Photovoltaic fault accurate identification and root cause diagnosis method based on steady-state profile characteristics

Abstract

The present application relates to photovoltaic power generation operation monitoring and fault diagnosis field, disclose a kind of photovoltaic fault accurate identification and root cause diagnosis method based on steady-state profile feature, including obtaining time synchronization target group string space initial thermal gradient field data and transient electrical characteristic data, address is established by spatial electrical topology hash table three-dimensional coordinate to bottom layer network node, extract reference group string output data to calculate electrical power residual error, solve adaptive convection heat compensation factor, the factor is applied to edge detection activation threshold to process thermal field data, extract morphological topology attribute and construct physical thermal characteristic vector, call historical reference data to perform difference operation on transient electrical data, obtain impedance distortion parameter and ac phase angle feature to construct multidimensional electrical characteristic vector, two kinds of vectors are input into feature discrete determination matrix to execute joint determination, output fault type and spatial coordinates.The present application reduces environmental heat dissipation interference, improves fault root cause diagnosis accuracy.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic power generation operation monitoring and fault diagnosis technology, specifically a method for accurate identification and root cause diagnosis of photovoltaic faults based on steady-state profile characteristics. Background Technology

[0002] With the continuous expansion of photovoltaic power plant installed capacity, monitoring the operational status and troubleshooting of photovoltaic modules have become crucial aspects of power plant operation and maintenance. Current photovoltaic fault diagnosis primarily relies on infrared thermal imaging inspections and conventional electrical data analysis of inverters. However, in actual power plant operation scenarios, infrared thermal image acquisition and underlying power grid data sampling are typically conducted in independent systems and at different time dimensions. Due to the lack of a mechanism for associating physical spatial coordinates with underlying electrical network nodes, the extracted thermal anomaly data and electrical response data are difficult to align in time and space, making it impossible to effectively cross-verify the apparent thermal characteristics with the underlying electrical properties.

[0003] In image-based thermal feature diagnosis, existing technologies are easily affected by natural environmental factors. Especially under complex micro-topographic conditions, local airflow changes and wind speed differences can lead to uneven convective heat dissipation on the surface of photovoltaic modules. Traditional diagnostic systems typically use fixed temperature gradient thresholds to determine thermal anomalies, lacking compensation mechanisms for localized environmental heat dissipation differences. This can easily lead to misjudging normal uneven heat dissipation as module hot spot faults, increasing the false alarm rate in operation and maintenance.

[0004] In electrical characteristic-based diagnostics, conventional methods primarily collect macroscopic operating data such as DC voltage, current, or output power during steady-state operation of the module. This type of steady-state data typically only reflects the overall decline in the power generation performance of the photovoltaic module and struggles to reveal microscopic physical changes in the internal circuitry. Because it doesn't delve into and analyze deeper electrical characteristics such as high-frequency AC impedance and phase distortion, existing technologies cannot effectively distinguish the specific physical causes of power degradation. For example, it's difficult to accurately identify whether the module performance degradation is caused by external shading, internal mechanical cracks in the cells, or material-level potential-induced degradation, resulting in insufficient precision in fault diagnosis. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method for accurate identification and root cause diagnosis of photovoltaic faults based on steady-state profile features. This method solves the problems in existing photovoltaic fault diagnosis, such as the difficulty in aligning thermoelectric features in time and space, the tendency for micro-topography heat dissipation differences to lead to false alarms of thermal anomalies, and the inability of conventional steady-state electrical parameters to accurately distinguish the specific physical root cause of the fault.

[0006] The first aspect of this invention provides a method for accurate identification and root cause diagnosis of photovoltaic faults based on steady-state profile features, comprising the following steps: Step S100: Obtain the initial spatial thermal gradient field data and transient electrical characteristic data of the target photovoltaic string that are synchronized with time, and establish direct addressing from physical three-dimensional coordinates to the underlying network nodes through a pre-loaded offline spatial electrical topology hash table. Step S200: Extract the high-frequency electrical output data of the reference string to calculate the electrical power residual, and inversely calculate the adaptive closed-loop convection heat dissipation compensation factor under the current micro-topography environment based on the electrical power residual. Step S300: Apply the adaptive closed-loop convection heat dissipation compensation factor to the activation threshold of the edge detection algorithm, perform dynamic threshold anti-disturbance processing on the initial spatial thermal gradient field data, and extract standardized morphological topological attributes to construct a physical thermal feature vector. Step S400: Call the historical fault-free steady-state operation reference data to perform reference differential operation on the transient electrical characteristic data, calculate the impedance distortion characteristic parameters, and fuse the impedance distortion characteristic parameters with the AC impedance phase angle characteristics to construct a multi-dimensional electrical characteristic vector; Step S500: Input the physical thermal feature vector and the multidimensional electrical feature vector into a preset feature discrete judgment matrix to perform orthogonal logic joint judgment, and output the fault type label of the photovoltaic module and the corresponding spatial coordinate data.

[0007] Furthermore, the process of establishing direct addressing and acquiring data includes: calculating the local temperature gradient amplitude based on the image pixel matrix acquired by the infrared thermal imager; when the local temperature gradient amplitude is greater than the judgment benchmark value, triggering the time capture channel to freeze the current timestamp and synchronously latching the corresponding physical three-dimensional coordinates; extracting the latitude and longitude data from the physical three-dimensional coordinates and using a spatial hash algorithm to reduce its dimension and map it into discrete spatial hash key values; using the spatial hash key value as an index to retrieve the offline spatial electrical topology hash table, extracting the network protocol address of the group control gateway in the area where the abnormal target point is located and the target inverter node identification code, and directly sending the underlying trigger frame carrying the frozen timestamp information through the local area network link to control the underlying inverter to deviate from steady-state operation and perform transient DC voltage linear scanning and high-frequency AC ripple extraction.

[0008] Furthermore, the process of calculating the electrical power residual includes: establishing a local spatial search domain centered on the three-dimensional coordinates of the target photovoltaic string; calculating the absolute elevation difference between neighboring candidate strings and the target photovoltaic string within the local spatial search domain; selecting strings whose absolute elevation difference is less than a preset elevation difference tolerance threshold and are in steady-state operation as benchmark strings; using the frozen timestamp information as the time reference anchor point, locking the historical data slices of the benchmark strings in the high-frequency ring buffer of the underlying group control gateway; extracting the instantaneous power corresponding to the two neighboring valid sampling nodes that enclose the frozen timestamp information; eliminating the discrete sampling frequency difference through linear interpolation operation of time domain approximation; calculating the benchmark actual electrical output power aligned with the frozen timestamp information time; using the transient environmental irradiance data and transient environmental temperature data aligned with the frozen timestamp information time sequence into the physical equivalent circuit model to calculate the theoretical output power of the benchmark strings; subtracting the theoretical output power from the benchmark actual electrical output power to obtain the electrical power residual characterizing the local wind cooling difference.

[0009] In a preferred embodiment of the present invention, the process of reverse calculation of the adaptive closed-loop convection heat dissipation compensation factor is as follows: when the theoretical output power is greater than the set lower limit of grid-connected power generation, based on the relative ratio of the electrical power residual to the theoretical output power, combined with the pre-calibrated empirical correlation coefficient of the photovoltaic module thermoelectric response, the macroscopic initial convection heat dissipation compensation factor introduced by the system is proportionally superimposed and corrected, and the local real convection heat dissipation compensation factor is calculated and obtained as the adaptive closed-loop convection heat dissipation compensation factor.

[0010] Furthermore, the dynamic threshold anti-disturbance processing includes: performing spatial Gaussian smoothing filtering on the acquired initial two-dimensional temperature distribution matrix; calculating the local temperature partial derivatives in the horizontal and vertical directions using discrete difference operators, and synthesizing vector magnitudes to construct the initial spatial thermal gradient field data; using a bidirectional limiter to clamp the boundary range of the adaptive closed-loop convection heat dissipation compensation factor between its maximum and minimum values; dividing the standard activation threshold of the standard morphological edge detection algorithm by the limit-constrained adaptive closed-loop convection heat dissipation compensation factor to obtain the adaptive activation threshold; and using the adaptive activation threshold to perform binarization traversal on the initial spatial thermal gradient field data, extracting pixel coordinates greater than or equal to the adaptive activation threshold as effective edge response points to generate a binarized thermal gradient edge image.

[0011] Preferably, the process of extracting standardized morphological topological attributes to construct physical thermal feature vectors includes: performing connected component analysis on the binarized thermal gradient edge image, calculating the actual equivalent physical heating area, boundary perimeter, and aspect ratio of the minimum bounding rectangle for each independent thermal anomaly connected region, and calculating the morphological compactness parameter using the ratio of pixel area to the square of the boundary perimeter; when the aspect ratio is greater than a set threshold, classifying the physical thermal topological attribute as a linear radial boundary label; when the aspect ratio is not greater than a set threshold, the actual equivalent physical heating area is less than a set first area constraint threshold, and the morphological compactness parameter is greater than a compactness limit threshold, classifying it as a closed sharp boundary label; when the actual equivalent physical heating area is greater than a second area constraint threshold and the morphological compactness parameter is less than a compactness limit threshold, classifying it as a large-area diffuse boundary label; and encapsulating each classification label into a multidimensional physical thermal feature vector using a unique thermal coding rule.

[0012] Furthermore, the process of obtaining impedance distortion characteristic parameters and AC impedance phase angle characteristics includes: substituting the voltage and current sampling points in the transient electrical characteristic data into the main equation of the standard photovoltaic single diode physical equivalent circuit model; using the extracted short-circuit current, open-circuit voltage, and maximum power characteristic points to calculate the initial iterative estimate; using the nonlinear least squares objective function and the Levenberg-Marquardt numerical optimization algorithm to iteratively approximate and solve the equivalent series resistance and equivalent parallel resistance of the photovoltaic module; using the real-time environmental irradiance parameter and the real-time backsheet temperature parameter as the retrieval matching primary key, and using the bilinear interpolation algorithm... Extract the historical reference equivalent series resistance and historical reference equivalent parallel resistance under the same meteorological conditions from the historical fault-free operation status feature database; calculate the deterioration drift rate of the equivalent series resistance and equivalent parallel resistance relative to their respective historical reference values, and introduce a sensitivity amplification factor for linear stretching to obtain the series differential amplification impedance distortion characteristic parameter and the parallel differential amplification impedance distortion characteristic parameter; traverse the high-frequency AC impedance phase angle data obtained by injecting high-frequency AC disturbance signals, and extract the AC impedance phase angle characteristic value corresponding to the characteristic resonant frequency point where the imaginary part reactance zero-crossing point or the extreme point of the first derivative of the phase angle abrupt change is located.

[0013] Preferably, the process of constructing the multidimensional electrical feature vector includes: retrieving the historical benchmark mean set and historical benchmark standard deviation set calculated statistically during the historical fault-free operation cycle; using the standard normal distribution drift transformation logic of extracting physical feature values, subtracting the benchmark statistical mean, and dividing by the benchmark statistical standard deviation to eliminate the heterogeneous dimensional differences between the series differential amplifier impedance distortion feature parameters, the parallel differential amplifier impedance distortion feature parameters, and the AC impedance phase angle feature values, respectively generating normalized series impedance distortion features, normalized parallel impedance distortion features, and normalized AC phase angle features; mapping and splicing the above three normalized features in a fixed channel arrangement order in a continuous memory address space to assemble a high-dimensional structured multidimensional electrical feature vector.

[0014] Furthermore, the preset feature discrete judgment matrix is ​​constructed as follows: a Gaussian mixture distribution model is used to fit the probability density of the health benchmark sample set in the early stage of production to obtain the benchmark drift mean and standard deviation of series impedance, parallel impedance and AC phase angle under healthy operating conditions; the basic confidence boundary is calculated based on the Laida criterion and the standard deviation, and the product of the total historical operating time of the photovoltaic module and the annualized attenuation sensitivity coefficient is superimposed to establish the positive drift threshold of series resistance, the negative drift threshold of parallel resistance and the phase angle deflection threshold of AC impedance with time dimension attenuation compensation; the normalized features in the multidimensional electrical feature vector are compared with the corresponding thresholds after dimensional alignment, and the electrical state judgment vector composed of Boolean discrete variables is generated according to the exceedance of the comparison and constructed as the feature discrete judgment matrix.

[0015] In another preferred embodiment of the present invention, the process of performing orthogonal logic joint determination includes: when the morphological label of the physical thermal feature vector is a closed sharp boundary, and the feature discrete determination matrix indicates that the parallel impedance has a significant negative drift and the series impedance and AC phase angle have not been distorted, it is determined to be an external shading recoverable fault; when the morphological label of the physical thermal feature vector is a linear radial boundary, and the feature discrete determination matrix indicates that the series impedance has a significant positive drift or the AC phase angle has a significant deflection, it is determined to be an internal irreversible damage fault caused by microcracks in the battery cell; when the morphological label of the physical thermal feature vector is a large-area diffuse boundary, and the feature discrete determination matrix indicates that the series impedance has a significant positive drift and the open-circuit voltage attenuation amplitude in the transient electrical feature data exceeds the voltage tolerance threshold, it is determined to be a potential-induced attenuation fault caused by material-level global degradation, and an automated maintenance instruction containing the target physical coordinates and high-voltage reverse DC bias parameters is generated and sent to the operation and maintenance management platform.

