Intelligent photovoltaic power station operation and maintenance management method and system

Abstract

The invention relates to the technical field of renewable energy sources, in particular to an intelligent photovoltaic power station operation and maintenance management method and system, and the method comprises the steps: obtaining multi-modal data, such as a hyperspectral image and an electroluminescent image, and carrying out the time-space registration; cross-modal feature fusion is realized through an attention mechanism; environmental stress and material degradation influence are separated by using a variational auto-encoder, and an aging feature map is generated in combination with a three-dimensional convolutional neural network; and generating a dynamic maintenance path based on the topology constraint graph and a reinforcement learning algorithm, and adjusting the priority of the path through real-time data. The system can comprehensively capture the aging characteristics of the photovoltaic module, improves the aging prediction precision and maintenance efficiency, reduces the operation and maintenance cost, and promotes the intelligent management of a photovoltaic power station.

Description

Technical Field

[0001] The present invention belongs to the technical field of photovoltaic power station operation and maintenance management, and specifically relates to an intelligent photovoltaic power station operation and maintenance management method and system. Background Art

[0002] The aging of photovoltaic modules is a key factor affecting the efficiency and economic benefits of photovoltaic power plants, but the traditional single data collection method is difficult to fully and accurately evaluate the aging status of modules. How to effectively fuse multimodal data to achieve accurate evaluation of the aging status of photovoltaic modules is a technical problem that needs to be solved urgently. This involves technical difficulties in many aspects: First, there are differences in the acquisition time and space of different modal data (such as hyperspectral images, electroluminescent images, etc.), and how to achieve accurate spatiotemporal registration is a basic problem. Secondly, different modal data reflect the characteristics of the module at different angles. How to effectively associate and fuse these feature information to fully characterize the status of the module is the core difficulty. Furthermore, the aging of the module is affected by multiple factors such as environmental stress and intrinsic degradation of the material. How to decouple these factors and construct an accurate three-dimensional aging gradient distribution model is the key. Finally, the optimal maintenance strategy based on the aging assessment results needs to comprehensively consider multiple factors such as the topological relationship of the module and the maintenance cost. How to design an efficient path planning algorithm is also an important technical difficulty. The solution to these problems is of great significance to improving the operation and maintenance efficiency and economic benefits of photovoltaic power plants. Summary of the invention

[0003] On the one hand, the present invention provides an intelligent photovoltaic power station operation and maintenance management method, which mainly includes:

[0004] Multimodal data of photovoltaic modules are collected by autonomous mobile robots. The multimodal data include hyperspectral images, electroluminescent images and module location information. The multimodal data are temporally and spatially aligned, and the hyperspectral chemical features and electroluminescent physical defect features are associated through a cross-modal attention fusion network. The environmental stress and material intrinsic degradation factors are decoupled based on the conditional generative adversarial network and variational autoencoder to generate a three-dimensional aging gradient distribution model. Combining the aging gradient distribution with the module topological relationship, an improved ant colony algorithm is used to output a maintenance path planning scheme with dynamic weights.

[0005] Furthermore, the specific analysis process of collecting multimodal data of photovoltaic modules by autonomous mobile robots includes: obtaining hyperspectral images of photovoltaic modules in the range of 400-2500nm, focusing on the 420-680nm band sensitive to material chemical degradation and the 950-1150nm band sensitive to carrier concentration; synchronously collecting electroluminescent images, using a pulse current mode to inject 80% to 120% of the short-circuit current to ensure clear visualization of defective areas; recording the location information of the modules, and forming a spatial distribution map of the modules based on the latitude and longitude coordinate system of the photovoltaic array.

[0006] Furthermore, the specific implementation process of the cross-modal attention fusion network includes: calculating the hyperspectral feature query vector Electroluminescence characteristic bond vector The attention weight is as follows: ;

[0007] in, Indicates The spectral channels and The correlation strength of the defect area, is the total number of defective areas, Represents the query vector for hyperspectral features The transposed matrix of .

[0008] Furthermore, the specific method for decoupling environmental stress and material intrinsic degradation factor based on conditional generative adversarial network and variational autoencoder includes: defining the loss function of variational autoencoder, the formula is as follows: ;

[0009] in, is a hyperparameter that controls the strength of the decoupling. Represents KL divergence, which is used to constrain the consistency of latent variable distribution and prior distribution; is the encoder output distribution; is the prior distribution, the latent variable Characterize the intrinsic degradation factor of materials; represents the log-likelihood of the decoder reconstruction, which is used to measure the accuracy of reconstructing the input data x from the latent variable z; represents the distribution of the output from the encoder Sample hidden variables z in the decoder and calculate the decoder reconstruction data expected value.

[0010] Furthermore, the dynamic volatility factor update rule of the improved ant colony algorithm is: ;

[0011] in, is the initial volatility coefficient, is the dynamic volatilization factor at the tth iteration, which controls the volatilization rate of pheromone, t is the current iteration number, is the total number of iterations, is a decay exponent that adjusts the nonlinear change of the volatilization rate.

[0012] Furthermore, the specific implementation process of spatiotemporal registration includes: aligning the hyperspectral image and the electroluminescent image by affine transformation based on the longitude and latitude coordinate system of the photovoltaic array; using the SIFT feature point matching algorithm to eliminate the spatial offset error of data collected at different times to ensure the spatial consistency of multi-source data.

[0013] Furthermore, the specific analysis process of the three-dimensional aging gradient distribution model includes: gridding the environmental stress parameters, splicing them with the intrinsic degradation factors into a three-dimensional input tensor, and generating the aging gradient distribution model in the temperature-humidity-irradiance space through three-dimensional convolutional neural network mapping.

