A method for identifying hot spots on rooftop photovoltaic modules based on multimodal remote sensing

Abstract

This invention discloses a hot spot identification method for rooftop photovoltaic modules based on multimodal remote sensing. By integrating the three-dimensional geometry of the roof with micro-meteorological field simulation, dual physical compensation for thermal infrared images is achieved, effectively eliminating temperature distortion caused by differences in observation angles and uneven local ventilation, thereby improving the accuracy of hot spot identification. In angle compensation, a nonlinear attenuation model is constructed based on the normal vector to normalize the radiation intensity, suppressing the radiation deviation of modules with large tilt angles. In heat dissipation correction, an airflow distribution map is generated using micro-meteorological field simulation, and the temperature field is dynamically corrected in combination with the convective heat transfer coefficient, weakening the non-faulty temperature rise caused by low-speed backflow in edge and corner areas.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic technology, and more specifically to a method for identifying hot spots on rooftop photovoltaic modules based on multimodal remote sensing. Background Technology

[0002] With the widespread application of distributed photovoltaic (PV) power generation systems, monitoring the operational status of rooftop PV modules has become a crucial aspect of ensuring power plant safety and power generation efficiency. Hot spot effects, a common failure mode in PV modules, cause abnormally high local temperatures, accelerating material aging. Therefore, timely and accurate identification of hot spots is essential for preventing equipment damage and improving maintenance efficiency. Multimodal remote sensing technology, due to its advantages such as wide coverage, fast response speed, and non-contact detection, is widely used in large-scale rooftop PV inspections. By fusing visible light images, thermal infrared images, and 3D point cloud data, it can achieve efficient monitoring of PV arrays in complex urban environments. However, existing technologies still face many challenges in practical applications. First, due to the diverse roof structures and varying installation tilt angles of PV modules, the observation angles of thermal infrared sensors differ, resulting in significant differences in radiation intensity for modules at the same temperature, affecting the accuracy of temperature interpretation. Second, the dense urban buildings and dynamic shadows cast by adjacent buildings or vegetation cause non-fault-related temperature anomalies in thermal infrared images, easily leading to misjudgments. Simultaneously, uneven wind distribution on the roof surface, especially in areas with localized low-speed backflow zones at the edges and corners, leads to differences in module heat dissipation conditions, further exacerbating the complexity of temperature distribution. Furthermore, the spatial registration accuracy between multi-source remote sensing data is limited by the diversity of roof materials and differences in surface reflectivity, making high-precision alignment difficult to achieve and thus affecting the reliability of subsequent fusion analysis. These issues make it difficult for current methods to meet the requirements of large-scale intelligent operation and maintenance in complex real-world scenarios regarding the accuracy and stability of hotspot identification. Summary of the Invention

[0003] Purpose of the invention: The purpose of this invention is to provide a hot spot identification method for rooftop photovoltaic modules based on multimodal remote sensing, which solves the technical problems of thermal radiation distortion caused by differences in observation angles, temperature misjudgment caused by uneven local wind fields in urban environments, insufficient registration accuracy of multi-source remote sensing data, and low identification accuracy caused by the lack of orientation adaptability of hot spot identification threshold.

[0004] Technical solution: The present invention provides a method for identifying hot spots on rooftop photovoltaic modules based on multimodal remote sensing, comprising the following steps:

[0005] (1) Acquire the three-dimensional point cloud data of the roof of the target area, the synchronously collected thermal infrared remote sensing images, visible light remote sensing images and the layout vector diagram of the photovoltaic modules on the roof;

[0006] (2) Based on the three-dimensional point cloud data, extract the normal vector of each photovoltaic module surface, construct a nonlinear attenuation model between the observation angle and the thermal radiation intensity. The model normalizes the radiation intensity by the angle between the sensor observation direction and the normal vector of the module surface, and compensates for the secondary scattering effect of the non-Lambertian surface. The radiation value of each pixel in the thermal infrared remote sensing image is normalized by angle, and the first temperature distribution map after angle compensation is output.

