Autonomous navigation and path planning method and system for photovoltaic panel cleaning robot

Abstract

The invention discloses an autonomous navigation and path planning method and system for a photovoltaic panel cleaning robot, and the method comprises the steps: estimating a dirt shielding rate SR (x, y) in real time through fusing an infrared thermal image, a visible light image and irradiation data, constructing a post-cleaning generation power gain deltaP model, and coupling the movement and cleaning energy consumption Etotal of the robot, thereby achieving the autonomous navigation and path planning of the photovoltaic panel cleaning robot. The unit energy consumption power generation gain EGR = delta P / Etotal is innovatively taken as a path cost function, online re-planning is carried out in a dynamic grid map by adopting an EGR-A * algorithm, and a high power generation income area is traversed preferentially, so that on-demand cleaning is realized. The system supports self-adaptive recharging of an electric quantity threshold value, a federated learning cloud end optimizes model parameters in a collaborative manner, invalid paths and energy consumption are remarkably reduced, the power generation capacity of a power station and the endurance utilization rate of a robot are improved, and the problems that traditional full-coverage paths are redundant, and low-pollution areas are excessively clean are solved.

Description

Technical Field

[0001] This application relates to the field of photovoltaic power generation technology, and in particular to an autonomous navigation and path planning method and system for photovoltaic panel cleaning robots. Background Technology

[0002] With the rapid growth of global photovoltaic power plant installed capacity, the loss of power generation efficiency caused by dust accumulation, bird droppings, and industrial dust on panel surfaces is becoming increasingly prominent. Manual or semi-automatic cleaning methods are no longer sufficient to meet the needs of large-scale, high-frequency, and low-cost operation and maintenance. Existing photovoltaic cleaning robots generally adopt fixed trajectories or simple full-coverage path planning (such as ox-plowing or spiral-type), and their path generation is decoupled from the actual degree of panel contamination. This results in over-cleaning of low-contamination areas and under-cleaning of high-contamination areas, leading to high operational redundancy and limited overall power generation gain.

[0003] Meanwhile, traditional systems have significant shortcomings in energy management: path planning only considers geometric distance or coverage order and does not dynamically link robot movement power consumption, cleaning power consumption and remaining battery power, often resulting in the inability to complete cleaning of high-value areas before the battery runs out; in addition, the evaluation of cleaning effect is mostly based on the visual dirt area and lacks quantitative modeling of the core economic indicator of "increased power generation after cleaning", making it difficult to achieve true optimal benefits.

[0004] In recent years, although some studies have attempted to introduce visual recognition of dirt and guide local path adjustments, they have remained at the level of "dirt priority," failing to establish a coupling relationship between power generation revenue and energy consumption, and lacking online replanning and adaptive termination capabilities. Therefore, there is an urgent need for an autonomous navigation and path planning system that can perceive pollution distribution in real time, quantify power generation revenue, couple energy consumption constraints, and dynamically optimize paths to solve the problems of low cleaning efficiency, high energy consumption, and poor economic efficiency in existing technologies. Summary of the Invention

[0005] This application provides an autonomous navigation and path planning method and system for a photovoltaic panel cleaning robot. Its advantages are that it can perceive the distribution of photovoltaic panel pollution in real time, quantify power generation benefits, couple energy consumption constraints and dynamically optimize the path, so as to solve the problems of low cleaning efficiency, high energy consumption and poor economic efficiency in the prior art.

[0006] The technical solution of this application is as follows: On the one hand, this application provides an autonomous navigation and path planning method for a photovoltaic panel cleaning robot, including the following steps: S1: Collect infrared thermal image data, RGB images and real-time irradiance of the photovoltaic panel surface, obtain the panel temperature based on the infrared thermal image data, and estimate the dirt shading rate of the photovoltaic panel. S2: Calculate the power generation gain of the grid after cleaning based on the dirt shading rate of each grid of the photovoltaic panel, the panel temperature and real-time irradiance; S3: Calculate the robot's movement energy consumption and cleaning energy consumption on the target path, and calculate the total energy consumption; S4: Calculate the power generation gain per unit energy consumption (EGR) based on the power generation gain and total energy consumption, and construct a dynamic cost map that is updated in real time using EGR as the path cost function. S5: Using EGR as the cost core on the grid map, the EGR-A* algorithm is used for online path search. When the remaining battery power is lower than the set safety threshold, the expected EGR of the subsequent path is compared with the historical average EGR. If the expected EGR is less than 80% of the historical average EGR, the recharge path planning is triggered.