[0014] Furthermore, the specific design process of the maintenance path planning scheme includes: generating a maintenance priority queue according to the component aging gradient value; assigning path weight coefficients to adjacent components based on the maintenance priority queue and the topological distance between components, and the weight between components with higher priorities and closer distances is greater; using the dynamic volatility factor to adjust the pheromone concentration distribution in real time according to the path weight coefficient, wherein the pheromone accumulation rate of high-weight paths is improved; outputting the maintenance path sequence through iterative optimization, and automatically matching the required maintenance tool list based on the component defect type.

[0015] On the other hand, the present invention provides an intelligent photovoltaic power station operation and maintenance management system that applies any of the above-mentioned intelligent photovoltaic power station operation and maintenance management methods, including an autonomous mobile robot, a data acquisition module, a data processing module, a path planning module and a human-computer interaction terminal; the autonomous mobile robot is equipped with a hyperspectral camera, an EL detector and a positioning module for performing multimodal data acquisition; the data processing module includes a spatiotemporal registration unit, a cross-modal attention fusion network and a degenerate decoupling model for deep correlation and decoupling analysis; the path planning module is configured to output a dynamic maintenance path using an improved ant colony algorithm; the human-computer interaction terminal is used to display a maintenance path plan; the autonomous mobile robot establishes a data connection with the data processing module and the path planning module through a wireless communication module.

[0016] The technical solution provided by the embodiment of the present invention may have the following beneficial effects:

[0017] The present invention discloses a method and system for intelligent maintenance of photovoltaic components. An autonomous mobile robot collects multimodal data such as hyperspectral images, electroluminescent images, and position information of photovoltaic components. After spatiotemporal registration of the data, a cross-modal attention fusion network is used to associate different modal features. Then, a three-dimensional aging gradient distribution model is generated based on a conditional generative adversarial network and a variational autoencoder to decouple environmental stress and material intrinsic degradation factors. The aging gradient distribution is combined with the component topological relationship, and an improved ant colony algorithm is used to output a maintenance path planning scheme with dynamic weights. The present invention realizes deep fusion analysis of multimodal data of photovoltaic components, accurately evaluates the aging status of components, and provides an optimized maintenance path, effectively improving the operation and maintenance efficiency and economic benefits of photovoltaic power stations. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 : The system architecture diagram of the present invention includes the connection relationship and data flow between the autonomous mobile robot, the data acquisition module, the data processing module, the path planning module and the human-computer interaction terminal.

[0019] Figure 2 : Schematic diagram of the dynamic maintenance path planning process based on the improved ant colony algorithm. DETAILED DESCRIPTION

[0020] In order to enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below in conjunction with the drawings in the embodiments of this specification. Obviously, the described embodiments are only part of the embodiments of this specification, not all of the embodiments. Based on the embodiments in this specification, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of this specification.

[0021] like Figure 1 In this embodiment, an intelligent photovoltaic power station operation and maintenance management method may specifically include:

[0022] Step S101, collect multimodal data of photovoltaic components through an autonomous mobile robot, the multimodal data including hyperspectral images, electroluminescent images and component location information; perform spatiotemporal registration on the multimodal data, and associate hyperspectral chemical features with electroluminescent physical defect features through a cross-modal attention fusion network; decouple environmental stress and material intrinsic degradation factors based on a conditional generative adversarial network and a variational autoencoder to generate a three-dimensional aging gradient distribution model; combine the aging gradient distribution with the component topology relationship, and use an improved ant colony algorithm to output a maintenance path planning scheme with dynamic weights.

[0023] The multimodal data of photovoltaic modules are collected by autonomous mobile robots. The multimodal data includes hyperspectral images, electroluminescence images and module location information. The autonomous mobile robot is equipped with a hyperspectral camera and an electroluminescence detector, which can efficiently collect data at the photovoltaic power station site. The hyperspectral camera can capture the spectral information of photovoltaic modules in the range of 400-2500nm, especially the 420-680nm band sensitive to material chemical degradation and the 950-1150nm band sensitive to carrier concentration. The electroluminescence detector is used to obtain the electroluminescence image of the photovoltaic module, which can reveal the physical defects inside the module. At the same time, the positioning module records the location information of the module based on the longitude and latitude coordinate system of the photovoltaic array to form a spatial distribution map of the module. This multimodal data collection method can fully capture the chemical degradation and physical defect characteristics of photovoltaic modules, and provide rich data support for the subsequent construction of a three-dimensional aging gradient distribution model. The multimodal data is spatiotemporally aligned, and the hyperspectral chemical features and electroluminescence physical defect features are associated through a cross-modal attention fusion network. The spatiotemporal registration unit first performs affine transformation to align the hyperspectral image with the electroluminescent image, and uses the SIFT feature point matching algorithm to eliminate the spatial offset error of data collected at different times to ensure the spatial consistency of multi-source data. Then, the cross-modal attention fusion network calculates the hyperspectral feature query vector Electroluminescence characteristic bond vector The attention weights are used to quantify the degree of correlation between different modal features. For example, a spectral channel in a hyperspectral image may be strongly correlated with a defective area in an electroluminescent image, and this correlation can help locate faults more accurately. Through this cross-modal deep correlation analysis, the coupling law between spectral features and physical defects can be effectively extracted to improve the accuracy of fault location. Based on the conditional generative adversarial network and variational autoencoder, the environmental stress and material intrinsic degradation factors are decoupled to generate a three-dimensional aging gradient distribution model.