[0007] (3) Input the normal vector into the micro-meteorological field simulation module. The module uses the lattice Boltzmann method to construct an incompressible fluid model, combines real-time wind speed and direction data to generate an airflow distribution map on the roof surface, calculates the convective heat transfer coefficient at each location, corrects the local heat dissipation deviation of the first temperature distribution map based on the convective heat transfer coefficient, and outputs the second temperature distribution map.

[0008] (4) Based on the semantic segmentation model, the roof material classification map is extracted from the visible light remote sensing image, and the registration process of the thermal infrared image and the visible light image is weighted based on the material classification map to achieve weighted mutual information maximization registration. Then, the second temperature distribution map and the layout vector map are spatially aligned and divided into the corresponding areas of each photovoltaic module.

[0009] (5) In the aligned temperature distribution data, extract the temperature value for each photovoltaic module area. When the temperature value of a certain area exceeds the dynamic reference temperature threshold of its orientation zone, it is marked as a candidate hot spot area. The dynamic reference temperature threshold is dynamically set by a function based on the orientation zone of the module, ambient temperature, cumulative solar energy and elevation change.

[0010] (6) Match the candidate hot spot regions with the preset electrical connection topology, retain only the continuous abnormal regions located in the same series circuit, remove isolated abnormal points across series, and generate hot spot identification results;

[0011] (7) Output a hot spot identification report containing location coordinates and the identifier of the component to which it belongs.

[0012] Furthermore, in step (2), the nonlinear attenuation model is as follows:

[0013] ,

[0014] in, This represents the radiation intensity of pixel P after angle normalization. This represents the radiation intensity of pixel P in the original thermal infrared image; The angle between the sensor's observation direction and the component's surface normal vector is obtained from three-dimensional point cloud computing. The material-dependent attenuation coefficient is used to compensate for the secondary scattering effect of non-Lambertian surfaces; the nonlinear attenuation model performs batch correction of the entire image through matrix element-wise operations.

[0015] Furthermore, step (1) also includes obtaining solar altitude angle and azimuth angle data to generate building shadow masks, and using a local statistical model to perform dynamic threshold segmentation on temperature data in the shadow transition zone to suppress misjudgment of the penumbra.

[0016] Furthermore, in step (4), the weight constraint is achieved through the weighted mutual information maximization criterion, in which the diffuse reflective material region is given a high weight and the specular reflective material region is given a low weight.

[0017] Furthermore, step (4) also includes obtaining the last cleaning time and historical rainfall data, calculating the additional temperature rise caused by dirt on the component surface through an empirical attenuation model, and subtracting the additional temperature rise from the second temperature distribution map before performing hot spot judgment.

[0018] Furthermore, in step (5), the dynamic reference temperature threshold is specifically as follows:

[0019] ;

[0020] in, Indicates the orientation as The reference temperature of the component area; Indicates ambient temperature; This refers to the thermal response gain coefficient of the photovoltaic module. Indicates time The solar irradiance at a location, with the integral term representing the cumulative solar energy; Let f(s) be the orientation response function, satisfying f(s) > f(west) > f(east) > f(north); This is the elevation temperature coefficient, which reflects the temperature change caused by every 1 meter increase in altitude. This indicates the change in local roof elevation relative to the reference surface.

[0021] Furthermore, in step (6), the spatial clustering degree is calculated based on the spatial distribution coordinates in the hot spot identification results. When the spatial clustering degree exceeds a preset threshold, it is determined to be a systemic fault risk area.

[0022] Furthermore, the degree of spatial clustering is calculated using Ripley's K function.

[0023] An electronic device according to the present invention includes a memory and a processor. The memory stores a computer program, and the processor executes the program to implement the steps of the method.

[0024] The present invention discloses a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the method.

