An intelligent water cleaning method and device for water photovoltaic panels based on the Internet of Things

Abstract

The application provides a water-based photovoltaic panel intelligent water cleaning method and device based on the Internet of Things, relates to the field of photovoltaic panel cleaning, and solves the technical problem that the method of determining the cleaning frequency only by relying on single data is prone to false cleaning or missed cleaning. The method comprises the following steps: collecting a photovoltaic panel surface stain image and an ambient light intensity of a location of the photovoltaic panel in real time, and obtaining meteorological prediction data of a future period at the location of the photovoltaic panel. Anti-reflection interference processing is performed on the photovoltaic panel surface stain image to obtain a dynamic stain index, and cross-modal fusion features are generated by fusing the ambient light intensity of the location of the photovoltaic panel, the meteorological data and the dynamic stain index. The cross-modal fusion features are input into a multi-modal decision model to obtain a predicted optimal cleaning window and fed back to a cleaning control module. The cleaning control module triggers a high-pressure nozzle mechanism to perform directional cleaning when a time sequence reaches the predicted optimal cleaning window. The application is used in the photovoltaic panel cleaning process.

Description

Technical Field

[0001] This application relates to the field of photovoltaic panel cleaning, and in particular to an intelligent water cleaning method and device for floating photovoltaic panels based on the Internet of Things. Background Technology

[0002] In the operation and maintenance of floating photovoltaic power plants, the accumulation of dirt on the surface of photovoltaic panels can significantly reduce photoelectric conversion efficiency. Therefore, cleaning is a crucial step in ensuring the power generation performance of the power plant. However, existing photovoltaic panel cleaning technologies rely on single-dimensional data to trigger cleaning operations. For example, using only meteorological data (such as rainfall probability and irradiance) combined with simple image processing (such as grayscale thresholding) to determine the state of dirt can easily lead to image misjudgment due to the aquatic environment, making it impossible to accurately identify the true dirt coverage. This often results in "false cleaning" or "missed cleaning." Alternatively, relying on manual inspections or fixed cleaning schedules (such as weekly regular cleaning) is problematic. Manual inspections are affected by the harsh aquatic environment and limited inspection range, making it difficult to grasp the pollution status of large-area photovoltaic panels in real time. Fixed schedules ignore the dynamic changes in the rate of dirt accumulation, which can lead to waste of water and energy during periods of slow pollution accumulation and failure to clean in time during periods of rapid pollution accumulation, resulting in increased power generation losses. Summary of the Invention

[0003] This application provides an intelligent water cleaning method and device for floating photovoltaic panels based on the Internet of Things, which solves the technical problem that the existing technology relies on a single data point to determine the cleaning frequency, which is prone to incorrect cleaning or missed cleaning.

[0004] To achieve the above objectives, this application adopts the following technical solution: Firstly, an IoT-based intelligent water cleaning method for floating photovoltaic panels includes: real-time acquisition of images of surface contaminants on the photovoltaic panel and the ambient light intensity at the location of the photovoltaic panel using environmental sensors; and obtaining future weather forecast data for the location of the photovoltaic panel by calling a meteorological API. The surface contaminant images of the photovoltaic panel are subjected to anti-reflection interference processing to obtain a dynamic contaminant index, which is a score of contaminant coverage on the photovoltaic panel surface after anti-reflection interference processing. The ambient light intensity, meteorological data, and dynamic contaminant index at the location of the photovoltaic panel are fused to generate a cross-modal fusion feature. This cross-modal fusion feature is input into a multimodal decision model to obtain the predicted optimal cleaning window, which is then fed back to the cleaning control module. When the time series reaches the predicted optimal cleaning window, the cleaning control module triggers a high-pressure nozzle mechanism to perform directional cleaning and adjusts the parameters of the multimodal decision model based on the post-cleaning images.

[0005] Based on the above technical solution, in the IoT-based intelligent water cleaning method for floating photovoltaic panels provided in this application, environmental sensors are used to collect images of dirt and ambient light intensity in real time. The collected dirt images are then processed to prevent reflection interference and eliminate misjudgments caused by water surface reflection. After improving the accuracy of the overall raw data, the real-time pollution index is calculated. The pollution index is then fused with future meteorological data and ambient light from the meteorological API to generate cross-modal fusion features. Finally, the features are processed through the cross-modal interaction layer and decision verification layer of the multimodal decision model to obtain the optimal cleaning window. This allows for accurate prediction of continuous periods with low meteorological risk, suitable lighting, and dirt deterioration, thus avoiding ineffective cleaning.

[0006] In conjunction with the first aspect mentioned above, one possible implementation involves performing anti-reflection interference processing on a photovoltaic panel surface stain image to obtain a dynamic stain index. This includes: converting the photovoltaic panel surface stain image to the HSV color space, separating the luminance and saturation channels, and reading the luminance and saturation values. Based on image processing threshold requirements, calculating the current image's maximum luminance threshold and maximum saturation threshold, filtering areas with luminance values ​​greater than the current image's maximum luminance threshold and saturation values ​​greater than the current image's maximum saturation threshold, marking suspected reflection interference areas, and marking the remaining areas as non-interference areas. Extracting local texture features in the RGB space of the non-interference areas using the gray-level co-occurrence matrix and gray-level gradient magnitude. Converting the suspected reflection interference areas to the Lab color space, extracting the chromaticity shift, and fusing the reflection intensity difference features in the near-infrared band to obtain interference features. Inputting the local texture features of the non-interference areas and the interference features into a pre-trained lightweight CNN, outputting a normalized stain coverage score. Dynamically adjusting the score weights based on ambient light intensity to generate an anti-interference dynamic stain index.

[0007] In conjunction with the first aspect mentioned above, one possible implementation involves extracting local texture features in the RGB space of the non-interference region using the gray-level co-occurrence matrix and gray-level gradient magnitude. This includes: converting the RGB image of the non-interference region into a grayscale image, calculating its gray-level co-occurrence matrix, and extracting four statistical measures from the gray-level co-occurrence matrix: energy, contrast, entropy, and inverse moment, to generate a texture feature vector. The Sobel operator is used to calculate the gradient magnitude of the grayscale image, and the mean and variance of the gradient magnitude within a sliding window are calculated to generate a gradient feature vector. The texture feature vector and the gradient feature vector are concatenated along the feature dimension to generate a local texture feature vector.

[0008] In conjunction with the first aspect mentioned above, one possible implementation involves fusing ambient light intensity, meteorological data, and dynamic stain index at the location of the photovoltaic panel to generate cross-modal fusion features. This includes: spatiotemporally aligning ambient light intensity, meteorological forecast data, and dynamic stain index to generate a time-synchronized multidimensional feature tensor; extracting rainfall probability, wind speed, and irradiance variation trends from the meteorological data using a bidirectional LSTM network to generate a meteorological time-series feature vector; extracting the temporal evolution features of the dynamic stain index using a lightweight temporal convolutional network to generate a stain dynamic feature vector; and concatenating the meteorological time-series feature vector, ambient light intensity, and stain dynamic feature vector along the feature dimensions to generate the cross-modal fusion features.

[0009] In conjunction with the first aspect mentioned above, one possible implementation involves inputting cross-modal fusion features into a multimodal decision model to obtain the predicted optimal cleaning window and feeding it back to the cleaning control module. This includes: inputting cross-modal fusion features into the multimodal decision model, which comprises a cascaded cross-modal interaction layer and a decision verification layer. The cross-modal interaction layer employs a scaled dot product attention mechanism, using ambient light intensity as the query vector, stain dynamic features as the key vector, and meteorological time series features as the value vector, to calculate dynamic weights and generate a weighted decision vector. The decision verification layer includes a preliminary screening layer, a fine screening layer, and a feedback verification layer. The preliminary screening layer determines, based on the weighted decision vector, the period in which the probability of rainfall within the next 24 hours is lower than a preset threshold and the ambient light intensity is within the workable range, and this period is designated as the initial cleaning period. The fine screening layer selects consecutive periods in the initial cleaning period where the dynamic stain index exceeds the pollution threshold and continues to rise as candidate windows. The feedback verification layer sends the start time and duration of the candidate windows to the cleaning control module and adjusts the weight parameters of the cross-modal interaction layer based on the cleaned image.

[0010] In conjunction with the first aspect mentioned above, one possible implementation also includes an energy management module: monitoring the photovoltaic panel's power generation and the energy storage unit's power status. Before the cleaning control module triggers the high-pressure nozzle mechanism, it determines whether the current energy storage unit's power level is lower than a preset operating threshold. If it is lower than the operating threshold, it calls a meteorological API to obtain irradiance prediction data for a preset future period and outputs a predicted power generation curve based on a multimodal decision model. By integrating the predicted power generation curve with the dynamic stain index, the optimal cleaning window is recalculated and fed back to the cleaning control module, ensuring that the high-pressure nozzle mechanism performs targeted cleaning only when the photovoltaic panel's power generation meets the cleaning energy consumption requirements.

[0011] In conjunction with the first aspect described above, one possible implementation further includes a high-pressure cleaning module comprising a nozzle rotation mechanism, a water pressure adaptive controller, a cleaning path planning unit, and a feedback calibration interface: The nozzle rotation mechanism adjusts the pitch and horizontal rotation angles of the high-pressure nozzle based on the distribution of dirt on the photovoltaic panel surface. The water pressure adaptive controller adjusts the water pressure intensity based on a dynamic dirt index, employing a low-pressure mode when the dirt index is below a first threshold and a high-pressure mode when it is above a second threshold. The cleaning path planning unit generates a rasterized cleaning path based on the spatial distribution of the dirt image, controlling the nozzle to cover the dirt area in a serpentine trajectory. The feedback calibration interface transmits the deviation between the actual cleaning path and the planned path to the cleaning control module in real time, used to adjust the output parameters of the multimodal decision model.

[0012] In conjunction with the first aspect mentioned above, one possible implementation also includes a self-diagnostic module comprising a real-time monitoring unit, a communication link diagnostic unit, a self-recovery mechanism unit, and a fault tracing interface: The real-time monitoring unit periodically scans the feedback signals from the water flow sensor, pressure sensor, and stepper motor of the nozzle rotation mechanism. If the water flow pressure is detected to be lower than a preset threshold or the motor torque is abnormally high for an extended period, a fault code is generated and an alarm level classification is triggered. The communication link diagnostic unit uses the packet loss rate statistics of the IoT communication module's handshake protocol to determine the communication stability between the environmental sensor network and the control module in real time. When the packet loss rate exceeds 10% for five consecutive cycles, it is marked as a communication link failure. The self-recovery mechanism unit automatically executes a reverse flushing pulse sequence for nozzle stall faults. If the blockage is not resolved, a backup nozzle is switched and the cleaning path planning is updated. The fault tracing interface compresses and stores the fault code, occurrence timestamp, and abnormal data stream to the edge computing node, and pushes it to the user interface via a mobile device interface to generate a visual fault report.

[0013] In conjunction with the first aspect mentioned above, one possible implementation also includes an IoT communication module comprising a user interface, a mobile device interface, a data synchronization unit, and a security authentication mechanism: The user interface displays real-time photovoltaic panel dirt index, ambient light intensity, weather forecast data, cleaning plans, fault alarm information, and historical operation records, and supports the generation of visualized data charts. The mobile device interface receives user commands via a mobile application, remotely starts or pauses cleaning operations, views real-time monitoring video, and pushes alarm notifications to the mobile terminal. The data synchronization unit uploads environmental sensor data, decision model outputs, and fault codes to the cloud platform in real time and connects to external systems via an API interface. The security authentication mechanism includes user authentication and access control to ensure the security and compliance of operating instructions.

[0014] Secondly, an IoT-based intelligent water cleaning device for floating photovoltaic panels is provided, comprising: a communication unit and a processing unit. The communication unit is used to collect real-time images of surface contaminants on the photovoltaic panel and the ambient light intensity at the location of the photovoltaic panel via environmental sensors, and to obtain future weather forecast data for the location of the photovoltaic panel by calling a meteorological API. The processing unit is used to perform anti-reflection interference processing on the surface contaminant images of the photovoltaic panel to obtain a dynamic contaminant index, which is a score of contaminant coverage on the photovoltaic panel surface after anti-reflection interference processing. The ambient light intensity, meteorological data, and dynamic contaminant index at the location of the photovoltaic panel are fused to generate cross-modal fusion features. These cross-modal fusion features are input into a multimodal decision model to obtain the predicted optimal cleaning window, which is then fed back to the cleaning control module. When the time series reaches the predicted optimal cleaning window, the cleaning control module triggers a high-pressure nozzle mechanism to perform directional cleaning and adjusts the parameters of the multimodal decision model based on the cleaned images.