[0024] Conditional generative adversarial networks and variational autoencoders can separate environmental stress (such as temperature and humidity fluctuations) from intrinsic material degradation (such as EVA yellowing and battery cracking) factors to generate a three-dimensional aging gradient distribution model. For example, by optimizing the loss function, a three-dimensional aging gradient distribution model can be generated to show the aging degree of each component and its impact on the overall performance. This model can provide a reliable basis for life prediction, helping operation and maintenance personnel to formulate maintenance plans in advance to avoid performance degradation caused by component aging. Combining the aging gradient distribution with the component topological relationship, the improved ant colony algorithm is used to output a maintenance path planning scheme with dynamic weights. The improved ant colony algorithm can comprehensively determine the optimal maintenance path based on the dynamic priority queue of the component health index, the path weight coefficient positively correlated with the aging gradient value, and the real-time updated pheromone concentration distribution map. For example, in the actual application of a large ground photovoltaic power station, after the autonomous mobile robot completes data collection, the data processing module generates a detailed three-dimensional aging gradient distribution map, showing the aging degree of each component and its impact on the overall performance. Based on this distribution map, the path planning module outputs an optimal maintenance path, giving priority to covering the component area with a higher degree of aging. This dynamic weighted maintenance path planning solution can optimize resource allocation, reduce path redundancy, and significantly improve operation and maintenance efficiency and economy.

[0025] Step S102, the specific analysis process of collecting multimodal data of photovoltaic modules by an autonomous mobile robot includes: obtaining hyperspectral images of photovoltaic modules in the range of 400-2500nm, focusing on the 420-680nm band sensitive to material chemical degradation and the 950-1150nm band sensitive to carrier concentration; synchronously collecting electroluminescent images, using a pulse current mode to inject 80% to 120% of the short-circuit current to ensure clear visualization of defective areas; recording the location information of the modules, and forming a spatial distribution map of the modules based on the latitude and longitude coordinate system of the photovoltaic array.

[0026] The specific analysis process of collecting multi-modal data of photovoltaic modules by an autonomous mobile robot includes: obtaining hyperspectral images of photovoltaic modules, with the acquisition range being 400 - 2500 nm, focusing on the 420 - 680 nm band sensitive to material chemical degradation and the 950 - 1150 nm band sensitive to carrier concentration. The hyperspectral image acquisition is achieved by a hyperspectral camera mounted on the robot, which can capture the reflection spectral information of the photovoltaic module at different wavelengths. In the 420 - 680 nm band, the hyperspectral image can reflect the characteristics of material chemical degradation, such as the aging of polymer encapsulation materials and the yellowing of backsheet materials. While in the 950 - 1150 nm band, the hyperspectral image can capture the change of carrier concentration, thus indirectly reflecting the electrical performance degradation of the cell. For example, when the carrier concentration decreases, the open-circuit voltage and short-circuit current of the cell will be affected, resulting in a decrease in power generation efficiency. By analyzing the spectral data of these bands, the aging degree of the photovoltaic module can be quantified, providing a basis for subsequent maintenance path planning.

[0027] Simultaneously collect electroluminescence images, injecting 80% to 120% of the short-circuit current in pulse current mode to ensure the clear manifestation of defect areas. Electroluminescence (EL) detection is achieved by injecting a pulse current into the photovoltaic module, making it emit a weak light signal under dark field conditions, thereby capturing the defects inside the cell. For example, when the injected current is 100% of the short-circuit current, the cell will emit a uniform light signal, while the cells with cracks, broken grids or hidden cracks will appear as dark or bright areas in the EL image. By adjusting the injection current range (80% to 120%), the contrast of the defect area can be enhanced, ensuring that even tiny defects can be clearly identified. For example, when the injection current is 120%, the luminescence intensity of the hidden crack area will be significantly reduced, making it easier to be detected. The combination of EL images and hyperspectral images can comprehensively evaluate the health status of photovoltaic modules, providing more accurate data support for maintenance decisions. Record the position information of the module, based on the longitude and latitude coordinate system of the photovoltaic array, to form a spatial distribution map of the modules. The acquisition of position information is achieved by a positioning module mounted on the robot, which can accurately record the longitude and latitude coordinates of each photovoltaic module.

[0028] For example, in a large-scale photovoltaic power station, the coordinate information of each module will be recorded and mapped into the global coordinate system of the power station to form a complete spatial distribution map. This map can not only show the position relationship of the modules, but also combine the aging gradient distribution and the module topological relationship to provide a spatial reference for maintenance path planning. For example, when the aging gradient value in a certain area is relatively high, the robot can preferentially plan the maintenance path in this area, thus improving the maintenance efficiency. By combining the position information with multi-modal data, the refined management of photovoltaic power station operation and maintenance can be realized, ensuring the efficient use of resources.

[0029] Step S103, the specific implementation process of the cross-modal attention fusion network includes: calculating the hyperspectral feature query vector and the electroluminescence feature key vector of the attention weight, the formula is as follows: ;

[0030] wherein, represents the association strength between the th spectral channel and the th defect region, is the total number of defect regions, represents the transposed matrix of the hyperspectral feature query vector .

[0031] The calculation process of the attention weight between the hyperspectral feature query vector and the electroluminescence feature key vector is essentially to mine the implicit relationship between the spectral features and physical defects through cross-modal association. Hyperspectral data contains spectral information of hundreds of continuous bands, and each band corresponds to the reflectance characteristics of different chemical components, while the electroluminescence image reflects the defect distribution caused by the carrier recombination inside the component. The attention mechanism quantifies the association strength between the spectral channel and the defect region by calculating the dot product similarity between the query vector and the key vector.

