Photovoltaic module intelligent cleaning timing judgment and path planning method

Abstract

This invention discloses a method for intelligent cleaning timing and path planning of photovoltaic modules, including: Step 1, multi-dimensional data acquisition and preprocessing; Step 2, dynamic dust accumulation loss assessment; Step 3, cleaning trigger condition determination; Step 4, multi-subarray cleaning priority ranking; Step 5, cleaning area topology map construction; Step 6, global traversal path planning; Step 7, local dynamic obstacle avoidance path correction; Step 8, cleaning closed-loop feedback and parameter optimization. This invention dynamically quantifies the power generation loss rate caused by dust accumulation through a photovoltaic physical model and a grey prediction algorithm, and combines future weather forecasts (rainfall probability, wind speed) and real-time grid-connected electricity prices to construct an adaptive dynamic threshold for economic loss, achieving "on-demand cleaning" rather than "timed cleaning."

Description

Technical Field

[0001] This invention relates to the field of photovoltaic power plant operation and maintenance technology, specifically to a method for determining the timing and path planning of intelligent cleaning of photovoltaic modules. Background Technology

[0002] Photovoltaic modules are exposed to the outdoor environment for extended periods, and their surfaces easily accumulate pollutants such as dust, bird droppings, and fallen leaves, leading to decreased light transmittance and a significant reduction in power generation efficiency. Traditional manual cleaning methods suffer from problems such as poor timeliness, high labor costs, and water waste, and it is often difficult to strike a balance between insufficient and excessive cleaning.

[0003] CN121761923A discloses a path optimization method for photovoltaic cleaning robots based on a hybrid algorithm. This method generates the optimal solution by constructing multiple objective optimization functions and performing non-dominated sorting and congestion calculation, offering advantages in high efficiency and energy saving. However, this method primarily focuses on path energy optimization for a single cleaning task, failing to integrate cleaning timing judgment and path planning into a unified closed-loop feedback framework. The cleaning decision does not adaptively adjust to factors such as fluctuations in the power plant's grid-connected electricity price, component degradation, and weather changes, resulting in limited matching between the cleaning strategy and the actual economic efficiency of the power plant's operation. Summary of the Invention

[0004] To address this, the present invention provides a method for intelligent cleaning timing and path planning of photovoltaic modules, which solves the problem that most existing cleaning path planning is based on preset fixed trajectories or simple reciprocating paths, lacking real-time adaptability to changes in photovoltaic array layout, dynamic obstacles (such as temporary construction equipment, bird gatherings) and the energy consumption of the cleaning equipment itself, resulting in low cleaning coverage, repeated cleaning or missed areas.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] The method for determining the timing and path planning for intelligent cleaning of photovoltaic modules includes the following steps:

[0007] Step 1: Multidimensional data acquisition and preprocessing: Real-time acquisition of DC output power, ambient light intensity, module backsheet temperature, historical dust accumulation attenuation rate, and weather forecast data for the first preset time in the photovoltaic power station for each string, and outlier removal and normalization of the acquired data.

[0008] Step 2, Dynamic Dust Accumulation Loss Assessment: Based on the data processed in Step 1, the instantaneous power loss rate under the current dust accumulation state is calculated using a photovoltaic physical model, and a gray prediction algorithm is used to construct a prediction function for the dust accumulation growth rate within the next preset time period.

[0009] Step 3: Determine the sweeping trigger condition: Compare the current power loss rate obtained in Step 2 with the preset dynamic threshold for economic loss, and at the same time verify the probability of future rainfall and the predicted wind speed. When the power loss rate continuously exceeds the threshold and there is no effective rainfall or strong wind warning in the next three preset time period, a sweeping trigger command is generated.

[0010] Step 4: Prioritizing cleaning of multiple subarrays: In response to the cleaning trigger command in Step 3, obtain the dust loss rate, installed capacity, last cleaning time and dust growth rate of all photovoltaic subarrays to be cleaned in the power station, calculate the comprehensive cleaning urgency index of each subarray, and sort them from high to low index.

[0011] Step 5: Construction of topology map of the cleaning area: For the target subarray with the highest priority in Step 4, obtain the component layout coordinates, spacing, support height and fixed obstacle positions, and construct a two-dimensional gridded cost map containing passable areas and obstacle markers.