[0007] Furthermore, in step S1, the dirt shading rate SR(x,y) of each spatial grid of the photovoltaic panel is estimated by using a lightweight CNN segmentation network.

[0008] Furthermore, in step S2, the power generation gain ΔP(x,y) is calculated using the following formula: ΔP(x,y)=Pstd(x,y)·ηclean·SR(x,y) Where Pstd(x,y) is the standard power generation after temperature derating correction, ηclean is the cleaning efficiency coefficient corresponding to the cleaning method, and SR(x,y) is the dirt cover rate.

[0009] Furthermore, in step S3, the robot's movement energy consumption (Emove) on the target path is obtained through the following piecewise model: Emove=k1·d+k2·Σθ+k3·ΣΔh Where d is the path length, Σθ is the cumulative steering angle, ΣΔh is the cumulative climb height, and k1, k2, and k3 are coefficients obtained through actual vehicle calibration; Total energy consumption (Etotal) is the sum of mobile energy consumption (Emove) and clean energy consumption (Eclean). Etotal = Emove + Eclean.

[0010] Furthermore, in step S4, the power generation gain per unit energy consumption (EGR) is obtained by the following formula: EGR=Σ[ΔP(x,y)·A(x,y)] / Etotal.

[0011] Furthermore, in step S5, the heuristic function of the EGR-A* algorithm is: f(n) = g(n) + λ•1 / hEGR(n) Where g(n) is the actual cumulative EGR value from the starting point to the current node n, hEGR(n) is the estimated EGR gain from node n to the target region, and λ is the dynamic coefficient.

[0012] Furthermore, in step S5, when the cleaning robot's battery SOC is ≤20%, the maximum EGR path length Lmax that the remaining power can support is calculated. If Lmax is less than 1.5 times the path length from the current point to the nearest charging station, the robot immediately switches to recharge mode and records the coordinates of the interrupted node and the list of uncleaned grids. After charging is completed, the EGR-A* planning is restarted based on the latest dirt heat map.

[0013] On another front, this application provides an autonomous navigation and path planning system for a photovoltaic panel cleaning robot, comprising: The multimodal sensing module is used to simultaneously collect infrared thermal image data, RGB images and real-time irradiance on the surface of the photovoltaic panel, obtain the panel temperature based on the infrared thermal image data, and estimate the dirt shading rate of the photovoltaic panel. The power generation gain modeling module is used to calculate the power generation gain of each grid after cleaning based on the dirt shading rate, panel temperature and real-time irradiance of each grid of the photovoltaic panel. The energy consumption modeling module is used to calculate the robot's movement energy consumption and cleaning energy consumption on the target path, and output the total energy consumption. The power generation revenue-energy consumption coupling optimization module is used to construct a dynamic cost map that is updated in real time, with the power generation gain per unit energy consumption (EGR) as the path cost function. The dynamic route replanning controller uses EGR as the cost core on the grid map and performs online route search using the EGR-A* algorithm. When the remaining battery power is lower than a set safety threshold, it compares the expected EGR of the subsequent route with the historical average EGR. If the expected EGR is less than 80% of the historical average EGR, it triggers recharging route planning.

[0014] In summary, the beneficial effects of this application are as follows: 1. It can sense the distribution of pollution on photovoltaic panels in real time, quantify power generation benefits, couple energy consumption constraints and dynamically optimize the path to solve the problems of low clean efficiency, high energy consumption and poor economic performance in existing technologies; 2. Couple "power generation revenue - energy consumption" into a single EGR cost function and embed it into A* online replanning to obtain the path related to power generation revenue - energy consumption through A* online replanning; 3. A three-dimensional joint ΔP model of temperature derating, irradiation, and fouling addresses the blind spot of existing patents that "only consider fouling and not temperature loss"; 4. Federated learning continuously updates ηclean, λ, and k1-k3, forming a system-level self-evolution, which is different from static calibration schemes. Attached Figure Description

[0015] Figure 1 This is a block diagram illustrating the principle of the present invention; Figure 2 This is a flowchart of the workflow of the present invention. Detailed Implementation

[0016] The specific embodiments of this application are described in detail below with reference to the accompanying drawings.

