A photovoltaic module dynamic local thermal protection method based on area array thermal imaging

Abstract

The present application relates to a kind of photovoltaic module dynamic local thermal protection method based on area array thermal imaging, the present application generates high-quality thermal field observation data set by multi-source sensor data space-time alignment and noise suppression, adopts adaptive threshold segmentation and superpixel clustering algorithm to extract semantic heat unit, constructs frame-by-frame thermal evolution atlas according to heat flow density gradient, and generates structure fingerprint vector by calculating topological stability index;Optimal control strategy template is matched using lightweight KNN-decision tree hybrid classifier, and the execution mechanism is driven to generate the concretization control instruction, and the feedback data closed-loop optimization strategy template library is combined.The present application realizes the accurate characterization of photovoltaic module thermal field evolution characteristics and the dynamic optimization of control strategy, improves the real-time control and action minimization effect.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic module thermal management and dynamic thermal control technology, and in particular to a method for dynamic local thermal protection of photovoltaic modules based on area array thermal imaging. Background Technology

[0002] Current thermal management and dynamic thermal control technologies for photovoltaic modules mainly revolve around the analysis of array thermal imaging data, combined with the detection of abnormal heat distribution areas, to intervene in and protect against localized overheating of the modules. Most mainstream solutions employ temperature control logic based on static threshold judgment or hysteresis feedback, automatically triggering control execution based on a pre-defined set of actions. For example, this could limit the upper limit of the thermal field temperature, or activate heat dissipation mechanisms, adjust the opening and closing of shading films, or optimize the operation of ventilation systems when the temperature change in a specific area exceeds a preset threshold. With the promotion of large-area photovoltaic power plants and highly integrated modules, the industry is also exploring the introduction of time-series prediction models (such as ARIMA and LSTM neural networks) based on historical thermal field records and environmental parameters, attempting to identify localized overheating trends in advance and issue pre-control commands before component damage occurs.

[0003] These technical approaches have reached a certain level of maturity and can adapt to the needs of controlling abnormal heat distribution under conditions such as normal illumination fluctuations and uniform shadow projection. Some literature also proposes to incorporate multi-objective optimization, finite element thermal resistance simulation, or use deep learning to automatically extract thermal map features to improve the accuracy and robustness of the control strategy.

[0004] Existing methods generally suffer from the following shortcomings: First, relying on static threshold detection and hysteresis feedback mechanisms, there is often an unavoidable lag between thermal field anomalies and control commands, making it difficult to respond promptly to rapid shadow migration and sudden occlusion-induced dramatic changes in heat distribution. Second, in complex dynamic shadow scenarios, time series regression or end-to-end deep learning prediction models mainly focus on the numerical continuity of temperature values, making it difficult to robustly capture the rapid evolution of the thermal field's spatial structure. This results in high sensitivity to absolute temperature calibration and the model's susceptibility to sensor noise instability. Third, existing deep prediction models involve large inference computations. When deployed at edge devices or embedded controllers, they struggle to achieve millisecond-level real-time policy generation, easily leading to high actual execution frequency, increased energy consumption, and delayed responses to local hot spots. Fourth, policy decisions often lack physical interpretability regarding the essential structural trends of thermal field evolution, making it difficult to dynamically identify the migration and interruption of main heat flow and the rapid shift of heat accumulation centers, thus limiting their adaptability to dynamic scenarios.

[0005] Therefore, existing strategy generation systems for dynamic local thermal protection of photovoltaic modules generally face core technical barriers such as insufficient real-time control response, weak ability to predict and actively intervene in drastic changes in the thermal field, redundancy and energy waste in execution actions, and decreased robustness under non-ideal operating conditions. Especially in complex environments, existing solutions cannot effectively balance prediction speed with minimizing the execution cost of the strategy, and it is also difficult to guarantee stable strategy output when sensor accuracy fluctuates, photovoltaic modules age, or even when environmental conditions change significantly.

[0006] Therefore, there is an urgent need to propose a new method for generating thermal control strategies that can break through static thresholds and numerical regression dependence, possess high real-time adaptability to dynamic heat distribution trends, have physical interpretability for the evolution of thermal field spatial structure, and can be lightweight and deployed in field edge systems. Summary of the Invention

[0007] This invention provides a dynamic local thermal protection method for photovoltaic modules based on area array thermal imaging, aiming to solve the problems of the prior art mentioned in the background section.