[0016] This invention provides a method for accurate identification and root cause diagnosis of photovoltaic faults based on steady-state profile features. It has the following beneficial effects:

[0017] 1. This invention establishes a direct addressing mechanism between physical three-dimensional coordinates and underlying network nodes by constructing an offline spatial electrical topology hash table. When a local temperature gradient anomaly is detected, the system synchronously freezes the timestamp and triggers the underlying inverter to extract transient electrical feature data. This mechanism ensures that infrared thermal field data and the underlying electrical response remain aligned in both time and space, overcoming the technical deficiency in traditional inspection methods where there is a time lag between thermal image acquisition and electrical data, making mutual verification difficult.

[0018] 2. This invention calculates the electrical power residual by comparing the theoretical and actual output power of a series of reference assemblies at the same elevation within a local spatial search domain. This residual is then used to calculate an adaptive closed-loop convection heat dissipation compensation factor, which is then used to dynamically adjust the activation threshold for thermal field edge detection. This method transforms the power deviation on the electrical side into a basis for local environmental heat dissipation compensation, eliminating the need for additional deployment of micro-meteorological equipment such as wind speed sensors. This reduces the misjudgment rate of thermal anomalies caused by uneven local airflow heat dissipation in complex micro-topography.

[0019] 3. This invention extracts the standardized morphological and topological attributes of the thermal gradient image, as well as the equivalent impedance and AC phase angle features based on benchmark differential operations. The physical thermal feature vector and the multi-dimensional electrical feature vector are then input into a preset matrix for joint orthogonal logic determination. Through cross-comparison of apparent thermal morphology and underlying impedance and electrical features, the system can distinguish the physical root causes of specific faults such as external shading, internal microcracks in the solar cell, and potential-induced attenuation, overcoming the limitation of single-dimensional data in determining specific fault types. Attached Figure Description

[0020] Figure 1 This is a system framework diagram of the present invention; Figure 2 This is a schematic diagram of the method flow of the present invention; Figure 3 This is a flowchart of the physical addressing and transient electrical feature synchronous extraction process of the present invention; Figure 4 This is a flowchart illustrating the calculation process of the adaptive closed-loop convection heat dissipation compensation factor of the present invention.

[0021] Among them, 100 is the spatial inspection and edge computing module; 200 is the underlying electrical operation and group control gateway module; and 300 is the cloud-based orthogonal decoupling and operation and maintenance scheduling module. Detailed Implementation

[0022] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] See attached document Figure 1 To be continued Figure 4 This invention provides a photovoltaic fault precision decoupling diagnosis system based on air-ground collaborative triggering and thermoelectric closed-loop calibration. The system includes: a space inspection and edge computing module 100, a bottom-level electrical operation and group control gateway module 200, and a cloud-based orthogonal decoupling and operation and maintenance scheduling module 300.

[0024] The spatial inspection and edge computing module 100 is mounted on an inspection UAV platform. The module includes an infrared thermal imager 110, a high-precision positioning unit 120, an airborne edge computing unit 130, and a local low-latency communication unit 140. The infrared thermal imager 110 collects spatial thermal distribution data of the photovoltaic power station. The high-precision positioning unit 120 provides the three-dimensional coordinates of the current acquisition location. The airborne edge computing unit 130 embeds an offline spatial electrical topology hash mapping table. It is configured to process the image data from the infrared thermal imager 110 to extract initial thermal anomaly areas and combine it with the data from the high-precision positioning unit 120 to perform local addressing matching from physical coordinates to network nodes. The local low-latency communication unit 140 sends low-level trigger frames to the underlying electrical operation and group control gateway module 200.

[0025] The underlying electrical operation and group control gateway module 200 is deployed at the photovoltaic power station site. The underlying electrical operation and group control gateway module 200 includes a string inverter array 210, an internal high-frequency sampling circuit 220, and a group control gateway 230. The string inverter array 210 is electrically connected to the photovoltaic strings, performing AC-DC conversion and maximum power point tracking operation. The internal high-frequency sampling circuit 220 is integrated inside the string inverter array 210 and is used for high-frequency acquisition of voltage and current operating data and AC ripple data. The group control gateway 230 has a built-in high-frequency ring buffer 231. The group control gateway 230 is configured to receive the underlying trigger frame broadcast by the local low-latency communication unit 140, respond to external interrupt signals to control the string inverter array 210 to perform transient DC voltage and current curve scanning, and simultaneously control the high-frequency ring buffer 231 to freeze and latch the microsecond-level electrical data acquired by the internal high-frequency sampling circuit 220.

[0026] The cloud-based orthogonal decoupling and operation and maintenance scheduling module 300 is deployed on a remote server. This module includes a historical steady-state database 310, a multi-dimensional feature fusion calculation engine 320, and a logic judgment matrix diagnostic engine 330. The historical steady-state database 310 stores historical fault-free steady-state operation baseline data for the entire photovoltaic string system. The multi-dimensional feature fusion calculation engine 320 is configured to call the data from the historical steady-state database 310 to perform feature difference calculations, fusing high-frequency AC impedance parameters to construct a multi-dimensional electrical feature vector. The logic judgment matrix diagnostic engine 330 is configured to receive the physical thermal feature vector output by the spatial inspection and edge computing module 100 and the multi-dimensional electrical feature vector output by the multi-dimensional feature fusion calculation engine 320, perform orthogonal matrix mapping judgment, and generate maintenance instructions with coordinate markers.

[0027] The method provided by this invention is based on the above-mentioned photovoltaic fault precision decoupling diagnosis system. The overall workflow of this method includes the following steps: In step S100, the spatial inspection and edge computing module 100 performs on-demand transient triggering of air-to-ground spatiotemporal handshake and edge direct connection, establishes direct addressing from physical three-dimensional coordinates to underlying network nodes through a pre-loaded offline spatial electrical topology hash mapping table, and obtains time-synchronized spatial initial thermal gradient field data and transient electrical characteristic data.

[0028] In step S200, the underlying electrical operation and group control gateway module 200 uses the high-frequency ring buffer 231 to extract the high-frequency electrical output data of the reference string that has not triggered the scanning command and calculate the power residual. The multi-dimensional feature fusion calculation engine 320 calculates the adaptive closed-loop convection heat dissipation compensation factor in the current local micro-terrain environment based on the power residual.

[0029] In step S300, the airborne edge computing unit 130 applies the adaptive closed-loop convection heat dissipation compensation factor to the activation threshold of the edge detection algorithm, performs dynamic threshold anti-disturbance processing on the initial spatial thermal gradient field data, and extracts standardized morphological topological attributes to construct a physical thermal feature vector.

[0030] In step S400, the multi-dimensional feature fusion calculation engine 320 calls the fault-free reference parameters in the historical steady-state database 310 to perform benchmark differential calculation, calculates the component impedance drift rate, and fuses the DC differential drift parameters with the AC impedance parameters to construct a multi-dimensional electrical feature vector.

[0031] In step S500, the logic judgment matrix diagnostic engine 330 inputs the physical thermal feature vector and the multi-dimensional electrical feature vector into the preset orthogonal logic judgment matrix to perform joint constraint judgment, and outputs the determined fault type label and the corresponding spatial coordinate data.

[0032] The edge computing unit 130 receives infrared radiation data of the photovoltaic module array surface collected by the infrared thermal imager 110 through the internal bus, and simultaneously receives the high-frequency three-dimensional coordinate signal sequence and the universal coordination timestamp signal sequence output by the positioning unit 120. The process of converting the infrared radiation data into the initial temperature two-dimensional distribution matrix can be carried out based on the existing infrared temperature measurement and imaging calibration technology.

[0033] The edge computing unit 130 performs a preliminary screening step S111 for local thermal anomalies based on the initial two-dimensional temperature distribution matrix. It characterizes the abrupt changes in thermal distribution on the photovoltaic module surface by calculating the temperature difference between adjacent pixels. Since a photovoltaic module generates local hot spots and causes a significant temperature jump between the defective area and the surrounding normal area when a fault occurs, the temperature gradient amplitude is used as a judgment criterion. Local temperature change features are extracted by traversing the image pixel matrix. During this process, the local temperature gradient amplitude is used as the detection criterion and combined with the following formula for numerical calculation. In the formula, Indicates the magnitude of the local temperature gradient. This represents the temperature value of the corresponding pixel in the initial two-dimensional temperature distribution matrix. and These represent the horizontal and vertical pixel coordinates in the image pixel matrix, respectively. This represents the partial derivative of the initial temperature two-dimensional distribution matrix in the horizontal direction. This represents the partial derivative of the initial temperature two-dimensional distribution matrix in the vertical direction.

[0034] When the calculated local temperature gradient amplitude is greater than the steady-state judgment benchmark value built into the system hardware, the edge computing unit 130 determines that there is an initial thermal anomaly in the current image block. As a result, a rising edge trigger signal with a time stamp is generated through the internal logic control circuit. If the calculated local temperature gradient amplitude is not greater than the steady-state judgment benchmark value, the next pixel block is traversed. The steady-state judgment benchmark value is obtained by collecting the maximum temperature fluctuation gradient amplitude under the same environmental conditions in the historical fault-free state of the photovoltaic power station and adding a preset safety margin calibration.

[0035] At the moment the rising edge trigger signal is generated, the edge computing unit 130 executes the spatiotemporal parameter hard latching sub-step S112, using the timer of the internal microcontroller to capture the universally coordinated timestamp at the moment the rising edge trigger signal is generated. At the same time, it extracts latitude and longitude data and elevation data that match the frozen universally coordinated timestamp from the high-frequency three-dimensional coordinate signal sequence output by the positioning unit 120. The extracted universally coordinated timestamp and three-dimensional coordinate parameters are jointly encapsulated and written into the independent address space of the local flash memory medium to complete the spatial information latching of the corresponding abnormal target point.

[0036] After the edge computing unit 130 completes the spatial information latching of the abnormal target points, it executes the spatial coordinate discretization sub-step S121. To achieve efficient conversion of three-dimensional spatial coordinates to network addresses in the computationally limited edge computing unit 130, a spatial hash algorithm is used to reduce the dimensionality of continuous geographic coordinates into discrete one-dimensional index keys, thereby accelerating the local table lookup and address matching process. Specifically, the edge computing unit 130 converts the latched three-dimensional coordinate parameters into discrete spatial block index keys for local table lookup. Since the electrical topology of the photovoltaic array has a two-dimensional array physical pattern on the horizontal projection plane, and the elevation data is independently retained and encapsulated by the edge computing unit 130 for subsequent adaptive calculation of micro-topography wind-cooling differences, the edge computing unit 130 extracts the latitude and longitude data from the three-dimensional coordinate parameters for two-dimensional dimensionality reduction mapping. The edge computing unit 130 pre-divides the physical space of the photovoltaic site into multiple standard grid blocks during the offline stage, and uses a set hash mapping function to map the extracted latitude and longitude data into unique spatial hash keys. The numerical calculation of the hash mapping function is performed using the following formula. In the formula, This represents the calculated spatial hash key value. This indicates the floor function. This represents the longitude data in the latched three-dimensional coordinate parameters. This represents the latitude data in the latched three-dimensional coordinate parameters. The longitude of the reference origin representing the physical space of a photovoltaic power station. The latitude of the reference origin representing the physical space of a photovoltaic power station. This represents the longitude span of a single standard grid block. This indicates the latitudinal span of a single standard grid block. This indicates the total number of grid columns in the latitudinal direction of the photovoltaic power station's physical space. The longitude, latitude, and total number of grid columns of the reference origin are set according to the actual physical mapping boundary data of the photovoltaic power station. The longitude span and latitude span of a single standard grid block are determined according to the actual spatial physical distribution range of a single inverter device in the underlying electrical operation and group control gateway module 200.

[0037] After obtaining the spatial hash key value, the edge computing unit 130 executes the local mapping matching sub-step S122. The edge computing unit 130 determines whether the spatial hash key value is within the preset range of valid key values. If the spatial hash key value exceeds the range of valid key values, it determines that the currently latched three-dimensional coordinate parameters are out of bounds and generates an out-of-bounds exception log, while terminating the current direct connection addressing process. If the spatial hash key value is within the range of valid key values, the edge computing unit 130 calls the pre-loaded offline spatial electrical topology hash table in the internal storage medium. The offline spatial electrical topology hash table stores the photovoltaic power station's global spatial block index and electrical pipes. The edge computing unit 130 uses spatial hash keys as a retrieval index to perform address matching in the offline spatial electrical topology hash table. It extracts the media access control address and network protocol address of the group control gateway 230 in the convergence area where the abnormal target point is located, which is locked by the spatial hash key value. It also extracts the target inverter node identification code corresponding to the abnormal target point. The hash collision during the offline spatial electrical topology hash table retrieval process can be handled according to the existing chaining method or open addressing method.