[0017] Example 1: An autonomous navigation and path planning method for a photovoltaic panel cleaning robot, referring to... Figure 1-2 This includes the following steps: S1: Collect infrared thermal image data, RGB images, and real-time irradiance Gpoa of the photovoltaic panel surface. Obtain the panel temperature Tpanel based on the infrared thermal image data and estimate the dirt shading rate of the photovoltaic panel. Estimate the dirt shading rate SR(x,y) of each spatial grid of the photovoltaic panel using a lightweight CNN segmentation network.

[0018] This embodiment uses a typical photovoltaic power station as an example. This station employs a general-purpose mobile chassis and cleaning head, eliminating the need for specific track / brush structures. The sensors used include: a visible light camera (1920×1080@30 fps), an infrared thermal imager (640×512@25 fps), a miniature radiometer (±2%), and an RTK+IMU+encoder, all outputting via ROS 2 Topic to maintain hardware neutrality. The difference between infrared and RGB timestamps is ≤1 ms, achieved via the PTP protocol. Mobile-U-Net segmentation: TensorRT INT8 quantization, GPU <1.5 GB, inference <25 ms, 15 FPS output binary mask, SR(x,y) estimation error ≤±5%. Temperature derating: Tpanel extracts the peak value of the infrared image histogram and corrects Pstd(x,y)=PSTC·[1-0.0045·(Tpanel-25)]·(Gpoa / 1000) in real time, which solves the defect of the background technology that "only considers dirt and not temperature loss".

[0019] S2: Calculate the power generation gain of each grid cell after cleaning based on the dirt shading rate, panel temperature, and real-time irradiance. The power generation gain ΔP(x,y) is calculated using the following formula: ΔP(x,y)=Pstd(x,y)·ηclean·SR(x,y) Where Pstd(x,y) is the standard power generation after temperature derating correction, ηclean is the cleaning efficiency coefficient corresponding to the cleaning method, and SR(x,y) is the dirt cover rate.

[0020] ηclean is updated weekly through federated learning: dry brush 0.78→0.81, water washing 0.92→0.94, solving the static ηclean drift problem.

[0021] A "profit threshold" ΔPmin = 0.5W is introduced. Areas below this value are marked as NO_INFORMATION in the cost map to avoid exploring invalid areas.

[0022] S3: Calculate the robot's energy consumption for movement and cleaning along the target path, and calculate the total energy consumption.

[0023] The robot's energy consumption (Emove) on the target path is obtained through the following piecewise model: Emove=k1·d+k2·Σθ+k3·ΣΔh Where d is the path length, Σθ is the cumulative steering angle, ΣΔh is the cumulative climb height, and k1, k2, and k3 are coefficients obtained through actual vehicle calibration. The coefficients k1-k3 are solved online using the least squares method in the actual vehicle 100 m × 100 m calibration field, with R² ≥ 0.96, and are stored in ~ / config / energy.yaml, supporting cloud OTA differential updates.

[0024] Eclean model: Eclean=(τ·ω·tbrush+Ppump·tpump) / ηbat, τ is dynamically adjusted according to SR level (0.6 N·m→1.4 N·m), Ppump is cleaning power, tpump is cleaning time, tbrush is given by the "SR→time" lookup table, ηbat=0.92.

[0025] Total energy consumption (Etotal) is the sum of mobile energy consumption (Emove) and clean energy consumption (Eclean). Etotal = Emove + Eclean.

[0026] S4: Calculate the power generation gain per unit energy consumption (EGR) based on the power generation gain and total energy consumption. Use EGR as the path cost function to construct a dynamically updated cost map in real time. The power generation gain per unit energy consumption (EGR) is obtained using the following formula: EGR=Σ[ΔP(x,y)·A(x,y)] / Etotal.