[0025] Beneficial Effects: Compared with existing technologies, this invention has the following significant advantages: By integrating the three-dimensional geometry of the roof with micro-meteorological field simulation, this invention achieves dual physical compensation for thermal infrared images, effectively eliminating temperature distortion caused by differences in observation angles and uneven local ventilation, thereby improving the accuracy of hot spot identification. A nonlinear attenuation model based on normal vectors is constructed to normalize radiation intensity, suppressing radiation deviation of large-tilt modules. Airflow distribution maps are generated using micro-meteorological field simulation, and the temperature field is dynamically corrected using the convective heat transfer coefficient, weakening the non-fault-related temperature rise caused by low-speed backflow in edge and corner areas. Through spatial alignment and electrical topology filtering mechanisms, temperature anomaly judgment is combined with the actual layout and series circuit structure of photovoltaic modules, avoiding misidentification of isolated noise points and non-circuit-related areas, improving the electrical rationality of the judgment. The dynamic reference temperature function introduces cumulative irradiance integral and orientation response factor, realizing differentiated threshold settings for modules with different orientations, enhancing the system's adaptability under complex lighting conditions. The multi-modal data collaborative processing framework improves the robustness and automation level of rooftop photovoltaic hot spot detection in dense urban environments. Attached Figure Description

[0026] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0027] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0028] like Figure 1 As shown in the figure, this embodiment of the invention provides a method for identifying hot spots on rooftop photovoltaic modules based on multimodal remote sensing, including the following steps:

[0029] S1: Acquire 3D point cloud data of the roof of the target area and synchronously acquired thermal infrared remote sensing images, and acquire a layout vector diagram of the photovoltaic modules on the roof; acquire visible light remote sensing images and extract the roof material classification map through a semantic segmentation model; acquire solar altitude angle and azimuth angle data, and generate a building shadow mask;

[0030] S2: Based on the three-dimensional point cloud data, extract the normal vector of the surface of each photovoltaic module, construct a nonlinear attenuation model between the observation angle and the thermal radiation intensity, use matrix transformation to perform angle normalization processing on the radiation value of each pixel in the thermal infrared remote sensing image, and output the first temperature distribution map after angle compensation.

[0031] S3: Input the normal vector into the preset micro-meteorological field simulation module, combine it with real-time wind speed and direction data to generate an airflow distribution map on the roof surface, calculate the convective heat transfer coefficient at each location, perform local heat dissipation deviation correction on the first temperature distribution map based on the convective heat transfer coefficient, and output a second temperature distribution map; apply weight constraints to the registration process of thermal infrared image and visible light image based on the material classification map to enhance the registration accuracy of tile and metal areas; use a local statistical model to perform dynamic threshold segmentation on temperature data in the shadow transition zone to suppress misjudgment of penumbra area;

[0032] S4: Spatially align the second temperature distribution map with the layout vector map, and divide the second temperature distribution map into the corresponding areas of each photovoltaic module;

[0033] S5: In the aligned temperature distribution data, extract the temperature value for each photovoltaic module area. When the temperature value of a certain area exceeds the dynamic reference temperature threshold of its orientation zone, it is marked as a candidate hot spot area. Obtain the last cleaning time and historical rainfall data, calculate the additional temperature rise caused by dirt on the module surface through the empirical attenuation model, and subtract the additional temperature rise from the second temperature distribution map before performing hot spot judgment.

[0034] S6: Match the candidate hot spot regions with the preset electrical connection topology map, retain only continuous abnormal regions located in the same series circuit, remove isolated abnormal points across series, and generate hot spot identification results; calculate Ripley's K function value based on the spatial distribution coordinates in the hot spot identification results, and determine the region as a systemic fault risk area when the K function value is greater than the preset aggregation threshold.

[0035] S7: Outputs a hotspot identification report containing location coordinates and the identifier of the component to which it belongs.

[0036] In one embodiment, a drone equipped with a lidar and thermal infrared sensor first acquires 3D point cloud data and thermal infrared remote sensing images of the rooftop in the target area simultaneously, ensuring a strict correspondence between spatial location and temperature information. Simultaneously, a pre-established building photovoltaic (PV) layout database is accessed to obtain the spatial arrangement and electrical connections of each PV module on the roof, forming a layout vector diagram.

[0037] The normal vectors of the photovoltaic module surfaces are extracted one by one from the point cloud data for subsequent angle correction. A nonlinear attenuation model is constructed using the angle between this normal vector and the sensor's observation direction to perform pixel-level radiation compensation on the original thermal infrared image, eliminating the underestimation or overestimation of temperature caused by tilted observation.