[0008] The technical solution of this invention is: a dynamic local thermal protection method for photovoltaic modules based on area array thermal imaging, comprising the following steps: S1: Collect the original thermal image sequence of photovoltaic modules, as well as multi-source environmental parameters such as wind speed, ambient temperature and light intensity; and perform alignment preprocessing to generate a spatiotemporally synchronized dynamic observation dataset of the thermal field. S2: Based on adaptive threshold segmentation and superpixel clustering algorithm, each frame of thermal image is divided into several semantic thermal units to form a set of semantic thermal units; S3: Using the set of semantic thermal units as nodes, construct directed weighted edges based on the direction and intensity of the heat flux density gradient between adjacent semantic thermal units to generate a frame-by-frame thermal evolution map; S4: Based on a continuous multi-frame thermal evolution map sequence, calculate the node connection stability, edge weight evolution robustness and subgraph structure persistence to generate a three-dimensional structural fingerprint vector. S5: Input the 3D structural fingerprint vector into a lightweight KNN-decision tree hybrid classifier trained offline, and match the control policy template with the highest similarity between the trigger condition and the structural fingerprint vector in the preset policy template library. S6: Based on the action abstract set and constraint priority chain in the matched control strategy template, and combined with the physical constraint parameters of the local actuator, generate a concrete control instruction. S7: Executes concrete control commands to drive the distributed thermal control actuator of photovoltaic modules, and dynamically updates the constraint priority chain in the strategy template library based on the thermal field feedback data collected after execution.

[0009] Furthermore, in step S1, the timestamps of each sensor in the original observation dataset are clock drift calibrated based on a high-precision GPS clock source to obtain a calibrated timestamp sequence. A linear interpolation algorithm is used to adjust the original thermal image sequence and multi-source environmental parameters in the calibration timestamp sequence to a uniform sampling rate benchmark, so as to generate an aligned data sequence with consistent temporal resolution; Median filtering based on a sliding window is performed on the aligned data sequence to form a complete and noise-suppressed data sequence; The complete data sequence is spatiotemporally registered according to the time index to generate a spatiotemporally synchronized dynamic observation dataset of thermal field containing thermal image frame sequences and corresponding environmental parameters.

[0010] Furthermore, in step S5, based on the historical thermal evolution map dataset and its labeled control strategy samples, the K-nearest neighbor algorithm and the decision tree model are used in a cascaded structure for offline training to generate lightweight KNN-decision tree hybrid classifier model parameters.

[0011] The beneficial technical effects of this invention are as follows: 1) By constructing a topological stability analysis framework based on semantic thermal units and thermal evolution maps, this invention effectively overcomes the high dependence of traditional photovoltaic thermal protection strategies on temperature numerical regression prediction and their sensitivity to dynamic shading interference. 2) This invention starts from the evolution law of thermal field spatial structure, uses adaptive segmentation and superpixel clustering to extract STUs with physical meaning, and uses a directed weighted graph to depict the thermal flow coupling relationship between them, so that the prediction logic focuses on the topological evolution of the heat conduction path rather than the temperature change trajectory of a single pixel. 3) This invention introduces three types of indicators: node connection stability, edge weight evolution robustness, and subgraph structure persistence, forming a three-dimensional structural fingerprint with clear physical interpretation. It can accurately identify the essential difference between instantaneous disturbances caused by cloud shadow movement, branch and leaf swaying, etc., and the actual local overheating risk, significantly improving the accuracy and anti-interference ability of thermal anomaly prediction. 4) This invention adopts a lightweight KNN-decision tree hybrid classifier to match pre-stored policy templates, realizing a predictive control generation mechanism with high real-time performance and low computational overhead, completely avoiding the rolling optimization solution, weight adjustment and online learning process required by traditional model predictive control or reinforcement learning methods. Attached Figure Description

[0012] Figure 1 A flowchart illustrating the method of this invention; Figure 2 A schematic diagram of the process for generating frame-by-frame thermal evolution maps according to an embodiment of the present invention; Figure 3A schematic diagram of the process for generating a three-dimensional structural fingerprint vector according to an embodiment of the present invention. Detailed Implementation