[0027] A(x,y) represents the area of ​​the photovoltaic panel.

[0028] S5: Using EGR as the cost core on the grid map, the EGR-A* algorithm is used for online path search. When the remaining battery power is lower than the set safety threshold, the expected EGR of the subsequent path is compared with the historical average EGR. If the expected EGR is less than 80% of the historical average EGR, the recharge path planning is triggered.

[0029] Cost map: Cost(x,y)=-EGR(x,y)=-ΔP(x,y)·A(x,y) / (Emove+Eclean), the negative sign guarantees that the minimum cost of A* search corresponds to the maximum EGR.

[0030] Cleaning priority weight: w(x,y)=ΔP(x,y) / tclean, which accelerates convergence to the high-yield region by adding negative costs. The weight coefficient α=0.1, and the convergence speed is improved by 27% according to Monte Carlo simulation experiments.

[0031] The heuristic function of the EGR-A* algorithm is: f(n) = g(n) + λ•1 / hEGR(n) Where g(n) is the actual cumulative EGR from the starting point to the current node n, hEGR(n) is the estimated EGR gain from node n to the target region, and λ is the dynamic coefficient. λ is initially set to 1.0, and after each operation, it is corrected with a step size of 0.05 according to the formula: Error = |Actual EGR - Estimated EGR| / Actual EGR, with a limit of [0.6, 1.4]. After four weeks, λ converges to 0.78, and the prediction error decreases from 7.8% to 2.9%.

[0032] Online replanning cycle: A local window of 5 m × 5 m is triggered for every 0.05 m × 0.05 m grid completed, with approximately 12 k nodes and a time of less than 80 ms, achieving "insertion immediately upon seeing a high-contamination area".

[0033] Battery Adaptive Termination: When the cleaning robot's battery SOC ≤ 20%, calculate the maximum EGR path length Lmax that the remaining battery power can support. If Lmax is less than 1.5 times the path length from the current point to the nearest charging station, immediately switch to recharge mode and record the coordinates of the interrupted node and the list of uncleaned grids. After charging is completed, restart EGR-A* planning based on the latest dirt heatmap. Specifically, when SOC ≤ 20%, calculate Lmax = (SOC - SOCsafe)·Ebat / (Etotal / Lstep), SOCsafe = 8%. If Lmax < 1.5·Lhome, save the list of uncleaned grids (approximately 1.2 kB) and generate a recharge path, with a cruising speed of 0.8 m / s to shorten the return trip.

[0034] The module includes a map management and task recording feature. This module employs a multi-layered semantic map: a dirt heatmap, cleaned areas, energy distribution, and obstacle layers. It uses the ROS 2 Costmap2D format with a resolution of 0.05 m and an update latency of <50ms. The cleaned area lifecycle is defined as follows: the timestamp label defaults to 72 hours, and after this timeout, the confidence level decays by 10% per hour. Areas with a confidence level below 0.5 become candidate clean grids again, addressing the issue of missed "secondary pollution after rain."

[0035] Federated learning cloud collaboration: JSON packets {job_id, EGR_real, ΣΔP, Etotal, λ_used, k1-k3, ηclean} are uploaded via MQTT, with a packet size <8 KB. Local caching is performed for 48 hours during 4G interruptions. The cloud uses FedAvg to aggregate data from hundreds of machines, issuing new parameters weekly. Hot replacement is performed after digital signature verification, requiring no downtime.

[0036] The following is a specific scenario to further illustrate the solution of this embodiment: This embodiment is implemented in a 50 MW centralized ground-mounted power station in Dalad Banner, Inner Mongolia. The single row of modules is 85 m long, with a tilt angle of 35°, a north-south spacing of 4.5 m, an altitude of 1450 m, and an annual minimum temperature of -20 ℃ and a maximum of 60 ℃. The experimental sub-array has a capacity of 0.8 MW, consisting of 2880 540 W monocrystalline modules arranged in a 4×11 configuration. The robot platform is a general-purpose tracked differential chassis with a weight of 42 kg, a 48 V 32.5Ah (1.56 kWh) battery, the ability to turn in place, a maximum speed of 0.8 m / s, and a cleaning head with a dual-mode design of a liftable roller brush and a low-pressure water brush, with a rated power of 350 W. The computing core uses an NVIDIA Jetson Orin Nano (8 GB), the operating system is ROS 2Humble, the middleware is DDS-Cyclone, the real-time loop is 10 ms, and the MCU is an STM32H743 responsible for the current loop at 16 kHz and the speed loop at 1 kHz. The following is a step-by-step explanation of the system closed-loop process.