[0038] The normal vector is then input into the micro-meteorological field simulation module. Combined with wind speed and direction data provided by the real-time weather station, the airflow distribution on the roof surface is simulated to identify low-speed recirculation zones at the edges and corners of the modules. Based on this, the convective heat transfer coefficient at each location is calculated to correct non-fault-related temperature rises caused by poor local heat dissipation. After completing the dual correction, the temperature distribution map and the layout vector map are aligned with the geographic coordinate system, accurately mapping the temperature data to each photovoltaic module area. The highest temperature value is extracted within each module area and compared with a reference temperature dynamically generated based on orientation and accumulated solar radiation. Areas exceeding the threshold are marked as candidate hotspots. These abnormal areas are further matched with the electrical connection topology map, retaining only continuous abnormal points within the same series loop and excluding isolated noise or non-circuit-related areas. Finally, a hotspot identification report containing geographic coordinates, module number, and anomaly level is generated, achieving a complete closed loop from raw data to operation and maintenance instructions.

[0039] The nonlinear decay model constructed in S2 is implemented through the following function:

[0040] ,

[0041] in, This represents the radiation intensity of pixel P after angle normalization. This represents the radiation intensity of pixel P in the original thermal infrared image; The angle between the sensor's observation direction and the component's surface normal vector is obtained from three-dimensional point cloud computing. The material-dependent attenuation coefficient is used to compensate for the secondary scattering effect of non-Lambertian surfaces; the nonlinear attenuation model performs batch correction of the entire image through matrix element-wise operations.

[0042] In this embodiment, during the angle compensation process, a nonlinear attenuation model incorporating a secondary scattering term is introduced to address the non-Lambertian radiation characteristics observed on the photovoltaic module's glass surface at large viewing angles. This model takes the angle between the sensor's line-of-sight direction and the module's normal vector as input, using a cosine term to dominate the main radiation attenuation, while also incorporating a squared sine term and material coefficients for correction, thus compensating for the enhanced reflection and edge scattering effects at the glass-air interface. The model is computed in parallel across the entire image using a matrix form, with each pixel independently applying the normal vector information of its corresponding module, achieving efficient batch processing.

[0043] This method overcomes the limitation of traditional cosine correction, which is only applicable to ideal diffuse reflective surfaces. It is particularly suitable for high-reflectivity glass encapsulation components commonly found in rooftop photovoltaic arrays, and significantly improves the temperature restoration accuracy in areas with large tilt angles or side views.

[0044] The dynamic reference temperature threshold in S5 is determined by the following function:

[0045] ;

[0046] in, Indicates the orientation as The reference temperature of the component area; Indicates ambient temperature; This refers to the thermal response gain coefficient of the photovoltaic module. Indicates time The solar irradiance at a location, with the integral term representing the cumulative solar energy; Let f(s) be the orientation response function, satisfying f(s) > f(west) > f(east) > f(north); This is the elevation temperature coefficient, which reflects the temperature change caused by every 1 meter increase in altitude. This indicates the change in local roof elevation relative to the reference surface.

[0047] In an embodiment, the orientation response function The specific values ​​are: 1.0 for south, 0.7 for east, 0.8 for west, and 0.4 for north.

[0048] The dynamic reference temperature setting comprehensively considers the orientation of the components, the cumulative solar irradiance of the day, and local elevation differences. Ambient air temperature serves as the basic reference, and solar irradiance is integrated over time to reflect the heat accumulation process of the components since morning, avoiding interference from instantaneous irradiance fluctuations. Components with different orientations are assigned different response weights due to differences in sunshine duration and intensity; south-facing components receive the most sunshine and have the highest weight, while north-facing components have the lowest. Simultaneously, elevation changes in different areas of the roof affect natural convection efficiency, with higher areas dissipating heat faster; an elevation temperature variation coefficient is introduced for negative compensation. This function is implemented through piecewise interpolation, ensuring adaptive adjustment of the judgment threshold under different geographical latitudes and climatic conditions, solving the problem of high false alarm rates with fixed thresholds in the morning / evening, seasonal, and urban microclimate conditions.