[0032] For example, when the reflectance of a certain component abnormally increases in the 900nm band, if the dot product value between the query vector of this band and the key vector of the hidden crack region in the electroluminescence image reaches 0.85 (the maximum normalized value is 1), it indicates that there is a strong correlation between the spectral features of this band and the hidden crack defect. In specific implementation, the hyperspectral query vector is extracted through a convolutional neural network, and its dimension is [256×1], and the electroluminescence key vector is extracted through the ResNet50 network, and the dimension is [256×N], where N is the number of defect regions. When calculating the attention weight, the query vector is transposed and multiplied by the key vector matrix, and then normalized by Softmax.

[0033] For example, a certain photovoltaic module detects 3 defect regions, and the dimension of its key vector matrix is [256×3]. After calculating with the query vector, the three attention weight values obtained are 0.6, 0.3, and 0.1 respectively, indicating that the first spectral channel has the strongest correlation with the first defect region. This weight distribution can be explained as: the absorption peak shift of silicon material in the 1200nm band is physically coupled with the carrier leakage caused by the hidden crack of the cell, so the weight of this band is significantly higher than other bands. The practical significance of the association strength is to locate the chemical cause of the defect.

[0034] For example, when a component shows local dark spots in an electroluminescent image, and the weight of the hyperspectral data reaches 0.9 in the 650 nm band, combined with the characteristic absorption peak of oxygen precipitates corresponding to this band, it can be inferred that the dark spots are caused by the aggregation of oxygen impurities in the silicon wafer. This cross-modal association avoids misjudgment of a single detection method: relying solely on the electroluminescent image may misjudge a contamination defect as a crack, but combining the spectral weight distribution can accurately distinguish between material intrinsic degradation and environmental factors. The dynamic change of the total number N of defect regions affects the attention weight resolution. When N = 10, the dimension of the weight matrix is [256×10], which can finely match spectral features with micro-defects (such as a single gate line break); while when N = 1, it degenerates into global association, which is suitable for large-area attenuation analysis. In practical applications, the value of N is automatically determined by a clustering algorithm: if 5 connected regions are detected in the electroluminescent image, then N = 5 is set. At this time, each attention weight corresponds to the chemical characteristics of a specific defect cluster.

[0035] For example, the weight distribution with N = 5 in an array shows that the second defect cluster has a weight of 0.7 in the 350 nm band, which corresponds to the degradation of the EVA film caused by ultraviolet aging, thus accurately locking that the backplane material in this area needs to be replaced. In terms of technical effects, this method realizes semantic alignment of multi-modal features through attention weights. Directly splicing hyperspectral and electroluminescent features by traditional methods will cause interference from modal differences, while the cross-modal attention mechanism retains their respective feature spaces and focuses on key associations through weights. Experimental data show that this method increases the defect classification accuracy by 12% and reduces the false detection rate by 8%, especially having a significant recognition effect on composite defects (such as PID effect superimposed with snail patterns).

[0036] Step S104, the specific method for decoupling environmental stress and material intrinsic degradation factors based on a conditional generative adversarial network and a variational autoencoder includes: defining the loss function of the variational autoencoder, and the formula is as follows: ;

[0037] Among them, is a hyperparameter for controlling the decoupling strength, represents the KL divergence, which is used to constrain the consistency between the latent variable distribution and the prior distribution; is the encoder output distribution; is the prior distribution, and the latent variable characterizes the material intrinsic degradation factor; represents the log-likelihood of the decoder reconstruction, which is used to measure the accuracy of reconstructing the input data x from the latent variable z; represents sampling the latent variable z from the distribution output by the encoder and calculating the expected value of the decoder reconstructed data .

[0038] In the specific method of decoupling environmental stress and material intrinsic degradation factors based on conditional generative adversarial networks and variational autoencoders, the loss function of the variational autoencoder is crucial. The design of the loss function aims to balance the difference between the encoder output distribution and the prior distribution by controlling the hyperparameter of the decoupling strength. KL divergence is used to measure the distance between these two distributions, ensuring that the encoder can effectively capture the latent structure of the data. By optimizing the loss function, the model can separate environmental stress and material intrinsic degradation factors, thus generating a more accurate three-dimensional aging gradient distribution model.

[0039] In specific implementation, assume that we have an aging dataset of photovoltaic modules, which contains the effects of environmental stress (such as temperature, humidity, irradiance) and material intrinsic degradation (such as material fatigue, chemical corrosion). The encoder part of the variational autoencoder maps the input data to a latent space, and the decoder reconstructs the data from the latent space. The KL divergence term in the loss function ensures that the distribution in the latent space is close to the prior distribution (usually a standard normal distribution), thus achieving the decoupling of environmental stress and material intrinsic degradation. Since it is impossible to enumerate all combinations of environmental stress (temperature, humidity, irradiance) in actual operation and maintenance to study their effects on module aging, the data from the aging dataset need to be used as input to train the conditional generative adversarial network. After training, the conditional generative adversarial network is used to simulate and generate a degradation dataset covering all working conditions, solving the problem of model overfitting caused by insufficient real data.

[0040] For example, assume that we set the hyperparameter β to 0.5, which means that in the optimization process, the weight of the KL divergence term is 0.5. By adjusting the value of β, the decoupling strength can be controlled. When β is larger, the model is more inclined to make the latent space distribution close to the prior distribution, thus more effectively separating environmental stress and material intrinsic degradation factors. On the contrary, when β is smaller, the model pays more attention to the reconstruction accuracy of the data, which may lead to a weakened decoupling effect. In practical applications, by optimizing the loss function, the model can generate a three-dimensional aging gradient distribution model, which can clearly show the effects of different environmental stress and material intrinsic degradation factors on the aging of photovoltaic modules.