[0038] After extracting the medium access control address, network protocol address, and target inverter node identification code of the group control gateway 230, the edge computing unit 130 controls the communication unit 140 to execute the edge direct broadcast sub-step S123. Based on the extracted medium access control address and network protocol address, the communication unit 140 sends a low-level trigger frame with frozen timestamp information directly to the group control gateway 230 pointed to by the network protocol address through the local area directional self-organizing network link. The edge computing unit 130 encapsulates the target inverter node identification code in the data payload of the low-level trigger frame. The direct broadcast communication interaction process between the communication unit 140 and the group control gateway 230 is limited to the edge end of the photovoltaic power station and does not go through the cloud server for wide area network relay. This effectively reduces the transmission delay and data congestion fluctuation risks common in public network communication links, enabling the low-level trigger frame to reach the low-level network node in a high real-time communication mode. This ensures that the spatial heat distribution data obtained by the inspection module 100 and the transient electrical feature data subsequently extracted by the electrical control module 200 have a high-precision alignment relationship at the physical time source.

[0039] After receiving the underlying trigger frame sent by the communication unit 140, the group control gateway 230 executes the hardware interrupt and steady-state deviation sub-step S131. The group control gateway 230 parses the target network protocol address carried in the underlying trigger frame and confirms that the target network protocol address matches. The group control gateway 230 further parses the data load of the underlying trigger frame to obtain the target inverter node identification code. The group control gateway 230 sends a high-priority hardware interrupt command to the unique target inverter corresponding to the target inverter node identification code in the inverter array 210 through the internal controller local area network bus. If the target network protocol address does not match, the group control gateway 230 actively discards the underlying trigger frame and maintains the original operating state. The microprocessor inside the target inverter responds to the high-priority hardware interrupt command and suspends the running maximum power point tracking algorithm process. The target inverter stops searching for the maximum power point and thus deviates from the steady-state operating point.

[0040] After the target inverter deviates from its steady-state operating point, the transient wide-scan sub-step S132 is executed. During normal grid-connected power generation of the photovoltaic modules, the system can only acquire steady-state electrical data for a single maximum power operating point. To obtain the broad impedance response characteristics of the photovoltaic string under multiple voltage conditions for subsequent hidden defect analysis, the digital signal processor inside the target inverter adjusts the duty cycle of the pulse width modulation signal in the boost chopper circuit to control the DC bus voltage at the input of the photovoltaic string to perform a linear voltage ramp-up scan from the initial operating voltage towards the open-circuit voltage. The dynamic voltage control command for the transient wide-scan is calculated using the following formula. In the formula, This indicates the DC scan voltage command value for the current control cycle. This represents the DC input voltage value of the photovoltaic string that the target inverter latches at the instant the maximum power point tracking algorithm process is suspended. This represents the theoretical upper limit of the open-circuit voltage of the photovoltaic string preset by the system. This indicates the total duration of the entire transient scan time window. This represents the cumulative time variable calculated from the start of the transient wide-scan. The theoretical upper limit of the open-circuit voltage of the photovoltaic string is calibrated according to the factory parameters on the nameplate of the photovoltaic modules connected to the grid in the photovoltaic power station and is stored in the internal memory. The system limits the wide-scan action to be completed within the range of 20 to 50 milliseconds according to the set transient scan time window to prevent drastic fluctuations in grid-connected power. The total duration of the entire transient scan time window is set according to the requirements of the grid dispatching regulations for the inverter's low-voltage ride-through timing. At the same time, the group control gateway 230 uses the frozen timestamp information carried in the underlying trigger frame as the starting point of the time zero point and strictly monitors the start and end boundaries of the entire transient scan time window to achieve synchronization between the ground and space. The specific control logic for adjusting the pulse width modulation signal of the boost chopper circuit inside the target inverter can be based on existing power electronic switching converter control technology.

[0041] Within the synchronous time window of the target inverter performing linear ramp-up scanning of the DC bus voltage, the sampling circuit 220 performs the high-frequency ripple parallel extraction sub-step S133. Since the insulated-gate bipolar transistor (IGBT) inside the target inverter naturally generates a high-frequency ripple current response when operating at a specific switching carrier frequency, using this high-frequency ripple current response as a passive AC excitation signal avoids additional hardware injection to demodulate and obtain the equivalent AC impedance characteristic parameters and impedance phase angle characteristics of the photovoltaic string. The sampling circuit 220 acquires the transient DC voltage sequence and transient DC current sequence of the photovoltaic string in real time at a sampling frequency in the kilohertz range. The sampling circuit 220 internally starts a parallel analog high-pass filter bypass circuit. The hardware cutoff frequency of the analog high-pass filter bypass circuit is preset according to the impedance matching of the specific switching carrier frequency of the IGBT. The analog high-pass filter bypass circuit filters out the DC bias component in the sampled signal, extracting the target inverter... The high-frequency switching action of the insulated gate bipolar transistor inside the inverter induces high-frequency AC voltage ripple data and high-frequency AC current ripple data on the DC side of the photovoltaic string. The group control gateway 230 aligns and packages the transient DC voltage sequence, transient DC current sequence, and high-frequency AC voltage ripple data and high-frequency AC current ripple data according to the time control cycle. If the DC bus voltage exceeds the grid safety threshold or the packaging waiting time exceeds the total duration of the set transient scanning time window during the transient wide-scan process, the group control gateway 230 actively triggers the abnormal protection interrupt mechanism. After completing the packaging action of the transient DC voltage sequence, transient DC current sequence, and high-frequency AC voltage ripple data and high-frequency AC current ripple data or triggering the abnormal protection interrupt mechanism, the group control gateway 230 cancels the high-priority hardware interrupt command and controls the target inverter to reactivate the maximum power point tracking algorithm process to restore the normal grid-connected operation state.

[0042] After receiving the data uploaded by the edge computing unit 130, the cloud decoupling module 300 executes the spatial reference anchoring sub-step S211. To eliminate the interference of uneven local wind cooling caused by micro-topographical undulations in the photovoltaic power station on the thermal imaging gradient, elevation undulations in complex physical terrain conditions such as mountains directly interfere with the near-surface wind speed gradient and cause significant local differences in thermal convection efficiency. The cloud decoupling module 300 needs to use the electrical output of normal strings that have not triggered abnormal scanning to characterize the reference heat dissipation environment of the current location in order to construct a reverse closed-loop calibration model. The cloud decoupling module 300 extracts the touch... The cloud decoupling module 300 establishes a local spatial search domain centered on the three-dimensional coordinates of the target array that caused the anomaly. The horizontal two-dimensional boundary of the local spatial search domain is set as a preset search radius based on the physical spacing between adjacent photovoltaic arrays of the photovoltaic power station and the cable routing topology. The cloud decoupling module 300 calibrates and sets the elevation difference tolerance threshold based on the contour map mapping data of the physical area where the photovoltaic power station is located and historical air convection wind speed test data. Elevation difference tolerance threshold Set as a meter-level physical height span value that will not cause significant abrupt changes in the wind field.

[0043] After establishing the local spatial search domain, the cloud decoupling module 300 executes the adjacency tolerance screening sub-step S212. The cloud decoupling module 300 traverses the photovoltaic power station equipment topology database to find neighboring candidate strings within the local spatial search domain. In order to effectively increase the probability that the selected neighboring candidate strings and the target string are in an equivalent micro-meteorological convection heat dissipation environment, the cloud decoupling module 300 calculates the absolute elevation difference between the elevation data of the neighboring candidate strings and the elevation data of the target string. The calculation of the absolute elevation difference is combined with the following formula. In the formula, This represents the calculated absolute elevation difference. This represents the elevation data of neighboring candidate strings. The cloud-based decoupling module 300 represents the elevation data of the target string, and combines the absolute elevation difference with the elevation difference tolerance threshold. Numerical comparison was performed, and the cloud-based decoupling module filtered out values ​​with absolute elevation differences exceeding the elevation difference tolerance threshold. Neighboring candidate strings, retaining those with absolute elevation differences less than or equal to the elevation difference tolerance threshold. The nearest candidate strings are used as the initial screening benchmark objects that meet the spatial constraints. If there is no absolute elevation difference less than or equal to the elevation difference tolerance threshold in the local spatial search domain... If the neighboring candidate strings are obtained, the cloud decoupling module 300 will gradually expand the search radius of the local spatial search domain according to the preset step size parameter until at least one preliminary screening benchmark object is obtained.

[0044] After obtaining the initial screening benchmark objects, the cloud decoupling module 300 executes the steady-state operation verification sub-step S213. Since subsequent steps require calling standard electrical characteristics under normal grid-connected operating conditions for residual comparison calculations, the cloud decoupling module 300 retrieves the current underlying control status flag of the initial screening benchmark objects from the electrical control module 200. The cloud decoupling module 300 checks the underlying control status flag and filters out initial screening benchmark objects in a hardware interrupt state or a transient wide-scan state. The cloud decoupling module 300 determines the initial screening benchmark objects whose underlying control status flag shows a steady-state operation state of maintaining maximum power point tracking as the final benchmark string. If all initial screening benchmarks are checked after checking the underlying control status flag... If all objects are eliminated, the cloud decoupling module 300 triggers the search compensation mechanism to re-execute the adjacency tolerance screening sub-step S212 and retrieves the more peripheral neighboring candidate strings to prevent deadlock in the addressing algorithm. If there are multiple final reference strings that meet the steady-state operating conditions, in order to avoid dimensional conflicts in subsequent formula calculations caused by multiple physical entities, the cloud decoupling module 300 performs arithmetic mean dimensionality reduction and fusion processing on the electrical parameters of multiple final reference strings to construct a unique equivalent reference string. By introducing multiple constraint mechanisms of spatial physical tolerance and electrical steady state, it is ensured that the subsequently extracted reference electrical reference data can reflect the reference operating parameters of the micro-topography environment where the target string is located under fault-free conditions with high confidence.

[0045] When maintaining daily communication tasks, the group control gateway 230 executes the high-frequency rolling cache sub-step S221. Under the condition of limited hardware resources at the bottom layer of the photovoltaic power station, in order to achieve efficient tracing of high-frequency historical data without consuming a large amount of memory space and to overcome the asynchronous sampling time difference caused by the internal crystal oscillator frequency drift of multiple bottom inverter node microcontrollers, the group control gateway 230 pre-allocates a continuous physical address space using its internal dynamic random access memory to construct a high-frequency ring buffer. The group control gateway 230 continuously collects transient voltage and transient current data of all photovoltaic strings under its jurisdiction at a fixed millisecond-level polling cycle. The group control gateway 230 encapsulates the collected transient voltage and transient current data and the local sampling time stamp generated synchronously by the system clock into a cached data frame. The group control gateway 230 writes the cached data frame into the high-frequency ring buffer according to the first-in-first-out queue rule, and after the high-frequency ring buffer address is full, it overwrites the oldest cached data frame.

[0046] After receiving the underlying trigger frame sent by the edge computing unit 130, the group control gateway 230 executes the hardware-level synchronization locking sub-step S222. The group control gateway 230 parses the underlying trigger frame and extracts the frozen timestamp information carried by the underlying trigger frame. At the concurrent moment when the group control gateway 230 issues a hardware interrupt command to the target inverter, the group control gateway 230 uses the direct memory access controller to issue a write protection command to the high-frequency ring buffer. The group control gateway 230 uses the frozen timestamp information as the time reference anchor point to search for and lock the historical data slices of the equivalent reference string at each of the preset buffer depths before and after the frozen timestamp information in the high-frequency ring buffer. The number of cached data frames included corresponding to the preset buffer depth is calculated and set by adding the maximum polling communication delay and the maximum network jitter duration of the photovoltaic power station's internal controller local area network bus and dividing by a fixed millisecond-level polling cycle. This effectively reduces the risk that the historical data slices of the equivalent reference string will be refreshed and covered by the regular sampled data during the transient wide-scan cycle. After the group control gateway 230 completes the extraction of the historical data slices, it actively releases the write protection command to restore the regular rolling refresh state of the high-frequency ring buffer.

[0047] After locking the historical data slice, the group control gateway 230 executes the frequency difference alignment and interpolation solution step S223. Because there is a millisecond-level phase misalignment between the global synchronization marker time generated by the underlying trigger frame and the discrete polling sampling period of the equivalent reference string, directly calling adjacent discrete sampled data will cause a truncation error at the energy input end of the subsequent thermoelectric decoupling model. The group control gateway 230 determines whether there are two adjacent valid sampling nodes containing the frozen timestamp information within the historical data slice. If two adjacent valid sampling nodes are missing from the historical data slice, the group control gateway 230 triggers a communication packet loss error and expands the retrieval span in the high-frequency ring buffer for retrying. If the test still fails, the group control gateway 230 calls the steady-state power average of the equivalent reference string before packet loss to perform data downgrading and replacement to prevent output distortion. If there are two adjacent valid sampling nodes that contain the frozen timestamp information in the historical data slice, the group control gateway 230 calls the internal digital signal processor to perform time-domain approximation linear interpolation on the instantaneous power corresponding to the two adjacent valid sampling nodes. The linear interpolation operation eliminates the data deviation caused by the discrete sampling frequency difference to obtain the equivalent reference actual electrical output power at the moment of time synchronization and alignment of the frozen timestamp information. The time-domain approximation linear interpolation operation is performed in combination with the following formula. In the formula, This represents the equivalent reference actual electrical output power at the moment when the time synchronization and freezing timestamp information obtained from the solution is captured. This indicates the freeze timestamp information carried by the underlying trigger frame. This represents the first valid sampling time point in the historical data slice that is prior to the frozen timestamp and has the smallest time span from the frozen timestamp. This represents the second valid sampling time point in the historical data slice that is located after the frozen timestamp and has the smallest time span from the frozen timestamp. This represents the equivalent reference string instantaneous power calculated by multiplying the transient voltage data and transient current data recorded in the buffered data frame at the first valid sampling time point. This represents the equivalent reference string instantaneous power calculated by multiplying the transient voltage data and transient current data recorded by the cached data frame at the second valid sampling time point. The group control gateway 230 feeds back the calculated equivalent reference actual electrical output power to the cloud decoupling module 300 via the uplink wide area network communication link of the photovoltaic power station as the basic reference parameter for subsequent calculation of the thermoelectric coupling compensation factor of the photovoltaic module.