[0013] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0014] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0015] like Figure 1 As shown, this embodiment provides a dynamic local thermal protection method for photovoltaic modules based on area array thermal imaging, specifically including the following steps: S1: Time-aligned preprocessing is performed on the original thermal image sequence acquired by the high-resolution area array thermal imaging sensor deployed on the surface of photovoltaic modules, as well as the multi-source environmental parameters of wind speed, ambient temperature and light intensity acquired simultaneously, in order to generate a spatiotemporally synchronized dynamic observation dataset of the thermal field. S2: Based on adaptive threshold segmentation and superpixel clustering algorithm, each frame of thermal image is divided into several semantic thermal units, where each semantic thermal unit represents a local region with similar thermal conduction properties and environmental coupling characteristics, so as to form a set of semantic thermal units; S3: Using the set of semantic thermal units as nodes, construct directed weighted edges based on the direction and intensity of the heat flux density gradient between adjacent semantic thermal units to generate a frame-by-frame thermal evolution map, where the thermal evolution map represents the temporal evolution relationship of the thermal field spatial structure. S4: Based on a continuous multi-frame thermal evolution map sequence, calculate three types of topological stability indices: node connection stability, edge weight evolution robustness, and subgraph structure persistence. Among them, node connection stability is the proportion of semantic thermal units that maintain adjacency within a preset number of frames. Edge weight evolution robustness is screened for low-volatility main heat flow paths through the coefficient of variation of a sliding window. Subgraph structure persistence is identified to identify recurring thermal aggregation subgraph patterns in order to generate a three-dimensional structural fingerprint vector. S5: Input the 3D structural fingerprint vector into a lightweight KNN-decision tree hybrid classifier trained offline, and match the control policy template with the highest similarity between the trigger condition and the structural fingerprint vector in the preset policy template library. The control policy template is encoded as a combination of trigger condition, action abstraction set and constraint priority chain. S6: Based on the action abstract set and constraint priority chain in the matched control strategy template, combined with the physical constraint parameters of the local actuator, the concrete control instructions are generated by parsing. The concrete control instructions include the fine-tuning light-shielding film response parameters and the local air duct opening and closing sequence. S7: Executes concrete control commands to drive the distributed thermal control actuator of photovoltaic modules, and dynamically updates the constraint priority chain in the strategy template library based on the thermal field feedback data collected after execution, so as to optimize the real-time performance and action minimization effect of subsequent control strategies.

[0016] In this embodiment, step S1: The original thermal image sequence acquired by the high-resolution area array thermal imaging sensor deployed on the surface of the photovoltaic module, along with synchronously acquired multi-source environmental parameters such as wind speed, ambient temperature, and light intensity, undergo time-aligned preprocessing to generate a spatiotemporally synchronized dynamic thermal field observation dataset. Specifically, this includes: S1.1: Perform real-time data stream acquisition processing on the raw thermal image sequence output by the high-resolution area array thermal imaging sensor deployed on the surface of the photovoltaic module, as well as the multi-source environmental parameters synchronously collected by the wind speed sensor, ambient temperature sensor and light intensity sensor, to form a raw observation data stream containing timestamps. S1.2: Based on a high-precision GPS clock source, clock drift calibration is performed on the timestamps of each sensor in the original observation data stream to eliminate timing deviations caused by hardware clock asynchrony between sensors and obtain a calibration timestamp sequence. S1.3: Use a linear interpolation algorithm to adjust the original thermal image sequence and multi-source environmental parameters in the calibration timestamp sequence to a unified sampling rate benchmark, so as to generate an aligned data sequence with consistent temporal resolution; S1.4: Perform median filtering based on a sliding window on the aligned data sequence to identify and fill missing data points and remove outliers, forming a complete and noise-suppressed data sequence; S1.5: Perform spatiotemporal registration processing on the complete data sequence according to the time index to generate a spatiotemporally synchronized dynamic observation dataset of thermal field containing thermal image frame sequences and corresponding environmental parameters.

[0017] In this embodiment, in step S2, based on the adaptive threshold segmentation and superpixel clustering algorithm, each frame of thermal image is divided into several semantic thermal units, where each semantic thermal unit represents a local region with similar thermal conduction properties and environmental coupling characteristics, to form a set of semantic thermal units. Specifically, the steps include the following: S2.1: Perform pixel intensity normalization processing on single-frame thermal images in the spatiotemporally synchronized thermal field dynamic observation dataset to eliminate temperature scaling bias caused by sensor nonlinear response and generate standardized thermal image data as input for the segmentation algorithm. S2.2: Based on the adaptive threshold segmentation algorithm, local threshold dynamic calculation processing is performed on the standardized thermal image data to identify the boundaries of temperature gradient change regions and generate an initial thermal anomaly region segmentation mask as the input object for superpixel clustering; S2.3: The initial thermal anomaly region segmentation mask is merged using a superpixel clustering algorithm to generate uniform thermal characteristic superpixel blocks based on the similarity of heat flux density between pixels, forming a superpixel topology structure as the basic unit for semantic annotation. S2.4: Perform semantic feature fusion processing based on the thermal conduction property parameters and environmental coupling characteristic parameters of the superpixel topology to quantitatively evaluate the consistency of thermal behavior in local areas and generate a semantic thermal unit definition set as the core output of the region representation; S2.5: Perform topological connectivity verification on the semantic thermal unit definition set to filter semantic thermal unit instances that conform to the physical heat conduction law, and output the semantic thermal unit set for node construction of thermal evolution graph; Furthermore, a thermal conductivity stability index calculation method is employed to evaluate the thermal conductivity stability of the candidate semantic thermal unit set and generate a stability index dataset. The thermal conductivity stability index is calculated using the following formula: in, Thermal conductivity of photovoltaic module materials, The temperature difference between two adjacent semantic thermal units. The distance between the centers of the nodes. This is the thermal conductivity stability index.