[0041] For example, the model may show that in a high-temperature and high-humidity environment, the rate of material intrinsic degradation accelerates, while in a low-temperature and low-humidity environment, the influence of environmental stress is smaller. This decoupling analysis provides a more reliable basis for the life prediction of photovoltaic modules, helping to formulate more effective maintenance strategies. Through this method, we can not only understand the aging mechanism of photovoltaic modules more accurately, but also optimize resource allocation and reduce unnecessary maintenance costs.

[0042] For example, in a certain photovoltaic power station, the model may predict that some components age faster under specific environmental conditions, and thus prioritize the maintenance of these components to ensure the long-term stable operation of the power station. This data-driven decoupling analysis method provides strong technical support for the intelligent operation and maintenance of photovoltaic power stations.

[0043] Step S105, improve the dynamic evaporation factor update rule of the ant colony algorithm as: ;

[0044] Among them, is the initial evaporation coefficient, is the dynamic evaporation factor at the t-th iteration, which controls the evaporation rate of pheromone, t is the current iteration number, is the total number of iterations, is the attenuation exponent, which adjusts the non-linear change of the evaporation rate.

[0045] The core of the dynamic evaporation factor update rule is to adaptively adjust the evaporation rate of pheromone through the iterative process. Set a relatively high evaporation coefficient (such as 0.8) at the initial stage. At this time, the algorithm tends to quickly explore the solution space and avoid falling into local optima prematurely. As the iteration number t increases, the evaporation coefficient decays according to the rule. When the total number of iterations is set to 100 and γ = 1.2, the at the 50th iteration will drop to 0.8×(1 - 50 / 100)^1.2 ≈ 0.34. At this time, the algorithm gradually turns to local fine search. This dynamic balance mechanism enables the algorithm to retain high-quality path pheromone information in the early stage and eliminate inefficient paths in a timely manner in the later stage. The value of the attenuation rate γ directly affects the algorithm performance. When γ = 0.5, the evaporation coefficient drops smoothly, which is suitable for dealing with scenarios where the component aging distribution is uniform; if there is a centralized defect area in the photovoltaic array (such as a batch of aging due to shadow in a certain quadrant), then γ needs to be increased to more than 1.5 to make the algorithm quickly focus on the key area. Experimental data shows that under the same number of iterations, the path planning scheme with γ = 1.5 has 23% fewer redundant paths than the fixed evaporation factor. The setting of the initial evaporation coefficient needs to be combined with the component space density. For a standard photovoltaic array with a spacing of 8 meters, taking 0.6 - 0.9 can effectively cover the neighborhood search; if the component spacing is compressed to less than 5 meters, then needs to be reduced to 0.4 - 0.6 to avoid excessive diffusion of pheromone. The measured case of a 2MW power station shows that after adopting the dynamic rule, the average moving distance of the maintenance robot is shortened by 37%. Among them, The combination of α = 0.7 and γ = 1.1 converges to the optimal solution within 200 iterations. In the time dimension, the dynamic adjustment of the evaporation factor is positively correlated with the component aging rate. When it is detected that the monthly change rate of the aging gradient in a certain area exceeds 5%, the system automatically increases the value of γ by 20% to accelerate the pheromone update frequency. For example, due to seasonal concentrated aging caused by salt spray corrosion in a coastal power station, the dynamic rule reduces the maintenance response time from 72 hours to 41 hours.

[0046] Step S106, the specific implementation process of spatio-temporal registration includes: performing affine transformation alignment on the hyperspectral image and the electroluminescence image based on the latitude and longitude coordinate system of the photovoltaic array; using the SIFT feature point matching algorithm to eliminate the spatial offset error of the data collected at different times and ensure the spatial consistency of multi-source data.

[0047] The specific implementation process of spatio-temporal registration first performs affine transformation alignment on the hyperspectral image and the electroluminescence image based on the latitude and longitude coordinate system of the photovoltaic array. The latitude and longitude coordinate system of the photovoltaic array is obtained through the GPS positioning module and can accurately describe the position of the photovoltaic components in the geographical space. The hyperspectral image and the electroluminescence image are collected by the hyperspectral camera and the EL detector respectively. However, due to the differences in the installation positions and angles of the devices, there may be offsets in the space of the two images. Affine transformation is a linear transformation method, including operations such as translation, rotation, scaling, and skew, which can align the spatial coordinate systems of the two images.

[0048] For example, assume that the center point coordinates of the hyperspectral image are , and the center point coordinates of the electroluminescence image are . By calculating the translation vector between the two, the electroluminescence image can be translated to align with the hyperspectral image. In addition, if there is a rotation difference between the two images, the rotation angle θ can be calculated to perform a rotation transformation on the electroluminescence image to ensure the spatial consistency of the two images. The SIFT (Scale-Invariant Feature Transform) algorithm is used to eliminate the spatial offset error of the data collected at different times and ensure the spatial consistency of multi-source data. The SIFT algorithm is an image matching method based on local features, which can extract key points in the image and generate feature descriptors. In the operation and maintenance of photovoltaic power stations, due to different data collection times, factors such as environmental light and device position may cause slight offsets between images.