[0048] After obtaining the equivalent reference actual electrical output power, the cloud decoupling module 300 executes the theoretical power reference solution step S231. The cloud decoupling module 300 extracts transient environmental irradiance data and transient environmental temperature data from the local meteorological monitoring stations deployed at the photovoltaic power station, which are time-aligned with the frozen timestamp information carried by the underlying trigger frame. Since the sampling frequency of local meteorological monitoring stations in the physical environment is usually much lower than the millisecond-level sampling rate of the electrical microcontroller, to avoid model input deviations caused by low-frequency meteorological jumps, the cloud decoupling module 300 calls the discrete meteorological sampling node data before and after the frozen timestamp information for time-domain linear interpolation reconstruction to fit and approximate a strictly time-aligned result. The transient environmental irradiance data and transient environmental temperature data are used. If a local meteorological monitoring station experiences a communication failure or data loss, the cloud decoupling module 300 retrieves the corresponding geographical area's macro-meteorological forecast data from the cloud server for downgraded replacement to ensure the continuity of the calculation process. Based on the transient environmental irradiance data and transient environmental temperature data, the cloud decoupling module 300 calls the photovoltaic module physical equivalent circuit model to calculate the theoretical output power of the equivalent reference string under the current environmental parameters. The specific numerical calculation of the theoretical output power derived from the photoelectric effect basic equation within the photovoltaic module physical equivalent circuit model is implemented based on existing photovoltaic theoretical power generation modeling technology.

[0049] After the cloud decoupling module 300 completes the theoretical output power calculation, it executes the electrical power residual evaluation sub-step S232. In actual mountainous micro-topography scenarios, spatial fluctuations in local wind direction and speed will change the thermal convection efficiency of the photovoltaic module backsheet surface and cause changes in the cell junction temperature inside the photovoltaic module. The change in cell junction temperature causes the actual electrical output power of the equivalent reference to deviate from the reference theoretical output power. When the actual local heat dissipation conditions of the photovoltaic module are worse than the macroscopic expectations, the actual operating junction temperature of the equivalent reference string is higher, thus causing a negative loss in output power. Since the equivalent reference string selected and constructed in the spatial reference anchoring sub-step S211 is physically... The system is highly adjacent to the abnormal target point in terms of spatial distance. The system is based on the premise that the cloud cover thickness and micro-meteorological irradiance distribution in the local area have macro-common mode characteristics. Therefore, the irradiance common mode interference can be effectively canceled and purified by the difference operation in mathematical space. This ensures that the polarity of the calculated residual is accurate and uniquely points to the heat accumulation fluctuation caused by the difference in micro-topography wind cooling. The cloud decoupling module 300 uses the theoretical output power and the equivalent reference actual electrical output power to perform direct difference operation to generate electrical power residual that characterizes the degree of deviation of actual operating temperature caused by micro-topography wind cooling conditions. The calculation of electrical power residual is combined with the following formula. In the formula, This represents the calculated electrical power residual. This represents the theoretical output power of the equivalent reference string obtained from the solution. The equivalent reference actual electrical output power at the time of the time synchronization alignment and interpolation solution sub-step S223 is obtained by solving the frequency difference alignment and interpolation solution sub-step. The cloud decoupling module 300 extracts the electrical power residual for subsequent use as a closed-loop negative feedback regulation and control parameter to reflect the differences in local meteorological convection environment.

[0050] After obtaining the electrical power residual, the cloud decoupling module 300 executes the compensation factor reverse derivation sub-step S233. The cloud decoupling module 300 retrieves the macroscopic initial convective heat dissipation compensation factor, which includes the entire photovoltaic power station area, from an external meteorological service server. The macroscopic initial convective heat dissipation compensation factor characterizes the average air convection heat exchange coefficient in a large-scale physical space. Since the macroscopic initial convective heat dissipation compensation factor cannot effectively characterize the spatial local heat exchange differences caused by micro-topographic undulations, to prevent division-by-zero errors in the reverse calculation logic, the cloud decoupling module 300 pre-determines whether the theoretical output power of the equivalent reference string obtained from the calculation is... If the theoretical output power is greater than the grid-connected power generation lower limit dead zone threshold set by the system, and the theoretical output power is not greater than the grid-connected power generation lower limit dead zone threshold, the cloud decoupling module 300 will stop the current calibration process and directly use the macroscopic initial convection heat dissipation compensation factor. If the theoretical output power is greater than the grid-connected power generation lower limit dead zone threshold, the cloud decoupling module 300 will use the electrical power residual as a negative feedback adjustment control parameter and substitute it into the reverse calculation logic to correct the macroscopic initial convection heat dissipation compensation factor, thereby calculating and obtaining the local real convection heat dissipation compensation factor after adaptive closed-loop calibration. The reverse closed-loop calculation of the local real convection heat dissipation compensation factor is combined with the following formula. In the formula, This represents the local real convection heat dissipation compensation factor after adaptive closed-loop calibration. This represents the initial macroscopic convective heat dissipation compensation factor introduced by the system. This represents the pre-calibrated empirical correlation coefficient of the photovoltaic module's thermoelectric response. The empirical correlation coefficient of the photovoltaic module's thermoelectric response is calculated and set based on the power temperature derating coefficient calibrated in the photovoltaic module's nameplate, and its physical value is limited to a positive constant. This ensures that when the electrical power residual is positive, the local real convection heat dissipation compensation factor can be proportionally amplified to compensate for the heat accumulation effect. The cloud decoupling module 300 outputs the local real convection heat dissipation compensation factor and sends it down to the subsequent data processing nodes for adaptive dynamic adjustment and compensation of the shallow thermal gradient edge detection activation threshold in the target string's location area.

[0051] After receiving the raw infrared thermal imaging data captured at the physical time corresponding to the frozen timestamp information uploaded by the edge computing unit 130, the cloud decoupling module 300 executes the temperature matrix mapping sub-step S311. In order to convert the visual features of the image pixel domain in which the raw infrared thermal imaging data is located into quantifiable physical thermodynamic parameters to support subsequent closed-loop calibration calculations, the cloud decoupling module 300 converts the digital radiation intensity value carried by each pixel coordinate point in the raw infrared thermal imaging data into the actual physical temperature value to generate an initial two-dimensional temperature distribution matrix. The specific numerical calculation process for converting the digital radiation intensity value in the raw infrared thermal imaging data into the actual physical temperature value can be based on the existing infrared temperature measurement calibration technology based on Stefan Boltzmann's law.

[0052] After obtaining the initial two-dimensional temperature distribution matrix, the cloud decoupling module 300 executes the spatial noise reduction filtering sub-step S312. In the actual operating environment of photovoltaic power plants, when micro-cracks occur in the internal cells of photovoltaic modules or when they are blocked by foreign objects, abnormal heat accumulation often occurs locally. This abnormal heat distribution forms a significant temperature difference with the surrounding normal heat dissipation area. By extracting the temperature change rate of the spatial distribution, the physical boundary of potential thermal anomalies can be made explicit. Due to the specular reflection characteristics of the glass panel on the surface of the photovoltaic module and the thermal noise of the infrared focal plane array sensor itself, high-frequency isolated noise points may be introduced into the initial two-dimensional temperature distribution matrix. If the initial two-dimensional temperature distribution matrix is ​​not filtered, Noise reduction processing may cause false edge activation in subsequent gradient calculation algorithms. The cloud decoupling module 300 calls a Gaussian smoothing filter to perform a two-dimensional spatial convolution operation on the initial two-dimensional temperature distribution matrix to filter out high-frequency isolated noise points. In a specific embodiment of the present invention, the convolution kernel size of the Gaussian smoothing filter is set according to the spatial physical resolution of the original infrared thermal imaging data in a proportional mapping. The standard deviation parameter of the Gaussian smoothing filter is adaptively calculated and matched according to the variance of the background thermal noise under the steady-state operation of the photovoltaic power station. The influence of abnormal spatial frequency domain components on the local thermodynamic characteristics of the initial two-dimensional temperature distribution matrix is ​​eliminated by the two-dimensional spatial convolution operation, and a smooth temperature distribution matrix is ​​generated.

[0053] After completing the spatial noise reduction filtering sub-step S312, the cloud decoupling module 300 executes the two-dimensional gradient field operator extraction sub-step S313. To prevent the matrix dimension from collapsing or even the algorithm from crashing due to missing neighboring pixels when calculating the local temperature partial derivatives of pixels in the edge region, the cloud decoupling module 300 performs edge mirroring filling operation on the four boundaries of the smooth temperature distribution matrix in advance to complete the spatial pixel dimension required for the difference calculation. The cloud decoupling module 300 uses the discrete difference operator to calculate the local temperature partial derivatives of the completed smooth temperature distribution matrix in the horizontal spatial pixel direction and the vertical spatial pixel direction respectively. The cloud decoupling module 300 performs vector magnitude synthesis calculation on the local temperature partial derivatives in the horizontal spatial pixel direction and the local temperature partial derivatives in the vertical spatial pixel direction to construct the two-dimensional initial temperature gradient field. The solution of the two-dimensional initial temperature gradient field is combined with the following formula. In the formula, This represents the magnitude of the temperature gradient at a specific pixel coordinate in the calculated initial temperature gradient field in two-dimensional space. This represents the actual physical temperature value at a specific pixel coordinate in the smoothed temperature distribution matrix. Represents the horizontal spatial pixel coordinate variable. Representing the vertical spatial pixel coordinate variable, the cloud decoupling module 300 stores the calculated two-dimensional initial temperature gradient field in the background database matrix as the reference feature input surface for subsequent introduction of micro-topography convection heat dissipation compensation factor for dynamic threshold adjustment.

[0054] After acquiring the initial temperature gradient field in two-dimensional space, the cloud decoupling module 300 executes the basic edge activation calibration sub-step S321. The cloud decoupling module 300 calls the built-in standard morphological edge detection algorithm. The specific underlying edge recognition logic of the standard morphological edge detection algorithm is implemented based on existing edge operator image processing technology. The cloud decoupling module 300 sets a standard activation threshold for the standard morphological edge detection algorithm. The standard activation threshold represents the lower limit of the baseline gradient response required to identify the hot spot edge induced by substantial electrical faults inside the photovoltaic module under ideal flat terrain and no meteorological wind speed disturbance conditions. The specific calibration value of the standard activation threshold is based on the statistical average of the baseline temperature gradient test data collected by the photovoltaic station during the historical steady-state operation period, which is stored as an a priori empirical constant in the background configuration database of the cloud decoupling module 300. The numerical dimension of the standard activation threshold is defined as the physical temperature pixel gradient dimension that is strictly consistent with the initial temperature gradient field in two-dimensional space to ensure the alignment of the comparison benchmark.

[0055] After completing the basic edge activation calibration sub-step S321, the cloud decoupling module 300 executes the dynamic division mapping compensation sub-step S322. During the actual operation of a micro-topographical mountain photovoltaic power station, local wind field convergence generates uneven convective heat dissipation physical space regions. These convective heat dissipation physical space regions cause shallow thermal gradient artifacts on the surface of the photovoltaic modules, similar to actual internal electrical faults. The cloud decoupling module 300 extracts the compensation factor and reversely derives the obtained local real convective heat dissipation compensation factor in the sub-step S233. To prevent the microprocessor's underlying floating-point division unit from experiencing a zero-overflow abnormal crash and to prevent the quotient from approaching zero due to an excessively large divisor, thus inducing background noise false activation, the cloud decoupling module 300 uses the system's built-in bidirectional limiter to constrain the range of the local real convective heat dissipation compensation factor. The cloud decoupling module 300 pre-determines whether the local real convective heat dissipation compensation factor is less than the system's underlying preset minimum protection limit or greater than the system's underlying preset maximum protection limit. (This is in the specific implementation scheme of the invention.) In this context, the lower limit of the mathematical calculation minimum protection is calibrated as a positive constant slightly greater than zero based on the processor's floating-point precision and the basic air-cooling attenuation coefficient. The upper limit of the mathematical calculation maximum protection is statistically calibrated based on the maximum heat accumulation coefficient under extreme high-temperature static windless conditions of the photovoltaic power station. If the local real convection heat dissipation compensation factor is less than the lower limit of the mathematical calculation minimum protection, the cloud decoupling module 300 will forcibly clamp the value of the local real convection heat dissipation compensation factor to the lower limit of the mathematical calculation minimum protection. If the local real convection heat dissipation compensation factor is greater than the upper limit of the mathematical calculation maximum protection, the cloud decoupling module 300 will forcibly clamp the value of the local real convection heat dissipation compensation factor to the upper limit of the mathematical calculation maximum protection. If the local real convection heat dissipation compensation factor is within the normal range between the lower limit of the mathematical calculation minimum protection and the upper limit of the mathematical calculation maximum protection, the cloud decoupling module 300 will apply the local real convection heat dissipation compensation factor to the standard activation threshold in a division mapping manner to perform dynamic adaptive correction calculation. The division mapping correction solution of the adaptive activation threshold is combined with the following formula. In the formula, This represents the adaptive activation threshold generated by the solution. This represents the standard activation threshold of the standard morphological edge detection algorithm. This indicates the local real convection heat dissipation compensation factor obtained by the reverse derivation sub-step S233 and processed by bidirectional amplitude limiting judgment. The cloud decoupling module 300 will adaptively activate the threshold cache for subsequent edge feature extraction processes.