[0015] Thirdly, this application provides an IoT-based intelligent water cleaning device for floating photovoltaic panels, comprising: a processor and a storage medium; the storage medium includes instructions, and the processor is used to execute the instructions to implement the method described in the first aspect and any possible implementation thereof. This IoT-based intelligent water cleaning device for floating photovoltaic panels can be an electronic device or a chip within an electronic device.

[0016] Fourthly, this application provides a computer-readable storage medium storing instructions that, when executed on the IoT-based smart water cleaning device for floating photovoltaic panels, cause the IoT-based smart water cleaning device for floating photovoltaic panels to perform the methods described in the first aspect and any possible implementation thereof.

[0017] Fifthly, this application provides a computer program product containing instructions that, when the computer program product is run on the IoT-based smart water cleaning device for floating photovoltaic panels, causes the IoT-based smart water cleaning device for floating photovoltaic panels to perform the methods described in the first aspect and any possible implementation thereof.

[0018] This application provides an intelligent water cleaning method and device for floating photovoltaic panels based on the Internet of Things. It can collect images of dirt and ambient light intensity in real time through environmental sensors, and perform anti-reflection interference processing on the collected dirt images to eliminate misjudgments caused by water surface reflection. After improving the accuracy of the overall raw data, it calculates the real-time pollution index. The pollution index can then be fused with future meteorological data and ambient light from a meteorological API to generate cross-modal fusion features. Finally, it is processed through the cross-modal interaction layer and decision verification layer of a multimodal decision model to obtain the optimal cleaning window. This allows for accurate prediction of continuous periods with low meteorological risk, suitable lighting, and dirt deterioration, thus avoiding ineffective cleaning.

[0019] It should be understood that the descriptions of technical features, technical solutions, beneficial effects, or similar language in this application do not imply that all features and advantages can be achieved in any single embodiment. Rather, it is understood that the description of a feature or beneficial effect means that a specific technical feature, technical solution, or beneficial effect is included in at least one embodiment. Therefore, the descriptions of technical features, technical solutions, or beneficial effects in this specification do not necessarily refer to the same embodiment. Furthermore, the technical features, technical solutions, and beneficial effects described in this embodiment can be combined in any suitable manner. Those skilled in the art will understand that embodiments can be implemented without one or more specific technical features, technical solutions, or beneficial effects of a particular embodiment. In other embodiments, additional technical features and beneficial effects may be identified in specific embodiments that do not embody all embodiments. Attached Figure Description

[0020] Figure 1 A system architecture diagram of an IoT-based intelligent water cleaning system for floating photovoltaic panels is provided in this application embodiment; Figure 2 A schematic flowchart illustrating an intelligent water cleaning method for floating photovoltaic panels based on the Internet of Things, provided in an embodiment of this application; Figure 3 A schematic flowchart of another IoT-based intelligent water cleaning method for floating photovoltaic panels provided in this application embodiment; Figure 4 A schematic flowchart of another IoT-based intelligent water cleaning method for floating photovoltaic panels provided in this application embodiment; Figure 5 A schematic diagram of a smart water cleaning method for floating photovoltaic panels based on the Internet of Things provided in this application embodiment; Figure 6 A schematic diagram of a smart water cleaning method for floating photovoltaic panels based on the Internet of Things provided in this application embodiment; Figure 7A schematic diagram of a smart water cleaning method for floating photovoltaic panels based on the Internet of Things provided in this application embodiment; Figure 8 A schematic diagram of a smart water cleaning method for floating photovoltaic panels based on the Internet of Things provided in this application embodiment; Figure 9 A schematic diagram of the hardware structure of an IoT-based intelligent water cleaning device for floating photovoltaic panels is provided in this application embodiment. Detailed Implementation

[0021] In the description of this application, unless otherwise stated, " / " means "or," for example, A / B can mean A or B. The "and / or" in this document is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, and B alone. Furthermore, "at least one" means one or more, and "multiple" means two or more. The terms "first," "second," etc., do not limit the quantity or order of execution, and "first," "second," etc., do not necessarily imply differences.

[0022] It should be noted that, in this application, the terms "exemplary" or "for example" are used to indicate that something is being described as an example, illustration, or illustration. Any embodiment or design described as "exemplary" or "for example" in this application should not be construed as being more preferred or advantageous than other embodiments or design solutions. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0023] To address the shortcomings of existing technologies that rely solely on meteorological data and simple image processing to determine contamination status, which are prone to misjudgment during image acquisition due to water surface reflections, and whose ineffectiveness with simple image processing (such as grayscale thresholding) hinders accurate identification of true contamination coverage, leading to either incorrect cleaning (mistaking reflections for contamination) or missed cleaning (contamination being obscured by reflections), and the lack of effective utilization of crucial environmental data such as ambient light intensity further impacts the accuracy of contamination assessment. Regarding cleaning decision-making mechanisms, some technologies rely on manual inspections or fixed cleaning schedules. Manual inspections are limited by harsh aquatic environments and limited inspection ranges, making it impossible to comprehensively and in real-time assess the contamination status of large-area photovoltaic panels. Fixed cleaning schedules ignore the dynamic changes in contamination accumulation rates, resulting in wasted water and energy during periods of slow contamination and exacerbating power generation losses during periods of rapid contamination accumulation. Meanwhile, existing technologies do not take into account the energy self-sufficiency of floating photovoltaic power stations. The energy consumption of cleaning equipment often depends on external power supply or fixed energy storage. The cleaning plan is not dynamically adjusted in combination with the real-time power generation of photovoltaic panels and the energy storage status, which makes it easy for cleaning to be interrupted or energy storage to be exhausted when power generation is insufficient, affecting the overall operational stability of the power station. Meanwhile, existing systems have weak fault handling capabilities and lack real-time monitoring and self-recovery mechanisms for cleaning mechanisms (such as nozzle blockage and abnormal water pressure) and communication links (such as sensor and control module disconnection). After a fault occurs, manual intervention is required, which prolongs the fault handling time and increases operation and maintenance costs and power generation losses. This application provides an IoT-based intelligent water cleaning method for floating photovoltaic panels. This method collects images of dirt and ambient light intensity in real time through environmental sensors, performs anti-reflection interference processing on the collected dirt images, eliminates misjudgments of water surface reflection, and improves the accuracy of the overall raw data before calculating the real-time pollution index. The pollution index can then be fused with future meteorological data and ambient light from the meteorological API to generate cross-modal fusion features. Finally, the cross-modal interaction layer and decision verification layer of the multimodal decision model are used for processing to obtain the optimal cleaning window, which can accurately predict continuous periods of low meteorological risk, suitable lighting, and dirt deterioration, thus avoiding ineffective cleaning.

[0024] like Figure 1 As shown in the embodiment of this application, a smart water cleaning method for floating photovoltaic panels based on the Internet of Things is provided, including: Step 101: Real-time acquisition of images of surface contaminants on the photovoltaic panel and the ambient light intensity at the location of the photovoltaic panel using environmental sensors; obtaining future weather forecast data for the photovoltaic panel location using a meteorological API. Anti-reflection interference processing is applied to the surface contaminant images of the photovoltaic panel to obtain a dynamic contaminant index, which is a score of contaminant coverage on the photovoltaic panel surface after anti-reflection interference processing. The ambient light intensity, meteorological data, and dynamic contaminant index at the location of the photovoltaic panel are fused to generate cross-modal fusion features.

[0025] Among them, environmental sensors refer to optical and optoelectronic devices that collect real-time images of surface contaminants on photovoltaic panels and ambient light intensity. The surface contaminant images are optical data of the contamination status of the photovoltaic panels captured by visual sensors. Ambient light intensity refers to the real-time illumination at the location of the photovoltaic panels. The meteorological API is a meteorological data service accessed via a network interface, providing weather forecasts for future periods (such as rainfall probability and irradiance). The dynamic contamination index is a quantitative score (0-100 points) generated after anti-reflection interference processing of the contamination images, reflecting the surface coverage of pollutants. Cross-modal fusion features refer to the spatiotemporally aligned multidimensional feature tensors generated by fusing ambient light intensity, meteorological data, and the dynamic contamination index.

[0026] In some implementations, environmental sensors (such as industrial cameras and photosensors) first acquire real-time images of dirt on the photovoltaic panel surface and ambient light intensity data. Simultaneously, a meteorological API is called to obtain weather forecast data such as the probability of rainfall, wind speed, and irradiance for the future period at the photovoltaic panel location. Then, the dirt images undergo anti-reflection interference processing, using color space conversion and threshold segmentation to identify reflective and non-reflective areas. Local texture features are then extracted from the anti-reflection interference processed photovoltaic panel surface dirt images and input into a lightweight CNN model. Combined with dynamic weighting based on ambient light intensity, a normalized dynamic dirt index is output. Finally, the ambient light intensity, weather forecast data, and dynamic dirt index are spatiotemporally aligned to construct a time-synchronized multidimensional feature tensor. Cross-modal fusion features are generated through feature concatenation to combine ambient light, meteorological data, and the dirt index, thereby solving the problems of reflection interference, multi-source data fusion, and energy optimization in floating photovoltaic systems, providing accurate data support for subsequent cleaning decisions.

[0027] It should be noted that cross-modal fusion relies on the spatiotemporal alignment of multi-source data (such as unified timestamps and geographic coordinates) to ensure the synchronization of weather forecasts and real-time data collection. Furthermore, the scoring weights of the dynamic stain index need to be dynamically adjusted according to ambient light intensity (e.g., reducing the weight of reflection interference under strong light) to improve anti-interference capabilities.

[0028] Step 102: Input the cross-modal fusion features into the multimodal decision model to obtain the predicted optimal cleaning window and feed it back to the cleaning control module.

[0029] The multimodal decision model is an intelligent decision-making architecture composed of a cascaded cross-modal interaction layer and a decision verification layer. The cross-modal interaction layer uses a scaled dot product attention mechanism to dynamically weight features. The optimal cleaning window refers to the continuous period predicted by the system that has a low probability of rainfall, suitable light intensity, and continuous deterioration of stains. The cleaning control module is a hardware control unit that receives decision commands and triggers the cleaning execution mechanism.

[0030] In some implementations, cross-modal fusion features are input into the multimodal decision model. First, in the cross-modal interaction layer, ambient light intensity is used as the query vector, dynamic stain features as the key vector, and meteorological time-series features as the value vector. Dynamic weights are calculated using a scaling dot product attention mechanism to generate a weighted decision vector. Subsequently, the decision verification layer processes the data in three levels: the initial screening layer selects periods with a probability of rainfall below a threshold (e.g., 15%) within the next 24 hours and an ambient light intensity within the workable range of 500-100,000 lux based on the weighted decision vector, marking these as initial cleaning periods to mitigate meteorological risks; the fine screening layer selects consecutive periods in the initial period where the dynamic stain index exceeds the pollution threshold (e.g., 60 points) and continues to rise as candidate windows to focus on the stain deterioration trend; the feedback verification layer sends the start time and duration of the candidate windows to the cleaning control module and dynamically adjusts the weight parameters of the cross-modal interaction layer based on subsequent cleaned images.

[0031] It should be noted that the dynamic weighting mechanism of the cross-modal interaction layer can adaptively adjust the decision contribution of ambient light, dirt, and weather (such as reducing the weight of reflection interference under strong light conditions).

[0032] Step 103: When the time series reaches the predicted optimal cleaning window, the cleaning control module triggers the high-pressure nozzle mechanism to perform directional cleaning and adjusts the parameters of the multimodal decision model based on the post-cleaning image.

[0033] Here, time series refers to the time-series data stream arranged chronologically within the system. The optimal cleaning window is a continuous period of time predicted by a multimodal decision model, characterized by low weather risk, suitable ambient light, and continuous deterioration of contaminants. The post-cleaning image is a real-time surface condition image acquired by environmental sensors after cleaning is completed.

[0034] In some implementations, the cleaning control module monitors time-series data in real time. When the system time reaches the predicted optimal cleaning window start point, it immediately triggers the high-pressure nozzle mechanism to start. At this time, the high-pressure nozzle performs cleaning actions according to a preset directional cleaning mode (such as gridded path planning), precisely covering the stained area by adjusting the water pressure intensity (such as high-pressure and low-pressure modes) and the nozzle rotation angle. After cleaning, the environmental sensor collects a new image of the photovoltaic panel surface. The cleaning control module compares and analyzes the post-cleaning image with the pre-cleaning data, calculates the stain removal rate (such as the percentage reduction in the area covered by pollutants), and dynamically adjusts the weight parameters of the cross-modal interaction layer in the multimodal decision model (such as the weight ratio of ambient light intensity to stain index) based on this removal rate.

