Photovoltaic module dusting state online detection method and system based on multi-modal fusion and edge computing

Abstract

This invention relates to the field of edge computing technology, specifically to an online detection method and system for the dust accumulation status of photovoltaic modules based on multimodal fusion and edge computing. The method includes the following steps: collecting real-time operating data of the photovoltaic modules, including electrical performance monitoring data, environmental meteorological data, and monitoring images of the module surface; transmitting the real-time operating data to an edge computing gateway; and using the edge computing gateway to perform timestamp alignment and outlier cleanup on the real-time operating data to obtain a standard multimodal input vector. In this invention, multi-source heterogeneous data is transmitted to the network edge for localized timestamp alignment and outlier cleanup. Time synchronization and noise filtering are directly performed at the data source, ensuring a strict time correspondence between multimodal data and reducing network bandwidth consumption and processing delays caused by invalid data uploads.

Description

Technical Field

[0001] This invention relates to the field of edge computing technology, and in particular to an online detection method and system for the dust accumulation status of photovoltaic modules based on multimodal fusion and edge computing. Background Technology

[0002] Edge computing is a distributed computing architecture that moves computing tasks, data storage, and application services from the centralized cloud to the network edge, closer to the data source. In other words, it performs localized data processing and analysis closer to the terminal device or the data generation source.

[0003] Current technologies generally employ a centralized cloud computing model to process photovoltaic (PV) operation and maintenance data. All terminals collect massive amounts of video streams and high-frequency electrical parameter data, which must be uploaded to a remote cloud server for centralized analysis via long links. This results in immense pressure on network bandwidth, and in the context of weak network environments common in remote PV power plants, data packet loss, transmission congestion, and even connection interruptions are prone to occur, causing significant delays in monitoring results. Furthermore, traditional detection methods often rely on single-dimensional indicators for judgment. For example, judging dust accumulation solely based on output power decline is difficult to accurately distinguish between power attenuation caused by cloud cover, building shadows, and actual dust accumulation. Alternatively, relying solely on image recognition technology is limited by changes in lighting, water reflections, and shooting angle deviations, leading to a high false alarm rate under complex weather conditions. Therefore, improvements are needed. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies that generally use a centralized cloud computing model to process photovoltaic operation and maintenance data. All terminals collect massive video streams and high-frequency electrical parameter data, which need to be uploaded to a remote cloud server for centralized analysis through long links, resulting in huge network transmission bandwidth pressure. The invention proposes an online detection method and system for the dust accumulation status of photovoltaic modules based on multimodal fusion and edge computing.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: an online detection method for the dust accumulation status of photovoltaic modules based on multimodal fusion and edge computing, comprising the following steps:

[0006] Collect real-time operating data of photovoltaic modules, including electrical performance monitoring data, environmental meteorological data, and monitoring images of the module surface;

[0007] The real-time running data is transmitted to an edge computing gateway, which performs timestamp alignment and outlier cleaning on the real-time running data to obtain a standard multimodal input vector.

[0008] The standard multimodal input vector is input into the multimodal feature extraction model deployed on the edge computing gateway, and the electrical feature map, environmental correlation feature and visual texture feature corresponding to the standard multimodal input vector are extracted by the multi-path parallel branches in the multimodal feature extraction model respectively.

[0009] The attention fusion layer in the multimodal feature extraction model is used to perform cross-modal interactive calculations on the electrical feature map, the environmental correlation features, and the visual texture features to generate a photovoltaic module dust accumulation health index.

[0010] The online dust accumulation status of the photovoltaic module is determined by comparing the dust accumulation health index of the photovoltaic module with a preset set of dust accumulation level thresholds.

[0011] Preferably, the steps for collecting real-time operating data of photovoltaic modules include:

[0012] The electrical performance monitoring data is acquired by a smart data acquisition device connected to the photovoltaic string. The electrical performance monitoring data includes string-level current and voltage curves, real-time output power, and fill factor parameters.

[0013] The environmental meteorological data is acquired through a meteorological station sensor array, including irradiance, ambient temperature, relative humidity, and wind speed.

[0014] A monitoring image of the photovoltaic module surface is acquired by a fixed-focus camera aimed at the module surface, and the monitoring image of the module surface includes dust distribution texture information.