[0056] After completing the dynamic division mapping compensation sub-step S322, the cloud decoupling module 300 executes the shallow artifact anti-interference assessment sub-step S323. From the perspective of actual thermodynamic physical mechanism, when the micro-topography wind cooling effect in the target area where the target string is located is enhanced, the local spatial convection heat dissipation rate is accelerated. The surface heat loss of the photovoltaic module leads to a low numerical response of the local real convection heat dissipation compensation factor. At this time, the division mapping mathematical logic drives the adaptive activation threshold to dynamically drift and increase to a higher value. The increase of the adaptive activation threshold is equivalent to directly raising the system's tolerance threshold for surface temperature gradient anomaly triggering alarm. Thus, the cloud decoupling module 300 effectively suppresses and filters out shallow thermal gradient artifacts caused by uneven external wind cooling by actively raising the system's tolerance threshold for surface temperature gradient anomaly triggering alarm. This reduces the probability that shallow thermal gradient artifacts are incorrectly judged by the upper-level logic of the system as deep thermal anomaly faults caused by substantial hardware damage or material aging degradation inside the photovoltaic module. This ensures the anti-interference robustness of subsequent micromorphological feature extraction operations to obtain micromorphological feature parameters.

[0057] After obtaining the adaptive activation threshold and the initial temperature gradient field in two-dimensional space, the cloud decoupling module 300 performs the anti-disturbance edge feature extraction sub-step S331. The cloud decoupling module 300 uses the adaptive activation threshold as a criterion and substitutes it into the standard morphological edge detection algorithm to perform binarized traversal of the pixels in the initial temperature gradient field in two-dimensional space. If the temperature gradient magnitude at the target pixel coordinates in the initial temperature gradient field in two-dimensional space is greater than or equal to the adaptive activation threshold, the cloud decoupling module 300 marks the target pixel coordinates as valid edge response points. If the temperature gradient magnitude at the target pixel coordinates in the initial temperature gradient field in two-dimensional space is less than the adaptive activation threshold, the cloud decoupling module 300 marks the target pixel coordinates as background noise. In the preferred embodiment of the present invention, after generating the binary thermal gradient edge image, the cloud decoupling module 300 traverses all pixels in the initial temperature gradient field in the two-dimensional space to generate a binary thermal gradient edge image. The cloud decoupling module 300 pre-counts the total number of valid edge response points after generating the binary thermal gradient edge image. If the total number of valid edge response points is zero, the cloud decoupling module 300 determines that the currently processed string has no significant abnormal thermal gradient and directly outputs the basic feature label representing the normal operating state and actively stops the subsequent morphological calculation steps. If the total number of valid edge response points is greater than zero, the cloud decoupling module 300 continues to execute the subsequent morphological contour parameter calculation process to prevent the underlying algorithm from crashing abnormally due to the empty pixel matrix.

[0058] After completing the anti-interference edge feature extraction sub-step S331, the cloud decoupling module 300 executes the morphological contour parameter calculation sub-step S332. Due to different external environmental interferences or internal physical defects of photovoltaic modules, heat distribution characteristics with specific geometric patterns will evolve on the surface of photovoltaic modules. For example, local point shading usually forms regular concentrated heat patches, while microcracks inside photovoltaic cells will form narrow heat accumulation bands along the internal metal grid lines. By extracting specific thermodynamic distribution features and transforming them into abstract geometric morphological parameters, a mapping relationship between physical faults and image visual features can be established. The cloud decoupling module 300 performs connected component analysis on the binarized thermal gradient edge image to locate discrete thermal anomaly connected regions. The specific algorithm operation for connected component analysis of the binarized thermal gradient edge image can be based on existing image connected component labeling technology. For each independent thermal anomaly connected region extracted from the binarized thermal gradient edge image... In the connected region, the cloud decoupling module 300 calculates the pixel area, the boundary perimeter, and the aspect ratio of the minimum bounding rectangle of each independent thermal anomaly connected region. The cloud decoupling module 300 retrieves the optical distance parameters and the focal length parameters of the infrared lens at the time of the original infrared thermal imaging data capture to calculate the equivalent real physical area corresponding to a single pixel. The cloud decoupling module 300 multiplies the equivalent real physical area corresponding to a single pixel by the pixel area of ​​each independent thermal anomaly connected region to calculate the actual equivalent physical heating area of ​​each independent thermal anomaly connected region. The cloud decoupling module 300 uses the pixel area and the boundary perimeter of each independent thermal anomaly connected region to calculate the morphological compactness parameter that characterizes the compactness of the hot spot shape. The calculation of the morphological compactness parameter is performed using the following formula. In the formula, This represents the morphological compactness parameter of the thermally anomalous connected regions obtained from the solution. The pixel area representing the thermally connected region. This represents the perimeter of the boundary of the thermal anomaly connected region. Representing the constant of pi, the cloud decoupling module 300 combines the morphological compactness parameter, the actual equivalent physical heating area of ​​each independent thermal anomaly connected region, and the aspect ratio of the minimum bounding rectangle of each independent thermal anomaly connected region into a topological feature dataset to support subsequent discretization classification and judgment.

[0059] After acquiring the topological feature dataset, the cloud decoupling module 300 executes the topological attribute discrete classification sub-step S333. The cloud decoupling module 300 pre-defines a first area constraint threshold, a second area constraint threshold, a compactness limit threshold, and an aspect ratio limit threshold within the system. The first and second area constraint thresholds are calculated and calibrated based on the standard physical area of ​​a single photovoltaic module cell, with the second area constraint threshold being greater than the first. The compactness limit threshold is set based on the theoretical compactness of an ideal circular patch with an adjusted error margin. The aspect ratio limit threshold is set based on the standard geometric proportion of the photovoltaic cell's main grid lines. The cloud decoupling module 300 performs specific classification of the topological features within the topological feature dataset corresponding to each independent thermal anomaly connected region. To avoid system output logic deadlock caused by overlapping values ​​of different morphological classification conditions, the cloud decoupling module 300 performs serial exclusive judgment according to the order of aspect ratio condition, small area compact condition, and large area diffuse condition. When the aspect ratio physical quantity in the topological feature dataset is greater than the aspect ratio threshold, the cloud decoupling module 300 determines that the physical fault morphology represented by the binarized thermal gradient edge image is a heat dissipation channel induced by the extension of microcracks along the lattice inside the battery cell. The cloud decoupling module 300 discretizes and classifies the physical thermal topological attributes of the thermal anomaly connected region into linear radial boundary labels. If the aspect ratio physical quantity in the topological feature dataset is not greater than the aspect ratio threshold and when the aspect ratio physical quantity in the topological feature dataset is greater than the aspect ratio threshold, the system output logic deadlock is determined. When the actual equivalent physical heating area is less than the first area constraint threshold and the morphological compactness parameter is greater than the compactness limit threshold, the cloud decoupling module 300 determines that the physical fault morphology represented by the binarized thermal gradient edge image is localized heat accumulation induced by external point occlusion. The cloud decoupling module 300 discretizes and classifies the physical thermal topology attributes of the thermally abnormal connected region into closed sharp boundary labels. If none of the executed conditional branch decisions are satisfied and the actual equivalent physical heating area in the topological feature dataset is greater than the second area constraint threshold and the morphological compactness parameter is less than the compactness limit threshold, the cloud decoupling module 300 determines that the physical fault morphology represented by the binarized thermal gradient edge image is a global thermal effect caused by large-scale material degradation. The cloud decoupling module 300 discretizes and classifies the physical thermal topological attributes of the thermally anomalous connected regions into large-area diffuse boundary labels. When the data in the topological feature dataset does not meet the criteria for closed sharp boundary labels, linear radial boundary labels, and large-area diffuse boundary labels, the cloud decoupling module 300 discretizes and classifies the physical thermal topological attributes of the thermally anomalous connected regions into undefined shape labels and triggers a manual review instruction to avoid dead zones in the classification algorithm. After completing the discretization and classification of the physical thermal topological attributes of all independent thermally anomalous connected regions, the cloud decoupling module 300 uses the digital unique thermal coding rule to encapsulate the multiple sets of physical thermal topological attribute classification labels corresponding to all independent thermally anomalous connected regions generated by the cloud decoupling module 300 into a multi-dimensional physical thermal feature vector matrix.The cloud-based decoupling module 300 outputs a multi-dimensional physical and thermal feature vector matrix, which serves as the direct visual input for subsequent multi-dimensional electrical feature vector joint analysis and cross-domain fault root cause diagnosis performed by the cloud-based decoupling module 300.

[0060] After receiving the transient voltage and current operating curve data of the photovoltaic module uploaded by the edge computing unit 130, the cloud decoupling module 300 executes the basic physical model construction sub-step S411. In order to transform the macroscopic external electrical output characteristics of the photovoltaic module into microscopic characterization parameters reflecting the internal physical aging and damage state of the photovoltaic module, the cloud decoupling module 300 loads the standard photovoltaic single diode physical equivalent circuit model into the system. The standard photovoltaic single diode physical equivalent circuit model uses nonlinear transcendental equations to describe the implicit mathematical constraint relationship between the output current and output voltage of the photovoltaic module. The cloud decoupling module 300 substitutes the discrete voltage sampling points and discrete current sampling points in the transient voltage and current operating curve data into the main equation of the standard photovoltaic single diode physical equivalent circuit model to construct the physical response function to be solved. The cloud decoupling module 300 retrieves the standard test condition parameters set on the photovoltaic module's nameplate and the absolute physical temperature value in the initial two-dimensional temperature distribution matrix to determine the non-unsolved environmental variables in the main equation of the standard photovoltaic single diode physical equivalent circuit model. The main equation of the standard photovoltaic single diode physical equivalent circuit model is calculated in combination with the following formula. In the formula, Indicates the output current of the photovoltaic module. Indicates the output voltage of the photovoltaic module. Represents the photocurrent generated by a photovoltaic module. This represents the reverse saturation current of the diodes inside the photovoltaic module. Represents the electron charge constant. The ideal factor represents the diodes inside a photovoltaic module. Represents the Boltzmann constant. Indicates the absolute operating temperature of the photovoltaic module. This represents the equivalent series resistance of the photovoltaic module to be solved. This represents the equivalent parallel resistance of the photovoltaic module to be solved.

[0061] After completing the basic physical model construction sub-step S411, the cloud decoupling module 300 executes the approximation solution initial value setting sub-step S412. Analyzing the underlying semiconductor physical degradation mechanism of photovoltaic modules, poor soldering of internal cell interconnects or detachment of metal grid lines directly leads to a significant increase in the equivalent series resistance. Conversely, microcracks or potential-induced degradation within the photovoltaic module exacerbate internal leakage current, resulting in a sharp decrease in the equivalent parallel resistance. Accurately solving for the equivalent series and equivalent parallel resistances allows for a quantitative assessment of the microscopic health of the photovoltaic module. Since the main equations of the standard photovoltaic single-diode physical equivalent circuit model are highly nonlinear transcendental equations and cannot be directly solved analytically, directly assigning random initial values ​​to the equivalent series and equivalent parallel resistances may cause subsequent numerical approximation algorithms to get stuck in local optima or experience non-convergence divergence errors. The cloud decoupling module 300 extracts short-circuit current characteristic points, open-circuit voltage characteristic points, and maximum power characteristic points from the transient voltage and current operating curve data. The cloud decoupling module 300 utilizes the transient voltage and current operating curve data... The initial iterative estimates of the equivalent series resistance and the equivalent parallel resistance are obtained by combining the short-circuit current characteristic points, the open-circuit voltage characteristic points in the transient voltage and current operating curve data, and the maximum power characteristic points in the transient voltage and current operating curve data with idealized boundary conditions. The specific analytical calculation operations for the initial iterative estimates of the equivalent series resistance and the equivalent parallel resistance calculated using boundary electrical characteristic points can be based on existing photovoltaic parameter analytical evaluation technology. In order to ensure that the subsequent nonlinear numerical iterative optimization algorithm has sufficient degrees of freedom to solve and does not fall into the unsolvable state of the underdetermined equation system, the cloud decoupling module 300 verifies whether the total number of discrete data points contained in the transient voltage and current operating curve data is greater than the total number of dimensions of the parameters to be solved. After the data point sufficiency verification is passed, the cloud decoupling module 300 constructs a nonlinear least squares objective function with the goal of minimizing the mean square error between the measured data points of the output current in the transient voltage and current operating curve data and the theoretical current data points calculated by the main equation of the standard photovoltaic single diode physical equivalent circuit model.