[0049] In S5, the dynamic reference temperature threshold is determined based on the orientation partition information in the layout vector diagram, with different reference temperature offsets corresponding to the south, east, west, and north directions.

[0050] Among them, the micro-meteorological field simulation module in S3 uses the lattice Boltzmann method to construct an incompressible fluid model, and the airflow distribution map is used to determine the low-speed recirculation zone at the edge and corner of the component.

[0051] In this embodiment, to improve the registration accuracy of multimodal images, a material-aware mechanism is introduced before aligning the thermal infrared and visible light images. A pre-trained semantic segmentation model is used to analyze the visible light images, identifying different material regions such as tiles, metal, glass, and vegetation. Because different materials have significantly different thermal radiation and optical reflection characteristics, the registration algorithm assigns higher weights to regions with strong diffuse reflection and clear textures, such as tiled roofs, during the optimization process. Conversely, it reduces the weights of regions with significant specular reflection and prone to flare, such as metal or glass, to prevent them from interfering with the overall transformation parameter solution. This strategy ensures sub-pixel-level registration even under complex roof surface conditions, guaranteeing accurate mapping of temperature and component positions in the subsequent process.

[0052] In this embodiment, the registration process employs a weighted mutual information maximization criterion as the optimization objective function. Mutual information measures the statistical correlation between corresponding regions of two images and is a commonly used metric for multimodal registration. This method introduces a spatial weight matrix, generated from a material classification map, assigning different weight coefficients to regions of different materials. When optimizing image transformation parameters, the algorithm prioritizes maximizing the mutual information of high-weight regions, thereby improving the alignment accuracy of key areas. This method does not rely on feature point extraction, is suitable for roof scenes with missing or repetitive textures, and effectively suppresses the misleading influence of specular reflection regions on registration.

[0053] The weight constraints are implemented through a weighted mutual information maximization criterion, in which diffuse reflective material regions are given higher weights and specular reflective material regions are given lower weights.

[0054] In this embodiment, after generating the hotspot identification results, their spatial distribution patterns are further analyzed. The spatial clustering degree of anomalies is assessed by calculating Ripley's K function. If multiple hotspots significantly cluster within a small spatial area, it indicates a potential systemic fault, such as DC-side grounding anomalies, string impedance mismatch, or localized shading design defects. This function identifies clustering phenomena exceeding expectations of random distribution by statistically analyzing point-pair densities at different distance scales, providing maintenance personnel with a basis for fault type identification, escalating from single component failures to system-level risk warnings.

[0055] In shadow transition regions where illumination changes drastically, conventional temperature thresholds are prone to false alarms. This method, after angle and heat dissipation corrections, introduces a shadow mask based on solar geometry to identify pixels in the penumbra. A local adaptive segmentation strategy is employed for these regions, dynamically setting the judgment threshold based on the mean and variance of the neighborhood temperature distribution to avoid false temperature rise alarms caused by gradual illumination changes. This processing is spatially localized and logically placed after physical correction, ensuring that it does not affect the global consistency of the main process.

[0056] Dust accumulation on component surfaces can form a heat insulation layer, leading to increased normal operating temperatures and interfering with hot spot detection. This method introduces a dirt temperature rise compensation step before temperature determination. Based on the system's recorded last cleaning time and historical rainfall data provided by the meteorological department, the current dust accumulation level is estimated using an empirical attenuation model. This model assumes that rainfall has a cleaning effect, which decays exponentially over time. The calculated additional temperature rise is subtracted from the corrected temperature field to restore the theoretical temperature of the component under clean conditions, thereby avoiding misjudgments caused by environmental factors and improving the stability of long-term monitoring.