[0018] In this embodiment, in step S3, using the semantic thermal unit set as nodes, directed weighted edges are constructed based on the direction and intensity of the heat flux density gradient between adjacent semantic thermal units to generate a frame-by-frame thermal evolution map. The thermal evolution map characterizes the temporal evolution relationship of the thermal field's spatial structure, such as... Figure 2 As shown, it includes the following steps: S3.1: Perform local thermal field state parameter extraction processing on the semantic thermal unit set to obtain the temperature distribution feature dataset of each semantic thermal unit; based on the spatiotemporal synchronous thermal field dynamic observation dataset output by the area array thermal imaging sensor, use an adaptive threshold segmentation algorithm to perform pixel-level temperature mean and variance statistics on the semantic thermal units of each frame of thermal image, and generate a semantic thermal unit feature vector set containing local thermal field state parameters, which serves as the input basis for heat flux density gradient calculation; S3.2: Based on the semantic thermal unit feature vector set of local thermal field state parameters, perform spatial adjacency relationship identification processing of adjacent semantic thermal unit pairs to generate a spatial adjacency relationship set; using the semantic thermal unit boundary coordinate information output by the superpixel clustering algorithm, calculate the Euclidean distance and azimuth angle between each pair of adjacent semantic thermal units, filter semantic thermal unit pairs within the distance threshold and with continuous azimuth angles to form a spatial adjacency relationship set, which serves as the input object for temperature gradient vector calculation; S3.3: Based on the set of spatial adjacency relationships, perform temperature gradient vector calculation to generate a temperature gradient vector set; based on the local thermal field state parameters of adjacent semantic thermal unit pairs, use the finite difference method to perform first-order partial derivative calculations on the temperature field at the boundary of the semantic thermal units, extract the temperature gradient direction angle and gradient magnitude, and construct a temperature gradient vector set containing direction and intensity as the input basis for the derivation of heat flux density gradient parameters; In this embodiment, by performing first-order partial derivative operations on the two-dimensional temperature distribution matrix, the independent extraction of the temperature change rates in the horizontal and vertical directions is achieved, and a directional component dataset is obtained. Temperature gradient direction angle is calculated using a direction angle calculation method. The derivation is as follows: In the formula, The horizontal temperature gradient The temperature gradient is in the vertical direction; The temperature gradient magnitude G is derived using a gradient magnitude calculation method, namely: In the formula, The temperature gradient amplitude reflects the strength of local heat flux density changes; By pairing the orientation angle with the gradient magnitude, the gradient component data from the previous step is transformed into a set of temperature gradient vectors containing direction and intensity, thereby achieving a quantitative representation of the heat flow trend between semantic thermal units.

[0019] S3.4: Based on the temperature gradient vector set, perform heat flux density gradient parameter derivation to generate a heat flux density gradient parameter set; according to Fourier's heat conduction law, linearly map the gradient magnitude in the temperature gradient vector set to the thermal conductivity parameter of the photovoltaic module material, calculate the heat flux direction vector and the heat flux intensity scalar, and form a heat flux density gradient parameter set containing the direction and intensity of the heat flux density gradient, which serves as the input condition for constructing the directed weighted edge; S3.5: The semantic thermal unit set is used as a node set, and directed weighted edge construction and graph structure integration processing are performed with the heat flux density gradient parameter set to generate a frame-by-frame thermal evolution map. The direction of the directed edges is determined according to the direction vector in the heat flux density gradient parameter set, the edge weight is quantized with the heat flux intensity scalar, a directed weighted edge set is constructed, and the node set and the directed weighted edge set are topologically integrated through the graph theory adjacency matrix representation method to output a frame-by-frame thermal evolution map data structure that represents the temporal evolution of the thermal field spatial structure.