[0015] Preferably, the step of using the edge computing gateway to perform timestamp alignment and outlier cleanup on the real-time running data includes:

[0016] Extract the first acquisition timestamp of the electrical performance monitoring data, the second acquisition timestamp of the environmental meteorological data, and the third acquisition timestamp of the component surface monitoring image;

[0017] Using the third acquisition timestamp as the reference time axis, a linear interpolation algorithm is used to map the electrical performance monitoring data and the environmental meteorological data to the reference time axis to generate a time-synchronized data frame;

[0018] The noise data in the time synchronization data frame is filtered using a statistical outlier removal algorithm, and the filtered electrical performance monitoring data, environmental meteorological data, and component surface monitoring images are vectorized and encapsulated to obtain the standard multimodal input vector.

[0019] Preferably, the step of extracting the electrical feature map, environmental correlation features, and visual texture features corresponding to the standard multimodal input vector using the multiple parallel branches in the multimodal feature extraction model includes:

[0020] The component surface monitoring image contained in the standard multimodal input vector is convolutionally pooled using a branch of a convolutional neural network to extract the visual texture features, which characterize the dust occlusion distribution on the component surface.

[0021] The electrical performance monitoring data contained in the standard multimodal input vector is analyzed for time-series dependence using a branch of a long short-term memory network, and the electrical feature map is extracted. The electrical feature map represents the trend information of power decay over time.

[0022] The environmental meteorological data contained in the standard multimodal input vector are nonlinearly mapped using a fully connected neural network branch to extract the environmental correlation features, which characterize the theoretical impact benchmark of environmental parameters on power generation efficiency.

[0023] Preferably, the step of performing cross-modal interactive computation on the electrical feature map, the environmental association features, and the visual texture features using the attention fusion layer in the multimodal feature extraction model includes:

[0024] A multi-head attention mechanism is used to calculate the cross-correlation weight matrix among the visual texture features, the electrical feature map, and the environmental correlation features;

[0025] The electrical feature map, the environmental correlation feature, and the visual texture feature are weighted and fused according to the cross-correlation weight matrix to obtain a multimodal fusion feature vector;

[0026] The multimodal fusion feature vector is input into the regression prediction layer, and the dust accumulation health index of the photovoltaic module is calculated through the mapping in the regression prediction layer.

[0027] Preferably, the formula for calculating the dust accumulation health index of the photovoltaic module by the regression prediction layer is: ,in, This indicates the dust accumulation health index of the photovoltaic module. This represents the normalized scalar value output by the regression prediction layer based on the visual texture feature map. This represents the mean of the feature response output by the regression prediction layer based on the electrical feature map mapping. This represents the intensity scalar value output by the regression prediction layer based on the environmental association feature mapping. Represents the visual feature weight coefficients. Represents the electrical characteristic weighting coefficient. Indicates the environmental characteristic weighting coefficient. Represents the bias constant. It represents the base of the natural logarithm.

[0028] Preferably, after comparing the photovoltaic module's dust accumulation health index with a preset set of dust accumulation level thresholds, the method further includes:

[0029] Calculate the rate of change of the dust accumulation health index of the photovoltaic module within a preset time window;

[0030] Determine whether the rate of change is greater than a preset ash accumulation rate threshold;

[0031] When the rate of change is greater than the dust accumulation rate threshold, an early dust accumulation warning signal is generated. The early dust accumulation warning signal is used to indicate that there is a thin layer of rapid dust accumulation on the surface of the photovoltaic module.

[0032] Preferably, the method further includes:

[0033] Collect vibration frequency data when the cleaning robotic arm comes into contact with the photovoltaic module;

[0034] The vibration frequency data is input into the edge computing gateway to extract vibration damping features;

[0035] The dust accumulation health index of the photovoltaic module is corrected by using the vibration damping characteristics to obtain the corrected dust accumulation health index.

[0036] Preferably, the formula for correcting the dust accumulation health index of the photovoltaic module using the vibration damping characteristics is as follows: ,in, This represents the corrected dust accumulation health index. This indicates the dust accumulation health index of the photovoltaic modules before the correction. Indicates the vibration correction factor. The normalized amplitude energy represents the vibration damping characteristic. Represents the logarithmic function with base 10;

[0037] The method further includes:

[0038] The historical distribution data of the dust accumulation health index of the photovoltaic modules are statistically analyzed in the edge computing gateway;

[0039] When the prediction deviation in the historical distribution data exceeds a preset deviation threshold, the cloud-edge collaborative update mechanism is triggered.

[0040] Receive updated model parameters from the cloud server and update the multimodal feature extraction model using the updated model parameters.