[0049] For example, there is a slight spatial misalignment between the hyperspectral image collected at a certain time and the electroluminescence image collected the previous day. The SIFT algorithm can be used to extract the feature points in both images, such as the edges and corners of photovoltaic modules, and calculate the matching relationships between these feature points. Suppose 100 feature points are extracted from the hyperspectral image and 120 feature points are extracted from the electroluminescence image. 80 pairs of matching points are found through the matching algorithm. Based on the coordinate offsets of these matching points, the electroluminescence image can be accurately corrected spatially to eliminate the offset error. Through the above steps, spatio-temporal registration can ensure that the hyperspectral image and the electroluminescence image are completely aligned spatially, providing a basis for subsequent multimodal data fusion.

[0050] For example, when analyzing the aging degree of photovoltaic modules, the hyperspectral image provides information on material chemical degradation, while the electroluminescence image reflects the change in carrier concentration. If the two images are not aligned spatially, it may lead to inaccurate analysis results. For example, the chemical degradation information of a certain module may be wrongly associated with another module. Through spatio-temporal registration, the spatial consistency of the two images can be ensured, thereby improving the accuracy and reliability of data analysis. In addition, spatio-temporal registration can also provide accurate spatial information for the path planning module. For example, when generating a maintenance path, it can accurately identify the location of the components that need to be repaired, avoiding waste of resources.

[0051] Step S107, the specific analysis process of the three-dimensional aging gradient distribution model includes: after gridifying the environmental stress parameters, splicing them with the intrinsic degradation factors to form a three-dimensional input tensor, and mapping through a three-dimensional convolutional neural network to generate an aging gradient distribution model in the temperature-humidity-irradiance space.

[0052] Obtain environmental stress parameter data, including temperature, humidity, and irradiance, and form a parameter matrix in the three-dimensional space coordinate system through gridification. Decouple the environmental stress and the material intrinsic degradation factors through a variational autoencoder to generate the feature vector of the intrinsic degradation factors. Use the splicing operation to combine the gridified environmental stress parameter matrix and the feature vector of the intrinsic degradation factors into a three-dimensional input tensor. Extract features from the three-dimensional input tensor through a three-dimensional convolutional neural network to obtain an aging feature map in the temperature-humidity-irradiance space. Calculate the degradation rate of each grid point in the three-dimensional space based on the aging feature map to generate an aging gradient distribution surface.

[0053] Obtain environmental stress parameter data, including temperature, humidity, and irradiance. Through grid processing, a parameter matrix in a three-dimensional space coordinate system is formed. The temperature range is from -20°C to 80°C, the humidity range is from 10% to 90%, the irradiance is from 0 to 1200 W / m², and the grid resolution is 1m×1m×1m. Decouple the environmental stress and the material intrinsic degradation factor by a variational autoencoder to generate the eigenvector of the intrinsic degradation factor. Use a concatenation operation to combine the grid environmental stress parameter matrix and the eigenvector of the intrinsic degradation factor into a three-dimensional input tensor, and the tensor dimension is 100×100×100×3. Extract features from the three-dimensional input tensor through a three-dimensional convolutional neural network, with a convolution kernel size of 3×3×3 and a stride of 1, to obtain an aging feature map in the temperature-humidity-irradiance space, and the resolution of the feature map is 50×50×50. According to the aging feature map, calculate the degradation rate of each grid point in the three-dimensional space to generate an aging gradient distribution surface, and the degradation rate range is from 0 to 0.1% / year.

[0054] Step S108, as Figure 2 shown, the specific design process of the maintenance path planning scheme includes: generating a maintenance priority queue according to the component aging gradient value; based on the maintenance priority queue and the topological distance between components, assigning path weight coefficients to adjacent components, and the higher the priority and the closer the distance, the greater the weight between components; using the dynamic evaporation factor to adjust the pheromone concentration distribution in real time according to the path weight coefficient, where the pheromone accumulation rate of the high-weight path is increased; outputting the maintenance path sequence through iterative optimization, and automatically matching the required maintenance tool list based on the component defect type.

[0055] According to the three-dimensional aging gradient distribution model, calculate the aging gradient value of each component, generate a maintenance priority queue based on the aging gradient value, and determine the maintenance priority order of the components. Obtain the longitude and latitude coordinate system data of the photovoltaic array, calculate the topological distance between components, and based on the maintenance priority queue and the topological distance, assign path weight coefficients to adjacent components to obtain high weight values for high-priority and short-distance components. Use an improved ant colony algorithm to adjust the pheromone concentration distribution in real time based on the path weight coefficient and the dynamic evaporation factor formula, and determine the pheromone accumulation rate of the high-weight path. Through iterative optimization of the pheromone concentration distribution, use the improved ant colony algorithm to generate a dynamic maintenance path sequence to obtain the optimal maintenance path. According to the defect features in the multi-modal data, use a classification algorithm to identify the defect type of each component and generate a maintenance tool list corresponding to the defect type. Through the visualization interface of the human-computer interaction terminal, based on the optimal maintenance path and the maintenance tool list, generate a component health status distribution map and a maintenance path plan, and determine the final operation and maintenance execution data.