[0020] In this embodiment, in step S4, based on a continuous multi-frame thermal evolution map sequence, three types of topological stability indices are calculated: node connection stability, edge weight evolution robustness, and subgraph structure persistence. Node connection stability is determined by the proportion of semantic thermal units maintaining adjacency within a preset number of frames. Edge weight evolution robustness is determined by filtering low-volatility main heat flow paths using a sliding window variation coefficient. Subgraph structure persistence identifies recurring thermal aggregation subgraph patterns to generate a three-dimensional structural fingerprint vector, such as... Figure 3 As shown, the specific steps include the following: S4.1: For each semantic thermal unit in a continuous multi-frame thermal evolution map sequence, obtain its historical trajectory of adjacent semantic thermal unit (STU) sets within a preset frame window; based on this trajectory, perform adjacency relationship change detection between adjacent frames, and calculate the number of consecutive frames in which adjacent STU groups remain consistent; divide this number of consecutive frames by the preset frame number to generate a node connection stability ratio value; normalize the stability ratio values ​​of all semantic thermal units to obtain a node connection stability index; this index quantifies the temporal consistency of node-level connectivity structures in the thermal field topology to characterize the stability of local heat conduction paths; S4.2: For each directed weighted edge in a continuous multi-frame thermal evolution map sequence, extract its weight time series data; based on this sequence, apply sliding window coefficient of variation analysis to calculate the ratio of the standard deviation to the mean of the edge weights within each window; filter edges with coefficients of variation below a preset threshold to form a low volatility edge set; define this set as the main heat flow path to identify stable heat flow transfer channels in the thermal field; generate edge weight evolution robustness. S4.3: Perform local thermal clustering subgraph detection on a continuous multi-frame thermal evolution map sequence to obtain a candidate subgraph pattern set; apply a dynamic graph isomorphism matching algorithm based on this set to compare the structural similarity of subgraphs between different frames; identify thermal clustering subgraph patterns that repeatedly appear within an M-frame window and have a similarity higher than a threshold; statistically analyze the occurrence frequency and duration of this pattern to generate a subgraph structure persistence metric; based on the metrics of all thermal clustering subgraph patterns, generate a subgraph structure persistence index to quantify the evolutionary persistence of the thermal field spatial structure; The system receives input data structures from a sequence of consecutive multi-frame thermal evolution maps and executes a local thermal clustering subgraph detection algorithm (parameters: heat flux density gradient threshold, node temperature uniformity threshold, and inter-frame adjacency variation tolerance) to achieve initial screening of candidate subgraph patterns based on graph structure features.

[0021] Furthermore, by matching the node and edge attributes in the candidate subgraph pattern set, and using the dynamic graph isomorphism matching algorithm, the accurate comparison of cross-frame subgraph structural similarity is achieved, and the similarity matrix of each pair of subgraphs is obtained.

[0022] Furthermore, the sliding window structure matching analysis method is used to extract the highly similar sub-graph patterns that repeatedly appear within the M-frame window. Patterns with similarity higher than a preset threshold are selected and their time intervals are marked as candidates for stable patterns of thermal aggregation evolution.

[0023] Furthermore, based on the time intervals of the marked patterns, frequency and duration statistics are performed, and the subgraph structure persistence metric is calculated using the following formula. : in, The frequency of the pattern's appearance within the window. This represents the average number of frames in a single consecutive occurrence. This is the window length.

[0024] Furthermore, by using a normalization mapping algorithm, the persistence metrics of all patterns are unified to the [0,1] interval, generating a set of subgraph structure persistence metrics.

[0025] S4.4: Data fusion processing is performed on the node connection stability index, edge weight evolution robustness index, and subgraph structure durability index; the three types of index values ​​are mapped to a unified scale based on the normalization algorithm; the normalized index values ​​are arranged in a preset dimension order; a three-dimensional structural fingerprint vector is generated, which represents the essential stability characteristics of the thermal field evolution trend; this vector serves as the structural fingerprint of the topological stability of the thermal evolution map to support the rapid matching of subsequent control strategy templates.

[0026] In this embodiment, in step S5, the three-dimensional structural fingerprint vector is input into a lightweight KNN-decision tree hybrid classifier trained offline, and the control policy template with the highest similarity between the trigger condition and the structural fingerprint vector in the preset policy template library is matched. The control policy template is encoded as a combination of trigger condition, action abstraction set and constraint priority chain, and specifically includes the following steps: S5.1: Based on the historical thermal evolution map dataset and its labeled control strategy samples, the K-nearest neighbor algorithm and the decision tree model are used for offline training to generate lightweight KNN-decision tree hybrid classifier model parameters. The output of the model parameters is a classifier structure with fast similarity matching capability. Furthermore, a hierarchical classification of regulatory strategy categories is achieved by generating rule sets through a decision tree model.