[0012] Step 6, Global Traversal Path Planning: On the two-dimensional gridded cost map constructed in Step 5, the improved A* algorithm with the introduction of a corner penalty term is adopted to generate a globally optimal traversal path covering the surface of all components of the target subarray with the joint optimization objective of minimizing the total cleaning distance and the number of turns.

[0013] Step 7, Local Dynamic Obstacle Avoidance Path Correction: During the movement of the cleaning equipment along the global path in Step 6, dynamic obstacles or unknown static obstacles appearing in front are detected in real time by LiDAR or depth camera. The dynamic window method is used to replan the local path in the velocity space, generate a smooth obstacle avoidance trajectory, bypass the obstacle, and return to the global path.

[0014] Step 8: Cleaning Closed-Loop Feedback and Parameter Optimization: After the cleaning equipment completes the path corrected in Step 7, it collects the power generation data after cleaning, calculates the actual net power generation gain brought by this cleaning, and uses this gain value to update the dynamic threshold of economic loss in Step 3 and the weight parameters of the improved A* algorithm in Step 6.

[0015] Preferably, the gray prediction algorithm in step two adopts the GM(1,1) model, using the average daily dust accumulation decay rate of the past 7 days as the original sequence, and calculates the time response function of the dust accumulation growth rate by accumulating and optimizing the background value.

[0016] Preferably, the dynamic threshold for economic loss in step three is adaptively adjusted based on the real-time grid-connected electricity price, the state of charge of the energy storage system, and the carbon trading price. When the grid-connected electricity price increases or the energy storage capacity is insufficient, the threshold is automatically reduced to prioritize the generation of electricity during high-value periods.

[0017] Preferably, the comprehensive cleaning urgency index in step four is calculated using the following formula:

[0018] Urgency = (Current loss rate × Installed capacity) × (1 + Time decay factor × Number of days since last cleaning)

[0019] The time decay factor is dynamically set according to the seasonal humidity curve, with a value of 0.3 in the rainy season and 0.8 in the dry season.

[0020] Preferably, the grid resolution of the two-dimensional gridded cost map in step five is set to 1 / 2 of the sweeping width of the cleaning device, and a passage cost higher than that of the component surface is assigned to the seam area between components, so as to guide the device to travel along the long side of the component first.

[0021] Preferably, the evaluation function of the improved A* algorithm in step six is ​​designed as f(n) = g(n) + h(n) + t(n), where g(n) is the actual travel distance from the starting point to the current node, h(n) is the Euclidean heuristic distance from the current node to the destination, and t(n) is a turning penalty term, which adds a preset cost constant when the path direction of the current node is different from that of the previous node.

[0022] Preferably, the dynamic window method in step seven specifically includes the following sub-steps:

[0023] Step 7a: Velocity sampling. Based on the kinematic constraints of the cleaning equipment and the acceleration and deceleration performance of the motor, discrete sampling is performed in the two-dimensional velocity space (v,ω) composed of the maximum and minimum linear velocities and angular velocities to generate a set of candidate velocity pairs.

[0024] Step 7b: Trajectory prediction. For each candidate velocity pair (v_i, ω_i), assuming that the device maintains the velocity unchanged within the prediction time window ΔT, a continuous motion trajectory is generated by simulating the circular motion model. The trajectory consists of a series of pose points (x, y, θ).

[0025] Step 7c: Trajectory safety assessment. Calculate the minimum Euclidean distance between each predicted trajectory and the nearest obstacle. If the distance is less than the device's safe radius threshold, the speed pair is directly eliminated.

[0026] Step 7d: Multi-target evaluation of the trajectory. The remaining trajectory is scored using the evaluation function G(v,ω)=α·heading(v,ω)+β·dist(v,ω)+γ·vel(v,ω), where heading represents the angular deviation between the trajectory endpoint and the global path target point, dist represents the distance between the trajectory and the nearest obstacle (saturation value is taken after exceeding the safe distance), vel represents the linear velocity, and α, β, and γ are preset weight coefficients.