[0035] Based on the above technical solution, environmental sensors are used to collect images of stains and ambient light intensity in real time. The collected stain images are then processed to prevent reflection interference and eliminate misjudgments caused by water surface reflection. After improving the accuracy of the overall raw data, the real-time pollution index is calculated. The pollution index can then be fused with future meteorological data and ambient light from the meteorological API to generate cross-modal fusion features. Finally, these features are processed through the cross-modal interaction layer and decision verification layer of the multimodal decision model to obtain the optimal cleaning window. This allows for accurate prediction of continuous periods with low meteorological risk, suitable lighting, and stain deterioration, thus avoiding ineffective cleaning.

[0036] In one possible implementation of the embodiments of this application, combined with Figure 1 ,like Figure 2 As shown, the dynamic stain index can be obtained by performing anti-reflection interference processing on the stain image on the photovoltaic panel surface through the following steps 201 to 204, which are explained in detail below: Step 201: Convert the image of the stains on the photovoltaic panel surface to the HSV color space, separate the luminance channel and saturation channel, and read the luminance and saturation values. Calculate the maximum luminance threshold and maximum saturation threshold of the current image based on image processing threshold requirements. Filter out areas where the luminance value is greater than the maximum luminance threshold and the saturation value is greater than the maximum saturation threshold, mark suspected reflection interference areas, and mark the remaining areas as non-interference areas.

[0037] The HSV color space is a color model that includes three channels: Hue, Saturation, and Value. The Value channel represents the lightness or darkness of a pixel; the Saturation channel represents the vividness or saturation of a color. The Value is the pixel's numerical value in the Value channel. The Saturation value is the pixel's numerical value in the Saturation channel. Image processing threshold requirements refer to dynamically determining the critical values ​​for distinguishing interference based on image characteristics. The current maximum Value threshold for a bright image is the highest distinguishing point calculated based on the overall brightness distribution of the image. The current maximum Value threshold for a saturation image is the highest distinguishing point calculated based on the overall saturation distribution of the image. Suspected reflection interference areas are regions where both brightness and saturation exceed the threshold. Non-interference areas are regions where either brightness or saturation does not exceed the threshold.

[0038] In some implementations, the collected image of the photovoltaic panel surface stains is first converted from the original RGB color space to the HSV color space to separate the hue, saturation, and brightness components. Next, the brightness and saturation channels are extracted from the converted HSV image and stored independently as two matrices. Then, the brightness value of each pixel in the brightness channel and the saturation value in the saturation channel are read. Subsequently, based on image processing requirements, the maximum brightness threshold and maximum saturation threshold of the current image are dynamically calculated by analyzing the global brightness and saturation distribution. Then, pixel regions with brightness values ​​greater than the maximum brightness threshold and pixel regions with saturation values ​​greater than the maximum saturation threshold are selected. Finally, regions that simultaneously meet the brightness and saturation thresholds are marked as suspected reflection interference areas, while other regions with brightness or saturation not exceeding the thresholds are marked as non-interference areas.

[0039] Step 202: Extract local texture features in the RGB space of the non-interference area using the gray-level co-occurrence matrix and gray-level gradient magnitude. Convert the suspected reflection interference area to the Lab color space, extract the chromaticity shift, and fuse the reflection intensity difference features in the near-infrared band to obtain the interference features.

[0040] The Gray-Level Co-occurrence Matrix (GLCM) is a mathematical model that describes texture features by statistically analyzing the spatial relationships of gray-level values. Gray-level gradient magnitude represents the intensity of change in pixel gray-level values ​​in space. Non-interference areas refer to smudged areas not marked as reflection interference. The RGB color space is a color model containing three channels: red (R), green (G), and blue (B). Local texture features are quantitative indicators describing the structural regularity of local regions in an image. The Lab color space contains luminance (L) and chromaticity (a, b) components, used to simulate human color perception. Chromaticity offset is the numerical difference between the a (red-green axis) and b (yellow-blue axis) channels in the Lab space. The near-infrared band refers to the spectral range of 700nm-2500nm; the reflection intensity difference feature represents the difference in reflectivity between near-infrared and visible light bands. Interference features are fusion vectors that comprehensively characterize reflection interference.

[0041] In some implementations, the smudge image marked as a non-interference area is first converted from the RGB color space to a grayscale image. Then, the gray-level co-occurrence matrix (GLCM) of this grayscale image is calculated, and four statistical measures—energy, contrast, entropy, and inverse moment—are extracted to generate a texture feature vector. Simultaneously, the Sobel operator is used to calculate the gradient magnitude of the grayscale image, and the mean and variance of the gradient magnitude within a sliding window are statistically analyzed to generate a gradient feature vector. Then, the texture feature vector and the gradient feature vector are concatenated along the feature dimension to form a local texture feature vector for the non-interference area. On the other hand, for suspected reflection interference areas, they are converted to the Lab color space, and the values ​​of the a channel (red-green hue) and b channel (yellow-blue hue) are extracted as chromaticity shifts. Simultaneously, the difference between the reflection intensity in the near-infrared band and the reflection intensity in the visible light band of this area is fused to form a reflection intensity difference feature. Finally, the chromaticity shift and the reflection intensity difference feature are combined to generate interference features characterizing reflection interference.

[0042] It should be noted that by separating the two types of regions through physical characteristics, the non-interference area is captured using grayscale texture statistics (GLCM energy, entropy) and gradient changes (Sobel mean, variance) to capture the stain morphology; while the interference area is identified based on Lab chromaticity shift (such as high reflectivity causing anomalous a and b channels) and near-infrared reflectance differences (the reflectivity of water surface reflection drops sharply in the near-infrared band). This allows GLCM to describe structural regularity, gradient to quantify edge sharpness, Lab chromaticity to distinguish color distortion, and near-infrared differences to verify reflectance characteristics, thus enabling multi-dimensional features to complement each other and jointly improve anti-interference robustness.

[0043] Step 203: Input the local texture features and interference features of the non-interference area into the pre-trained lightweight CNN and output the normalized stain coverage score.

[0044] Here, pre-trained lightweight CNN refers to a small convolutional neural network pre-trained on a stain dataset. Normalized stain coverage score is a standardized (0-100) quantification of the degree of contaminant surface coverage.

[0045] In some implementations, the local texture feature vectors extracted from non-interference areas (including GLCM texture statistics and gradient magnitude statistics) are aligned with the interference feature vectors extracted from suspected reflection interference areas (including Lab chromaticity shift and near-infrared reflectance intensity differences) in terms of feature dimensions. Then, these two types of feature vectors are used as input data and fed into a pre-trained lightweight CNN model. At this point, the CNN model performs deep fusion and abstract representation of the features through multi-layer convolution and pooling operations. The original stain coverage prediction value is then output at the fully connected layer of the CNN model and normalized to a standardized score of 0-100 using the Sigmoid function.

[0046] It should be noted that texture features in the non-interference area (such as GLCM energy and gradient variance) describe the physical morphology of the stain, while chromaticity and near-infrared features in the interference area identify optical artifacts. The two complement each other and can eliminate misjudgment of water surface reflection. Meanwhile, lightweight CNNs ensure computational efficiency through pre-trained parameters (non-real-time training).

[0047] Step 204: Combine the ambient light intensity to dynamically adjust the scoring weights and generate an interference-resistant dynamic stain index.

[0048] In some implementations, the ambient light intensity sensor collects values ​​in real time; then, according to a preset weight mapping rule (e.g., weight coefficient = 0.6 when ambient light intensity > 80,000 lux, = 0.8 when 5,000-80,000 lux, = 1.0 when < 5,000 lux), the scoring weight coefficient under the current lighting conditions is dynamically calculated; then, this weight coefficient is multiplied by the normalized stain coverage score output by the pre-trained lightweight CNN; an interference-resistant dynamic stain index corrected for ambient light intensity is generated, with the formula: Dynamic stain index = original score × weight coefficient.

[0049] It should be noted that strong light intensifies water surface reflection, necessitating a reduction in scoring weight to avoid misjudgment; conversely, weak light reduces reflection interference, allowing for a high weight to ensure scoring reliability. Furthermore, the dynamic mapping rule is implemented through a piecewise function (not a fixed threshold), enabling the system to adapt to changes in day / night cycles and weather conditions. In addition, the final index incorporates physical environmental factors, overcoming the limitations of traditional static scoring.

[0050] Local texture features in the RGB space of the non-interference region are extracted using the gray-level co-occurrence matrix and gray-level gradient magnitude, including: The RGB image of the non-interference area is converted into a grayscale image, and its gray-level co-occurrence matrix is ​​calculated. Four types of statistics are extracted from the gray-level co-occurrence matrix: energy, contrast, entropy, and inverse difference moment, to generate texture feature vectors.

[0051] The Sobel operator is used to calculate the gradient magnitude of the grayscale image, and the mean and variance of the gradient magnitude within a window are statistically analyzed by a sliding window to generate a gradient feature vector.

[0052] The texture feature vector and the gradient feature vector are concatenated along the feature dimension to generate a local texture feature vector.

[0053] Based on the above technical solution, the reflection interference area and the non-interference area are segmented by HSV spatial dynamic thresholding. Complementary feature extraction paths with complementary physical properties are then used. This allows the interference area to quantify color distortion using Lab chromaticity shift and capture water surface reflectivity using near-infrared reflectivity differences. Meanwhile, the non-interference area is described by gray-level co-occurrence matrices (energy, entropy, etc.) to describe the stain structure and gradient magnitudes (mean, variance) to quantify edge sharpness. This accurately isolates the water surface reflectivity area (interference area) from the real stain area (non-interference area), physically isolating the features of the interference and non-interference areas, ensuring that stain detection is not affected by water surface reflection and avoiding false positives. Simultaneously, the texture features of the non-interference area and the optical features of the interference area are input into a pre-trained lightweight CNN to achieve deep cross-modal feature fusion. This generates a normalized stain score to eliminate scoring bias caused by regional differences. After providing a consistency quantification index, the stain score weights are dynamically adjusted based on ambient light intensity. This reduces the contribution of reflection false positives in strong light and retains real stain information in weak light, improving all-weather robustness and eliminating the impact of lighting changes on the detection results.

[0054] In one possible implementation of this application embodiment, the generation of cross-modal fusion features by integrating the ambient light intensity, meteorological data, and dynamic pollution index of the photovoltaic panel location can be achieved through the following steps 301 to 304, which are described in detail below: Step 301: Perform spatiotemporal alignment of ambient light intensity, meteorological forecast data and dynamic stain index to generate a time-synchronized multidimensional feature tensor.

[0055] Spatiotemporal alignment involves synchronizing ambient light intensity, weather forecast data, and dynamic pollution indices in both time and space to ensure all data have consistent timestamps and geographic coordinates. Temporal synchronization refers to aligning data points across time series to create a continuous and synchronized time series dataset. A multidimensional feature tensor is a data structure containing temporal, spatial, and feature dimensions, used to integrate aligned multi-source data.

[0056] In some implementations, the timestamps and geographic location information of ambient light intensity, meteorological forecast data, and dynamic stain index are unified to complete spatiotemporal alignment processing. This ensures that the ambient light intensity (collected in real time by a photosensitive sensor), meteorological forecast data (future rainfall probability, wind speed, and irradiance obtained from a meteorological API), and dynamic stain index (coverage score generated after stain image processing) remain consistent at the same time point (e.g., every minute or hour) and the same photovoltaic panel location. Then, the ambient light intensity value, meteorological data value (including rainfall probability, wind speed, and irradiance), and dynamic stain index value at each time point are integrated into a three-dimensional data structure to generate a time-synchronized multidimensional feature tensor. The time dimension represents a continuous time series, the spatial dimension corresponds to the fixed location of the photovoltaic panel, and the feature dimension covers specific values ​​such as illumination, meteorology, and stain score. Finally, a synchronized multidimensional feature tensor is formed for subsequent cross-modal fusion.

[0057] Step 302: Extract the trends of rainfall probability, wind speed and irradiance from meteorological data using a bidirectional LSTM network to generate meteorological time series feature vectors.