[0037] S1 Multimodal Data Acquisition and SR(x,y) Generation A visible light camera (Hikvision MV-CA013-20GM, 1 / 3" global shutter, 1920×1080, 30 fps) and an infrared thermal imager (FLIR Lepton 3.5, 640×512, 25 fps) achieve sub-millisecond synchronization via a hard-triggered board. A radiometer is mounted on the top horizontal plane of the robot, with a cosine error ≤ ±2%, and outputs GpoA. The two images are hardware timestamped and then fed into Mobile-U-Net (encoder MobileNetV3-small, decoder 4-level upsampling). The network uses TensorRT INT8 quantization at the edge, with GPU usage <1.5 GB, a single inference time of 22 ms, and a stable frame rate of 15 FPS. After outputting a binary mask, the grid dirt occlusion rate SR(x,y) = Ssoil / Stotal is calculated. The grid resolution is 0.05 m, and the estimation error is verified to be ±4.3% by 2000 manually labeled samples, meeting the ±5% requirement. Indicators. The infrared image synchronously provides the panel surface temperature Tpanel, and the histogram peak method is used to remove border and sky reflections, with an accuracy of ±1.5 ℃.

[0038] S2 ΔP(x,y) and energy consumption are calculated in parallel. The power generation gain modeling module subscribes to the topics / soiling_grid, / ir_temp, and / irradiance, and then... Pstd(x,y)=PSTC·[1-0.0045·(Tpanel-25)]·(Gpoa / 1000) A dual derating process involving temperature and irradiation was implemented, with PSTC = 540 W. Cleaning efficiency ηclean was distributed weekly from the cloud: 0.81 for dry brush mode and 0.94 for wet wash mode, representing a 3-4% improvement over the static value. The final ΔP(x,y) = Pstd(x,y)·ηclean·SR(x,y) was published at 15 Hz via the / gain_grid topic.

[0039] The energy consumption modeling module simultaneously receives chassis CAN messages (wheel speed, steering angle, current) and IMU elevation data, integrates them in real time to obtain d, Σθ, and ΣΔh, and then... Emove=k1·d+k2·Σθ+k3·ΣΔh Calculate mobile energy consumption. During the calibration phase, three sets of experiments were conducted on a 100 m × 100 m concrete surface: constant speed, turning, and climbing. Least square fitting yielded k1 = 76 J·m⁻¹, k2 = 1.96 J·rad⁻¹, and k3 = 558 J·m⁻¹ (elevation), with R² > 0.96. Clean energy consumption. Eclean=(τ·ω·tbrush+Ppump·tpump) / ηbat, τ is based on a linear interpolation of 0.6–1.4 N·m according to SR levels, ω = 83.7 rad / s (800 rpm), tpump is given by the “SR→time” lookup table, and ηbat = 0.92. Etotal = Emove + Eclean is published at 15 Hz on the / energy_cost_grid topic, spatially aligned with / gain_grid.

[0040] S3 Cost Map and Weight Overlay The power generation revenue-energy consumption coupling optimization module aligns the two topic grids and then calculates... Cost(x,y)=-ΔP(x,y)·A(x,y) / (Emove+Eclean), A(x,y) = 0.05 m × 0.05 m = 0.0025 m². To avoid A* getting trapped in low-yield regions, a cleaning priority weight w(x,y) = ΔP(x,y) / tclean is added, written with an additional negative cost -α·w(x,y), where α = 0.1. This was determined through 1000 Monte Carlo simulations and can improve convergence speed by 27% without sacrificing optimality. For grids with SR < 0.05 and ΔP < 0.5 W, NO_INFORMATION is used to reduce invalid node expansion by 18%. The updated / egr_costmap is published at a frequency of 15 Hz with a latency of < 3 ms.

