Markov decision process model, and a dynamic path planning sequence considering the balance between clean energy consumption and power generation revenue is generated, including: Constructing the expected long-term return function Used to measure the initial state of contamination. Execution strategy The cumulative returns that can be obtained. Specifically: Modeling the expected long-term return function based on the Bellman optimality equation : ; in, This is the discount factor, and its value range is... This indicates the importance of future returns; the closer the value is to 1, the more the model values ​​long-term returns; the closer it is to 0, the more it favors short-term gains. It is typically set as... It is 0.9 or 0.95; The reward function represents the state. Take action below Instant rewards; Let be the state transition probability, representing the probability in state . Take action below Then transferred to state The probability of; Indicates all possible actions Choose the action that yields the greatest benefit, i.e., find the optimal strategy; This represents all possible next states. Summation is performed to reflect the expected value calculation of the state transition; Indicates the next state The expected long-term return value is used for backtracking calculations of the current state value.

[0063] The Bellman optimality equation is solved using either value iteration or policy iteration algorithms to obtain the optimal policy. This refers to the optimal combination of cleaning actions to be performed under various pollution conditions. The optimal strategy It is transformed into a dynamic path planning sequence with time series characteristics, which is used to guide the photovoltaic robot to perform the optimal cleaning action.

[0064] For example, a grid area of ​​a photovoltaic field In the process, the system collected the pollution status at time point t as follows: The corresponding pollution level is "severe pollution". The robot's current set of selectable actions is: Perform cleaning immediately; Cleaning is temporarily suspended, pending the next round of scheduling; Skip the current area.

[0065] Then, calculate the reward function, if selected (Cleaning) This component is expected to restore increased power generation. At the same time, it requires the consumption of clean energy. kWh; Set the power generation revenue weighting factor Energy consumption weighting coefficient The immediate benefit is: . Then, obtain the state transition probability: Based on the dynamic pollution state transition matrix corrected by pollution trend prediction and cleaning effect feedback In execution Subsequently, the probability of pollution state transition is (obtained through statistical empirical data or predictions from AI models): =0.6, downgraded to light pollution; =0.3, downgraded to moderate pollution; =0.1, still indicating severe pollution.

[0066] Set the current discount factor Then the Bellman expected long-term return function is:

[0067] With action For example, if the system currently estimates: ; ; Substituting this into the equation, we get:

[0068] Assuming other actions , The profit calculation results are as follows: ; Then we have: ; Final action selection The reason for immediately performing the cleaning operation is that, in the current state, the overall reward for performing this action is the highest (6.65 > 5.2 > 4.1). This not only considers immediate rewards but also the expected benefits of future state transitions, thereby ensuring that the system achieves optimal returns in the long run.

[0069] This decision is written into the route scheduling table to generate a route planning sequence with time-series characteristics. ,like:

[0070] Ultimately, the robot executes the cleaning tasks of each grid area in sequence according to the strategy, achieving a dynamic balance between maximizing energy benefits and controlling energy consumption costs.

[0071] In summary, in step S4, by constructing the expected long-term benefit function based on the Bellman optimal equation and solving for the optimal strategy using value iteration or strategy iteration methods, it is possible to flexibly select the most comprehensive combination of cleaning actions under different pollution conditions. This allows the system to dynamically generate a time-series optimal path planning sequence while fully balancing power generation gain and energy consumption cost. This precisely guides the photovoltaic robot to carry out cleaning operations in stages and in a strategic manner, thereby improving overall operating efficiency and intelligence.

[0072] S5: Control the cleaning robot to execute the dynamic path planning sequence and obtain real-time feedback data on the pollution removal effect during the operation; It should be noted that feedback data refers to the data set collected in real time by the photovoltaic cleaning robot during the cleaning process, which includes, but is not limited to: Pre-cleaning contamination image data, used to record the original contamination state; Contaminated image data after cleaning is used to determine the cleaning effect; The actual pollution level change value is calculated based on the difference in pollution scores between images before and after cleaning. Record data on job actions, including but not limited to the current action (such as path, cleaning mode), cleaning time, cleaning energy consumption, and power consumption; Environmental disturbance factors are recorded, and environmental data such as wind speed, light intensity, and PM concentration are collected during the cleaning process to identify abnormal impacts.