Claims

1. A method for identifying hot spot of roof photovoltaic module based on multi-modal remote sensing, characterized in that, The method comprises the following steps: (1) obtaining the three-dimensional point cloud data of the roof of the target area, the synchronously collected thermal infrared remote sensing image, the visible light remote sensing image, and the layout vector diagram of the photovoltaic modules on the roof; (2) extracting the normal vector of the surface of each photovoltaic module based on the three-dimensional point cloud data, constructing a nonlinear attenuation model between the observation angle and the thermal radiation intensity, normalizing the radiation intensity through the angle between the sensor observation direction and the normal vector of the module surface in the model, compensating for the secondary scattering effect of the non-Lambertian surface, and performing angle normalization processing on the radiation value of each pixel in the thermal infrared remote sensing image to output a first temperature distribution diagram after angle compensation; The nonlinear attenuation model is as follows: ， wherein, represents the radiance intensity of the pixel P after angle normalization; represents the radiance intensity of the pixel P in the original thermal infrared image; represents the included angle between the sensor observation direction and the surface normal vector of the component, which is calculated by a three-dimensional point cloud; is a material-dependent attenuation coefficient, which is used to compensate for the secondary scattering effect of non-Lambertian surfaces; the non-linear attenuation model is corrected in batch for the entire image through matrix element-by-element operation; (3) inputting the normal vector into a micro-meteorological field simulation module, constructing an incompressible fluid model by using the lattice Boltzmann method, generating a roof surface airflow distribution diagram combined with real-time wind speed and direction data, calculating the convective heat transfer coefficient at each position, performing local heat dissipation deviation correction on the first temperature distribution diagram based on the convective heat transfer coefficient, and outputting a second temperature distribution diagram; (4) extracting a roof material classification diagram from the visible light remote sensing image based on a semantic segmentation model, and imposing a weight constraint on the registration process of the thermal infrared image and the visible light image based on the material classification diagram to realize weighted mutual information maximization registration, and then spatially aligning the second temperature distribution diagram with the layout vector diagram and dividing it into the corresponding regions of each photovoltaic module; (5) in the aligned temperature distribution data, extracting the temperature value for each photovoltaic module region, and marking a candidate hot spot region when the temperature value of a certain region exceeds the dynamic reference temperature threshold value of the orientation partition where the component is located; the dynamic reference temperature threshold value is dynamically set according to the orientation partition where the component is located, the ambient temperature, the cumulative solar energy, and the elevation change through a function; the dynamic reference temperature threshold value is as follows: ； wherein, represents the reference temperature of the component area with the orientation ; represents the ambient air temperature; is the thermal response gain coefficient of the photovoltaic component; represents the solar irradiance at time , and the integral term represents the cumulative solar energy; is the orientation response function, satisfying f(south) > f(west) > f(east) > f(north); is the elevation temperature variation coefficient, reflecting the temperature change caused by each 1-meter increase in altitude; represents the change amount of the local elevation of the roof relative to the reference surface; (6) matching the candidate hot spot region with a pre-set electrical connection topology diagram, retaining only the continuous abnormal regions located in the same string loop, removing the cross-string isolated abnormal points, and generating a hot spot recognition result; (7) outputting a hot spot recognition report containing position coordinates and belonging component identifiers.

2. The method according to claim 1, wherein, Step (1) further comprises obtaining solar elevation angle and azimuth angle data to generate a building shadow mask, and using a local statistical model for dynamic threshold segmentation on the temperature data in the shadow transition zone to suppress misjudgment in the half-shadow area.

3. The method of claim 1, wherein, In step (4), the weight constraint is realized through the weighted mutual information maximization criterion, wherein the diffuse reflection material region is given a high weight, and the specular reflection material region is given a low weight.

4. The method of claim 1, wherein, In step (4), it further comprises obtaining the last cleaning time and historical rainfall data, calculating the additional temperature rise caused by the dirt on the surface of the component through an empirical attenuation model, and subtracting the additional temperature rise from the second temperature distribution diagram before performing hot spot judgment.

5. The method of claim 1, wherein, In step (6), it further comprises calculating the spatial aggregation degree based on the spatial distribution coordinates in the hot spot recognition result, and determining a systematic failure risk area when the spatial aggregation degree exceeds a pre-set threshold.

6. The method of claim 5, wherein, The spatial aggregation degree is calculated by Ripley's K function.

7. An electronic device, comprising: A computer program product comprising a memory storing a computer program and a processor, which, when executing the program, implements the steps of the method according to any one of claims 1-5.

8. A computer-readable storage medium, characterized in that, A computer program product comprising a memory storing a computer program and a processor, which, when executing the program, implements the steps of the method according to any one of claims 1-5.

Images