[0041] S4 EGR-A* Online Replanning The dynamic path replanning controller operates EGR-A* with 8-connectivity and a diagonal cost of 1.414 using a local window of 5 m × 5 m (100 × 100 grid). Heuristic function... f(n)=g(n)+λ·1 / hEGR(n), g(n) is the cumulative EGR from the starting point to node n, hEGR(n) is the pre-stored upper bound graph, λ has an initial value of 1.0, and is corrected after each operation by step size of 0.05 according to the formula: Error = |Actual EGR - Estimated EGR| / Actual EGR, with a limit of [0.6, 1.4]. After four weeks of actual operation, λ converged to 0.78, and the EGR prediction error decreased from 7.8% to 2.9%. The average number of nodes for a single replanning is 12 k, and the time taken is 65 ms, which meets the 10 ms real-time margin. If a high-pollution area with SR>0.3 is newly detected, the target is immediately inserted and the search is restarted, with a recalculation delay of 72 ms, achieving "see and turn". Motion control adopts PID trajectory tracking, with a linear velocity of 0.5 m / s, an angular velocity limit of 0.8 rad / s, and a lateral error of <2 cm.

[0042] S5 adaptive charging The robot checks the battery SOC every 1 second, and calculates the value when SOC is ≤20%. Lmax=(SOC-SOCsafe)·Ebat / (Etotal / Lstep), SOCsafe = 8%, Ebat = 1.56 kWh, Lstep is the average step length traveled. If Lmax < 1.5·Lhome (Lhome is the straight-line distance to the nearest charging station), the list of uncleaned grid cells (approximately 1.2 kB) is saved to SQLite, a return-to-charge path is generated, and the cruise speed is 0.8 m / s to shorten the return journey. Automatic recovery occurs when SOC reaches ≥ 80%, resuming the EGR-A* search from the breakpoint to avoid duplicate overwriting. In actual testing, the daily average number of "invalid return trips" due to insufficient battery power decreased from 2.3 times to 0.4 times.

[0043] S6 Federated Learning Cloud Evolution Within 5 minutes of the task completion, the map management module will upload a JSON packet via MQTT. {job_id, EGR_real, ΣΔP, Etotal, λ_used, k1-k3, ηclean}, The packet size is <8 KB, the average latency on 4G is 180 ms, and the local cache is maintained for 48 hours without network access. FedAvg aggregates one week's worth of data from 100 robots in the cloud, with the target loss being |EGR_real - EGR_pred|. After training, differential packets (approximately 260 kB) are distributed, and hot-swap is performed after digital signature verification, requiring no system downtime. After four consecutive weeks, ηclean dry brushing improved from 0.78 to 0.81, wet washing improved from 0.92 to 0.94, k1-k3 fine-tuning was <3%, and λ improved from 1.00 to 0.78, completing the system-level self-evolutionary closed loop.

[0044] S7 Quantitative Effects and Economy 30-day control trial (same 0.8 MW subarray): Cleaning frequency: Traditional full coverage: 18 times → This system: 11 times (-38.9%). Total power consumption: 26.4 kWh → 15.1 kWh (-42.8%). Single-cycle work area with the same power consumption: +65%; Average daily power generation: 3.18 MWh → 3.34 MWh (+5.0%); An additional 58 MWh of electricity will be generated annually, which, at RMB 0.6 / kWh, amounts to RMB 348,000, with a static investment payback period of less than 1.1 years. Secondary cleaning rate: 4.7% (100% for the traditional fixed two-times strategy); EGR prediction error: 7.8% → 2.9%; Invalid return trips: -82%.

[0045] S8 Environmental Adaptability Temperature: After running continuously for 8 hours at -20℃ to 60℃, the average CPU temperature on the Jetson side was 68℃, and the GPU temperature was 65℃, with no frequency reduction. - Dust: With an IP65 protective housing and a fluorine-coated protective glass in front of the lens, the SR error is still within ±5% after 2 hours of continuous operation in a dust environment of 20 g / m³. Communication: MQTT retransmission rate <1% when 4G signal strength is -105 dBm, no packet loss in local cache after 48 hours of network outage.