[0058] Among them, the bidirectional LSTM network is a time-series modeling neural network containing forward and backward recurrent units, capturing long-term dependencies through bidirectional information flow. The changing trend refers to the directional characteristics of meteorological parameters fluctuating over time, such as the upward / downward trend of rainfall probability, the fluctuation amplitude of wind speed, and the periodic changes in irradiance. The meteorological time-series feature vector is a low-dimensional dense vector containing the above trend information, used to characterize the dynamic evolution of meteorological data.

[0059] In some implementations, spatiotemporally aligned meteorological data (including rainfall probability, wind speed, and irradiance) are input into a bidirectional LSTM network in time series. The forward LSTM layer processes the data in forward time order to learn the dependencies from history to the present, while the backward LSTM layer processes the data in reverse time order to capture potential impacts from the future to the present. Then, the bidirectional LSTM network uses gating mechanisms (input gate, forget gate, and output gate) to filter and memorize key meteorological features step by step (e.g., the forget gate suppresses non-periodic wind speed fluctuations), the input gate strengthens the abrupt change signal of rainfall probability, and the output gate controls the transmission intensity of the diurnal trend of irradiance. Next, the bidirectional LSTM network fuses bidirectional information in the hidden state layer to extract the cumulative change of rainfall probability, the fluctuation frequency of wind speed, and the periodic rise and fall patterns of irradiance. Finally, the bidirectional hidden states of each time step are concatenated to generate a fixed-length meteorological time series feature vector, which compresses and stores the multidimensional dynamic trend information of meteorological parameters.

[0060] It should be noted that bidirectional LSTM can simultaneously learn the positive evolution of meteorological parameters (such as the increase of irradiance with sunrise) and the negative correlation (such as the impact of sudden changes in rainfall probability on preceding wind speed), overcoming the temporal limitations of unidirectional networks. The hidden state gating mechanism can adaptively distinguish between noise fluctuations and true trends (such as filtering instantaneous wind speed interference), but requires presetting network hyperparameters (such as a hidden layer dimension of 64); the output feature vector needs to be aligned with the feature dimensions of other modalities (such as 128 dimensions) to avoid affecting the cross-modal fusion effect.

[0061] Step 303: Extract the temporal evolution features of the dynamic stain index through a lightweight temporal convolutional network to generate a dynamic stain feature vector.

[0062] Lightweight temporal convolutional networks are causal convolutional neural networks with few parameters and high computational efficiency, specifically designed for temporal data. Temporal evolution characteristics refer to the dynamic patterns of changes in the dynamic stain index over time, such as short-term accumulation, abrupt changes, or periodic fluctuations. The stain dynamic feature vector is a low-dimensional dense vector characterizing these evolutionary patterns, used to compress and store the temporal patterns of stain changes.

[0063] In some implementations, the spatiotemporally aligned dynamic stain index time-series data is input into a lightweight temporal convolutional network. This network consists of multiple layers of causal convolutions. Each layer scans the input sequence in chronological order using a one-dimensional convolutional kernel (e.g., kernel size = 3) to ensure that the output depends only on the current and historical data. Subsequently, the network uses dilated convolutions to expand the receptive field (e.g., dilation coefficients [1, 2, 4]) to enable high-level features to capture long-term evolution patterns (e.g., the 24-hour cumulative trend of the stain index). Then, key features are filtered through gated activation units (e.g., GLU gates) to suppress noise fluctuations (e.g., score jumps caused by brief glare interference). Finally, a 1×1 convolution is applied to the output layer to compress the channel dimension, generating a fixed-length dynamic stain feature vector (e.g., 128-dimensional).

[0064] Step 304: Concatenate the meteorological time series feature vector, ambient light intensity, and stain dynamic feature vector along the feature dimension to generate cross-modal fusion features.

[0065] In some implementations, ambient light intensity is expanded into a new vector of the same dimension as the feature vector through a one-dimensional linear transformation (e.g., using a fully connected layer to map a scalar to a 128-dimensional vector). This allows for dimension alignment preprocessing of the meteorological time-series feature vector (e.g., a 128-dimensional vector), ambient light intensity (single numerical value), and stain dynamic feature vector (e.g., a 128-dimensional vector). The 128 feature values ​​of the meteorological time-series feature vector, the expanded ambient light intensity vector, and the stain dynamic feature vector are then sequentially concatenated to form a 384-dimensional cross-modal fusion feature vector, completing the horizontal concatenation operation along the feature dimension direction of the three vectors.

[0066] Based on the above technical solution, a spatiotemporal alignment engine is constructed by unifying timestamps and geographic coordinates. This engine integrates ambient light intensity (real-time illumination), meteorological forecast data (rainfall probability / wind speed / irradiance), and dynamic stain index (stain score after anti-reflection interference) into a time-synchronized three-dimensional tensor (time × space × feature dimension). This solves the spatiotemporal consistency problem of multi-source heterogeneous data, forming a structured input basis and avoiding data misalignment caused by sensor acquisition delays and differences in meteorological API update frequencies (such as mismatch between light intensity data and rainfall forecast time). Simultaneously, a bidirectional LSTM network, combined with forward and backward gating mechanisms (forget gate to suppress noise, input gate to strengthen abrupt signal change), extracts the evolutionary patterns of cumulative rainfall probability, wind speed fluctuation frequency, and periodic irradiance changes from meteorological data, outputting a 128-dimensional dense vector. This allows for the simultaneous capture of the positive evolution of meteorological parameters (such as daily irradiance increase) and negative correlations (such as the impact of rainfall on previous wind speed), improving the robustness of trend prediction. In addition, a lightweight temporal convolutional network (≤4 layers, ≤64 channels) uses causal convolution + dilated convolution (dilation coefficients [1,2,4]) to scan the stain index sequence. GLU gating filters out instantaneous reflective interference, generating a 128-dimensional vector encoding the stain accumulation rate and abrupt change features. This adapts to the resource limitations of edge computing devices, enabling real-time extraction of long-term stain evolution patterns (e.g., 24-hour accumulation). Finally, meteorological temporal features (128-dimensional), ambient light intensity (expanded to 128-dimensional), and stain dynamic features (128-dimensional) are horizontally concatenated along the feature dimensions to generate a 384-dimensional unified vector. This preserves the original modal semantic independence, avoids information loss from feature weighted fusion, and provides an uncompressed cross-modal interaction foundation for multimodal decision models.

[0067] In one possible implementation of this application embodiment, inputting cross-modal fusion features into a multimodal decision model to obtain the predicted optimal cleaning window and feeding it back to the cleaning control module can be achieved through the following steps 401 to 404, which are described in detail below: Step 401: Input the cross-modal fusion features into the multimodal decision model. The multimodal decision model includes a cascaded cross-modal interaction layer and a decision verification layer. The cross-modal interaction layer adopts a scaled dot product attention mechanism, using ambient light intensity as the query vector, stain dynamic features as the key vector, and meteorological time series features as the value vector, to calculate dynamic weights and generate a weighted decision vector.

[0068] The cross-modal interaction layer is the first-level structure of the multimodal decision model, employing a scaled dot product attention mechanism to dynamically fuse multi-source features. The decision validation layer comprises cascaded initial screening, fine screening, and feedback validation layers, optimizing the cleaning window decision through a hierarchical screening mechanism. The query vector, composed of ambient light intensity data, serves as the retrieval benchmark for the attention mechanism to evaluate lighting constraints. The key vector originates from the dynamic features of the stains, characterizing the temporal evolution of pollutant accumulation. The value vector, based on meteorological time-series features, encodes meteorological risk parameters such as rainfall probability and wind speed. The weighted decision vector, the dynamically weighted fusion result output from the cross-modal interaction layer, drives subsequent decision-making processes.

[0069] In some implementations, cross-modal fusion features are input into a cascaded multimodal decision model. At this point, the cross-modal fusion features enter the first-level cross-modal interaction layer of the multimodal decision model. The cross-modal interaction layer uses a scaled dot product attention mechanism for calculation: using ambient light intensity as the query vector, stain dynamic features as the key vector, and meteorological time series features as the value vector, the similarity between the query vector and the key vector is calculated through dot product operation, and then normalized by a scaling factor to generate an attention weight distribution; subsequently, the weight distribution is applied to the value vector for weighted summation to generate a dynamically weighted decision vector.

[0070] It should be noted that the dynamic weight allocation in the scaling dot product attention mechanism will adaptively adjust the contribution ratio of ambient light, dirt, and weather conditions. For example, it will automatically reduce the weight ratio of reflection interference in strong light environments.

[0071] Step 402: The decision verification layer includes a preliminary screening layer, a fine screening layer, and a feedback verification layer. The preliminary screening layer determines, based on the weighted decision vector, the period in which the probability of rainfall in the next 24 hours is lower than a preset threshold and the ambient light intensity is within the workable range, and this period is recorded as the initial cleaning period.

[0072] The decision validation layer is a cascaded structure in the multimodal decision model used to optimize the cleaning window prediction. The initial screening layer uses weighted decision vectors to select periods with low meteorological risk and suitable lighting conditions, with a preset rainfall probability threshold typically ≤15%. The operable range refers to the ambient light intensity range of 500 to 100,000 lux. The initial cleaning period is the set of candidate periods output by the initial screening layer that meet the meteorological and lighting constraints. The fine screening layer selects consecutive periods from the initial period where the stains have worsened and exceeded the threshold.

[0073] In some implementations, after receiving the weighted decision vector generated by the cross-modal interaction layer, the decision verification layer first uses the initial screening layer to determine whether the probability of rainfall in the next 24 hours is lower than a preset threshold (e.g., ≤15%) based on the meteorological time-series feature components in the weighted decision vector, and at the same time verifies whether the ambient light intensity is within the workable range of 500-100,000 lux; the time period that simultaneously meets the requirements of low rainfall risk and suitable light is marked as the initial cleaning period.

[0074] Step 403: In the fine screening stage, the continuous period when the dynamic stain index exceeds the contamination threshold and continues to rise during the initial cleaning period is selected as the candidate window.

[0075] The fine screening stage layer is the module responsible for refined screening within the decision verification layer. The contamination threshold is a preset stain index critical value (e.g., ≥60 points); exceeding this value indicates severe contamination requiring cleaning. "Continuously rising" refers to the stain index monotonically increasing over time. "Continuous period" refers to a seamlessly connected time interval. The candidate window represents the period with a high risk of contamination deterioration, output by the fine screening layer.

[0076] In some implementations, the fine screening stage layer receives the initial cleaning period set output by the primary screening layer, then extracts the time series data of the dynamic stain index within the initial cleaning period set and analyzes its changing trend; next, the fine screening layer determines whether the stain index exceeds the preset pollution threshold (e.g., ≥60 points) and verifies whether the index is in a continuous upward state (e.g., monotonically increasing for three consecutive time periods); finally, the fine screening layer selects the continuous period that simultaneously meets the threshold and the continuous upward condition, marks it as a candidate window, and passes the start time and duration to the subsequent feedback verification layer.

[0077] It should be noted that the pollution threshold is set according to the actual pollution tolerance (typically 60 points); the judgment of a continuous increase depends on the strict monotonicity of the stain index (to avoid fluctuation interference); Cui Hou, the continuity of the candidate window requires no interruption within the time period to ensure the continuous and efficient cleaning operation.

[0078] Step 404: The feedback verification layer sends the start time and duration of the candidate window to the cleaning control module, and adjusts the weight parameters of the cross-modal interaction layer based on the cleaned image.

[0079] The feedback validation layer is the terminal module of the decision validation layer, responsible for outputting the final decision instruction and optimizing the model parameters. The weight parameters of the cross-modal interaction layer refer to the dynamic weight coefficients of the query, key, and value vectors in the scaling dot product attention mechanism.

[0080] In some implementations, the feedback verification layer first sends the start time and duration of the candidate window output by the fine screening layer to the cleaning control module. At this time, the cleaning control module will trigger it to perform targeted cleaning during the specified time period. After the cleaning is completed, the environmental sensor collects a new image of the photovoltaic panel surface and transmits it to the feedback verification layer. Then, the feedback verification layer calculates the stain coverage score of the image after cleaning and compares it with the dynamic stain index before cleaning to generate the stain removal rate (such as the proportion of reduction in the area of ​​pollutant removal). Finally, based on the removal rate deviation (the difference between the target removal rate and the actual removal rate), the gradient descent method is used to dynamically adjust the weight distribution ratio of the three factors in the cross-modal interaction layer: ambient light intensity (query vector), stain dynamic features (key vector), and meteorological time series features (value vector), to optimize the accuracy of subsequent decisions.