[0073] The feedback data is primarily used to: verify the prediction accuracy of the edge AI prediction model; make practical corrections to the pollution state transition matrix; and support incremental updates of the subsequent edge AI model and optimization adjustments to the Markov policy model.

[0074] S6: Update the parameters of the edge AI prediction model and the probability distribution of the state transition matrix online based on the feedback data.

[0075] In some specific implementations, the parameters of the edge AI prediction model are updated, and the specific process is as follows: After the cleaning robot completes its cleaning operation, the system will re-collect the pollution status information of the same photovoltaic module or area after cleaning. Combining the pollution status recorded before the operation with the actual actions performed, a set of "pre-state-action-post-state" triplet samples (S→A→S′) will be formed to characterize the transition and change of the pollution level under specific actions. The system will store several sets of continuously collected state transition samples in the update cache and trigger the update operation according to the set time period or sample number threshold. During the update process, the system first counts the frequency of occurrence of the successor state S′ corresponding to the same state-action pair (S, A), and constructs an empirical transition frequency matrix. Then, this matrix is ​​compared with the currently used state transition matrix. To achieve fusion, weighted average, exponential moving average, or Bayesian update methods are used to dynamically adjust the transition probabilities, specifically as follows: ; in, To update the weighting coefficients, The parameters can be set by the user or automatically and dynamically adjusted by the system based on factors such as sample size and prediction bias, ensuring that new data has a reasonable impact on model adjustments while avoiding drastic fluctuations. The parameters are given by the system's original model or obtained through offline training; The transition frequency of all (S,A,S′) triplet samples within a certain time window is calculated. The above-mentioned updated formula continuously optimizes the accuracy and timeliness of the state transition model by integrating current knowledge and the latest collected pollution transfer sample data, enabling the Markov model to dynamically adapt to the changing characteristics of the actual pollution environment, thereby improving the effectiveness of path planning and strategy selection.

[0076] Furthermore, during the update process, the system pays special attention to the actual impact of cleaning actions on pollution improvement: if the system observes that the pollution level does not decrease significantly after multiple "cleaning" actions (e.g., it remains at the original state or only slightly improves), it will automatically reduce the corresponding level of that action. Medium to low pollution status (e.g.) or The probability of state transition is adjusted to reduce the expected cleaning effect of the action, thereby suppressing inefficient cleaning strategies. In this way, the system continuously optimizes the convergence and adaptability of the state transition model, making subsequent strategy decisions more consistent with the pollution evolution trend and cleaning execution feedback in the actual scenario, effectively improving the system's intelligent planning and precise control capabilities.

[0077] Step S6 above updates the probability distribution of the parameters and state transition matrix of the edge AI prediction model online based on feedback data. This enables the system to continuously correct the prediction bias of pollution trends and the modeling error of pollution evolution laws. This not only improves the accuracy and generalization ability of pollution state prediction, but also enhances the response capability and long-term decision-making effect of the Markov decision model on which the cleaning strategy is based to complex environmental changes, thereby achieving more intelligent, efficient and dynamically adaptive cleaning path optimization control.

[0078] In summary, this invention constructs a spatiotemporal pollution state sequence that integrates pollution image data, environmental perception data, GPS location information, and timestamps. This achieves a deep fusion representation of pollution states in both time and space dimensions. Inputting this into an edge AI model significantly improves the accuracy of pollution trend prediction and the ability to judge local area spread, providing precise predictive basis for subsequent path planning. Based on pollution trend prediction values, this invention constructs a dynamic pollution state transition matrix and introduces a Markov decision process model with pollution level as the state, clean-up behavior as the action, and power generation gain and energy consumption as reward functions. By solving the Bellman optimal equation, the optimal strategy is obtained, enabling the output of the optimal path planning sequence that balances clean energy consumption and power generation benefits under each pollution state, achieving intelligent, profit-maximizing clean operation scheduling. This invention utilizes feedback data acquired during cleaning operations, such as before-and-after image comparisons, pollution level changes, and action and environmental records, to compare the deviation between the actual pollution state and the predicted results. This enables online incremental updates of the edge AI model parameters and probabilistic correction of the state transition matrix based on the actual evolution trajectory of the pollution state. Consequently, it achieves continuous closed-loop optimization of model prediction and path strategy, improving the system's adaptability and long-term performance in dynamic environments.