[0062] After completing the initial value setting sub-step S412 for approximation solution, the cloud-based decoupling module 300 executes the nonlinear numerical iterative optimization sub-step S413. The cloud-based decoupling module 300 calls the built-in Levenberg-Marquardt numerical optimization algorithm to iteratively reduce the dimensionality of the nonlinear least squares objective function. The cloud-based decoupling module 300 inputs the initial iterative estimates of the equivalent series resistance and the equivalent parallel resistance into the Levenberg-Marquardt numerical optimization algorithm as the starting point for the optimization iteration. In each iterative calculation loop, the cloud-based decoupling module 300 dynamically calculates the Jacobian moment. The system updates the current step values ​​of the equivalent series resistance and the equivalent parallel resistance. The cloud decoupling module 300 monitors the current output error value of the nonlinear least squares objective function in real time. When the current output error value of the nonlinear least squares objective function is less than the system's preset lower limit of convergence tolerance or the number of iterations of the Levenburg-Marquardt numerical optimization algorithm reaches the preset maximum cycle limit, in a specific embodiment of the present invention, the system's preset lower limit of convergence tolerance is set to an empirical constant between 10 to the power of -4 and 10 to the power of -6, and the preset maximum cycle limit is set to 100. For integer iterations between one and one thousand, the cloud-based decoupling module 300 determines that the numerical approximation solution process has converged and exits the iterative calculation loop. The cloud-based decoupling module 300 extracts the approximate solutions for the equivalent series resistance and equivalent parallel resistance from the final output of the Levenberg-Marquardt numerical optimization algorithm. To prevent the numerical solution algorithm from outputting negative resistance values ​​that violate objective physical laws due to abnormal fluctuations, the cloud-based decoupling module 300 performs positive value constraint verification on the approximate solutions for the equivalent series resistance and equivalent parallel resistance. If the approximate solution for the equivalent series resistance or the approximate solution for the equivalent parallel resistance... If the approximate solution is less than or equal to zero, the cloud decoupling module 300 forcibly rolls back and resets the approximate solutions of the equivalent series resistance and the equivalent parallel resistance to the initial iterative estimate of the equivalent series resistance and the initial iterative estimate of the equivalent parallel resistance to eliminate the algorithm logic dead zone. After the positive value constraint verification is passed, the cloud decoupling module 300 encapsulates the approximate solutions of the equivalent series resistance and the equivalent parallel resistance into a one-dimensional electrical parameter feature vector as the direct circuit dimension input basis for the subsequent multi-dimensional electrical feature vector joint analysis and cross-domain fault root cause diagnosis executed by the cloud decoupling module 300.

[0063] After obtaining the approximate solutions for the equivalent series resistance and the equivalent parallel resistance, the cloud decoupling module 300 executes the environmental parameter matching and benchmark retrieval sub-step S421. Since the equivalent series resistance and equivalent parallel resistance of the photovoltaic module are affected not only by the internal physical aging of the photovoltaic module but also by nonlinear shifts with fluctuations in the external meteorological environment, directly using the absolute resistance values ​​of the approximate solutions for the equivalent series resistance and equivalent parallel resistance for fault determination can easily lead to the early, minute degradation characteristics of the photovoltaic module being submerged in the system's background meteorological noise. From the perspective of the aging physical mechanism of photovoltaic module materials, the contact resistance of the photovoltaic module... Increased mechanical fatigue of the internal metal solder strips leads to a monotonically increasing trend in the true series impedance of the photovoltaic module, while potential-induced degradation and damage to the insulation layer structure of the encapsulation material cause a monotonically decreasing trend in the true parallel impedance. By differentially comparing the real-time resistance operation state of the photovoltaic module with its healthy initial state under the same meteorological conditions, meteorological variables can be decoupled and the pure material degradation increment can be extracted. The cloud decoupling module 300 obtains the real-time environmental irradiance parameter and the real-time backsheet temperature parameter corresponding to the transient voltage and current operation curve data acquisition time of the photovoltaic module. The cloud decoupling module 300 will then... The ambient irradiance parameter and the real-time backplane temperature parameter are used as the primary key inputs to the pre-built historical fault-free operation status feature library of the retrieval system. This library records the baseline equivalent circuit parameters mapped to different meteorological conditions under the initial commissioning state of the photovoltaic power station without any physical defects. If the real-time ambient irradiance parameter and the real-time backplane temperature parameter exceed the physical record boundary of the historical fault-free operation status feature library, the cloud decoupling module 300 forcibly clamps the real-time ambient irradiance parameter and the real-time backplane temperature parameter to the outermost effective boundary extreme point of the historical fault-free operation status feature library to prevent matching addressing out-of-bounds collapse. This out-of-bounds prevention and limiting mechanism is then implemented. Subsequently, to avoid the real-time ambient irradiance parameter and the real-time backplane temperature parameter falling into the sampling blind zone between discrete record points in the historical fault-free operation state feature library, which would cause a benchmark matching step error, the cloud decoupling module 300 searches the historical fault-free operation state feature library for four steady-state data entries that are spatially adjacent to the real-time ambient irradiance parameter and the real-time backplane temperature parameter. The cloud decoupling module 300 uses a bilinear interpolation algorithm based on the real-time ambient irradiance parameter and the real-time backplane temperature parameter to perform a two-dimensional spatial numerical weighted average calculation on the four steady-state data entries to extract the corresponding historical benchmark equivalent series resistance and historical benchmark equivalent parallel resistance.

[0064] After completing the environmental parameter matching and benchmark retrieval sub-step S421, the cloud decoupling module 300 executes the small distortion difference decomposition operator step S422. The cloud decoupling module 300 calculates the relative degradation drift rate of the series impedance using the approximate solution of the equivalent series resistance and the historical benchmark equivalent series resistance. The cloud decoupling module 300 also calculates the relative degradation drift rate of the parallel impedance using the approximate solution of the equivalent parallel resistance and the historical benchmark equivalent parallel resistance. This considers that the physical degradation of the photovoltaic module's parallel impedance manifests as a continuous decrease in absolute value. To reduce distortion, the cloud decoupling module 300 employs a reverse subtraction logic—subtracting the approximate solution of the equivalent parallel resistance from the historical reference—when calculating the relative deterioration drift rate of the parallel impedance. This ensures that the final output of the algorithm shows a positive distortion index for the relative deterioration drift rate of the parallel impedance. Furthermore, to prevent the underlying floating-point division unit from approaching zero and overflowing due to extremely small values ​​of the historical reference equivalent series resistance or equivalent parallel resistance, the cloud decoupling module 300 forcibly compares the historical reference equivalent series resistance with the historical reference equivalent parallel resistance before performing the division operation. The reference equivalent parallel resistance is compared and clamped with the minimum positive real protection constant set at the system's underlying layer. To significantly extract the faint early aging signs at the string level where the photovoltaic modules are located from the background meteorological noise, the cloud decoupling module 300 introduces a pre-determined sensitivity amplification factor into the system to perform linear stretching transformation on the relative degradation drift rate of the series impedance and the relative degradation drift rate of the parallel impedance, respectively, to generate differential amplified impedance distortion characteristic parameters. In a specific embodiment of the present invention, the pre-determined series sensitive amplification gain coefficient and the pre-determined parallel sensitive amplification gain coefficient are derived and set based on the digital resolution constant of the analog-to-digital converter of the photovoltaic power station data acquisition terminal and the variance constant of the background thermal noise during historical steady-state operation. The value range of the pre-determined series sensitive amplification gain coefficient and the pre-determined parallel sensitive amplification gain coefficient is set to an empirical proportionality constant between ten and one hundred. The numerical calculation of the differential amplified impedance distortion characteristic parameters is performed using the following formula. In the formula, This represents the impedance distortion characteristic parameter of the series differential amplifier generated by the solution. This represents the impedance distortion characteristic parameter of the parallel differential amplifier generated by the solution. This represents the pre-determined series sensitive amplifier gain coefficient within the system. This represents the pre-determined gain coefficient of the parallel sensitive amplifier within the system. This represents the approximate solution for the equivalent series resistance. This represents the historical baseline equivalent series resistance extracted from the historical fault-free operating state feature library. This represents the approximate solution for the equivalent parallel resistance. This represents the historical baseline equivalent parallel resistance extracted from the historical fault-free operation state feature library. The cloud decoupling module 300 merges the series differential amplifier impedance distortion feature parameter and the parallel differential amplifier impedance distortion feature parameter into the one-dimensional electrical parameter feature vector output by the cloud decoupling module 300 in order to complete the joint construction of multi-dimensional electrical feature vectors. The cloud decoupling module 300 outputs multi-dimensional electrical feature parameters to support the multi-modal data joint analysis and cross-domain fault root cause diagnosis subsequently executed by the cloud decoupling module 300.

[0065] After completing the micro-distortion difference decomposition operator step S422, the cloud decoupling module 300 executes the high-frequency AC phase angle extraction sub-step S431. In the actual physical degradation process of photovoltaic modules, besides the change in the DC equivalent resistance, changes in the dielectric constant of the internal encapsulation materials and the parasitic capacitance introduced by physical cracks in the photovoltaic cells cause significant impedance phase shifts in the AC frequency domain. The cloud decoupling module 300 receives the high-frequency AC impedance phase angle data of the photovoltaic module uploaded by the edge computing unit 130. This high-frequency AC impedance phase angle data is obtained by the edge computing unit 130 controlling the underlying hardware to inject a continuous sweep frequency high-frequency AC disturbance signal covering the parasitic parameter response range of the photovoltaic module into the photovoltaic module and measuring the voltage and current phase difference fed back by the photovoltaic module. This data is then used for control... The specific hardware-driven operation of injecting continuous frequency-sweeping high-frequency AC disturbance signals into the photovoltaic module and measuring the voltage and current phase difference fed back by the photovoltaic module can be based on existing electrochemical impedance spectroscopy measurement technology. The cloud decoupling module 300 extracts the AC impedance phase angle characteristic value of the photovoltaic module at the characteristic resonant frequency point from the high-frequency AC impedance phase angle data. The specific selection logic of the characteristic resonant frequency point is that the cloud decoupling module 300 traverses the entire test frequency band corresponding to the high-frequency AC impedance phase angle data to find the frequency physical coordinates where the imaginary reactance zero-crossing point or the extreme point of the first derivative of the phase angle change occurs. The cloud decoupling module 300 uses the frequency physical coordinates where the imaginary reactance zero-crossing point or the extreme point of the first derivative of the phase angle change occurs as the reference alignment frequency point for the essential change of parasitic parameters inside the photovoltaic module to perform feature extraction.

[0066] After extracting the AC impedance phase angle characteristic value, the cloud decoupling module 300 executes the heterogeneous dimension unification normalization sub-step S432. Since the series differential amplifier impedance distortion characteristic parameters and parallel differential amplifier impedance distortion characteristic parameters output by the cloud decoupling module 300 in the preceding steps are relative ratio dimensionless values ​​in the DC time domain, while the AC impedance phase angle characteristic value is an absolute angle physical dimension in the AC frequency domain, directly mathematically concatenating the relative ratio dimensionless values ​​in the DC time domain with the absolute angle physical dimensions in the AC frequency domain will cause the multi-modal data joint analysis algorithm subsequently executed by the cloud decoupling module 300 to fall into a local optimization gradient vanishing dead zone due to excessive differences in input feature scales. Block 300 retrieves the system's internally preset historical benchmark mean and standard deviation sets. These sets are obtained offline based on the statistical distribution of large data samples accumulated during the photovoltaic power station's historical fault-free operation cycles. The cloud-based decoupling module 300 performs standard normal distribution drift conversion calculations on the series differential amplifier impedance distortion characteristic parameters, parallel differential amplifier impedance distortion characteristic parameters, and AC impedance phase angle characteristic values ​​to eliminate the dimensional differences between these parameters. The standard normal distribution drift conversion calculation is performed using the following formula. In the formula, This represents the dimensionlessly normalized characteristic value obtained from the solution. This represents the original characteristic value input to the standard normal distribution drift conversion calculation logic, i.e., the impedance distortion characteristic parameter of the series differential amplifier, the impedance distortion characteristic parameter of the parallel differential amplifier, or the phase angle characteristic value of the AC impedance. This represents the baseline statistical mean retrieved from the historical baseline mean set, which matches the physical attributes of the original feature values. This represents the baseline statistical standard deviation retrieved from the historical baseline standard deviation set to match the physical properties of the original feature values. This represents a minimal positive real number compensation constant preset at the system's underlying level to prevent abnormal crashes due to division by zero overflow. The specific value of this minimal positive real number compensation constant preset at the system's underlying level to prevent abnormal crashes due to division by zero overflow is an empirical constant between 10 to the power of -8 and 10 to the power of -6, based on the lower limit of the processor's floating-point calculation precision. The cloud decoupling module 300 obtains the normalized series impedance distortion characteristics, normalized parallel impedance distortion characteristics, and normalized AC phase angle characteristics respectively through standard normal distribution drift conversion calculation.