[0041] This invention provides a system comprising:

[0042] The data acquisition module is used to collect real-time operating data of the photovoltaic module, including electrical performance monitoring data, environmental meteorological data, and monitoring images of the module surface.

[0043] The data preprocessing module is used to transmit the real-time running data to the edge computing gateway, and use the edge computing gateway to perform timestamp alignment and outlier cleaning on the real-time running data to obtain a standard multimodal input vector.

[0044] The feature extraction module is used to input the standard multimodal input vector into the multimodal feature extraction model deployed on the edge computing gateway, and use the multi-parallel branches in the multimodal feature extraction model to extract the electrical feature map, environmental correlation features and visual texture features corresponding to the standard multimodal input vector respectively;

[0045] The feature fusion module is used to perform cross-modal interactive calculations on the electrical feature map, the environmental correlation features, and the visual texture features using the attention fusion layer in the multimodal feature extraction model to generate a photovoltaic module dust accumulation health index.

[0046] The status determination module is used to compare the dust accumulation health index of the photovoltaic module with a preset set of dust accumulation level thresholds to determine the online dust accumulation status of the photovoltaic module.

[0047] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0048] This invention captures the real-time operating status of photovoltaic modules from a multi-dimensional physical perspective by collecting electrical performance monitoring data, environmental meteorological data, and module surface monitoring images, overcoming the limitations and biases of information from a single data source. Multi-source heterogeneous data is transmitted to the network edge for localized timestamp alignment and outlier removal, achieving time synchronization and noise filtering directly at the data source. This ensures strict time correspondence between multi-modal data while reducing network bandwidth consumption and processing delays caused by invalid data uploads. Parallel computing branches are used to extract visual texture, electrical trends, and environmental correlation features, mining deep semantic information from different modal data. An attention mechanism is used for cross-modal interactive computation, achieving complementary enhancement and dynamic weight allocation between features, eliminating interference from shadow occlusion, sudden changes in illumination, or sensor drift. By generating a multi-dimensional feature-integrated dust accumulation health index and comparing it with a preset threshold set, high-precision quantitative assessment and real-time response to module dust accumulation status are achieved. This ensures detection accuracy while improving the independent decision-making capability and response speed of the photovoltaic operation and maintenance system at the edge, ensuring efficient dust accumulation monitoring and early warning services even in weak or offline environments. Attached Figure Description

[0049] Figure 1 This is a schematic diagram of the steps of the present invention. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0051] Please see Figure 1 This invention provides a technical solution: an online detection method for the dust accumulation status of photovoltaic modules based on multimodal fusion and edge computing, comprising the following steps:

[0052] Collect real-time operating data of photovoltaic modules, including electrical performance monitoring data, environmental meteorological data, and monitoring images of the module surface;

[0053] The real-time running data is transmitted to the edge computing gateway, which performs timestamp alignment and outlier cleaning on the real-time running data to obtain a standard multimodal input vector.

[0054] The standard multimodal input vector is input into the multimodal feature extraction model deployed on the edge computing gateway. The multimodal feature extraction model uses multiple parallel branches to extract the electrical feature map, environmental correlation features and visual texture features corresponding to the standard multimodal input vector.

[0055] The attention fusion layer in the multimodal feature extraction model is used to perform cross-modal interactive calculations on electrical feature maps, environmental correlation features, and visual texture features to generate a dust accumulation health index for photovoltaic modules.

[0056] The online dust accumulation status of photovoltaic modules is determined by comparing the dust accumulation health index of the photovoltaic modules with a preset set of dust accumulation level thresholds.

[0057] In this embodiment, the step of collecting real-time operating data of photovoltaic modules includes: acquiring electrical performance monitoring data through an intelligent data acquisition device connected to the photovoltaic string, the electrical performance monitoring data including string-level current and voltage curves, real-time output power and fill factor parameters; acquiring environmental meteorological data through a meteorological station sensor array, the environmental meteorological data including irradiance value, ambient temperature value, relative humidity value and wind speed value; and acquiring a monitoring image of the module surface through a fixed-focus camera aimed at the surface of the photovoltaic module, the monitoring image of the module surface including dust distribution texture information.