[0027] Furthermore, by utilizing a cascaded structure to fuse the K-nearest neighbor index and the decision tree rule set, a two-stage determination mechanism for the classifier is achieved: the first stage retrieves several candidate samples with the highest similarity, and the second stage accurately determines the category of the candidate samples based on the decision tree rules.

[0028] Through offline training, the historical thermal evolution map dataset and regulation strategy samples are transformed into lightweight KNN-decision tree hybrid classifier model parameters, enabling fast similarity matching and policy category determination when inputting structural fingerprint vectors.

[0029] S5.2: Standardize the 3D structural fingerprint vector generated in S4 to eliminate dimensional differences and generate a normalized structural fingerprint feature vector, which serves as the input feature of the lightweight KNN-decision tree hybrid classifier. S5.3: Based on the normalized structural fingerprint feature vector, the K-nearest neighbor module of the lightweight KNN-decision tree hybrid classifier is used to calculate the Euclidean distance similarity with each triggering condition in the policy template library, and generate a similarity score vector. For the normalized structural fingerprint feature vector, a lightweight KNN module is used to calculate the Euclidean distance between the feature vector and each trigger condition in the policy template library, thereby realizing the spatial similarity assessment between local topological stability features and template conditions.

[0030] S5.4: Based on the similarity score vector, the classification rule matching is performed by the decision tree module of the lightweight KNN-decision tree hybrid classifier to identify the regulation strategy template with the highest similarity between the trigger condition and the structural fingerprint vector in the policy template library. The regulation strategy template is encoded as a combination of trigger condition, action abstract set and constraint priority chain. Furthermore, the similarity score vector is processed by the splitting function of the decision tree node to perform feature-dominant ordering, thereby realizing the dynamic adjustment of the weights of different similarity dimensions and obtaining the adjusted similarity feature sequence.

[0031] S5.5: Outputs a matching control strategy template to provide combined information on trigger conditions, action abstraction sets, and constraint priority chains. This serves as the input for subsequent parsing and generation of concrete control instructions, ensuring the real-time performance of the control strategy and minimizing the action effect.

[0032] In this embodiment, in step S6, based on the action abstract set and constraint priority chain in the matched control strategy template, and combined with the physical constraint parameters of the local actuator, a concrete control command is generated. The concrete control command includes the fine-tuning light-shielding film response parameters and the local air duct opening and closing sequence, specifically including the following steps: S6.1: Based on the matching control strategy template, extract the action abstract set and constraint priority chain to obtain the abstract description of the control action and the execution priority sequence as the basis for parsing; S6.2: Obtain local actuator physical constraint parameters, including the response delay of the fine-tuning shading film and the opening and closing time of the local air duct, to provide physical characteristic constraint information of the actuator; S6.3: Based on the action abstract set and constraint priority chain, parse and generate an abstract instruction sequence of control actions to determine the type of control actions and the execution logic order; S6.4: Combine the physical constraint parameters of the local actuator with the abstract instruction sequence to visualize the execution parameters in order to calculate the response parameters of the fine-tuning shading film and the opening and closing sequence of the local air duct.

[0033] Furthermore, based on the thermal behavior prediction model, the optimal response time adjustment amount under the current thermal field trend is calculated using a predictive control objective function. The formula is as follows: in, This is the thermal regulation coefficient. For local heat flow, For material density, For specific heat capacity, The area is the controlled region.

[0034] Furthermore, by superimposing the optimal response time adjustment amount with the effective set of execution parameters, the final parameter synthesis process is performed to form a concrete execution parameter matrix that includes the response parameters of the shading film and the opening and closing sequence of the air duct.

[0035] S6.5: Generate and output concrete control commands to drive the distributed thermal control actuator of the photovoltaic module to perform thermal control operations.