[0067] After acquiring the normalized series impedance distortion characteristics, normalized parallel impedance distortion characteristics, and normalized AC phase angle characteristics, the cloud decoupling module 300 executes the orthogonal feature matrix encapsulation sub-step S433. From the perspective of physical information science analysis of the multi-dimensional state characterization of photovoltaic modules, the normalized series impedance distortion characteristics reflect the ohmic loss state of the conductive path of the photovoltaic module, the normalized parallel impedance distortion characteristics reflect the interlayer insulation leakage current state of the photovoltaic module, and the normalized AC phase angle characteristics reflect the parasitic capacitive reactance distribution state of the space charge depletion layer inside the photovoltaic module. The normalized series impedance distortion characteristics, normalized parallel impedance distortion characteristics, and normalized AC phase angle characteristics theoretically exhibit mutually independent physical evolution laws in the physical characterization dimension, thereby minimizing linear collinear data redundancy. The cloud decoupling module 300 aligns the memory data according to the system's preset settings. The rules map and splice the normalized series impedance distortion features, normalized parallel impedance distortion features, and normalized AC phase angle features into the same continuous memory address space in a fixed channel order, such as the cloud decoupling module 300 sequentially arranging the series feature channels, parallel feature channels, and AC phase angle feature channels in a fixed order. The cloud decoupling module 300 assembles and encapsulates all feature data in the memory address space into a high-dimensional structured multidimensional electrical orthogonal feature vector. The multidimensional electrical orthogonal feature vector is specifically represented as a 3x1 single-column feature matrix in the underlying data structure. The cloud decoupling module 300 pushes the multidimensional electrical orthogonal feature vector into the system message queue as the core electrical modal input tensor that drives the subsequent execution of the fault root cause diagnosis model based on the cross-domain attention mechanism by the cloud decoupling module 300.

[0068] The cloud-based decoupling module 300 executes historical benchmark statistical distribution solution step S511. The cloud-based decoupling module 300 retrieves high-frequency AC impedance phase angle data from the photovoltaic power station during a preset time period in the initial stage of commissioning, along with approximate solutions for the equivalent parallel resistance and equivalent series resistance corresponding to the high-frequency AC impedance phase angle data, as a health benchmark sample set. Considering that the underlying impedance operation characteristics of the photovoltaic power station under complex weather alternation environments exhibit typical multi-peak probability clustering, the cloud-based decoupling module 300 uses a Gaussian mixture distribution model containing several independent Gaussian distribution kernel functions to analyze the health benchmark sample set. Probability density fitting is performed to calculate the mean series impedance drift, standard deviation of series impedance drift, mean parallel impedance drift, standard deviation of parallel impedance drift, mean AC phase angle deflection, and standard deviation of AC phase angle deflection of photovoltaic modules under healthy operating conditions. The specific mathematical and statistical operations for probability density fitting of the healthy benchmark sample set using the Gaussian mixture distribution model can be based on existing data probability distribution parameter estimation techniques.

[0069] After completing the historical baseline statistical distribution calculation sub-step S511, the cloud decoupling module 300 executes the drift deflection threshold adaptive setting sub-step S512. Analyzing the physical evolution of the photovoltaic module throughout its entire lifecycle, the reference impedance of the photovoltaic module will slowly drift due to the accumulation of ultraviolet radiation and alternating thermal stress. If a fixed static threshold is used, it will inevitably cause large-scale false alarm dead zones in the later stages of the photovoltaic module's service life. Therefore, a time-dimensional attenuation compensation mechanism must be introduced. Based on the Raida criterion and the annualized attenuation rate constant specified at the photovoltaic module's factory, the cloud decoupling module 300 calculates and establishes the series resistance positive drift threshold, the parallel resistance negative drift threshold, and the AC impedance phase angle deflection threshold. The setting of the series resistance positive drift threshold, the parallel resistance negative drift threshold, and the AC impedance phase angle deflection threshold is calculated using the following formula. In the formula, This represents the positive drift threshold of the series resistance calculated and output by the cloud-based decoupling module 300. This represents the negative drift threshold of the parallel resistance calculated and output by the cloud-based decoupling module 300. This represents the AC impedance phase angle deflection threshold calculated and output by the cloud-based decoupling module 300. This represents the average series impedance drift of a photovoltaic module under healthy operating conditions. This represents the standard deviation of the series impedance drift of a photovoltaic module under healthy operating conditions. This represents the average parallel impedance drift of the photovoltaic module under healthy operating conditions. This represents the standard deviation of the parallel impedance drift of photovoltaic modules under healthy operating conditions. This represents the average AC phase angle deflection of the photovoltaic module under healthy operating conditions. represents the standard deviation of the AC phase angle deflection of the photovoltaic module under healthy operating conditions, and λ represents the system's preset confidence expansion coefficient constant for the Laida criterion. The specific value of the Laida criterion confidence expansion coefficient constant is fixed at the system's underlying level as a dimensionless constant between two and three to cover the confidence boundary of the system's operating error. This represents the system's preset annualized degradation sensitivity coefficient for photovoltaic modules connected in series. This represents the system's preset annualized degradation sensitivity coefficient for parallel photovoltaic modules. The system's preset annualized degradation sensitivity coefficients for series and parallel photovoltaic modules are calibrated based on the factory parameters provided on the nameplate by the photovoltaic module manufacturer. The specific value range for both the system's preset annualized degradation sensitivity coefficients for series and parallel photovoltaic modules is set to a real number between 0.2% and 0.8% per year. This represents the historical total operating time constant of the photovoltaic module from the time it is put into operation and connected to the grid until the time of data acquisition of the transient voltage and current operating curves.

[0070] After completing the adaptive setting sub-step S512 for the drift deflection threshold, the cloud decoupling module 300 executes the state discretization judgment matrix construction sub-step S513. Since the multi-dimensional electrical feature vector output by the cloud decoupling module 300 has already undergone dimensional normalization, in order to ensure the dimensional consistency of the multi-dimensional features in the numerical comparison logic and eliminate the dead zone of dimensional misalignment comparison, the cloud decoupling module 300 substitutes the series resistance positive drift threshold, the parallel resistance negative drift threshold, and the AC impedance phase angle deflection threshold into the standard normal distribution drift transformation equation to calculate and obtain the multi-dimensional electrical feature vector. After dimensional alignment of the comparison benchmarks, the cloud decoupling module 300 extracts the normalized series impedance distortion features, normalized parallel impedance distortion features, and normalized AC phase angle features from the multi-dimensional electrical feature vector output by the cloud decoupling module 300. The cloud decoupling module 300 then performs a data amplitude comparison operation between the normalized series impedance distortion features and the normalized series resistance positive drift judgment threshold. The 300-level cloud decoupling module performs a data amplitude comparison calculation between the normalized parallel impedance distortion characteristics and the normalized parallel resistance negative drift judgment threshold. The 300-level cloud decoupling module performs a data amplitude comparison calculation between the normalized AC phase angle characteristics and the normalized AC impedance phase angle deflection judgment threshold. Based on the data amplitude comparison calculation results of the normalized series impedance distortion characteristics and the normalized series resistance positive drift judgment threshold, the normalized parallel impedance distortion characteristics and the normalized parallel resistance negative drift judgment threshold, and the normalized AC phase angle characteristics and the normalized AC resistance... The data amplitude comparison operation result of the anti-phase angle deflection judgment threshold generates an electrical state judgment vector composed of Boolean discrete variables. The cloud decoupling module 300 maps and concatenates the electrical state judgment vector as a column vector into the pre-allocated global operating state memory of the system bottom layer to construct and continuously update the feature discrete judgment matrix. The feature discrete judgment matrix is ​​used to convert the continuous multi-dimensional electrical feature vector into a deterministic physical failure mode classification index of photovoltaic modules that supports discrete logic gate judgment, thereby driving the data interaction and inference of the cross-domain diagnostic level executed by the cloud decoupling module 300.

[0071] After extracting the feature discrete judgment matrix, the cloud decoupling module 300 performs the cross-domain modal feature alignment step S521. Since photovoltaic modules often face external recoverable interferences such as bird droppings or localized tree shadows in complex outdoor environments, relying solely on the electrical impedance change or infrared thermal imaging temperature rise characteristics of a single photovoltaic module for fault diagnosis can easily lead to misjudgments caused by weather fluctuations. Analyzing the electrothermal multi-physics coupling mechanism of photovoltaic modules, external physical shading causes clearly defined localized hot spots on the photovoltaic module surface. Furthermore, external physical shading is equivalent to a reverse dissipation diode connected in parallel inside the photovoltaic module, thus altering the insulation leakage characteristics of the photovoltaic module. Meanwhile, microcracks in the cells or poor solder joints inside the photovoltaic module generate large-area diffuse Joule heating along the internal metal grid lines, and physical fractures directly... By cutting off the current transmission path and introducing additional space charge layer parasitic capacitance, the cloud decoupling module 300 retrieves the morphological labels of the physical thermal feature vectors of the infrared thermal imaging spectrum of the photovoltaic module stored at the bottom layer of the system. The morphological labels of the physical thermal feature vectors of the infrared thermal imaging spectrum are obtained by the image processing module at the front end of the system through morphological edge extraction and region discretization calibration of the infrared thermal imaging image of the photovoltaic module. The specific image processing operations for morphological edge extraction and region discretization calibration of the infrared thermal imaging image of the photovoltaic module can be carried out using existing computer vision contour extraction technology. The cloud decoupling module 300 performs absolute timestamp matching and data frame alignment of the morphological labels of the physical thermal feature vectors of the infrared thermal imaging spectrum with the electrical state judgment vectors contained in the feature discretization judgment matrix in the time dimension to construct a thermoelectric coupling cross-domain judgment data structure.

[0072] After completing the cross-domain modal feature alignment sub-step S521, the cloud decoupling module 300 executes the external occlusion physical mapping determination sub-step S522. The cloud decoupling module 300 reads the morphological labels of the physical thermal feature vectors of the infrared thermal imaging spectrum and the electrical state determination vector from the thermoelectric coupling cross-domain determination data structure. The cloud decoupling module 300 determines whether the morphological labels of the physical thermal feature vectors of the infrared thermal imaging spectrum exhibit closed sharp boundaries. Simultaneously, the cloud decoupling module 300 checks whether the electrical state determination vector contains a Boolean logic truth value indicating that the normalized parallel impedance distortion feature exceeds the normalized parallel resistance negative drift judgment threshold, and whether the electrical state determination vector contains an indication that the normalized series impedance distortion feature does not exceed the normalized series resistance positive drift judgment threshold and indicates that the normalized series impedance distortion feature does not exceed the normalized series resistance positive drift judgment threshold. If the phase angle characteristic does not exceed the normalized AC impedance phase angle deflection judgment threshold (Boolean logic false value), and the morphological label of the physical thermal feature vector of the infrared thermal imaging spectrum shows a closed sharp boundary, and the electrical state judgment vector indicates a significant negative drift in the parallel impedance of the photovoltaic module, while the electrical state judgment vector indicates that the series impedance of the photovoltaic module and the AC phase angle remain in a steady state without significant distortion, the cloud decoupling module 300 triggers the execution of the judgment instruction that the external shading condition is met in the system's underlying logic controller. The cloud decoupling module 300 confirms the current physical state of the photovoltaic module as a non-permanent, recoverable fault state with the accumulation of external opaque shading objects. To quantify the severity of external shading on the surface of the photovoltaic module, the cloud decoupling module 300 calculates the external shading damage probability index using the following formula. In the formula, This represents the external occlusion damage probability index calculated and output by the cloud decoupling module 300. This represents the system's preset image edge gradient weight coefficients. This represents the system's preset parallel impedance drift weighting factor. This represents the actual pixel gradient value of the infrared thermal image. This represents the system-preset reference pixel gradient constant used for calibrating sharp boundaries. This represents the impedance distortion characteristic parameter of the parallel differential amplifier extracted from the multidimensional electrical eigenvector. The cloud decoupling module 300 calculates the negative drift threshold of the parallel resistance. To ensure the completeness of the calculation of the external shading damage probability index, the sum of the system-preset image edge gradient weight coefficient and the system-preset parallel impedance drift weight coefficient is always equal to the system-preset image edge gradient weight coefficient. The specific value of the system-preset image edge gradient weight coefficient is set to a real number between 0.4 and 0.6. The system-preset reference pixel gradient constant used to calibrate sharp boundaries is calibrated to an empirical constant of 30 to 50 gray levels per pixel distance based on the spatial resolution of the infrared thermal imaging sensor at the front end of the system. The cloud decoupling module 300 writes the external shading damage probability index into the system's operation and maintenance scheduling log to generate a cleaning and maintenance work order for the photovoltaic module surface.