[0027] Step 7e: Optimal speed selection. Select the speed pair with the largest evaluation function value as the motion command for the current control cycle and send it to the chassis controller of the sweeping equipment so that the equipment travels along the optimal local trajectory.

[0028] Step 7f: Rolling time-domain update. Repeat steps 7a to 7e every δt time interval until it is detected that the device has completely bypassed the obstacle and the angle between the current heading and the global path direction is less than the preset angle threshold. Then, exit local replanning and resume global path tracking.

[0029] Preferably, the closed-loop feedback and parameter optimization in step eight are performed using a reinforcement learning framework based on deep Q-networks: a Markov decision process model of the cleaning strategy is constructed, wherein the state space includes the current dust loss rate, the probability of rainfall in the next 6 hours, the remaining power of the equipment, the grid electricity price, and the cumulative irradiance since the last cleaning; the action space consists of a combination of cleaning trigger decisions and path planning parameters, wherein the combination of path planning parameters includes the turning penalty constant in step six and the evaluation function weight coefficient vector (α,β,γ) in step seven; the reward function is designed as the net benefit obtained after a single cleaning task is executed, where net benefit = post-cleaning power generation gain - cleaning equipment energy consumption cost - downtime opportunity cost, wherein the post-cleaning power generation gain is determined by the pre-cleaning power generation gain. The difference in actual power generation is multiplied by the electricity price to calculate the result. After each cleaning task is completed, the status of the task, the selected action, the reward obtained, and the status at the next moment are stored as an experience sample in the experience replay pool. When the number of samples in the experience replay pool reaches the preset batch size, a small batch of samples is randomly sampled, and the network parameters of the deep Q network are updated by minimizing the temporal difference error loss function. The calculation of the target Q value adopts a fixed target network to stabilize the training process. Every preset number of cleaning cycles, the trained Q network parameters are synchronized to the online decision network, thereby dynamically adjusting the dynamic threshold of economic loss in step three and the weight of the path planning evaluation function in step six, realizing the joint adaptive optimization of cleaning timing and path planning.

[0030] The present invention has the following advantages: The present invention dynamically quantifies the power generation loss rate caused by dust accumulation through a photovoltaic physical model and a gray prediction algorithm, and constructs an adaptive dynamic threshold for economic loss by combining future weather forecasts (probability of rainfall, wind speed) and real-time grid-connected electricity prices, so as to achieve "on-demand cleaning" rather than "timed cleaning".

[0031] This invention employs a two-layer strategy combining global path planning and local dynamic correction: the improved A* algorithm is used to generate the shortest global traversal path with the fewest turns, while the dynamic window method is used to perform local obstacle avoidance replanning when dynamic obstacles (such as construction workers and birds) appear in real time, ensuring the safety and continuity of the cleaning process.

[0032] By constructing a closed-loop feedback mechanism through a deep Q-network, the actual net power generation gain of each cleaning is used as a reward signal to update the cleaning trigger threshold and path planning weight parameters in reverse, so that the cleaning strategy can continuously adapt to long-term dynamic factors such as seasonal changes in power plants, component degradation, and electricity price fluctuations. Attached Figure Description

[0033] To more intuitively illustrate the prior art and this application, exemplary drawings are provided below. It should be understood that the specific shapes and structures shown in the drawings should not generally be regarded as limiting conditions for implementing this application; for example, based on the technical concept disclosed in this application and the exemplary drawings, those skilled in the art are able to easily make conventional adjustments or further optimizations to the addition / reduction / classification, specific shapes, positional relationships, connection methods, size ratios, etc. of certain units (components).

[0034] Figure 1 A flowchart of the photovoltaic module intelligent cleaning timing judgment and path planning method provided in the embodiments of this application. Detailed Implementation

[0035] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. It should be understood that these embodiments are merely for further explanation of the present invention and should not be construed as limiting the scope of protection of the present invention. Those skilled in the art can make some non-essential improvements and adjustments to the present invention based on the above-described content.

[0036] Please see Figure 1 The method for determining the timing and path planning for intelligent cleaning of photovoltaic modules includes the following steps:

[0037] Step 1: Multidimensional data acquisition and preprocessing: Real-time acquisition of DC output power, ambient light intensity, module backsheet temperature, historical dust accumulation attenuation rate, and weather forecast data for the next 72 hours for each string in the photovoltaic power station, and outlier removal and normalization of the acquired data.