[0081] Based on the above technical solution, weights can be dynamically allocated by using ambient light intensity as the query vector, stain dynamic features as the key vector, and meteorological time series features as the value vector. This integrates three modal features: ambient light, stain status, and meteorological risk, solving the problem of data isolation. Simultaneously, the three-level cascaded screening decision verification layer effectively avoids meteorological risks through the initial screening layer, then focuses on pollution deterioration through the fine screening layer, and finally performs closed-loop optimization through the feedback verification layer. The attention weight allocation is adaptively optimized based on cleaning effect feedback (such as cleaning rate deviation), improving the accuracy of window prediction.

[0082] In one possible implementation of this application embodiment, the energy management module can be implemented through the following steps 501 to 502, which are described in detail below: Step 501: Monitor the photovoltaic panel's power generation and the energy storage unit's power status. Before the cleaning control module triggers the high-pressure nozzle mechanism, determine whether the current energy storage unit's power is lower than the preset operating threshold. If it is lower than the operating threshold, call the meteorological API to obtain the irradiance prediction data for the preset future period, and output the predicted power generation curve based on the multimodal decision model.

[0083] Among them, photovoltaic panel power generation refers to the output power value of photovoltaic panels monitored in real time by environmental sensors; energy storage unit power status refers to the real-time remaining capacity percentage of the built-in energy storage equipment in the system; preset operation threshold refers to the minimum energy storage safety value set by the system to trigger the cleaning operation; irradiance prediction data refers to the forecast of solar radiation intensity for future periods obtained through API; and predicted power generation curve refers to the time series distribution map of future power generation of photovoltaic panels generated by the multimodal decision model based on irradiance data.

[0084] In some implementations, the energy management module first collects the photovoltaic panel's power generation in real time using a current sensor, while simultaneously monitoring the state of charge of the energy storage unit. Just before the cleaning control module triggers the high-pressure nozzle mechanism, it automatically compares the current energy storage capacity with a preset operating threshold (e.g., 30% capacity). If the capacity is detected to be below the threshold, it immediately calls the meteorological API to obtain irradiance prediction data for a preset time period (e.g., the next 6 hours). Then, the irradiance data is input into a multimodal decision model, which, combined with historical photovoltaic panel conversion efficiency data, outputs a predicted power generation curve with minute-level accuracy. Finally, the predicted power generation curve is fused with the real-time dynamic stain index to recalculate the optimal cleaning window that meets the cleaning energy consumption requirements, and the corrected execution command is fed back to the cleaning control module.

[0085] It should be noted that the prediction accuracy of the multimodal decision model must ensure spatiotemporal alignment between meteorological API data and local irradiance sensors. Furthermore, during comparison, the preset operating threshold needs to be dynamically adjusted based on the water pressure power of the high-pressure nozzle mechanism; for example, the threshold can be lowered in low-pressure mode.

[0086] Step 502: Integrate the predicted power generation curve and dynamic stain index, recalculate the optimal cleaning window and feed it back to the cleaning control module, so that the high-pressure nozzle mechanism performs targeted cleaning only when the photovoltaic power generation meets the cleaning energy consumption requirements.

[0087] Among them, the high-pressure nozzle mechanism refers to the actuator that performs targeted cleaning based on the distribution of dirt; the cleaning energy consumption requirement refers to the minimum energy consumption required for the high-pressure nozzle mechanism to perform the cleaning operation.

[0088] In some implementations, the predicted power generation curve output by the multimodal decision model is spatiotemporally aligned and fused with the real-time dynamic stain index to generate a fused feature vector. Then, by verifying whether the cumulative power generation in the predicted power generation curve is higher than the preset cleaning energy consumption demand threshold, and combining this with whether the dynamic stain index exceeds the pollution threshold and continues to rise, a continuous period of time that meets the energy consumption and stain conditions is selected and recorded as the recalculated optimal cleaning window. Next, the start time and duration of the recalculated optimal cleaning window are fed back to the cleaning control module. The cleaning control module triggers the high-pressure nozzle mechanism to perform directional cleaning only when the time series reaches this window.

[0089] Based on the above technical solution, the irradiance prediction data from the meteorological API is converted into a power generation curve using a multimodal decision model. This curve is then aligned with the dynamic stain index, which is processed to resist reflection interference. This allows for the verification of the dual conditions of "cumulative power generation > preset energy consumption" and "stain index > pollution threshold" based on fused features. This enables the cleaning control module to directly embed energy consumption verification logic, physically blocking cleaning commands that do not meet the conditions. This improves the energy self-sufficiency rate of the cleaning operation, ensuring that cleaning energy consumption comes from concurrent photovoltaic power generation while reducing the frequency of ineffective cleaning. Furthermore, it reduces power generation losses due to pollution.

[0090] In one possible implementation of this application embodiment, a high-pressure cleaning module including a nozzle rotation mechanism, a water pressure adaptive controller, a cleaning path planning unit, and a feedback calibration interface is further included. This module can be implemented through the following steps 601 to 604, which are described in detail below: Step 601: The nozzle rotation mechanism is used to adjust the pitch angle and horizontal rotation angle of the high-pressure nozzle according to the distribution of stains on the photovoltaic panel surface.

[0091] Among them, the nozzle rotation mechanism is a mechanical device used to adjust the angle of the high-pressure nozzle to cover the stained area; the pitch angle refers to the vertical tilt angle of the nozzle in the vertical plane, controlling the change of the water spray height; the horizontal rotation angle refers to the horizontal rotation angle of the nozzle in the horizontal plane, achieving lateral coverage; the stain distribution location refers to the specific spatial coordinates of the contaminants on the photovoltaic panel surface, determined based on image analysis.

[0092] In some implementations, the cleaning path planning unit first generates a rasterized cleaning path based on the spatial distribution of the stain image. The path adopts a serpentine trajectory design to ensure full coverage of the stain area. Then, the nozzle rotation mechanism automatically adjusts the pitch angle (vertical tilt angle) and horizontal rotation angle (left and right turn angle) of the high-pressure nozzle according to the planned path, for example, by using a stepper motor drive mechanism to perform precise angle changes. Then, the nozzle is controlled to move along the generated serpentine trajectory to achieve directional cleaning.

[0093] Step 602: Water pressure adaptive controller adjusts water pressure intensity according to dynamic stain index. When the stain index is lower than the first threshold, low pressure mode is used, and when it is higher than the second threshold, high pressure mode is activated.

[0094] The water pressure adaptive controller is a device that automatically adjusts the water spray pressure. Water pressure refers to the pressure applied during spraying. The first threshold is a preset low threshold for the stain index. The second threshold is a preset high threshold for the stain index. Low-pressure mode uses lower water pressure for cleaning. High-pressure mode uses higher water pressure for cleaning.

[0095] In some implementations, the dynamic stain index is first acquired in real time as an input parameter. Then, the controller compares the index with a preset first threshold and a second threshold. Based on the comparison result, the controller automatically switches modes. When the dynamic stain index is lower than the first threshold, the controller starts a low-pressure mode for energy saving and gentle cleaning. When the dynamic stain index is higher than the second threshold, the controller switches to a high-pressure mode to enhance the cleaning power. Finally, the controller adjusts the water pressure output to the nozzle mechanism in real time to ensure that the water pressure matches the selected mode and completes adaptive control.

[0096] Step 603: The cleaning path planning unit generates a rasterized cleaning path based on the spatial distribution of the stain image and controls the nozzle to cover the stain area in a serpentine trajectory.

[0097] In some implementations, the specific coordinates of contaminants on the photovoltaic panel are analyzed based on the spatial distribution data of the stain images. Then, the surface of the photovoltaic panel is divided into uniform grid cells, and a nozzle movement sequence is planned to generate a gridded cleaning path. The nozzle is then controlled to move strictly according to a serpentine trajectory, scanning each grid cell from the starting point along a continuous S-shaped path, ultimately covering all stain areas and ensuring that the cleaning process is efficient and thorough.

[0098] It should be noted that the grid size must be proportionally adapted to the size of the photovoltaic panel.

[0099] Step 604: Feedback calibration interface, which transmits the deviation between the actual cleaning path and the planned path to the cleaning control module in real time, to adjust the output parameters of the multimodal decision model.

[0100] The feedback calibration interface serves as the data channel connecting the cleaning execution and decision-making modules, used to transmit path deviation signals. The actual cleaning path refers to the real-time movement trajectory coordinate data recorded during the execution of the high-pressure nozzle mechanism. The planned path is a pre-generated rasterized serpentine trajectory coordinate sequence from the cleaning path planning unit. Deviation refers to the spatial difference between the actual trajectory coordinates and the planned trajectory coordinates.

[0101] In some implementations, the actual movement trajectory coordinates of the high-pressure nozzle mechanism are first collected in real time and compared point by point with the gridded serpentine trajectory generated by the cleaning path planning unit. Then, the spatial position deviation value (such as Euclidean distance) between the actual trajectory and the planned trajectory is calculated. The deviation data is then transmitted to the cleaning control module in real time through the Internet of Things communication unit. Finally, the cleaning control module dynamically adjusts the output parameters of the decision verification layer in the multimodal decision model (such as the pollution threshold weight of the fine screening layer) based on the deviation data to optimize the generation logic of subsequent cleaning path planning.

[0102] Based on the above technical solution, the water pressure adaptive controller receives the dynamic stain index in real time, compares it with a preset first threshold (low-pressure mode threshold) and a second threshold (high-pressure mode threshold), and automatically switches the water pressure output through a solenoid valve or pressure regulator. This allows for energy and water conservation by activating a low-pressure mode (e.g., 5MPa) when the stain index is below the first threshold, and a high-pressure mode (e.g., 15MPa) for powerful cleaning when the index is above the second threshold, ensuring a high stain removal rate while reducing water consumption. Simultaneously, the cleaning path planning unit analyzes the stain image to obtain spatial distribution coordinates, divides the photovoltaic panel surface into uniform grid units, generates a rasterized serpentine trajectory path (e.g., a continuous S-shaped movement sequence), and controls the nozzles to scan along this trajectory. This serpentine trajectory optimizes the movement path, reduces nozzle idle time and reciprocation, shortens cleaning time, and ensures that all stained areas are systematically covered. Finally, the actual cleaning path coordinates of the high-pressure nozzles are collected in real time and compared with the serpentine trajectory coordinates of the planned path point by point to calculate the deviation (such as Euclidean distance). The deviation is then transmitted to the cleaning control module through the Internet of Things communication unit to dynamically adjust the output parameters of the multimodal decision model. This allows the multimodal decision model to adaptively optimize and improve the accuracy of subsequent cleaning paths, forming a continuous learning mechanism to enhance the robustness of the system.

[0103] In one possible implementation of this application embodiment, a self-diagnosis module comprising a real-time monitoring unit, a communication link diagnosis unit, a self-recovery mechanism unit, and a fault tracing interface is further included. This module can be implemented through steps 701 to 704, which are described in detail below: Step 701: The real-time monitoring unit periodically scans the water flow sensor, pressure sensor, and stepper motor feedback signals of the nozzle rotation mechanism. If the water flow pressure is detected to be lower than the preset threshold or the abnormal motor torque is detected to last for more than the set time, a fault code is generated and an alarm level classification is triggered.

[0104] In this context, a real-time monitoring unit refers to equipment responsible for continuously monitoring key components of the system. Periodic scanning refers to the process of checking sensor data at fixed time intervals. A water flow sensor is a device that detects water flow parameters. A pressure sensor is a device that measures water pressure. A stepper motor feedback signal refers to the operating status feedback information provided by the stepper motor. Water flow pressure refers to the pressure value of flowing water. A preset threshold refers to a pre-set minimum pressure critical value. Abnormal motor torque refers to a state where the stepper motor torque deviates from the normal range. A set duration refers to a pre-defined limit on the duration of the abnormality. A fault code refers to a coded sequence that identifies a specific fault type. Alarm hierarchy classification refers to a mechanism for classifying alarms according to the severity of the fault.

[0105] In some implementations, the real-time monitoring unit periodically scans the water flow sensor of the high-pressure nozzle mechanism to obtain real-time water flow data, while simultaneously scanning the pressure sensor to monitor the water pressure value and scanning the feedback signal of the stepper motor of the nozzle rotation mechanism to check the motor's operating status. If the water pressure is detected to be lower than a preset threshold, or if the abnormal motor torque persists for more than a set time, a corresponding fault code is generated. Then, an alarm hierarchy classification mechanism is triggered to classify alarm levels according to the type and duration of the anomaly. For example, abnormal water pressure is classified as a medium alarm, while abnormal motor torque that persists for more than a set time is classified as a high alarm, ensuring timely and efficient fault response.