[0058] Specifically, the steps for collecting real-time operating data of photovoltaic modules include configuring the Modbus-RTU communication protocol and setting the baud rate to 9600bps; polling the intelligent data acquisition unit connected to the photovoltaic string; reading the values ​​in the register address at a preset sampling frequency, such as once every 5 minutes, to obtain electrical performance monitoring data. This data specifically includes the string-level current and voltage curve with 256 sampling points obtained by scanning string characteristics; the real-time output power obtained by calculating the product of voltage and current; and the fill factor parameter calculated by the ratio of the maximum power point power to the product of open-circuit voltage and short-circuit current. Simultaneously, the meteorological station sensor array deployed within the photovoltaic power station is activated, and the data is collected via a high-precision total radiation meter. Solar irradiance values ​​are collected at a frequency of 1Hz. Ambient temperature values ​​are collected through a platinum resistance thermometer, relative humidity values ​​are collected through a humidity-sensitive capacitive sensor, and wind speed values ​​are collected through a three-cup anemometer. The above analog signals are converted into digital signals as environmental meteorological data by a 16-bit A / D converter. A fixed-focus camera installed on the top of the photovoltaic support with its lens optical axis perpendicular to the surface of the module is activated. The shooting resolution is set to 1920×1080 pixels, and the aperture value is adjusted to F / 2.8 to ensure the amount of light entering. The shutter is triggered at the same moment in each data acquisition cycle to capture a monitoring image of the module surface that reflects the cleanliness of the module glass cover. The pixel grayscale distribution of this image directly corresponds to the dust distribution texture information.

[0059] In this embodiment, the steps of using an edge computing gateway to perform timestamp alignment and outlier cleaning on real-time running data include: extracting the first acquisition timestamp of electrical performance monitoring data, the second acquisition timestamp of environmental meteorological data, and the third acquisition timestamp of component surface monitoring images; using the third acquisition timestamp as a reference time axis, mapping the electrical performance monitoring data and environmental meteorological data to the reference time axis using a linear interpolation algorithm to generate a time synchronization data frame; filtering the noise data in the time synchronization data frame using a statistical outlier removal algorithm; and vectorizing and encapsulating the filtered electrical performance monitoring data, environmental meteorological data, and component surface monitoring images to obtain a standard multimodal input vector.

[0060] Specifically, the process involves using an edge computing gateway to perform timestamp alignment and outlier cleanup on real-time operational data. This includes parsing the header of received data packets and reading the first acquisition timestamp (when electrical performance monitoring data was generated), the second acquisition timestamp (when environmental meteorological data was generated), and the third acquisition timestamp (when component surface monitoring images were captured). Considering the high latency of image transmission, the third acquisition timestamp is selected as the unique reference timeline. The process then iterates through the first and second acquisition timestamps to find the data point preceding each time point on the reference timeline. Data points at the next time step Applying linear interpolation formula ,in The mapped value, As the reference time point, and The timestamps of the preceding and following data are used to uniformly map the electrical performance monitoring data and environmental meteorological data to the reference time axis in the time dimension, generating time-synchronized data frames that strictly correspond in time for each modality of data. Then, the Raida criterion, i.e., 3... The criteria validate the values ​​within the time-synchronized data frame and calculate the average value of each parameter over a sliding window over the past hour. with standard deviation Determine if the current value falls within the interval The data outside of this range is marked as noise and replaced using the mean of the sliding window. Finally, the cleaned numerical electrical performance data and environmental data are processed by Min-Max normalization and mapped to the 0 to 1 range. The image data is kept in three-dimensional tensor format. All processed data are spliced ​​or structuredly combined according to a predetermined channel order to obtain a standard multimodal input vector.

[0061] In this embodiment, the steps of extracting electrical feature maps, environmental correlation features, and visual texture features corresponding to the standard multimodal input vector using multiple parallel branches in the multimodal feature extraction model include: using a convolutional neural network branch to perform convolutional pooling on the component surface monitoring image contained in the standard multimodal input vector to extract visual texture features, which characterize the dust occlusion distribution on the component surface; using a long short-term memory network branch to perform time-series dependency analysis on the electrical performance monitoring data contained in the standard multimodal input vector to extract electrical feature maps, which characterize the trend information of power decay over time; and using a fully connected neural network branch to perform nonlinear mapping on the environmental meteorological data contained in the standard multimodal input vector to extract environmental correlation features, which characterize the theoretical influence benchmark of environmental parameters on power generation efficiency.