[0046] S9 Scalability All algorithms and parameters in this embodiment are implemented based on the ROS 2 standard topic and Costmap2D plugin. The chassis, cleaning head, and camera models can be replaced. Only k1-k3 and ηclean need to be recalibrated. No modification to the upper-level logic is required. It has the ability to be directly ported to different power plants and different robot platforms.

[0047] The above solutions improve the cleanliness of photovoltaic power plants as follows: Quantitative analysis of the effects of the technology (30 days, comparison of 0.8 MW subarrays) Cleaning frequency: Traditional 18 times → This system 11 times (-38.9%); Total power consumption: 26.4 kWh → 15.1 kWh (-42.8%). Single-cycle work area with the same power consumption: +65%; Average daily power generation: 3.18 MWh → 3.34 MWh (+5.0%); An additional 58 MWh of electricity will be generated annually, equivalent to RMB 348,000 (RMB 0.6 / kWh), with a static payback period of less than 1.1 years; EGR prediction error: 7.8% → 2.9%; Invalid return trips: 2.3 times / day → 0.4 times / day (-82%).

[0048] The solution in this embodiment can achieve the following: ① Clean as needed: Prioritize cleaning areas that generate the most electricity per kilowatt-hour; ②Optimal energy consumption: Cleaning area increased by ≥65% with the same amount of electricity; ③ Power Adaptive: Even when SOC≤20%, it can still ensure that the high EGR region is completed first; ④ Cloud-based self-evolution: The model parameters are continuously updated through federated learning, becoming more accurate with use.

[0049] Example 2: An autonomous navigation and path planning system for a photovoltaic panel cleaning robot is provided to implement the steps of the method described in Example 1, including: The multimodal sensing module is used to simultaneously collect infrared thermal image data, RGB images and real-time irradiance on the surface of the photovoltaic panel, obtain the panel temperature based on the infrared thermal image data, and estimate the dirt shading rate of the photovoltaic panel. The power generation gain modeling module is used to calculate the power generation gain of each grid after cleaning based on the dirt shading rate, panel temperature and real-time irradiance of each grid of the photovoltaic panel. The energy consumption modeling module is used to calculate the robot's movement energy consumption and cleaning energy consumption on the target path, and output the total energy consumption. The power generation revenue-energy consumption coupling optimization module is used to construct a dynamic cost map that is updated in real time, with the power generation gain per unit energy consumption (EGR) as the path cost function. The dynamic route replanning controller uses EGR as the cost core on the grid map and performs online route search using the EGR-A* algorithm. When the remaining battery power is lower than a set safety threshold, it compares the expected EGR of the subsequent route with the historical average EGR. If the expected EGR is less than 80% of the historical average EGR, it triggers recharging route planning.

[0050] The overall system operation process is as follows: S1. The multimodal sensing module publishes SR(x,y), Tpanel, and Gpoa at a frequency of ≥15 Hz; S2. The power generation gain modeling module and the energy consumption modeling module calculate ΔP(x,y), Emove, and Eclean→Etotal in parallel. S3. The power generation revenue-energy consumption coupling optimization module writes EGR(x,y) into the cost map and superimposes the weight of w(x,y); S4. The dynamic path replanning controller runs EGR-A* to output the next target pose, and the PID tracks the motion; S5. After each grid is cleaned, the map management module refreshes the four layers, forming a single-cycle closed loop, which continues until the task ends or the battery is triggered to recharge.

[0051] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several modifications and improvements can be made without departing from the inventive concept of this application, and these all fall within the protection scope of this application.

Claims

1. An autonomous navigation and path planning method for a photovoltaic panel cleaning robot, characterized in that, Includes the following steps: S1: Collect infrared thermal image data, RGB images and real-time irradiance of the photovoltaic panel surface, obtain the panel temperature based on the infrared thermal image data, and estimate the dirt shading rate of the photovoltaic panel. S2: Calculate the power generation gain of the grid after cleaning based on the dirt shading rate of each grid of the photovoltaic panel, the panel temperature and real-time irradiance; S3: Calculate the robot's movement energy consumption and cleaning energy consumption on the target path, and calculate the total energy consumption; S4: Calculate the power generation gain per unit energy consumption (EGR) based on the power generation gain and total energy consumption, and construct a dynamic cost map that is updated in real time using EGR as the path cost function. S5: Using EGR as the cost core on the grid map, the EGR-A* algorithm is used for online path search. When the remaining battery power is lower than the set safety threshold, the expected EGR of the subsequent path is compared with the historical average EGR. If the expected EGR is less than 80% of the historical average EGR, the recharge path planning is triggered.