[0079] Example 2 Reference Figure 2-3This is a second embodiment of the present invention, providing a photovoltaic cleaning robot applied to the photovoltaic cleaning robot operation method in Embodiment 1, comprising: The robot body 100 includes a self-driven moving device 101 for driving the robot to move on the surface of the photovoltaic module and a cleaning actuator 102 for removing contaminants from the surface of the photovoltaic module. A control device, which is communicatively connected to the robot body 100, includes: The image and environmental perception acquisition module is used to acquire pollution image data and environmental perception data on the surface of photovoltaic modules in real time, and to construct a spatiotemporally correlated pollution state sequence for pollution trend prediction. The edge AI prediction module is used to predict the changes in pollution level, diffusion range, and pollution probability distribution within a preset time period based on the pollution state sequence. The path decision and control module is used to construct a Markov decision process model based on pollution trend prediction results and generate the optimal clean path; The feedback perception module is used to collect information on changes in pollution status before and after cleaning, form feedback data, and use it to update the edge AI prediction module and state transition probability online.

[0080] The edge AI prediction module includes: A spatiotemporal convolutional unit is used to extract spatial structural features and environmental factor distributions in polluted images; Long Short-Term Memory (LSTM) network units are used to model the dynamic evolution trend of pollution over time series. An online fine-tuning submodule is used to periodically and incrementally update the parameters of the AI ​​prediction model based on feedback data.

[0081] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for operating a photovoltaic cleaning robot, characterized in that, Includes the following steps: Real-time acquisition of surface contamination image data and environmental sensing data of photovoltaic modules to construct a spatiotemporally correlated contamination state sequence; The pollution state sequence is input into an edge AI prediction model deployed locally on the robot, and the predicted pollution trend value for a future preset time period is output. A dynamic pollution state transition matrix is ​​constructed based on the pollution trend prediction values, and a Markov decision process model is constructed based on the dynamic pollution state transition matrix. Solving the Markov decision process model yields the optimal clean energy strategy and generates a dynamic path planning sequence that considers the balance between clean energy consumption and power generation revenue. The cleaning robot is controlled to execute the dynamic path planning sequence, and feedback data on the pollution removal effect is obtained in real time during the operation. The parameters of the edge AI prediction model and the probability distribution of the state transition matrix are updated online based on the feedback data.

2. The photovoltaic cleaning robot operation method as described in claim 1, characterized in that: Constructing the spatiotemporally correlated pollution state sequence includes the following steps: Pollution images of photovoltaic modules and corresponding environmental sensing data were collected at multiple time points. By associating the pollution status at different points in time with the spatial coordinates of the corresponding photovoltaic modules using spatial location information, a pollution status time series matrix with temporal continuity and spatial distribution is formed.

3. The photovoltaic cleaning robot operation method as described in claim 1, characterized in that: The edge AI prediction model adopts a hybrid network architecture that integrates spatiotemporal convolutional neural networks and long short-term memory networks. The spatiotemporal convolutional neural network is used to extract multi-scale spatial features from polluted images and environmental perception features, while the long short-term memory network is used to model the dynamic evolution of pollution features at various spatial locations in the time dimension.

4. The photovoltaic cleaning robot operation method as described in claim 1, characterized in that: The pollution trend prediction values ​​include, but are not limited to, changes in pollution level, pollution spread range, and pollution probability distribution.