[0062] Specifically, the multi-parallel branches of the multimodal feature extraction model are used to extract electrical feature maps, environmental correlation features, and visual texture features corresponding to the standard multimodal input vector. A multi-branch architecture for the deep learning model is constructed. In the visual branch, the component surface monitoring image contained in the standard multimodal input vector is input into a convolutional neural network. This network consists of three consecutive convolutional blocks, each containing a 3×3 convolutional kernel layer, a ReLU activation layer, and a 2×2 max-pooling layer. Edge, speckle, and texture gradient information in the image are extracted through layer-by-layer convolution operations. The final feature map is flattened into a one-dimensional vector to extract visual texture features. These visual texture features represent the dust occlusion distribution on the component surface in a high-dimensional vector form. In the electrical branch, the past 12 data points contained in the standard multimodal input vector are extracted. The electrical performance monitoring data sequence at each time step is input into a Long Short-Term Memory (LSTM) network containing two hidden layers. The number of hidden layer units is set to 64. A forget gate is used to control the retention of historical information. The current electrical state is updated using an input gate. The output of the hidden state at the last time step is taken as the extracted electrical feature map. This electrical feature map represents the trend of power decay over time in the form of a time-series vector. In the environmental branch, the current environmental meteorological data contained in the standard multimodal input vector is input into a fully connected neural network. This network contains an input layer, two hidden layers with 32 and 16 neurons respectively, and an output layer. Nonlinear mapping is performed through weight matrix multiplication and bias addition operations of the fully connected layer to extract environmental correlation features. These environmental correlation features represent the theoretical impact benchmark of environmental parameters on power generation efficiency in the form of a numerical vector.

[0063] In this embodiment, the step of using the attention fusion layer in the multimodal feature extraction model to perform cross-modal interactive calculations on electrical feature maps, environmental correlation features, and visual texture features includes: calculating the cross-correlation weight matrix between visual texture features, electrical feature maps, and environmental correlation features using a multi-head attention mechanism; weighting and fusing the electrical feature maps, environmental correlation features, and visual texture features according to the cross-correlation weight matrix to obtain a multimodal fusion feature vector; inputting the multimodal fusion feature vector into the regression prediction layer; and calculating the photovoltaic module dust accumulation health index through mapping within the regression prediction layer.

[0064] Specifically, by utilizing the attention fusion layer in the multimodal feature extraction model to perform cross-modal interactive computation on electrical feature maps, environmental association features, and visual texture features, a multi-head attention mechanism module is constructed to project visual texture features, electrical feature maps, and environmental association features into query vectors, respectively. Key vector Sum value vector ,calculate and dot product and divide by scaling factor ,in If the dimension of the key vector is set to 64, and then normalized using the Softmax function, a cross-correlation weight matrix is ​​obtained. This matrix reflects the dependencies and importance between modal features. For example, the weight of visual features is reduced in extremely low light conditions. Based on the cross-correlation weight matrix, the value vector is adjusted... Weighted summation is performed, that is, adaptive weighted fusion of electrical feature map, environmental correlation feature and visual texture feature using the calculated weight coefficients. The weighted results of the three modalities are concatenated to obtain a multimodal fusion feature vector containing global information. Finally, the multimodal fusion feature vector is input into the regression prediction layer, which consists of a fully connected layer and a sigmoid activation function, mapping the high-dimensional feature vector to a scalar value between 0 and 1. The dust accumulation health index of photovoltaic module is calculated through the mapping in the regression prediction layer.

[0065] In this embodiment, the formula for calculating the dust accumulation health index of the photovoltaic module by the regression prediction layer is as follows: ,in, Indicates the dust accumulation health index of photovoltaic modules. This represents the normalized scalar value output by the regression prediction layer based on visual texture feature mapping. This represents the mean of the feature responses output by the regression prediction layer based on the electrical feature map mapping. This represents the intensity scalar value output by the regression prediction layer based on the environmental association feature mapping. Represents the visual feature weight coefficients. Represents the electrical characteristic weighting coefficient. Indicates the environmental characteristic weighting coefficient. Represents the bias constant. It represents the base of the natural logarithm.