[0038] Step 2, Dynamic Dust Accumulation Loss Assessment: Based on the data processed in Step 1, the instantaneous power loss rate under the current dust accumulation state is calculated using the photovoltaic physical model, and the dust accumulation growth rate prediction function for the next 24 hours is constructed using the grey prediction algorithm. The grey prediction algorithm adopts the GM(1,1) model, using the average daily dust accumulation decay rate of the past 7 days as the original sequence, and calculates the time response function of the dust accumulation growth rate by accumulating and optimizing the background value.

[0039] Step 3: Determine the trigger condition for cleaning: Compare the current power loss rate obtained in Step 2 with the preset dynamic threshold for economic loss (the dynamic threshold for economic loss is adaptively adjusted based on the real-time grid-connected electricity price, the state of charge of the energy storage system, and the carbon trading price. When the grid-connected electricity price increases or the energy storage capacity is insufficient, the threshold is automatically reduced to prioritize the generation of power during high-value periods). At the same time, check the probability of future rainfall and the wind speed forecast. When the power loss rate continuously exceeds the threshold and there is no effective rainfall or strong wind warning in the next 6 hours, a cleaning trigger command is generated.

[0040] Step 4: Prioritizing cleaning of multiple subarrays: In response to the cleaning trigger command in Step 3, obtain the dust loss rate, installed capacity, last cleaning time and dust growth rate of all photovoltaic subarrays to be cleaned in the power station, calculate the comprehensive cleaning urgency index of each subarray, and sort them from high to low index.

[0041] Step 5: Construction of the topology map of the cleaning area: For the target subarray with the highest priority in Step 4, obtain its component layout coordinates, spacing, bracket height and fixed obstacle positions, and construct a two-dimensional gridded cost map containing passable areas and obstacle markers. The grid resolution of the two-dimensional gridded cost map is set to 1 / 2 of the sweeping width of the cleaning device, and a passable cost higher than the component surface is assigned to the seam area between components to guide the device to travel along the long side of the component first.

[0042] Step 6, Global Traversal Path Planning: On the two-dimensional gridded cost map constructed in Step 5, the improved A* algorithm with the introduction of a corner penalty term is adopted to generate a globally optimal traversal path covering the surface of all components of the target subarray with the joint optimization objective of minimizing the total cleaning distance and the number of turns.

[0043] Step 7, Local Dynamic Obstacle Avoidance Path Correction: During the movement of the cleaning equipment along the global path in Step 6, dynamic obstacles or unknown static obstacles appearing in front are detected in real time by LiDAR or depth camera. The dynamic window method is used to replan the local path in the velocity space, generate a smooth obstacle avoidance trajectory, bypass the obstacle, and return to the global path.

[0044] Step 8: Cleaning Closed-Loop Feedback and Parameter Optimization: After the cleaning equipment completes the path corrected in Step 7, it collects the power generation data after cleaning, calculates the actual net power generation gain brought by this cleaning, and uses this gain value to update the dynamic threshold of economic loss in Step 3 and the weight parameters of the improved A* algorithm in Step 6.

[0045] This invention first collects multi-dimensional data in real time, including photovoltaic string power, irradiance, temperature, historical dust accumulation, and weather forecasts, and then removes outliers. This preprocessing step ensures the accuracy of subsequent judgments and avoids misjudgments caused by sensor noise. Next, a photovoltaic physical model is used to calculate the power loss rate caused by current dust accumulation, and a grey prediction algorithm is used to predict the dust growth trend over the next 24 hours. Compared to a simple method that only considers current losses, this step achieves "predictability" in determining the timing of cleaning, preventing continuous power generation losses caused by delayed cleaning.

[0046] The system then compares the current loss rate with a dynamic economic threshold, while simultaneously verifying future rainfall and wind speed conditions: a cleaning command is triggered only when the loss continuously exceeds the threshold and there is no effective rainfall or strong wind within the next 6 hours. This joint decision utilizes natural rainfall to achieve free cleaning, avoids ineffective cleaning, and ensures the safety of equipment in windy weather, significantly improving the economic efficiency of cleaning decisions.