5. The photovoltaic cleaning robot operation method as described in claim 1, characterized in that: The construction of the dynamic pollution state transition matrix includes: The pollution trend prediction values ​​are discretized to define a pollution level state set. There are multiple pollution levels, each corresponding to a different level of pollution. In each preset prediction time period, based on the probability distribution of pollution level output by the edge AI model, multiple pollution state transition sample pairs are statistically analyzed to form a pollution state transition sample set. By combining historical pollution state evolution data with current prediction results, the transition probability values ​​between pollution states are updated using weighted average or Bayesian estimation methods. Constructing a time-varying pollution state transition matrix The element in the i-th row and j-th column of the matrix represents the state at the current time t. Transition to state The probability of; Disturbance factor calculated based on real-time environmental perception data. For the time-varying pollution state transition matrix Dynamic corrections are made to form an environment-adaptive state transition matrix.

6. The photovoltaic cleaning robot operation method as described in claim 1, characterized in that: The construction of the Markov decision process model includes the following steps: Define a state space S, which is the set of states after the pollution level is discretized. ; Define action space Action space A includes a variety of optional cleaning behaviors and scheduling strategies; Design reward function Consider the power generation gain and clean energy consumption after the robot performs action A; Using dynamic pollution state transition matrix Construct state transition function .

7. The photovoltaic cleaning robot operation method as described in claim 6, characterized in that: Solving the Markov decision process model yields the optimal clean energy strategy, generating a dynamic path planning sequence that considers the balance between clean energy consumption and power generation revenue, including: Constructing the expected long-term return function based on Bellman optimality equation modeling : ； Used to measure the initial state of contamination. Execution strategy The cumulative returns that can be obtained; in: This is a discount factor, indicating the importance of future returns; The reward function represents the state. Take action below Instant rewards; Let be the state transition probability, representing the probability in state . Take action below Then transition to state The probability of; Indicates all possible actions Choose the action that yields the greatest benefit, i.e., find the optimal strategy; This represents all possible next states. Summation is performed to reflect the expected value calculation of the state transition; Indicates the next state The expected long-term return value is used for backtracking calculation of the current state value; The Bellman optimality equation is solved using either value iteration or policy iteration algorithms to obtain the optimal policy. This refers to the optimal combination of cleaning actions to be performed under various pollution conditions. The optimal strategy It is transformed into a dynamic path planning sequence with time series characteristics, which is used to guide the photovoltaic robot to perform the optimal cleaning action.

8. The photovoltaic cleaning robot operation method as described in claim 1, characterized in that: The feedback data refers to the data set of the photovoltaic cleaning robot that is used to perceive and record the pollution removal effect in real time during the cleaning process, including but not limited to pollution image data before cleaning, pollution image data after cleaning, actual pollution level change value, operation action record data, and environmental interference factor record. The feedback data is used to verify the prediction accuracy of the edge AI prediction model, to make actual corrections to the pollution state transition matrix, and to support the incremental updates of the subsequent edge AI model and the optimization and adjustment of the Markov policy model.

9. A photovoltaic cleaning robot, applied to the photovoltaic cleaning robot operation method as described in any one of claims 1-8, characterized in that, include: The robot body (100) includes a self-driving mobile device (101) for driving the robot to move on the surface of the photovoltaic module and a cleaning actuator (102) for removing contaminants from the surface of the photovoltaic module. A control device, which is communicatively connected to the robot body (100), includes: The image and environmental perception acquisition module is used to acquire pollution image data and environmental perception data on the surface of photovoltaic modules in real time, and to construct a spatiotemporally correlated pollution state sequence for pollution trend prediction. The edge AI prediction module is used to predict the changes in pollution level, diffusion range, and pollution probability distribution within a preset time period based on the pollution state sequence. The path decision and control module is used to construct a Markov decision process model based on pollution trend prediction results and generate the optimal clean path; The feedback perception module is used to collect information on changes in pollution status before and after cleaning, form feedback data, and use it to update the edge AI prediction module and state transition probability online.

10. The photovoltaic cleaning robot as described in claim 9, characterized in that: The edge AI prediction module includes: A spatiotemporal convolutional unit is used to extract spatial structural features and environmental factor distributions in polluted images; Long Short-Term Memory (LSTM) network units are used to model the dynamic evolution trend of pollution over time series. An online fine-tuning submodule is used to periodically and incrementally update the parameters of the AI ​​prediction model based on feedback data.

Images