[0036] In this embodiment, step S7 involves executing a concrete control command to drive the distributed thermal control actuator of the photovoltaic module, and dynamically updating the constraint priority chain in the strategy template library based on the thermal field feedback data collected after execution, in order to optimize the real-time performance and action minimization effect of subsequent control strategies. Specifically, this includes the following steps: S7.1: Perform actuator parameter parsing processing on the concrete control command to extract the response parameters of the fine-tuning shading film and the opening and closing sequence of the local air duct, execute the thermal control action based on the physical constraints of the distributed thermal control actuator, and generate the execution action completion signal; S7.2: Based on the execution action completion signal triggering the high-resolution area array thermal imaging sensor, real-time data acquisition and processing of the thermal field state on the surface of the photovoltaic module is performed to obtain the thermal field feedback dataset. The thermal field feedback data is synchronized with the execution time of the concrete control command using the timestamp alignment algorithm to generate a spatiotemporally aligned thermal field observation sequence. S7.3: Perform thermal evolution map reconstruction on the spatiotemporally aligned thermal field observation sequence to extract node connection stability index and subgraph structure persistence index. Calculate the real-time deviation value of the control strategy and the action frequency of the execution module based on the sliding window coefficient of variation analysis algorithm, and generate strategy effect evaluation results. S7.4: Based on the strategy effect evaluation results, execute the constraint priority chain dynamic optimization algorithm to adjust the weights of the constraint priority chains in the strategy template library in order to identify the main heat flow path interruption threshold and the heat aggregation subgraph mode triggering conditions, and generate an updated constraint priority chain. S7.5: Perform closed-loop verification on the updated constraint priority chain to execute pre-matching tests for subsequent control strategies. Verify the real-time performance improvement and action minimization metrics based on the similarity calculation results of the three-dimensional structural fingerprint vectors, and output the optimized strategy template library version.

[0037] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

[0038] Unless otherwise defined, the technical or scientific terms used herein should be understood in their ordinary sense by one of ordinary skill in the art to which this invention pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the elements or objects preceding “comprising” or “including” encompass the elements or objects listed following “comprising” or “including” and their equivalents, and do not exclude other elements or objects. The “multiple” involved in the embodiments of this invention refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.

[0039] The above description is merely an exemplary embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for dynamic local thermal protection of photovoltaic modules based on area array thermal imaging, characterized in that, Includes the following steps: S1: Collect the original thermal image sequence of photovoltaic modules, as well as multi-source environmental parameters such as wind speed, ambient temperature and light intensity; and perform alignment preprocessing to generate a spatiotemporally synchronized dynamic observation dataset of the thermal field. S2: Based on adaptive threshold segmentation and superpixel clustering algorithm, each frame of thermal image is divided into several semantic thermal units to form a set of semantic thermal units; S3: Using the set of semantic thermal units as nodes, construct directed weighted edges based on the direction and intensity of the heat flux density gradient between adjacent semantic thermal units to generate a frame-by-frame thermal evolution map; S4: Based on a continuous multi-frame thermal evolution map sequence, calculate the node connection stability, edge weight evolution robustness and subgraph structure persistence to generate a three-dimensional structural fingerprint vector. S5: Input the 3D structural fingerprint vector into a lightweight KNN-decision tree hybrid classifier trained offline, and match the control policy template with the highest similarity between the trigger condition and the structural fingerprint vector in the preset policy template library. S6: Based on the action abstract set and constraint priority chain in the matched control strategy template, and combined with the physical constraint parameters of the local actuator, generate a concrete control instruction. S7: Executes concrete control commands to drive the distributed thermal control actuator of photovoltaic modules, and dynamically updates the constraint priority chain in the strategy template library based on the thermal field feedback data collected after execution.

2. The method for dynamic local thermal protection of photovoltaic modules based on area array thermal imaging according to claim 1, characterized in that, In step S1, the timestamps of each sensor in the original observation dataset are clock drift calibrated based on a high-precision GPS clock source to obtain a calibrated timestamp sequence. A linear interpolation algorithm is used to adjust the original thermal image sequence and multi-source environmental parameters in the calibration timestamp sequence to a uniform sampling rate benchmark, so as to generate an aligned data sequence with consistent temporal resolution; Median filtering based on a sliding window is performed on the aligned data sequence to form a complete and noise-suppressed data sequence; The complete data sequence is spatiotemporally registered according to the time index to generate a spatiotemporally synchronized dynamic observation dataset of thermal field containing thermal image frame sequences and corresponding environmental parameters.

3. The method for dynamic local thermal protection of photovoltaic modules based on area array thermal imaging according to claim 1, characterized in that, Step S2 specifically includes: The adaptive threshold segmentation algorithm is used to perform local threshold dynamic calculation processing on standardized thermal image data to identify the boundaries of temperature gradient change regions and generate an initial thermal anomaly region segmentation mask. The initial thermal anomaly region segmentation mask is merged using a superpixel clustering algorithm to generate uniform thermal characteristic superpixel blocks based on the similarity of heat flux density between pixels, forming a superpixel topology structure as the basic unit for semantic annotation. Semantic feature fusion processing is performed based on the thermal conductivity parameters and environmental coupling characteristics parameters of the superpixel topology to generate a set of semantic thermal unit definitions.