[0047] After triggering cleaning, the system sorts each subarray according to its urgency index, prioritizing the cleaning of areas with the highest loss and greatest benefit, thus achieving "value-oriented" scheduling of cleaning resources. A two-dimensional cost map is constructed for the target subarray, and an improved A* algorithm incorporating a corner penalty term is used to generate the globally optimal traversal path, combining the total travel distance and the number of turns, which can reduce cleaning energy consumption by 15%–25%. During cleaning, dynamic obstacles are detected in real time using LiDAR, and a dynamic window method is used for local obstacle avoidance and replanning. After bypassing obstacles, the system automatically returns to the global path, ensuring cleaning continuity and equipment safety.

[0048] Finally, the system dynamically updates the threshold and path planning weights based on the actual power generation gain before and after cleaning, forming a closed-loop self-optimization mechanism that enables the cleaning strategy to continuously adapt to electricity price fluctuations and component degradation, and to approach optimal economic efficiency in the long term.

[0049] The comprehensive cleaning urgency index in step four is calculated using the following formula:

[0050] Urgency = (Current loss rate × Installed capacity) × (1 + Time decay factor × Number of days since last cleaning)

[0051] The time decay factor is dynamically set according to the seasonal humidity curve, with a value of 0.3 in the rainy season and 0.8 in the dry season. The core of this formula is to quantify the instantaneous loss caused by dust accumulation, the power generation scale of the subarray, the time decay after the last cleaning, and the influence of seasonal humidity into a comparable value, thereby achieving a scientific ranking of the cleaning priorities of different subarrays.

[0052] In some embodiments, a photovoltaic power station includes two subarrays, A and B. Subarray A currently has a loss rate of 8%, an installed capacity of 2MW, is located in a dry season region (time decay factor 0.8), and was last cleaned 10 days ago. Subarray B currently has a loss rate of 7%, an installed capacity of 5MW, is located in a rainy season region (time decay factor 0.3), and was last cleaned 5 days ago. If we disregard cleaning history and seasonal factors and simply calculate based on the instantaneous total power generation loss (i.e., loss rate × installed capacity): the total loss of subarray A is 0.16MW equivalent, and the total loss of subarray B is 0.35MW equivalent, so B would be cleaned first. However, this calculation ignores the dust accumulation trend and natural cleaning opportunities: subarray A is located in the dry season, where dust accumulation is rapid and there is no rain for cleaning; it has also not been cleaned for 10 days, resulting in a high risk of dust deterioration. Subarray B is located in the rainy season, where the probability of future rainfall is high, and it was only cleaned 5 days ago, so the dust accumulation is still relatively light. The comprehensive urgency formula of this invention is used: Urgency = (Current Loss Rate × Installed Capacity) × (1 + Time Decay Factor × Number of Days Since Last Cleaning). The calculation yields: Subarray A urgency = (8% × 2) × (1 + 0.8 × 10) = 0.16 × 9 = 1.44; Subarray B urgency = (7% × 5) × (1 + 0.3 × 5) = 0.35 × 2.5 = 0.875. Clearly, subarray A has a much higher urgency than B and should be cleaned first. This formula incorporates a time decay factor and a seasonal factor (0.3 in the rainy season, 0.8 in the dry season) to include the dust accumulation rate and natural cleaning opportunities in the evaluation. This prioritizes cleaning resources for subarrays with "high instantaneous total loss, accelerating future loss growth, and no short-term rainfall relief," thus avoiding the problem of "wasting cleaning resources on subarrays about to be rained on" that might occur with decisions based solely on instantaneous total loss. Meanwhile, the high factor during the dry season increases the urgency of long-neglected subarrays, while the low factor during the rainy season automatically reduces the cleaning priority of areas with frequent rainfall. This design makes the allocation of cleaning resources more in line with power generation economics, and actual measurements show that it can increase the net benefit of priority cleaning decisions by approximately 20% to 30%.