[0066] Specifically, the formula for calculating the dust accumulation health index of photovoltaic modules using the regression prediction layer is as follows: ,in, The dust accumulation health index of photovoltaic modules ranges from 0 to 1, with values ​​closer to 1 indicating cleaner modules. This represents the normalized scalar value output by the regression prediction layer based on visual texture feature mapping, representing the cleanliness score of the image dimension. This represents the mean of the feature responses output by the regression prediction layer based on the electrical feature map mapping, which indicates the health score in the electrical performance dimension. This represents the intensity scalar value output by the regression prediction layer based on the environmental correlation feature mapping, signifying the working condition score in the environmental dimension. This represents the visual feature weight coefficient, which is automatically optimized using gradient descent during model training. For example, the initial value is set to 0.4. This represents the electrical characteristic weighting coefficient, with an initial value set to 0.35. This represents the environmental characteristic weighting coefficient, with an initial value set to 0.25. This represents the bias constant, used to adjust the center position of the activation function; for example, it can be set to 0.1. It represents the base of the natural logarithm, approximately equal to 2.71828.

[0067] In this embodiment, after comparing the photovoltaic module's dust accumulation health index with a preset set of dust accumulation level thresholds, the method further includes: calculating the rate of change of the photovoltaic module's dust accumulation health index within a preset time window, determining whether the rate of change is greater than a preset dust accumulation rate threshold, and generating an early dust accumulation warning signal when the rate of change is greater than the dust accumulation rate threshold. The early dust accumulation warning signal is used to indicate that there is a thin layer of rapid dust accumulation on the surface of the photovoltaic module.

[0068] Specifically, after comparing the photovoltaic module dust accumulation health index with a preset set of dust accumulation level thresholds, a time window of 24 hours is set. The time series data of the photovoltaic module dust accumulation health index within this window is extracted, and the slope is calculated by linear fitting of the time series using the least squares method. The absolute value of the slope is taken as the rate of change of the photovoltaic module dust accumulation health index within the preset time window. The dust accumulation rate threshold is set to 0.05 per hour. This threshold is based on the statistical analysis of dust accumulation data under historical sandstorm weather. It is determined whether the rate of change is greater than the preset dust accumulation rate threshold. If the rate of change is less than or equal to the threshold, monitoring continues. When the rate of change is greater than the dust accumulation rate threshold, it indicates that the cleanliness of the module surface has dropped sharply in a short period of time, generating an early dust accumulation warning signal. The early dust accumulation warning signal is used to indicate that there is a thin layer of rapid dust accumulation on the surface of the photovoltaic module.

[0069] In this embodiment, vibration frequency data when the cleaning robot arm contacts the photovoltaic module is collected, the vibration frequency data is input to the edge computing gateway, vibration damping features are extracted, and the dust accumulation health index of the photovoltaic module is corrected using the vibration damping features to obtain the corrected dust accumulation health index.

[0070] Specifically, vibration frequency data is collected when the robotic arm contacts the photovoltaic modules. A triaxial accelerometer is installed at the end of the robotic arm to record the time-domain acceleration signal of the robotic arm when performing the cleaning action and contacting the module surface at a sampling rate of 1000Hz. The time-domain signal is processed by Fast Fourier Transform (FFT) to obtain the frequency domain spectrum. The vibration frequency data is input to the edge computing gateway to identify the main resonance peak frequency in the frequency domain spectrum. The damping ratio is calculated using the half-power bandwidth method, i.e., the frequency at which the resonance peak amplitude is located. The corresponding frequency width is divided by twice the resonant frequency, and the calculated damping ratio is used as the extracted vibration damping feature. Considering that dust accumulation will increase the viscous damping of the contact surface, the vibration damping feature is used to correct the dust accumulation health index of the photovoltaic module, and the corrected dust accumulation health index is obtained.

[0071] In this embodiment, the formula for correcting the dust accumulation health index of photovoltaic modules using vibration damping characteristics is as follows: ,in, This indicates the corrected health index for dust accumulation. This indicates the dust accumulation health index of photovoltaic modules before the correction. Indicates the vibration correction factor. The normalized amplitude energy represents the vibration damping characteristic. This represents a logarithmic function with base 10. Historical distribution data of the dust accumulation health index of photovoltaic modules are statistically analyzed in the edge computing gateway. When the prediction deviation in the historical distribution data exceeds the preset deviation threshold, the cloud-edge collaborative update mechanism is triggered. The updated model parameters are received from the cloud server and the multimodal feature extraction model is updated using the updated model parameters.