4. The method for dynamic local thermal protection of photovoltaic modules based on area array thermal imaging according to claim 1, characterized in that, In step S3, local thermal field state parameter extraction processing is performed on the semantic thermal unit set to obtain the temperature distribution feature dataset of each semantic thermal unit; Based on the spatiotemporal synchronous dynamic observation dataset of thermal field output by the area array thermal imaging sensor, an adaptive threshold segmentation algorithm is used to perform pixel-level temperature mean and variance statistics on the semantic thermal units of each frame of thermal image, generating a semantic thermal unit feature vector set containing local thermal field state parameters.

5. A method for dynamic local thermal protection of photovoltaic modules based on area array thermal imaging according to claim 4, characterized in that, In step S3, based on the semantic thermal unit feature vector set of local thermal field state parameters, spatial adjacency relationship identification processing of adjacent semantic thermal unit pairs is performed to generate a spatial adjacency relationship set. Using the semantic hot unit boundary coordinate information output by the superpixel clustering algorithm, the Euclidean distance and azimuth angle between each pair of adjacent semantic hot units are calculated. Semantic hot unit pairs within the distance threshold and with continuous azimuth angles are selected to form a set of spatial adjacency relationships. Based on the set of spatial adjacency relationships, perform temperature gradient vector calculation to generate a set of temperature gradient vectors; Based on the local thermal field state parameters of adjacent semantic thermal unit pairs, the finite difference method is used to perform first-order partial derivative calculations on the temperature field at the boundary of the semantic thermal units, extract the temperature gradient direction angle and gradient magnitude, and construct a temperature gradient vector set containing direction and intensity.

6. The method for dynamic local thermal protection of photovoltaic modules based on area array thermal imaging according to claim 5, characterized in that, Step S3 further includes: performing heat flux density gradient parameter derivation processing based on the temperature gradient vector set to generate a heat flux density gradient parameter set; Based on Fourier's law of heat conduction, the gradient magnitude in the temperature gradient vector set is linearly mapped to the thermal conductivity parameter of the photovoltaic module material, and the heat flow direction vector and heat flow intensity scalar are calculated to form a heat flow density gradient parameter set. The semantic thermal unit set is used as a node set, and directed weighted edge construction and graph structure integration processing are performed with the heat flux density gradient parameter set to generate a frame-by-frame thermal evolution map. The orientation of directed edges is determined based on the direction vector in the heat flux density gradient parameter set. The edge weights are quantized using heat flux intensity scalars to construct a set of directed weighted edges. The node set and the set of directed weighted edges are then topologically integrated using graph theory adjacency matrix representation to generate a frame-by-frame thermal evolution map.

7. A method for dynamic local thermal protection of photovoltaic modules based on area array thermal imaging according to claim 1, characterized in that, In step S4, the node connection stability statistical semantic hot unit maintains the continuous proportion of adjacency relationships within a preset number of frames, the edge weight evolution robustness is screened by the sliding window variation coefficient to select low volatility main heat flow paths, and the subgraph structure persistence is identified by identifying recurring heat clustering subgraph patterns.

8. A method for dynamic local thermal protection of photovoltaic modules based on area array thermal imaging according to claim 1, characterized in that, In step S5, based on the historical thermal evolution map dataset and its labeled control strategy samples, the K-nearest neighbor algorithm and the decision tree model are used for offline training to generate lightweight KNN-decision tree hybrid classifier model parameters.

9. A method for dynamic local thermal protection of photovoltaic modules based on area array thermal imaging according to claim 8, characterized in that, In step S5, based on the normalized structural fingerprint feature vector, the K-nearest neighbor module of the lightweight KNN-decision tree hybrid classifier is used to calculate the Euclidean distance similarity with each triggering condition in the policy template library, generating a similarity score vector; and the decision tree module of the lightweight KNN-decision tree hybrid classifier is used to perform classification rule matching to identify the control policy template in the policy template library with the highest similarity between the triggering condition and the structural fingerprint vector.

10. A method for dynamic local thermal protection of photovoltaic modules based on area array thermal imaging according to claim 1, characterized in that, In step S5, the control strategy template is encoded as a combination of triggering conditions, action abstraction sets, and constraint priority chains.

Images