[0053] The evaluation function of the improved A* algorithm in step six is ​​designed as f(n) = g(n) + h(n) + t(n), where g(n) is the actual travel distance from the starting point to the current node, h(n) is the Euclidean heuristic distance from the current node to the destination, and t(n) is a turning penalty term. This penalty term adds a preset cost constant when the path direction of the current node is different from that of the previous node. Thus, when planning the global traversal path, it not only pursues the shortest travel distance, but also minimizes the frequent turning of the cleaning equipment, thereby reducing energy consumption, reducing mechanical wear, and shortening the total cleaning time.

[0054] The dynamic window method in step seven specifically includes the following sub-steps:

[0055] Step 7a: Velocity sampling. Based on the kinematic constraints of the cleaning equipment and the acceleration and deceleration performance of the motor, discrete sampling is performed in the two-dimensional velocity space (v,ω) composed of the maximum and minimum linear velocities and angular velocities to generate a set of candidate velocity pairs.

[0056] Step 7b: Trajectory prediction. For each candidate velocity pair (v_i, ω_i), assuming that the device maintains the velocity unchanged within the prediction time window ΔT, a continuous motion trajectory is generated by simulating the circular motion model. The trajectory consists of a series of pose points (x, y, θ).

[0057] Step 7c: Trajectory safety assessment. Calculate the minimum Euclidean distance between each predicted trajectory and the nearest obstacle. If the distance is less than the device's safe radius threshold, the speed pair is directly eliminated.

[0058] Step 7d: Multi-target evaluation of the trajectory. The remaining trajectory is scored using the evaluation function G(v,ω)=α·heading(v,ω)+β·dist(v,ω)+γ·vel(v,ω), where heading represents the angular deviation between the trajectory endpoint and the global path target point, dist represents the distance between the trajectory and the nearest obstacle (saturation value is taken after exceeding the safe distance), vel represents the linear velocity, and α, β, and γ are preset weight coefficients.

[0059] Step 7e: Optimal speed selection. Select the speed pair with the largest evaluation function value as the motion command for the current control cycle and send it to the chassis controller of the sweeping equipment so that the equipment travels along the optimal local trajectory.

[0060] Step 7f: Rolling time-domain update. Repeat steps 7a to 7e every δt time interval until it is detected that the device has completely bypassed the obstacle and the angle between the current heading and the global path direction is less than the preset angle threshold. Then, exit local replanning and resume global path tracking.

[0061] To explain the above scheme in more detail, the following embodiments are provided:

[0062] The cleaning equipment is traveling at a constant speed along the global path (linear velocity 0.5 m / s, angular velocity 0). Suddenly, the lidar detects a temporarily placed sandbag (dynamic obstacle) 3 meters ahead. The system initiates the following steps:

[0063] 7a Velocity sampling: Based on the equipment's maximum acceleration / deceleration (linear acceleration 0.3 m / s²). 2 Angular acceleration 0.5 rad / s² 2 Discrete sampling is performed within the range of linear velocity [0, 0.8] m / s and angular velocity [-0.6, 0.6] rad / s to generate approximately 60 candidate velocity pairs.

[0064] 7b Trajectory Prediction: For each velocity pair (v, ω), assuming a future prediction time window ΔT = 1.5 seconds, a trajectory is generated using a circular motion model. For example, the velocity pair (0.4 m / s, 0.3 rad / s) will produce a rightward-curving arc.

[0065] 7c Safety Assessment: Calculate the minimum distance between each predicted trajectory and the sandbag outline. If a trajectory's distance is less than the safe radius of 0.2 meters, it is directly discarded. For example, a straight trajectory (ω=0) will collide with the sandbag and be discarded; while a right-curving trajectory with a distance greater than 0.25 meters is retained.

[0066] 7D Multi-Objective Evaluation: The evaluation function G(v,ω) = α·heading + β·dist + γ·vel is calculated for the retained trajectories. Let α = 0.5, β = 0.4, and γ = 0.1. A candidate trajectory has an endpoint orientation that deviates 10° from the global path target direction (heading = 0.9), a distance of 0.3 meters from the sandbag (dist = 0.3; after exceeding the safe distance, a saturation value of 0.5 is used, so the actual input is 0.5), and a linear velocity of 0.4 m / s (vel = 0.4). Therefore, G = 0.5 × 0.9 + 0.4 × 0.5 + 0.1 × 0.4 = 0.45 + 0.2 + 0.04 = 0.69. Another trajectory has a slower but safer velocity, and may receive similar scores. The system selects the velocity pair corresponding to the highest score.