[0072] Specifically, the formula for correcting the dust accumulation health index of photovoltaic modules using vibration damping characteristics is as follows: ,in, This indicates the corrected health index for dust accumulation. This indicates the dust accumulation health index of photovoltaic modules before the correction. This represents the vibration correction factor, for example, a value of 0.15, used to control the weight of the vibration characteristics on the final index. The normalized amplitude energy, representing the vibration damping characteristic, is a dimensionless value obtained by dividing the calculated damping ratio by a preset maximum standard damping ratio. This represents a logarithmic function with base 10. A local database is established in the edge computing gateway to store and statistically analyze the historical distribution data of the dust accumulation health index of photovoltaic modules in chronological order. The mean square error (MSE) between the historical predicted value and the true cleanliness value fed back after actual cleaning is calculated. When the prediction deviation in the historical distribution data exceeds the preset deviation threshold, such as MSE being greater than 0.1, an update request is sent to the cloud to trigger the cloud-edge collaborative update mechanism. The updated model parameters are received from the cloud server, which has retrained and distributed the updated model parameters using massive amounts of data. The updated model parameters are used to replace the original weights and biases in the multimodal feature extraction model, thus updating the multimodal feature extraction model.

[0073] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. An online detection method for dust accumulation status of photovoltaic modules based on multimodal fusion and edge computing, characterized in that, Includes the following steps: Collect real-time operating data of photovoltaic modules, including electrical performance monitoring data, environmental meteorological data, and monitoring images of the module surface; The real-time running data is transmitted to an edge computing gateway, which performs timestamp alignment and outlier cleaning on the real-time running data to obtain a standard multimodal input vector. The standard multimodal input vector is input into the multimodal feature extraction model deployed on the edge computing gateway, and the electrical feature map, environmental correlation feature and visual texture feature corresponding to the standard multimodal input vector are extracted by the multi-path parallel branches in the multimodal feature extraction model respectively. The attention fusion layer in the multimodal feature extraction model is used to perform cross-modal interactive calculations on the electrical feature map, the environmental correlation features, and the visual texture features to generate a photovoltaic module dust accumulation health index. The online dust accumulation status of the photovoltaic module is determined by comparing the dust accumulation health index of the photovoltaic module with a preset set of dust accumulation level thresholds.

2. The online detection method for dust accumulation status of photovoltaic modules based on multimodal fusion and edge computing according to claim 1, characterized in that, The steps for collecting real-time operating data of photovoltaic modules include: The electrical performance monitoring data is acquired by a smart data acquisition device connected to the photovoltaic string. The electrical performance monitoring data includes string-level current and voltage curves, real-time output power, and fill factor parameters. The environmental meteorological data is acquired through a meteorological station sensor array, including irradiance, ambient temperature, relative humidity, and wind speed. A monitoring image of the photovoltaic module surface is acquired by a fixed-focus camera aimed at the module surface, and the monitoring image of the module surface includes dust distribution texture information.

3. The online detection method for dust accumulation status of photovoltaic modules based on multimodal fusion and edge computing according to claim 2, characterized in that, The steps of using the edge computing gateway to perform timestamp alignment and outlier cleanup on the real-time running data include: Extract the first acquisition timestamp of the electrical performance monitoring data, the second acquisition timestamp of the environmental meteorological data, and the third acquisition timestamp of the component surface monitoring image; Using the third acquisition timestamp as the reference time axis, a linear interpolation algorithm is used to map the electrical performance monitoring data and the environmental meteorological data to the reference time axis to generate a time-synchronized data frame; The noise data in the time synchronization data frame is filtered using a statistical outlier removal algorithm, and the filtered electrical performance monitoring data, environmental meteorological data, and component surface monitoring images are vectorized and encapsulated to obtain the standard multimodal input vector.

4. The online detection method for dust accumulation status of photovoltaic modules based on multimodal fusion and edge computing according to claim 3, characterized in that, The steps of extracting the electrical feature map, environmental correlation features, and visual texture features corresponding to the standard multimodal input vector using the multi-path parallel branches in the multimodal feature extraction model include: The component surface monitoring image contained in the standard multimodal input vector is convolutionally pooled using a branch of a convolutional neural network to extract the visual texture features, which characterize the dust occlusion distribution on the component surface. The electrical performance monitoring data contained in the standard multimodal input vector is analyzed for time-series dependence using a branch of a long short-term memory network, and the electrical feature map is extracted. The electrical feature map represents the trend information of power decay over time. The environmental meteorological data contained in the standard multimodal input vector are nonlinearly mapped using a fully connected neural network branch to extract the environmental correlation features, which characterize the theoretical impact benchmark of environmental parameters on power generation efficiency.