[0056] According to the three-dimensional aging gradient distribution model, calculate the aging gradient values of each component, generate a maintenance priority queue based on the aging gradient values, and determine the maintenance priority order of the components. For example, the priority of components with an aging gradient value greater than 0.8 is 1, the priority of those between 0.5 and 0.8 is 2, and the priority of those less than 0.5 is 3. Obtain the longitude and latitude coordinate system data of the photovoltaic array, calculate the topological distance between components, and based on the maintenance priority queue and topological distance, assign path weight coefficients to adjacent components to obtain high weight values for high-priority and close-distance components. For example, the weight coefficient between components with a priority of 1 and a distance less than 10 meters is 0.9, and the weight coefficient between components with a priority of 2 and a distance less than 10 meters is 0.7. Adopt an improved ant colony algorithm to adjust the pheromone concentration distribution in real time based on the path weight coefficient and the dynamic evaporation factor formula, and determine the pheromone accumulation rate of the high-weight path. For example, the dynamic evaporation factor formula is ;

[0057] where is the initial evaporation coefficient, t is the current iteration number, is the total number of iterations, controls the attenuation rate and adjusts the pheromone concentration distribution. By iteratively optimizing the pheromone concentration distribution, use the improved ant colony algorithm to generate a dynamic maintenance path sequence and obtain the optimal maintenance path. For example, after 100 iterations, an optimal path covering all high-priority components is generated. According to the defect characteristics in the multimodal data, use the clustering analysis algorithm to identify the defect types of each component and generate a maintenance tool list corresponding to the defect types. For example, if the identified component defect type is a crack, the generated maintenance tool list includes glue and repair film. Through the visual interface of the human-computer interaction terminal, based on the optimal maintenance path and the maintenance tool list, generate a component health status distribution map and a maintenance path plan, and determine the final operation and maintenance execution data. For example, display the component health status distribution map in the visual interface and mark the optimal maintenance path and the required tool list.

[0058] An embodiment of the present invention also provides an intelligent photovoltaic power station operation and maintenance management system applying any one of the above embodiments, including an autonomous mobile robot, a data acquisition module, a data processing module, a path planning module, and a human-computer interaction terminal; the autonomous mobile robot is equipped with a hyperspectral camera, an EL detector, and a positioning module for performing multimodal data acquisition; the data processing module includes a spatio-temporal registration unit, a cross-modal attention fusion network, and a degradation decoupling model for in-depth correlation and decoupling analysis; the path planning module is configured to output a dynamic maintenance path using an improved ant colony algorithm; the human-computer interaction terminal is used to display the maintenance path plan; the autonomous mobile robot establishes a data connection with the data processing module and the path planning module through a wireless communication module.

[0059] The autonomous mobile robot is the core execution unit of the intelligent photovoltaic power station operation and maintenance management system. The hyperspectral camera it carries can capture the characteristic information of photovoltaic modules in the 420 - 680 nm and 950 - 1150 nm bands, which respectively correspond to the sensitive regions of material chemical degradation and carrier concentration. The EL detector is used to obtain the electroluminescence image of the photovoltaic module, which can effectively identify physical defects such as hidden cracks and broken grids in the solar cells. The positioning module realizes the precise positioning of the robot through GPS or lidar technology to ensure the spatial accuracy of data collection. For example, during a certain inspection in a photovoltaic power station, the autonomous mobile robot detected an abnormal decrease in the spectral reflectance of a certain module in the 680 nm band through the hyperspectral camera. At the same time, the EL detector found an obvious hidden crack in this module. Combining the data of the positioning module, the system can accurately mark the position and fault type of this module.

[0060] The data acquisition module is responsible for coordinating the autonomous mobile robot to execute multi-modal data acquisition tasks. This module interacts with the robot in real time through wireless communication to ensure the synchronous acquisition of hyperspectral images, EL images, and positioning data. For example, during the daily inspection of a photovoltaic power station, the data acquisition module will send instructions to the autonomous mobile robot, asking it to scan all the modules in the power station according to the preset path and transmit the collected data to the data processing module for subsequent analysis. The data processing module includes a spatio-temporal registration unit, a cross-modal attention fusion network, and a degradation decoupling model. The spatio-temporal registration unit is responsible for spatially aligning the hyperspectral images, EL images, and positioning data to eliminate errors caused by differences in acquisition time or location.

[0061] The cross-modal attention fusion network quantifies the correlation degree between different modal features by calculating the attention weights of the hyperspectral feature query vector and the electroluminescence feature key vector. For example, when the hyperspectral image shows an abnormal reflectance of a certain module in the 950 - 1150 nm band, while the EL image does not show obvious defects, the cross-modal attention fusion network will calculate the attention weights to determine whether this abnormality is related to the change in carrier concentration, thus assisting in fault diagnosis.

[0062] The degradation decoupling model is used to analyze the interaction between environmental stress and material intrinsic degradation. For example, by analyzing the relationship between temperature and humidity fluctuations and EVA yellowing and cell hidden cracks, a three-dimensional gradient degradation characterization is constructed to provide a basis for life prediction.

[0063] The path planning module uses an improved ant colony algorithm to output a dynamic maintenance path. This algorithm is based on the dynamic priority queue and aging gradient value of the component health index, adjusts the path weight coefficient in real time, and combines with the pheromone concentration distribution map to generate the optimal maintenance path. For example, in a certain maintenance task of a photovoltaic power station, the path planning module preferentially planned the maintenance paths of components with lower health indices and higher aging gradient values according to the component health index and aging gradient value. At the same time, through the real-time updated pheromone concentration distribution map, path redundancy and resource waste were avoided, and the maintenance efficiency was significantly improved.

[0064] The human-machine interaction terminal is used to display the maintenance path plan. Maintenance personnel can view the maintenance path generated by the system through this terminal for on-site navigation. For example, in a certain maintenance task of a photovoltaic power station, the maintenance personnel viewed the maintenance path recommended by the system through the human-machine interaction terminal for on-site navigation, accurately found the location of the faulty component, and improved the accuracy and efficiency of the maintenance.