[0067] 7e Optimal Speed ​​Selection: Send the selected (0.4m / s, 0.3rad / s) to the chassis controller, and the equipment will turn right to avoid the obstacle.

[0068] 7f Rolling Time Domain Update: The above process is repeated every δt = 0.1 seconds. When the device bypasses the sandbag and detects that the angle between the current heading and the global path direction is less than 5°, it automatically exits local replanning and resumes traveling along the global path.

[0069] The closed-loop feedback and parameter optimization in step eight employ a reinforcement learning framework based on deep Q-networks: a Markov decision process model for the cleaning strategy is constructed, where the state space includes the current dust loss rate, the probability of rainfall in the next 6 hours, the remaining power of the equipment, the grid electricity price, and the cumulative irradiance since the last cleaning. The action space consists of a combination of cleaning trigger decisions and path planning parameters, where the path planning parameter combination includes the turning penalty constant from step six and the evaluation function weight coefficient vector (α, β, γ) from step seven. The reward function is designed as the net benefit obtained after a single cleaning task, where net benefit = post-cleaning power generation gain - cleaning equipment energy consumption cost - downtime opportunity cost, where the post-cleaning power generation gain is determined by the actual power generation before and after cleaning. The difference in actual power generation is multiplied by the electricity price to calculate the result. After each cleaning task is completed, the status of this task, the selected action, the reward obtained, and the status at the next moment are stored as an experience sample in the experience replay pool. When the number of samples in the experience replay pool reaches the preset batch size, a small batch of samples is randomly sampled, and the network parameters of the deep Q network are updated by minimizing the temporal difference error loss function. The calculation of the target Q value adopts a fixed target network to stabilize the training process. Every preset number of cleaning cycles, the trained Q network parameters are synchronized to the online decision network, thereby dynamically adjusting the dynamic threshold of economic loss in step three and the weight of the path planning evaluation function in step six, realizing the joint adaptive optimization of cleaning timing and path planning.

[0070] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for judging the cleaning time and planning the cleaning path of a photovoltaic module, characterized in that, Includes the following steps: Step 1: Multidimensional data acquisition and preprocessing: Real-time acquisition of DC output power, ambient light intensity, module backsheet temperature, historical dust accumulation attenuation rate, and weather forecast data for the first preset time in the photovoltaic power station for each string, and outlier removal and normalization of the acquired data. Step 2, Dynamic Dust Accumulation Loss Assessment: Based on the data processed in Step 1, the instantaneous power loss rate under the current dust accumulation state is calculated using a photovoltaic physical model, and a gray prediction algorithm is used to construct a prediction function for the dust accumulation growth rate within the next preset time period. Step 3: Determine the sweeping trigger condition: Compare the current power loss rate obtained in Step 2 with the preset dynamic threshold for economic loss, and at the same time verify the probability of future rainfall and the predicted wind speed. When the power loss rate continuously exceeds the threshold and there is no effective rainfall or strong wind warning in the next three preset time period, a sweeping trigger command is generated. Step 4: Prioritizing cleaning of multiple subarrays: In response to the cleaning trigger command in Step 3, obtain the dust loss rate, installed capacity, last cleaning time and dust growth rate of all photovoltaic subarrays to be cleaned in the power station, calculate the comprehensive cleaning urgency index of each subarray, and sort them from high to low index. Step 5: Construction of topology map of the cleaning area: For the target subarray with the highest priority in Step 4, obtain the component layout coordinates, spacing, support height and fixed obstacle positions, and construct a two-dimensional gridded cost map containing passable areas and obstacle markers. Step 6, Global Traversal Path Planning: On the two-dimensional gridded cost map constructed in Step 5, the improved A* algorithm with the introduction of a corner penalty term is adopted to generate a globally optimal traversal path covering the surface of all components of the target subarray with the joint optimization objective of minimizing the total cleaning distance and the number of turns. Step 7, Local Dynamic Obstacle Avoidance Path Correction: During the movement of the cleaning equipment along the global path in Step 6, dynamic obstacles or unknown static obstacles appearing in front are detected in real time by LiDAR or depth camera. The dynamic window method is used to replan the local path in the velocity space, generate a smooth obstacle avoidance trajectory, bypass the obstacle, and return to the global path. Step 8: Cleaning Closed-Loop Feedback and Parameter Optimization: After the cleaning equipment completes the path corrected in Step 7, it collects the power generation data after cleaning, calculates the actual net power generation gain brought by this cleaning, and uses this gain value to update the dynamic threshold of economic loss in Step 3 and the weight parameters of the improved A* algorithm in Step 6. 2.The method of claim 1, wherein, The gray prediction algorithm in step two adopts the GM(1,1) model, which uses the average daily dust decay rate of the past few days as the original sequence, and calculates the time response function of the dust growth rate by accumulating and optimizing the background value. 3.The method of claim 2, wherein, The dynamic threshold for economic loss in step three is adaptively adjusted based on the real-time grid-connected electricity price, the state of charge of the energy storage system, and the carbon trading price. When the grid-connected electricity price increases or the energy storage capacity is insufficient, the threshold is automatically reduced to prioritize the generation of electricity during high-value periods. 4.The method of claim 3, wherein, The comprehensive cleaning urgency index in step four is calculated using the following formula: Urgency = (Current loss rate × Installed capacity) × (1 + Time decay factor × Number of days since last cleaning) The time decay factor is dynamically set according to the seasonal humidity curve, with a value of 0.3 in the rainy season and 0.8 in the dry season.