[0106] It should be noted that the interval of periodic scanning needs to be optimized according to the system load to avoid excessively high frequency leading to resource waste or excessively low frequency causing missed anomalies; preset thresholds and set durations should be dynamically adjusted according to equipment specifications, for example, water flow pressure thresholds need to take into account the differences in nozzle working modes; fault code generation needs to ensure uniqueness to facilitate subsequent traceability; alarm level classification needs to be mapped to actual operation and maintenance priorities, such as high-level alarms triggering immediate notifications, and medium-level alarms recording logs.

[0107] Step 702: The communication link diagnosis unit uses the packet loss rate statistics of the handshake protocol of the IoT communication module to determine the communication stability between the environmental sensor network and the control module in real time. When the packet loss rate is higher than 10% for 5 consecutive cycles, it is marked as a communication link failure.

[0108] The communication link diagnostic unit refers to an independent functional module responsible for assessing the communication connection status. The IoT communication module refers to the hardware unit that implements remote data transmission. The handshake protocol refers to the set of rules for signal exchange between communicating parties when establishing a connection. Packet loss rate statistics refer to the calculation process of the percentage of data packets lost during data transmission. An environmental sensor network refers to a network composed of multiple sensors monitoring environmental parameters. The control module refers to the unit responsible for receiving instructions and executing control. Communication stability refers to the ability of the communication link to maintain reliable transmission. Packet loss rate refers to the proportion of data packets lost during data transmission out of the total number of packets sent. Five consecutive cycles refer to five uninterrupted detection time intervals. Above 10% refers to a packet loss rate exceeding the critical value of 10%. Communication link failure refers to a state where the communication connection is completely interrupted or unavailable.

[0109] In some implementations, the communication link diagnostic unit first sends test data packets periodically through the handshake protocol of the IoT communication module and counts the loss rate of response data packets; then, based on the packet loss rate statistics, it determines the communication stability between the environmental sensor network and the control module in real time; when the packet loss rate is detected to be higher than 10% for 5 consecutive cycles, it is determined that the communication quality has seriously deteriorated; finally, the communication link is marked as a failure state, and the corresponding status flag is updated.

[0110] It should be noted that the frequency of the handshake protocol needs to be matched with the period length to avoid excessive bandwidth consumption; packet loss rate statistics need to exclude instantaneous interference (such as brief signal fluctuations) and only focus on persistent anomalies; the setting of 5 consecutive periods needs to be adjusted according to the actual network environment. Too long a period may delay fault detection, while too short a period may lead to false alarms; the 10% threshold needs to be adjusted according to the technical specifications of the communication module. For example, it can be appropriately relaxed in low-bandwidth networks; after the mark fails, the state can only be reset after the link is restored and verified.

[0111] Step 703: Self-recovery mechanism unit, used to automatically execute reverse flushing pulse sequence in response to nozzle clogging failure. If the blockage is not cleared, switch to backup nozzle and update cleaning path planning.

[0112] The self-recovery mechanism unit refers to the module within the system responsible for automatically handling faults and attempting to restore functionality. A nozzle stall fault refers to the state where the nozzle rotation mechanism is blocked and cannot rotate. A backflushing pulse sequence refers to the set of operation instructions used to clear blockages through short-term reverse water flow impacts. Blockage not resolved means the fault condition persists after backflushing. Switching to a backup nozzle refers to the operation of activating a redundant nozzle to replace the faulty one. Updating the cleaning path planning refers to the process of recalculating the optimal cleaning path based on available nozzles.

[0113] In some implementations, the self-recovery mechanism unit is automatically triggered when a nozzle stalling fault is detected. It immediately executes a preset reverse flushing pulse sequence, which generates an impact force to clear the blockage by briefly reversing the water flow direction. Then, it checks whether the stalling state is cleared. If the blockage is not cleared, it controls the actuator to switch to the backup nozzle. Finally, it updates the cleaning path planning according to the layout of the currently available nozzles to ensure the continuity of the cleaning operation.

[0114] Step 704: Fault tracing interface, used to compress and store fault codes, occurrence timestamps and abnormal data streams to edge computing nodes, and push them to the user interface through the mobile device interface to generate a visual fault report.

[0115] Among these, the fault tracing interface refers to the interface module within the system specifically used for tracking and recording fault data. A fault code is a unique coded sequence identifying a specific fault type. Compressed storage refers to the process of saving data after reducing its size using algorithms. Edge computing nodes refer to local computing devices deployed at the network edge. Mobile device interfaces refer to interface units that communicate with mobile terminals.

[0116] In some implementations, the fault tracing interface automatically receives the generated fault code, occurrence timestamp, and abnormal data stream when the system detects a fault; then it compresses this data in real time to reduce storage resource consumption; next, it stores the compressed data packet to the edge computing node; then it pushes the stored data to the user terminal through the mobile device interface; finally, it automatically generates and displays a visual fault report on the user interface, integrating key information such as fault code, timestamp, and abnormal data stream.

[0117] Based on the above technical solution, the real-time monitoring unit uses a preset threshold comparison and duration monitoring mechanism to accurately detect abnormal water flow pressure or motor torque failure, generating a unique fault code and triggering hierarchical alarms. This achieves real-time identification of high-risk issues and reduces response latency to the second level. Simultaneously, the communication link diagnostic unit employs continuous cycle counting and threshold judgment modules. When the failure rate exceeds 10% for five consecutive cycles, it automatically marks the system as faulty, eliminating transient interference and ensuring accurate communication stability assessment, thus avoiding control interruptions caused by misjudgments. This works in conjunction with the self-recovery mechanism unit to execute a reverse flushing pulse sequence to clear blockages. If the blockage persists, it automatically switches to a backup nozzle and uses a path planning algorithm to update the trajectory, quickly restoring operation and maintaining cleaning coverage while minimizing the need for manual intervention. Finally, the fault source tracing interface processes fault codes, timestamps, and abnormal data streams using a compression algorithm, stores them at edge nodes, and pushes them to mobile devices to generate visual reports. This integrates scattered data into graphical analysis, accelerating the location of the root cause of the fault. The system forms an intelligent closed-loop management system for detection, diagnosis, recovery, and traceability, thereby achieving the core objective of improving system operating efficiency and reliability, reducing downtime losses caused by faults, optimizing the allocation of maintenance resources, ensuring the stable operation of photovoltaic power generation, and ultimately supporting the sustainable operation and maintenance of floating photovoltaic power stations.

[0118] In one possible implementation of this application embodiment, an IoT communication module including a user interface, a mobile device interface, a data synchronization unit, and a security authentication mechanism is also included. This can be achieved through steps 801 to 804, which are described in detail below: Step 801: User interface, used to display photovoltaic panel stain index, ambient light intensity, weather forecast data, cleaning plan, fault alarm information and historical operation records in real time, and supports the generation of visual data charts.

[0119] Among these, the cleaning plan refers to the optimal cleaning window time arrangement output by the multimodal decision model. Fault alarm information refers to the fault codes and hierarchical classification alarms generated after the self-diagnosis module detects anomalies. Historical operation records refer to the historical logs stored in the system, including remote command execution and cleaning triggers. Visualized data chart generation refers to converting data into graphical display formats such as line charts and bar charts.

[0120] In some implementations, the user interface first acquires the photovoltaic panel stain index and ambient light intensity collected by environmental sensors in real time through the data synchronization unit of the IoT communication module. At the same time, it integrates meteorological forecast data such as rainfall probability and wind speed provided by the meteorological API. Then, it combines the cleaning plan (including start time and duration) generated by the cleaning control module and dynamically displays the above information on the main panel of the interface. In addition, when the fault self-diagnosis unit detects an anomaly, the user interface immediately pops up fault alarm information (such as insufficient water flow pressure or excessive communication packet loss rate) and allows users to query historical operation records (such as the timestamp of the remote start cleaning command). Finally, the interface has a built-in visualization engine that automatically generates charts (such as a time series line chart of stain index or a bar chart comparing meteorological data) to present the data intuitively.

[0121] Step 802: Mobile device interface, used to receive user commands, remotely start or pause cleaning operations, view real-time monitoring video, and push alarm notifications to mobile terminals via mobile applications.

[0122] In this context, the mobile device interface refers to the communication channel between the system and the mobile terminal (such as a mobile phone or tablet), used for data interaction and command transmission. The mobile application refers to the software program installed on the mobile device, providing the user interface and functional logic. User commands refer to remote control commands entered through the mobile application, such as starting or pausing the cleaning operation. Remotely starting or pausing the cleaning operation refers to triggering or terminating the cleaning process of the high-pressure nozzle mechanism through the mobile device interface. The mobile terminal refers to the user's smartphone or tablet.

[0123] In some implementations, the mobile device interface first establishes a secure connection with the mobile application through the security authentication mechanism (such as user authentication) of the IoT communication module. Then, it receives instructions triggered by the user through the mobile application interface (such as clicking the "Start Cleaning" button) and immediately forwards the instruction to the cleaning control module to remotely start or stop the cleaning operation. At the same time, the interface acquires real-time monitoring video streams collected by environmental sensors through the data synchronization unit, compresses and transmits them to the mobile application for the user to view the current status of the photovoltaic panels. In addition, when the fault self-diagnosis unit detects an anomaly (such as nozzle blockage or communication packet loss), the mobile device interface actively pushes an alarm notification (including fault codes and hierarchical classification information) to the application interface of the mobile terminal to ensure that the user is informed of the system status in a timely manner.

[0124] Step 803: Data synchronization unit, used to upload environmental sensor data, decision model output and fault codes to the cloud platform in real time, and connect to external systems through API interface.

[0125] The data synchronization unit refers to the module within the system responsible for data transmission, used to implement data uploading and external connection functions. The cloud platform refers to the service platform for external data storage and processing. The API interface refers to the application programming interface, used for data interaction between systems.

[0126] In some implementations, the data synchronization unit first establishes a secure connection through the secure authentication mechanism of the IoT communication module. Then, it receives stain images and ambient light intensity data collected by environmental sensors in real time, and simultaneously obtains cleaning plan information output by the multimodal decision model, and integrates fault codes generated by the self-diagnosis module. Next, the unit compresses these three types of data and uploads them to the cloud platform for storage and analysis. At the same time, it connects with external systems (such as meteorological services or energy management systems) through integrated API interfaces, allowing external systems to call or push data, thereby realizing cross-platform data sharing and collaborative control.

[0127] Step 804: Security authentication mechanism, including user authentication and access control, to ensure the security and compliance of operation instructions.

[0128] Among these, security authentication mechanism refers to the set of security protocols used by the system to verify user identity and control access permissions. User authentication refers to the process of confirming the authenticity of a user's identity through credentials (such as username and password). Access control refers to the system function of assigning different operating permissions (such as administrator or ordinary user) according to user roles. Operation instructions refer to control commands sent through the system, such as remotely starting or pausing the cleaning process. Security refers to the ability to protect the system from unauthorized access and attacks. Compliance refers to ensuring that all operations comply with the requirements of relevant laws and standards.

[0129] In some implementations, the security authentication mechanism is first implemented in the IoT communication module. When a user logs into the system through the user interface or mobile device interface, a user authentication process is triggered, requiring the input of a preset username and password for credential matching. After successful verification, the permission management module dynamically assigns operation permissions based on user roles (e.g., administrators have full permissions, operators can only view), ensuring that users can only execute instructions within their authorized scope. Subsequently, before the user sends an operation instruction (e.g., remotely start cleaning), the mechanism checks the permission status and instruction content again, encrypts all instructions for transmission, prevents man-in-the-middle attacks or data tampering, and ultimately ensures the security and compliance of the instruction execution process.

[0130] Based on the above technical solution, the data synchronization unit of the IoT communication module transmits data collected by environmental sensors, cleaning plans output by the decision model, and fault alarm information to the user interface in real time. The built-in visualization engine automatically converts the data into line charts or bar charts, enabling real-time status visualization through unified data integration and dynamic display. This helps users quickly identify stain deterioration trends and cleaning windows, improving decision-making efficiency. Simultaneously, the mobile device interface is integrated into the IoT communication module. After verifying the user's identity through a secure authentication mechanism, it receives instructions from the mobile application and calls the cleaning control module to execute operations. At the same time, it acquires real-time monitoring video streams through the data synchronization unit, compresses them, and transmits them to the mobile terminal. When the fault self-diagnosis unit triggers an alarm, the interface actively pushes notifications, enabling deep coupling between the mobile application and the IoT to achieve remote control and instant response anytime, anywhere, improving operational flexibility and fault handling speed. Furthermore, embedding the data synchronization unit into the IoT communication module allows for real-time uploading to the cloud platform for storage. Simultaneously, it connects with external systems (such as weather services) through standardized API interfaces (such as RESTful protocols), allowing bidirectional data calls. Through cloud platform integration and API interoperability, data sharing and cross-system collaboration can be achieved, improving overall intelligent operation and maintenance. In addition, the security authentication mechanism can prevent unauthorized access and instruction tampering through layered verification and encrypted transmission, ensuring the security and regulatory compliance of system operations.