[0065] The autonomous mobile robot establishes a data connection with the data processing module and the path planning module through the wireless communication module to achieve full-process intelligent operation and maintenance management. For example, in a certain inspection of a photovoltaic power station, the autonomous mobile robot transmits the collected hyperspectral images, EL images, and positioning data to the data processing module through the wireless communication module. After the data processing module completes the data analysis, the results are fed back to the path planning module. The path planning module generates a maintenance path and sends it to the human-machine interaction terminal, and the maintenance personnel perform maintenance according to the path displayed on the terminal. The whole process realizes full-process intelligent management.

[0066] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the various embodiments of the present application.

Claims

1. An intelligent photovoltaic power station operation and maintenance management method, characterized in that: The following steps are involved: Collect multimodal data of photovoltaic modules through autonomous mobile robots, including hyperspectral images, electroluminescent images and module location information; Perform spatiotemporal registration of multimodal data and associate hyperspectral chemical features with electroluminescent physical defect features through a cross-modal attention fusion network; Based on the conditional generative adversarial network and variational autoencoder, environmental stress and material intrinsic degradation factors are decoupled to generate a three-dimensional aging gradient distribution model; Combining the aging gradient distribution with the component topological relationship, an improved ant colony algorithm is used to output a maintenance path planning scheme with dynamic weights.

2. The intelligent photovoltaic power station operation and maintenance management method according to claim 1 is characterized in that: The specific analysis process of collecting multimodal data of photovoltaic panels through autonomous mobile robots includes: Acquire hyperspectral images of photovoltaic modules in the range of 400-2500nm, focusing on the 420-680nm band sensitive to material chemical degradation and the 950-1150nm band sensitive to carrier concentration; Synchronously collect electroluminescent images and use pulse current mode to inject 80% to 120% of the short-circuit current to ensure clear visualization of the defect area; The location information of the components is recorded, and a spatial distribution map of the components is formed based on the longitude and latitude coordinate system of the photovoltaic array.

3. The intelligent photovoltaic power station operation and maintenance management method according to claim 1 is characterized in that: The specific implementation process of the cross-modal attention fusion network includes: calculating the hyperspectral feature query vector Electroluminescence characteristic bond vector The attention weight is as follows: ； in, Indicates The spectral channels and The correlation strength of the defect area, is the total number of defective areas, Represents the query vector for hyperspectral features The transposed matrix of .

4. The intelligent photovoltaic power station operation and maintenance management method according to claim 1 is characterized in that: The specific method for decoupling environmental stress and material intrinsic degradation factor based on conditional generative adversarial network and variational autoencoder includes: defining the loss function of variational autoencoder, the formula is as follows: ； in, is a hyperparameter that controls the strength of the decoupling. Represents KL divergence, which is used to constrain the consistency of latent variable distribution and prior distribution; is the encoder output distribution; is the prior distribution, the latent variable Characterize the intrinsic degradation factor of materials; represents the log-likelihood of the decoder reconstruction, which is used to measure the accuracy of reconstructing the input data x from the latent variable z; represents the distribution of the output from the encoder Sample the hidden variable z in the decoder and calculate the decoder reconstruction data expected value.

5. The intelligent photovoltaic power station operation and maintenance management method according to claim 1, characterized in that: The dynamic volatility factor update rule of the improved ant colony algorithm is: ； in, is the initial volatility coefficient, is the dynamic volatilization factor at the tth iteration, which controls the volatilization rate of pheromone, t is the current iteration number, is the total number of iterations, is a decay exponent that adjusts the nonlinear change of the volatilization rate.

6. The intelligent photovoltaic power station operation and maintenance management method according to claim 1, characterized in that: The specific implementation process of spatiotemporal registration includes: aligning the hyperspectral image and the electroluminescent image by affine transformation based on the longitude and latitude coordinate system of the photovoltaic array; using the SIFT feature point matching algorithm to eliminate the spatial offset error of data collected at different times to ensure the spatial consistency of multi-source data.

7. The intelligent photovoltaic power station operation and maintenance management method according to claim 1, characterized in that: The specific analysis process of the three-dimensional aging gradient distribution model includes: gridding the environmental stress parameters, splicing them with the intrinsic degradation factors into a three-dimensional input tensor, and generating an aging gradient distribution model in the temperature-humidity-irradiance space through three-dimensional convolutional neural network mapping.

8. The intelligent photovoltaic power station operation and maintenance management method according to claim 5, characterized in that: The specific design process of the maintenance path planning solution includes: Generate maintenance priority queue according to component aging gradient value; Assign path weight coefficients to adjacent components based on the maintenance priority queue and the topological distance between components; The dynamic volatilization factor is used to adjust the pheromone concentration distribution in real time according to the path weight coefficient; The repair path sequence is output through iterative optimization, and the required repair tool list is automatically matched based on the component defect type.

9. An intelligent photovoltaic power station operation and maintenance management system, applying the intelligent photovoltaic power station operation and maintenance management method according to any one of claims 1 to 8, characterized in that: It includes an autonomous mobile robot, a data acquisition module, a data processing module, a path planning module and a human-computer interaction terminal; The autonomous mobile robot is equipped with a hyperspectral camera, EL detector, and positioning module to perform multimodal data acquisition; The data processing module includes a spatiotemporal registration unit, a cross-modal attention fusion network, and a degenerate decoupling model for deep correlation and decoupling analysis; The path planning module is configured to output a dynamic maintenance path using an improved ant colony algorithm; The human-computer interaction terminal is used to display the maintenance path plan; The autonomous mobile robot establishes data connection with the data processing module and the path planning module through the wireless communication module.

Images