2. The autonomous navigation and path planning method for a photovoltaic panel cleaning robot according to claim 1, characterized in that, In step S1, the dirt shading rate SR(x,y) of each spatial grid of the photovoltaic panel is estimated by using a lightweight CNN segmentation network.

3. The autonomous navigation and path planning method for a photovoltaic panel cleaning robot according to claim 1, characterized in that, In step S2, the power generation gain ΔP(x,y) is calculated using the following formula: ΔP(x,y)=Pstd(x,y)·ηclean·SR(x,y) Where Pstd(x,y) is the standard power generation after temperature derating correction, ηclean is the cleaning efficiency coefficient corresponding to the cleaning method, and SR(x,y) is the dirt cover rate.

4. The autonomous navigation and path planning method for a photovoltaic panel cleaning robot according to claim 1, characterized in that, In step S3, the robot's movement energy consumption (Emove) on the target path is obtained through the following piecewise model: Emove=k1·d+k2·Σθ+k3·ΣΔh Where d is the path length, Σθ is the cumulative steering angle, ΣΔh is the cumulative climb height, and k1, k2, and k3 are coefficients obtained through actual vehicle calibration; Total energy consumption (Etotal) is the sum of mobile energy consumption (Emove) and clean energy consumption (Eclean). Etotal = Emove + Eclean.

5. The autonomous navigation and path planning method for a photovoltaic panel cleaning robot according to claim 1, characterized in that, In step S4, the power generation gain per unit energy consumption (EGR) is obtained by the following formula: EGR=Σ[ΔP(x,y)·A(x,y)] / Etotal.

6. The autonomous navigation and path planning method for a photovoltaic panel cleaning robot according to claim 1, characterized in that, In step S5, the heuristic function of the EGR-A* algorithm is: f(n) = g(n) + λ•1 / hEGR(n) Where g(n) is the actual cumulative EGR value from the starting point to the current node n, hEGR(n) is the estimated EGR gain from node n to the target region, and λ is the dynamic coefficient.

7. The autonomous navigation and path planning method for a photovoltaic panel cleaning robot according to claim 1, characterized in that, In step S5, when the cleaning robot's battery SOC is ≤20%, the maximum EGR path length Lmax that the remaining power can support is calculated. If Lmax is less than 1.5 times the path length from the current point to the nearest charging station, the robot immediately switches to recharge mode and records the coordinates of the interrupted node and the list of uncleaned grids. After charging is completed, the EGR-A* planning is restarted based on the latest dirt heat map.

8. An autonomous navigation and path planning system for a photovoltaic panel cleaning robot, characterized in that, include: The multimodal sensing module is used to simultaneously collect infrared thermal image data, RGB images and real-time irradiance on the surface of the photovoltaic panel, obtain the panel temperature based on the infrared thermal image data, and estimate the dirt shading rate of the photovoltaic panel. The power generation gain modeling module is used to calculate the power generation gain of each grid after cleaning based on the dirt shading rate, panel temperature and real-time irradiance of each grid of the photovoltaic panel. The energy consumption modeling module is used to calculate the robot's movement energy consumption and cleaning energy consumption on the target path, and output the total energy consumption. The power generation revenue-energy consumption coupling optimization module is used to construct a dynamic cost map that is updated in real time, with the power generation gain per unit energy consumption (EGR) as the path cost function. The dynamic route replanning controller uses EGR as the cost core on the grid map and performs online route search using the EGR-A* algorithm. When the remaining battery power is lower than a set safety threshold, it compares the expected EGR of the subsequent route with the historical average EGR. If the expected EGR is less than 80% of the historical average EGR, it triggers recharging route planning.

Images