5. The online detection method for dust accumulation status of photovoltaic modules based on multimodal fusion and edge computing according to claim 4, characterized in that, The steps of performing cross-modal interactive computation on the electrical feature map, the environmental association features, and the visual texture features using the attention fusion layer in the multimodal feature extraction model include: A multi-head attention mechanism is used to calculate the cross-correlation weight matrix among the visual texture features, the electrical feature map, and the environmental correlation features; The electrical feature map, the environmental correlation feature, and the visual texture feature are weighted and fused according to the cross-correlation weight matrix to obtain a multimodal fusion feature vector; The multimodal fusion feature vector is input into the regression prediction layer, and the dust accumulation health index of the photovoltaic module is calculated through the mapping in the regression prediction layer.

6. The online detection method for dust accumulation status of photovoltaic modules based on multimodal fusion and edge computing according to claim 5, characterized in that, The formula for calculating the dust accumulation health index of the photovoltaic module by the regression prediction layer is as follows: ,in, This indicates the dust accumulation health index of the photovoltaic module. This represents the normalized scalar value output by the regression prediction layer based on the visual texture feature map. This represents the mean of the feature response output by the regression prediction layer based on the electrical feature map mapping. This represents the intensity scalar value output by the regression prediction layer based on the environmental association feature mapping. Represents the visual feature weight coefficients. Represents the electrical characteristic weighting coefficient. Indicates the environmental characteristic weighting coefficient. Represents the bias constant. It represents the base of the natural logarithm.

7. The online detection method for dust accumulation status of photovoltaic modules based on multimodal fusion and edge computing according to claim 1, characterized in that, After comparing the photovoltaic module's dust accumulation health index with a preset set of dust accumulation level thresholds, the method further includes: Calculate the rate of change of the dust accumulation health index of the photovoltaic module within a preset time window; Determine whether the rate of change is greater than a preset ash accumulation rate threshold; When the rate of change is greater than the dust accumulation rate threshold, an early dust accumulation warning signal is generated. The early dust accumulation warning signal is used to indicate that there is a thin layer of rapid dust accumulation on the surface of the photovoltaic module.

8. The online detection method for dust accumulation status of photovoltaic modules based on multimodal fusion and edge computing according to claim 1, characterized in that, The method further includes: Collect vibration frequency data when the cleaning robotic arm comes into contact with the photovoltaic module; The vibration frequency data is input into the edge computing gateway to extract vibration damping features; The dust accumulation health index of the photovoltaic module is corrected by using the vibration damping characteristics to obtain the corrected dust accumulation health index.

9. The online detection method for dust accumulation status of photovoltaic modules based on multimodal fusion and edge computing according to claim 8, characterized in that, The formula for correcting the dust accumulation health index of the photovoltaic module using the vibration damping characteristics is as follows: ,in, This represents the corrected dust accumulation health index. This indicates the dust accumulation health index of the photovoltaic modules before the correction. Indicates the vibration correction factor. The normalized amplitude energy represents the vibration damping characteristic. Represents the logarithmic function with base 10; The method further includes: The historical distribution data of the dust accumulation health index of the photovoltaic modules are statistically analyzed in the edge computing gateway; When the prediction deviation in the historical distribution data exceeds a preset deviation threshold, the cloud-edge collaborative update mechanism is triggered. Receive updated model parameters from the cloud server and update the multimodal feature extraction model using the updated model parameters.

10. The system for online detection of dust accumulation status of photovoltaic modules based on multimodal fusion and edge computing according to any one of claims 1-9, characterized in that, include: The data acquisition module is used to collect real-time operating data of the photovoltaic module, including electrical performance monitoring data, environmental meteorological data, and monitoring images of the module surface. The data preprocessing module is used to transmit the real-time running data to the edge computing gateway, and use the edge computing gateway to perform timestamp alignment and outlier cleaning on the real-time running data to obtain a standard multimodal input vector. The feature extraction module is used to input the standard multimodal input vector into the multimodal feature extraction model deployed on the edge computing gateway, and use the multi-parallel branches in the multimodal feature extraction model to extract the electrical feature map, environmental correlation features and visual texture features corresponding to the standard multimodal input vector respectively; The feature fusion module is used to perform cross-modal interactive calculations on the electrical feature map, the environmental correlation features, and the visual texture features using the attention fusion layer in the multimodal feature extraction model to generate a photovoltaic module dust accumulation health index. The status determination module is used to compare the dust accumulation health index of the photovoltaic module with a preset set of dust accumulation level thresholds to determine the online dust accumulation status of the photovoltaic module.

Images