5. The method for intelligent cleaning timing and path planning of photovoltaic modules according to claim 4, characterized in that, In step five, the grid resolution of the two-dimensional gridded cost map is set to 1 / 2 of the sweeping width of the cleaning device, and a passage cost higher than that of the component surface is assigned to the seam area between components to guide the device to travel along the long side of the component first.

6. The method for intelligent cleaning timing and path planning of photovoltaic modules according to claim 5, characterized in that, The evaluation function of the improved A* algorithm in step six is ​​designed as f(n)=g(n)+h(n)+t(n), where g(n) is the actual travel distance from the starting point to the current node, h(n) is the Euclidean heuristic distance from the current node to the destination, and t(n) is the turning penalty term, which increases the preset cost constant when the path direction of the current node is different from that of the previous node.

7. The method for intelligent cleaning timing and path planning of photovoltaic modules according to claim 6, characterized in that, The dynamic window method in step seven specifically includes the following sub-steps: Step 7a: Velocity sampling. Based on the kinematic constraints of the cleaning equipment and the acceleration and deceleration performance of the motor, discrete sampling is performed in the two-dimensional velocity space (v,ω) composed of the maximum and minimum linear velocities and angular velocities to generate a set of candidate velocity pairs. Step 7b: Trajectory prediction. For each candidate velocity pair (v_i, ω_i), assuming that the device maintains the velocity unchanged within the prediction time window ΔT, a continuous motion trajectory is generated by simulating the circular motion model. The trajectory consists of a series of pose points (x, y, θ). Step 7c: Trajectory safety assessment. Calculate the minimum Euclidean distance between each predicted trajectory and the nearest obstacle. If the distance is less than the device's safe radius threshold, the speed pair is directly eliminated. Step 7d: Multi-target evaluation of the trajectory. The remaining trajectory is scored using the evaluation function G(v,ω)=α·heading(v,ω)+β·dist(v,ω)+γ·vel(v,ω), where heading represents the angular deviation between the trajectory endpoint and the global path target point, dist represents the distance between the trajectory and the nearest obstacle (saturation value is taken after exceeding the safe distance), vel represents the linear velocity, and α, β, and γ are preset weight coefficients. Step 7e: Optimal speed selection. Select the speed pair with the largest evaluation function value as the motion command for the current control cycle and send it to the chassis controller of the sweeping equipment so that the equipment travels along the optimal local trajectory. Step 7f: Rolling time-domain update. Repeat steps 7a to 7e every δt time interval until it is detected that the device has completely bypassed the obstacle and the angle between the current heading and the global path direction is less than the preset angle threshold. Then, exit local replanning and resume global path tracking.

Images