[0131] The above primarily describes the solutions of the embodiments of this application from the perspective of device implementation. It is understood that each device, for example, an IoT-based intelligent water cleaning device for floating photovoltaic panels, includes at least one of the hardware structures and software modules corresponding to the execution of each function in order to achieve the above-mentioned functions. Those skilled in the art should readily recognize that, in conjunction with the units and algorithm steps of the various examples described in the embodiments disclosed herein, this application can be implemented in hardware or a combination of hardware and computer software. Whether a function is executed by hardware or by computer software driving hardware depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0132] This application embodiment can divide the IoT-based smart water cleaning device for floating photovoltaic panels into functional units based on the above method example. For example, each function can be divided into separate functional units, or two or more functions can be integrated into the same processing unit. The integrated unit can be implemented in hardware or as a software functional unit. It should be noted that the unit division in this application embodiment is illustrative and only represents one logical functional division; other division methods may be used in actual implementation.

[0133] When using integrated units, Figure 9 A possible structural schematic diagram of an IoT-based smart water cleaning device for floating photovoltaic panels (referred to as an IoT-based smart water cleaning device for floating photovoltaic panels 90) is shown in the above embodiments. The IoT-based smart water cleaning device for floating photovoltaic panels 90 includes a processing unit 901 and a communication unit 902, and may also include a storage unit 903. Figure 9 The structural diagram shown can be used to illustrate the structure of the IoT-based smart water cleaning device for floating photovoltaic panels involved in the above embodiments.

[0134] when Figure 9 The structural diagram shown illustrates the structure of the IoT-based smart water cleaning device for floating photovoltaic panels involved in the above embodiments. The processing unit 901 is used to control and manage the operation of the IoT-based smart water cleaning device for floating photovoltaic panels, the communication unit 902 is used for the IoT-based smart water cleaning device for floating photovoltaic panels to communicate with other devices, and the storage unit 903 is used to store the program code and data of the IoT-based smart water cleaning device for floating photovoltaic panels.

[0135] For example, the communication unit 902 is used to collect images of dirt on the surface of the photovoltaic panel and the ambient light intensity at the location of the photovoltaic panel in real time through environmental sensors, and call the meteorological API to obtain meteorological forecast data for the future period at the location of the photovoltaic panel; Processing unit 901 performs anti-reflection interference processing on the surface dirt image of the photovoltaic panel to obtain a dynamic dirt index. The dynamic dirt index refers to the pollutant coverage score of the photovoltaic panel surface after anti-reflection interference processing. It integrates ambient light intensity, meteorological data, and the dynamic dirt index at the location of the photovoltaic panel to generate cross-modal fusion features. These cross-modal fusion features are input into a multimodal decision model to obtain the predicted optimal cleaning window, which is then fed back to the cleaning control module. When the time series reaches the predicted optimal cleaning window, the cleaning control module triggers the high-pressure nozzle mechanism to perform directional cleaning and adjusts the parameters of the multimodal decision model based on the post-cleaning image.

[0136] In one possible implementation, the processing unit 901 is further configured to perform anti-reflection interference processing on the photovoltaic panel surface stain image to obtain a dynamic stain index, including: converting the photovoltaic panel surface stain image to the HSV color space, separating the luminance channel and saturation channel, and reading the luminance and saturation values. Based on image processing threshold requirements, the maximum luminance threshold and the maximum saturation threshold of the current image are calculated, and regions with luminance values ​​greater than the current image's maximum luminance threshold and saturation values ​​greater than the current image's maximum saturation threshold are selected and marked as suspected reflection interference areas; the remaining regions are marked as non-interference areas. Local texture features in the RGB space of the non-interference areas are extracted using the gray-level co-occurrence matrix and gray-level gradient magnitude. The suspected reflection interference areas are converted to the Lab color space, the chromaticity shift is extracted, and the reflection intensity difference features of the near-infrared band are fused to obtain interference features. The local texture features of the non-interference areas and the interference features are input into a pre-trained lightweight CNN, outputting a normalized stain coverage score. The scoring weights are dynamically adjusted based on ambient light intensity to generate an anti-interference dynamic stain index.

[0137] In one possible implementation, processing unit 901 is further configured to extract local texture features in the RGB space of the non-interference region using the gray-level co-occurrence matrix and gray-level gradient magnitude. This includes: converting the RGB image of the non-interference region into a gray-level image, calculating its gray-level co-occurrence matrix, and extracting four statistical measures from the gray-level co-occurrence matrix: energy, contrast, entropy, and inverse moment, to generate a texture feature vector. The Sobel operator is used to calculate the gradient magnitude of the gray-level image, and the mean and variance of the gradient magnitude within a sliding window are calculated to generate a gradient feature vector. The texture feature vector and the gradient feature vector are concatenated along the feature dimension to generate a local texture feature vector.

[0138] In conjunction with the first aspect mentioned above, one possible implementation involves fusing ambient light intensity, meteorological data, and dynamic stain index at the location of the photovoltaic panel to generate cross-modal fusion features. This includes: spatiotemporally aligning ambient light intensity, meteorological forecast data, and dynamic stain index to generate a time-synchronized multidimensional feature tensor; extracting rainfall probability, wind speed, and irradiance variation trends from the meteorological data using a bidirectional LSTM network to generate a meteorological time-series feature vector; extracting the temporal evolution features of the dynamic stain index using a lightweight temporal convolutional network to generate a stain dynamic feature vector; and concatenating the meteorological time-series feature vector, ambient light intensity, and stain dynamic feature vector along the feature dimensions to generate the cross-modal fusion features.

[0139] In one possible implementation, processing unit 901 is further configured to input cross-modal fusion features into a multimodal decision model to obtain the predicted optimal cleaning window and feed it back to the cleaning control module. This includes: inputting cross-modal fusion features into the multimodal decision model, which comprises a cascaded cross-modal interaction layer and a decision verification layer. The cross-modal interaction layer employs a scaled dot product attention mechanism, using ambient light intensity as the query vector, stain dynamic features as the key vector, and meteorological time series features as the value vector, to calculate dynamic weights and generate a weighted decision vector. The decision verification layer includes a preliminary screening layer, a fine screening layer, and a feedback verification layer. The preliminary screening layer determines, based on the weighted decision vector, the period in which the probability of rainfall within the next 24 hours is lower than a preset threshold and the ambient light intensity is within the workable range, and this period is designated as the initial cleaning period. The fine screening layer selects consecutive periods in the initial cleaning period where the dynamic stain index exceeds the pollution threshold and continues to rise as candidate windows. The feedback verification layer sends the start time and duration of the candidate windows to the cleaning control module and adjusts the weight parameters of the cross-modal interaction layer based on the cleaned image.

[0140] In one possible implementation, the processing unit 901 further includes an energy management module: monitoring the photovoltaic panel's power generation and the energy storage unit's power status. Before the cleaning control module triggers the high-pressure nozzle mechanism, it determines whether the current energy storage unit's power level is lower than a preset operating threshold. If it is lower than the operating threshold, it calls a meteorological API to obtain irradiance prediction data for a preset future period and outputs a predicted power generation curve based on a multimodal decision model. The predicted power generation curve is then fused with the dynamic stain index to recalculate the optimal cleaning window and feed it back to the cleaning control module, ensuring that the high-pressure nozzle mechanism performs targeted cleaning only when the photovoltaic panel's power generation meets the cleaning energy consumption requirements.

[0141] In one possible implementation, the processing unit 901 further includes a high-pressure cleaning module comprising a nozzle rotation mechanism, a water pressure adaptive controller, a cleaning path planning unit, and a feedback calibration interface: The nozzle rotation mechanism adjusts the pitch and horizontal rotation angle of the high-pressure nozzle based on the distribution of dirt on the photovoltaic panel surface. The water pressure adaptive controller adjusts the water pressure intensity based on a dynamic dirt index, using a low-pressure mode when the dirt index is below a first threshold and a high-pressure mode when it is above a second threshold. The cleaning path planning unit generates a rasterized cleaning path based on the spatial distribution of the dirt image, controlling the nozzle to cover the dirt area in a serpentine trajectory. The feedback calibration interface transmits the deviation between the actual cleaning path and the planned path to the cleaning control module in real time, used to adjust the output parameters of the multimodal decision model.

[0142] In one possible implementation, the processing unit 901 further includes a self-diagnostic module comprising a real-time monitoring unit, a communication link diagnostic unit, a self-recovery mechanism unit, and a fault tracing interface: The real-time monitoring unit periodically scans the water flow sensor, pressure sensor, and stepper motor feedback signals of the nozzle rotation mechanism. If the water flow pressure is detected to be lower than a preset threshold or the motor torque is abnormally high for an extended period, a fault code is generated and an alarm level classification is triggered. The communication link diagnostic unit uses the packet loss rate statistics of the IoT communication module's handshake protocol to determine the communication stability between the environmental sensor network and the control module in real time. When the packet loss rate is higher than 10% for nine consecutive cycles, it is marked as a communication link failure. The self-recovery mechanism unit automatically executes a reverse flushing pulse sequence for nozzle stall faults. If the blockage is not resolved, a backup nozzle is switched and the cleaning path planning is updated. The fault tracing interface compresses and stores the fault code, occurrence timestamp, and abnormal data stream to the edge computing node, and pushes it to the user interface via the mobile device interface to generate a visual fault report. In one possible implementation, the processing unit 901 further includes an IoT communication module comprising a user interface, a mobile device interface, a data synchronization unit, and a security authentication mechanism: The user interface displays real-time photovoltaic panel dirt index, ambient light intensity, weather forecast data, cleaning plans, fault alarm information, and historical operation records, and supports the generation of visualized data charts. The mobile device interface receives user commands via a mobile application, remotely starts or pauses cleaning operations, views real-time monitoring video, and pushes alarm notifications to the mobile terminal. The data synchronization unit uploads environmental sensor data, decision model outputs, and fault codes to the cloud platform in real time and connects to external systems via an API interface. The security authentication mechanism includes user authentication and access control to ensure the security and compliance of operating instructions.

[0143] The processing unit 901 can be a processor or a controller, and the communication unit 902 can be a communication interface, transceiver, transceiver circuit, transceiver device, etc. The term "communication interface" is a general term and may include one or more interfaces. The storage unit 903 can be a memory. When the IoT-based smart water cleaning device 90 for floating photovoltaic panels is a chip, the processing unit 901 can be a processor or a controller, and the communication unit 902 can be an input interface and / or an output interface, pins, or circuits, etc. The storage unit 903 can be a storage unit within the chip (e.g., a register, cache, etc.) or a storage unit located outside the chip (e.g., read-only memory (ROM), random access memory (RAM, etc.)).

[0144] The communication unit can also be called a transceiver unit. The antenna and control circuit with transceiver functions in the IoT-based smart water cleaning device 90 for floating solar panels can be considered as the communication unit 902 of the IoT-based smart water cleaning device 90 for floating solar panels, and the processor with processing functions can be considered as the processing unit 901 of the IoT-based smart water cleaning device 90 for floating solar panels. Optionally, the device in the communication unit 902 that implements the receiving function can be considered as the communication unit, which is used to execute the receiving steps in the embodiments of this application. The communication unit can be a receiver, a receiver circuit, etc. The device in the communication unit 902 that implements the transmitting function can be considered as the transmitting unit, which is used to execute the transmitting steps in the embodiments of this application. The transmitting unit can be a transmitter, a transmitter, a transmitting circuit, etc.

[0145] Figure 9 If the integrated units in the process are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this application, in essence, or the parts that contribute to the prior art, or all or part of the technical solutions, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods described in the various embodiments of this application. Storage media for storing computer software products include various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory, random access memory, magnetic disks, or optical disks.

[0146] In implementation, each step of the method provided in this embodiment can be completed by integrated logic circuits in the processor or by instructions in software form. The steps of the method disclosed in the embodiments of this application can be directly manifested as being executed by a hardware processor, or being executed by a combination of hardware and software modules in the processor.

[0147] The processor in this application may include, but is not limited to, at least one of the following: a central processing unit (CPU), a microprocessor, a digital signal processor (DSP), a microcontroller unit (MCU), or an artificial intelligence processor, etc., and other computing devices that run software. Each computing device may include one or more cores for executing software instructions to perform calculations or processing. The processor may be a standalone semiconductor chip or integrated with other circuits into a single semiconductor chip. For example, it may be integrated with other circuits (such as encoding / decoding circuits, hardware acceleration circuits, or various bus and interface circuits) to form a System-on-a-Chip (SoC), or it may be integrated as a built-in processor within an ASIC. The ASIC with the integrated processor may be packaged separately or together with other circuits. In addition to the cores for executing software instructions to perform calculations or processing, the processor may further include necessary hardware accelerators, such as field-programmable gate arrays (FPGAs), programmable logic devices (PLDs), or logic circuits that implement dedicated logic operations.

[0148] The memory in the embodiments of this application may include at least one of the following types: read-only memory (ROM) or other types of static storage devices capable of storing static information and instructions; random access memory (RAM) or other types of dynamic storage devices capable of storing information and instructions; or electrically erasable programmable-only memory (EEPROM). In some scenarios, the memory may also be compact disc read-only memory (CD-ROM) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disk storage media, or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but is not limited thereto.

[0149] This application also provides a computer-readable storage medium including instructions that, when run on a computer, cause the computer to perform any of the methods described above.

[0150] This application also provides a computer program product containing instructions that, when run on a computer, cause the computer to perform any of the methods described above.

[0151] This application also provides a chip including a processor and an interface circuit. The interface circuit is coupled to the processor. The processor is used to run computer programs or instructions to implement the above-described method. The interface circuit is used to communicate with other modules outside the chip.

[0152] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented using software programs, implementation can be, in whole or in part, in the form of a computer program product. This computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium accessible to a computer or a data storage device containing one or more servers, data centers, etc., that can be integrated with the medium. The available media can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media (e.g., solid-state drives (SSDs)).

[0153] Although this application has been described herein in conjunction with various embodiments, those skilled in the art, by reviewing the accompanying drawings, the disclosure, and the appended claims, will understand and implement other variations of the disclosed embodiments in carrying out the claimed application. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude multiple components. A single processor or other unit can implement several functions listed in the claims. While different dependent claims may recite certain measures, this does not mean that these measures cannot be combined to produce good results.

[0154] Although this application has been described in conjunction with specific features and embodiments, it is obvious that various modifications and combinations can be made thereto without departing from the spirit and scope of this application. Accordingly, this specification and drawings are merely exemplary illustrations of this application as defined by the appended claims, and are considered to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from the spirit and scope of this application. Thus, if such modifications and variations of this application fall within the scope of the claims of this application and their equivalents, this application is also intended to include such modifications and variations.

Claims

1. A smart water cleaning method for floating photovoltaic panels based on the Internet of Things, characterized in that, include: Real-time images of surface stains on photovoltaic panels and ambient light intensity at the location of the photovoltaic panels are collected using environmental sensors, and meteorological APIs are called to obtain weather forecast data for the location of the photovoltaic panels for future periods. The image of the surface stains of the photovoltaic panel is subjected to anti-reflection interference processing to obtain a dynamic stain index, which is a score of the contaminant coverage of the photovoltaic panel surface after the anti-reflection interference processing. By integrating the ambient light intensity, meteorological data, and dynamic stain index of the location of the photovoltaic panel, cross-modal fusion features are generated; The cross-modal fusion features are input into the multimodal decision model to obtain the predicted optimal cleaning window and then fed back to the cleaning control module. When the time series reaches the predicted optimal cleaning window, the cleaning control module triggers the high-pressure nozzle mechanism to perform directional cleaning and adjusts the parameters of the multimodal decision model based on the post-cleaning image.

2. The intelligent water cleaning method for floating photovoltaic panels based on the Internet of Things according to claim 1, characterized in that, The process of performing anti-reflection interference processing on the surface stain image of the photovoltaic panel to obtain a dynamic stain index includes: The image of the stains on the photovoltaic panel surface is converted to the HSV color space, the luminance channel and saturation channel are separated, and the luminance and saturation values ​​are read. Based on the image processing threshold requirements, calculate the maximum brightness threshold and the maximum saturation threshold of the current image, filter out areas where the brightness value is greater than the maximum brightness threshold and areas where the saturation value is greater than the maximum saturation threshold, mark suspected reflection interference areas, and mark the remaining areas as non-interference areas; Local texture features in the RGB space of the non-interference region are extracted using the gray-level co-occurrence matrix and gray-level gradient magnitude. The suspected reflection interference area is converted to the Lab color space, the chromaticity shift is extracted, and the reflection intensity difference features in the near-infrared band are fused to obtain the interference features; The local texture features and interference features of the non-interference area are input into a pre-trained lightweight CNN to output a normalized stain coverage score. By dynamically adjusting the scoring weights based on ambient light intensity, an interference-resistant dynamic stain index is generated.

3. The intelligent water cleaning method for floating photovoltaic panels based on the Internet of Things according to claim 2, characterized in that, The extraction of local texture features in the RGB space of the non-interference region through the gray-level co-occurrence matrix and gray-level gradient magnitude includes: The RGB image of the non-interference region is converted into a grayscale image, and its grayscale co-occurrence matrix is ​​calculated. Four types of statistics are extracted from the grayscale co-occurrence matrix: energy, contrast, entropy, and inverse difference moment, to generate a texture feature vector. The gradient magnitude of the grayscale image is calculated using the Sobel operator, and the mean and variance of the gradient magnitude within a window are statistically analyzed using a sliding window to generate a gradient feature vector. The texture feature vector and the gradient feature vector are concatenated along the feature dimension to generate a local texture feature vector.

4. The intelligent water cleaning method for floating photovoltaic panels based on the Internet of Things according to claim 3, characterized in that, The system integrates ambient light intensity, meteorological data, and dynamic pollution index at the location of the photovoltaic panels to generate cross-modal fusion features, including: The ambient light intensity, meteorological forecast data and dynamic stain index are spatiotemporally aligned to generate a time-synchronized multidimensional feature tensor. A bidirectional LSTM network is used to extract the trends of rainfall probability, wind speed and irradiance in meteorological data to generate meteorological time series feature vectors. A lightweight temporal convolutional network is used to extract the temporal evolution features of the dynamic stain index, generating a dynamic stain feature vector. The meteorological time-series feature vector, ambient light intensity, and stain dynamic feature vector are concatenated along the feature dimension to generate cross-modal fusion features.

5. The intelligent water cleaning method for floating photovoltaic panels based on the Internet of Things according to claim 4, characterized in that, The step of inputting the cross-modal fusion features into the multimodal decision model to obtain the predicted optimal cleaning window and feeding it back to the cleaning control module includes: The cross-modal fusion features are input into a multimodal decision model, which includes a cascaded cross-modal interaction layer and a decision verification layer. The cross-modal interaction layer adopts a scaled dot product attention mechanism, using ambient light intensity as the query vector, stain dynamic features as the key vector, and meteorological time series features as the value vector to calculate dynamic weights and generate a weighted decision vector. The decision verification layer includes a preliminary screening layer, a fine screening layer, and a feedback verification layer. The preliminary screening layer determines, based on a weighted decision vector, the period during which the probability of rainfall in the next 24 hours is lower than a preset threshold and the ambient light intensity is within the workable range, and this period is recorded as the initial cleaning period. The fine screening stage layer selects a continuous period of time during the initial cleaning period in which the dynamic stain index exceeds the contamination threshold and continues to rise as a candidate window. The feedback verification layer sends the start time and duration of the candidate window to the cleaning control module, and adjusts the weight parameters of the cross-modal interaction layer based on the cleaned image.

6. The intelligent water cleaning method for floating photovoltaic panels based on the Internet of Things according to claim 5, characterized in that, It also includes an energy management module: Monitor the power generation of photovoltaic panels and the energy storage unit's power status; Before the cleaning control module triggers the high-pressure nozzle mechanism, it determines whether the current energy storage unit power is lower than the preset operating threshold; if it is lower than the operating threshold, it calls the meteorological API to obtain the irradiance prediction data for the preset period in the future, and outputs the predicted power generation curve based on the multimodal decision model. By integrating the predicted power generation curve and the dynamic stain index, the optimal cleaning window is recalculated and fed back to the cleaning control module, so that the high-pressure nozzle mechanism performs targeted cleaning only when the photovoltaic power generation meets the cleaning energy consumption requirements.

7. The intelligent water cleaning method for floating photovoltaic panels based on the Internet of Things according to claim 6, characterized in that, It also includes a high-pressure cleaning module comprising a nozzle rotation mechanism, a water pressure adaptive controller, a cleaning path planning unit, and a feedback calibration interface. The nozzle rotation mechanism is used to adjust the pitch angle and horizontal rotation angle of the high-pressure nozzle according to the distribution position of the stains on the photovoltaic panel surface. The water pressure adaptive controller adjusts the water pressure intensity according to the dynamic stain index. When the stain index is lower than the first threshold, a low-pressure mode is used, and when it is higher than the second threshold, a high-pressure mode is activated. The cleaning path planning unit generates a rasterized cleaning path based on the spatial distribution of the stain image and controls the nozzle to cover the stain area in a serpentine trajectory. The feedback calibration interface transmits the deviation between the actual cleaning path and the planned path to the cleaning control module in real time, which is used to adjust the output parameters of the multimodal decision model.

8. The intelligent water cleaning method for floating photovoltaic panels based on the Internet of Things according to claim 7, characterized in that, It also includes a self-diagnostic module that incorporates a real-time monitoring unit, a communication link diagnostic unit, a self-recovery mechanism unit, and a fault tracing interface. The real-time monitoring unit periodically scans the water flow sensor, pressure sensor, and stepper motor feedback signals of the nozzle rotation mechanism. If the water flow pressure is detected to be lower than the preset threshold or the abnormal motor torque is detected to last for more than the set time, a fault code is generated and an alarm level classification is triggered. The communication link diagnostic unit determines the communication stability between the environmental sensor network and the control module in real time by statistically analyzing the packet loss rate of the handshake protocol of the IoT communication module. When the packet loss rate is higher than 10% for 5 consecutive cycles, it is marked as a communication link failure. The self-recovery mechanism unit is used to automatically execute a reverse flushing pulse sequence in response to nozzle clogging failure. If the blockage is not cleared, it switches to a backup nozzle and updates the cleaning path plan. The fault tracing interface is used to compress and store fault codes, occurrence timestamps, and abnormal data streams to edge computing nodes, and push them to the user interface through the mobile device interface to generate a visual fault report.

9. A smart water cleaning method for floating photovoltaic panels based on the Internet of Things according to claim 8, characterized in that, It also includes an IoT communication module that incorporates a user interface, mobile device interface, data synchronization unit, and security authentication mechanism: The user interface is used to display the photovoltaic panel stain index, ambient light intensity, weather forecast data, cleaning plan, fault alarm information and historical operation records in real time, and supports the generation of visual data charts; The mobile device interface is used to receive user commands via a mobile application, remotely start or pause cleaning operations, view real-time monitoring video, and push alarm notifications to the mobile terminal. The data synchronization unit is used to upload environmental sensor data, decision model output and fault codes to the cloud platform in real time, and connect to external systems through API interface; A security authentication mechanism, including user authentication and access control, ensures the security and compliance of operational instructions.

10. A smart water cleaning device for floating photovoltaic panels based on the Internet of Things, characterized in that, The device includes: a communication unit and a processing unit; The communication unit is used to collect images of dirt on the surface of the photovoltaic panel and the ambient light intensity at the location of the photovoltaic panel in real time through environmental sensors, and to call the meteorological API to obtain meteorological forecast data for the future period at the location of the photovoltaic panel. The processing unit is used to perform anti-reflection interference processing on the surface stain image of the photovoltaic panel to obtain a dynamic stain index. The dynamic stain index refers to the pollutant coverage score of the surface stain image of the photovoltaic panel after anti-reflection interference processing. By integrating the ambient light intensity, meteorological data, and dynamic stain index of the location of the photovoltaic panel, cross-modal fusion features are generated; The cross-modal fusion features are input into the multimodal decision model to obtain the predicted optimal cleaning window and then fed back to the cleaning control module. When the time series reaches the predicted optimal cleaning window, the cleaning control module triggers the high-pressure nozzle mechanism to perform directional cleaning and adjusts the parameters of the multimodal decision model based on the post-cleaning image.

Images