[0073] After executing the external shading physical mapping judgment sub-step S522, the cloud decoupling module 300 executes the internal irreversible damage prevention and avoidance mapping judgment sub-step S523. In order to accurately identify substantial physical damage caused by mechanical stress or material aging inside the photovoltaic module, the cloud decoupling module 300 judges whether the morphological label of the physical thermal feature vector of the infrared thermal imaging spectrum presents a linear radial boundary. At the same time, the cloud decoupling module 300 checks whether the electrical state judgment vector contains a Boolean logic truth value indicating that the normalized series impedance distortion feature exceeds the normalized series resistance positive drift judgment threshold or an indicator of the normalized AC phase angle. The Boolean logic truth value of the feature exceeding the normalized AC impedance phase angle deflection judgment threshold is crucial. Since microcracks inside photovoltaic modules are often accompanied by significant phase angle deflection caused by extremely weak early parasitic capacitance, relying solely on a single DC series resistance feature for fault diagnosis easily leads to missed detection dead zones. Performing a logical OR operation between the DC ohmic loss index and the AC parasitic capacitive reactance deflection index can capture the microscopic physical fracture features in the high-frequency domain to the greatest extent. When the morphological label of the physical thermal feature vector of the infrared thermal imaging spectrum presents a linear radial boundary and the electrical state judgment vector indicates a significant positive drift in the series impedance of the photovoltaic module or a significant positive drift in the AC impedance of the photovoltaic module... When a significant deflection occurs in the impedance phase angle, the cloud decoupling module 300 triggers a judgment instruction in the system's underlying logic controller to determine if the internal irreversible damage condition has been met. The cloud decoupling module 300 confirms that the current physical state of the photovoltaic module is a permanent damage fault state caused by internal irreversible material breakage, such as microcracks in the solar cells or poor soldering of the busbars. Based on the permanent damage fault state, the cloud decoupling module 300 activates the bypass protection diode conduction control program connected to the photovoltaic module at the system's underlying level to electrically isolate the photovoltaic module with microcracks from the global power generation array, thereby effectively suppressing the photovoltaic module-level damage caused by internal microcracks. To mitigate the risk of thermal runaway and fire, and to prevent algorithmic dead zone crashes caused by the physical thermal feature vector and multidimensional electrical feature vector of the infrared thermal imaging spectrum in the thermoelectric coupling cross-domain determination data structure not satisfying either the external shading condition or the internal irreversible damage condition, those skilled in the art can understand that the cloud decoupling module 300 will default to downgrading the current physical state of the photovoltaic module to a pending observation period when neither the external shading condition nor the internal irreversible damage condition is met, and will trigger an adaptive monitoring instruction to increase the frequency of transient voltage and current operating curve data acquisition to ensure the closed-loop completeness of the system's cross-domain determination logic.

[0074] After executing the internal irreversible damage prevention and avoidance mapping judgment sub-step S523, the cloud decoupling module 300 executes the multi-source global degradation feature alignment sub-step S531. From the perspective of the physical mechanism of potential-induced degradation of photovoltaic modules, under the long-term high-voltage bias of the system, sodium ions inside the glass panel will migrate and accumulate to the surface of the photovoltaic cell, causing a surface polarization effect. Ion migration and surface polarization effects will not only cause the equivalent series resistance at the global level inside the photovoltaic module to gradually increase, but also directly weaken the built-in electric field of the photovoltaic module, thus manifesting as a significant decrease in macroscopic open-circuit voltage. Since potential-induced degradation is a global material-level degradation phenomenon spanning the entire area of ​​the photovoltaic module, In the infrared thermal imaging spectrum, the photovoltaic module no longer exhibits localized high-temperature abrupt spots, but rather a diffuse temperature rise area with no obvious edge pixel gradient and a large area coverage. Those skilled in the art can understand that the cloud decoupling module 300 retrieves the morphological label of the physical thermal feature vector of the infrared thermal imaging spectrum in the thermoelectric coupling cross-domain determination data structure. The cloud decoupling module 300 evaluates whether the morphological label of the physical thermal feature vector of the infrared thermal imaging spectrum presents a large-area diffuse boundary. The cloud decoupling module 300 simultaneously extracts the component open-circuit voltage feature parameter from the multi-dimensional electrical feature vector to combine with the electrical state determination vector contained in the feature discrete determination matrix in preparation for entering the multi-modal matrix mapping deduction process.

[0075] After completing the multi-source global degradation feature alignment sub-step S531, the cloud decoupling module 300 executes the potential-induced attenuation matrix mapping judgment sub-step S532. After determining that the morphological label of the physical thermal feature vector of the infrared thermal imaging spectrum presents as a large-area diffuse boundary, the cloud decoupling module 300 checks whether the electrical state judgment vector contains a Boolean logic truth value indicating that the normalized series impedance distortion feature exceeds the normalized series resistance positive drift judgment threshold. Simultaneously, the cloud decoupling module 300 performs differential calculations between the module open-circuit voltage feature parameter and the standard open-circuit voltage reference value specified on the photovoltaic module's nameplate to evaluate the open-circuit voltage attenuation amplitude of the photovoltaic module. When the morphological label of the physical thermal feature vector of the infrared thermal imaging spectrum presents... When a large-area diffuse boundary is formed and the electrical state determination vector indicates a significant positive drift in the series impedance of the photovoltaic module, and the open-circuit voltage attenuation exceeds the system's preset voltage tolerance threshold, the specific value of the system's preset voltage tolerance threshold is set to a real number between 5% and 10% of the standard open-circuit voltage reference value, based on the voltage fluctuation tolerance under standard test conditions for the photovoltaic module. This is to eliminate voltage drop noise caused by normal temperature drift. The cloud decoupling module 300 assesses that the photovoltaic module conforms to the characteristics of potential-induced degradation effect caused by global material-level degradation mechanism. In order to quantitatively evaluate the evolution depth of potential-induced degradation of the photovoltaic module, the cloud decoupling module 300 calculates the material-level global degradation depth index using the following formula. In the formula, This represents the material-level global degradation depth index calculated and output by the cloud-based decoupling module 300. This represents the system's preset open-circuit voltage attenuation weighting coefficient. This represents the system's preset series impedance drift weighting factor. This indicates the standard open-circuit voltage reference value specified on the photovoltaic module's nameplate. This represents the component open-circuit voltage characteristic parameter in the multidimensional electrical feature vector. This represents the series differential amplifier impedance distortion characteristic parameter extracted by the cloud decoupling module 300 from the multi-dimensional electrical feature vector. This indicates the positive drift threshold of the series resistance calculated and output by the cloud decoupling module 300. The sum of the system-preset open-circuit voltage attenuation weight coefficient and the system-preset series impedance drift weight coefficient is always equal to one. The specific value of the system-preset open-circuit voltage attenuation weight coefficient is fixed at the system bottom layer as a real number between 0.5 and 0.8 to highlight the core dominant role of open-circuit voltage drop in the physical determination of potential-induced attenuation. If the morphological label of the physical thermal feature vector of the infrared thermal image spectrum shows a large-area diffuse boundary, but the electrical state determination vector does not indicate that the series impedance of the photovoltaic module has a significant positive drift or the open-circuit voltage attenuation amplitude does not exceed the system-preset voltage tolerance threshold, then the cloud decoupling module 300 determines that the photovoltaic module has not experienced a potential-induced attenuation effect and archives the current physical state of the photovoltaic module as a normal thermal fluctuation state caused by uniform environmental heating to eliminate the logic dead zone of the multi-mode determination algorithm.

[0076] After executing the potential-induced degradation matrix mapping judgment sub-step S532, the cloud decoupling module 300 executes the global maintenance instruction encapsulation and output sub-step S533. After confirming the existence of potential-induced degradation effect in the photovoltaic module and completing the material-level global degradation depth index calculation, the cloud decoupling module 300 needs to issue a targeted recovery intervention instruction to the operation and maintenance management platform of the photovoltaic power station. Since potential-induced degradation can achieve a certain degree of performance recovery in the early stage by applying a reverse bias voltage to the photovoltaic module, the cloud decoupling module 300 generates BeiDou spatial three-dimensional coordinate parameters corresponding to the photovoltaic module. Standardized maintenance instruction data packages including potential-induced decay fault type identification codes, material-level global degradation depth index, and reverse recovery voltage amplitude parameters, as well as BeiDou spatial three-dimensional coordinate parameters. The high-precision positioning module, carried by the drone at the system's front end, synchronously captures and transmits the data to the cloud decoupling module 300 via the underlying communication link at the moment of triggering transient acquisition. The specific value of the reverse recovery voltage amplitude parameter is dynamically and adaptively matched and set by the cloud decoupling module 300 based on 1.1 to 1.5 times the maximum operating voltage of the photovoltaic array system to ensure sufficient ion reverse dissipation electric field energy without damaging the insulation layer of the photovoltaic module. The cloud decoupling module 300 assembles and encapsulates the standardized maintenance instruction data packet by performing header and tail data frame check codes according to the system's preset Ethernet communication protocol rules. The cloud decoupling module 300 then pushes the assembled and encapsulated standardized maintenance instruction data packet into the work order scheduling queue of the operation and maintenance management platform via the wireless communication link to instruct the underlying physical operation and maintenance equipment to use the BeiDou spatial three-dimensional coordinate parameters during the nighttime period when there is no sunlight. The system precisely navigates to and locates the photovoltaic module that is experiencing potential-induced degradation and applies a high-voltage reverse DC bias to disperse sodium ions accumulated on the cell surface. The cloud decoupling module 300 establishes an automated fault management closed loop from multimodal electrical feature extraction to precise positioning and intervention of underlying physical equipment through the distribution of standardized maintenance instruction data packets.

[0077] The following is a specific application example of a photovoltaic power station in a mountainous area with complex micro-topography.

[0078] The application scenario is set for deployment in mountainous photovoltaic power generation bases with significant elevation variations and local canyon wind gap effects. The photovoltaic modules are distributed on slopes with different elevation gradients according to array topology.

[0079] The inspection module is installed on the inspection drone platform to perform daily patrols. An infrared thermal imager collects raw infrared thermal imaging data of the photovoltaic module surface. The edge computing unit receives the raw infrared thermal imaging data and converts it into an initial two-dimensional temperature distribution matrix, calculating the local temperature gradient amplitude. When the edge computing unit detects that the local temperature gradient amplitude in a certain area exceeds the steady-state judgment benchmark value, the edge computing unit triggers a time synchronization latch instruction, recording the current general coordination timestamp and the three-dimensional coordinate parameters output by the positioning unit.

[0080] The edge computing unit extracts latitude and longitude data from the three-dimensional coordinate parameters, substitutes it into the spatial hash mapping function to calculate the spatial hash key value, and matches it with the target inverter node identification code and the communication address of the group control gateway in the offline spatial electrical topology hash mapping table. The communication unit sends a low-level trigger frame carrying the target inverter node identification code and freeze timestamp information to the target group control gateway through the local directional ad hoc network link.

[0081] The group control gateway deployed at the photovoltaic power plant receives the underlying trigger frame and issues a hardware interrupt command to the target inverter. The target inverter suspends the maximum power point tracking process and performs a linear voltage ramp-up scan of the DC bus voltage from the initial operating voltage to the open circuit voltage. Within the synchronization time window, the group control gateway uses the direct memory access controller to lock the valid sampling node data recorded in the high-frequency ring buffer, and uses a simulated high-pass filter bypass circuit to intercept the high-frequency AC voltage ripple data and high-frequency AC current ripple data induced by the high-frequency switching action of the insulated gate bipolar transistor inside the target inverter.

[0082] The cloud-based decoupling module receives uploaded positioning data and electrical status parameters. It establishes a local spatial search domain centered on the three-dimensional coordinates of the target string. By comparing absolute elevation differences, it filters preliminary benchmark objects within an equivalent microclimate convection cooling environment and verifies the underlying control status flags to determine the benchmark string maintaining steady-state operation at maximum power point tracking. The cloud-based decoupling module then uses the transient voltage and current data of the benchmark string for interpolation calculations to obtain the benchmark actual electrical output power aligned with the frozen timestamp information. It then subtracts the benchmark actual electrical output power from the theoretical output power to generate an electrical power residual. Finally, the cloud-based decoupling module substitutes this electrical power residual into the inverse calculation logic to obtain the local true convection cooling compensation factor.

[0083] The cloud-based decoupling module corrects the standard activation threshold of the standard morphological edge detection algorithm by using a division mapping method to adjust the local real convection heat dissipation compensation factor, generating an adaptive activation threshold. The cloud-based decoupling module then uses this adaptive activation threshold to binarize the smoothed and denoised initial temperature gradient field in two-dimensional space, extracting effective edge response points and calculating the morphological compactness parameter, aspect ratio, and actual equivalent physical heat-generating area of ​​the thermally anomalous connected region. Based on the set area constraint threshold and compactness limit threshold, the cloud-based decoupling module performs conditional branching decisions and outputs a physical thermal topology attribute classification label, such as classifying it as a linear radial boundary or a closed sharp boundary.

[0084] The cloud-based decoupling module substitutes voltage and current scanning data into the standard photovoltaic single-diode physical equivalent circuit model and uses the Levenberg-Marquardt numerical optimization algorithm to approximate and solve for the equivalent series resistance and equivalent parallel resistance. Combining real-time environmental irradiance and backplane temperature parameters, the module extracts historical baseline equivalent series and parallel resistances from a historical fault-free operating state feature library, and calculates the impedance distortion characteristic parameters of the series differential amplifier and the parallel differential amplifier. Finally, the module performs a standard normal distribution drift transformation on the AC impedance phase angle characteristic values ​​and assembles them with the DC impedance characteristics into a multi-dimensional electrically orthogonal characteristic vector.

[0085] The cloud-based decoupling module aligns the physical thermal topology attribute classification labels with the multidimensional electrical orthogonal feature vectors in the data frames. When the morphological label shows a linear radial boundary and the multidimensional electrical orthogonal feature vectors indicate that the normalized series impedance distortion feature exceeds the judgment threshold or the normalized AC phase angle feature exceeds the judgment threshold, the cloud-based decoupling module determines that the photovoltaic module has suffered irreversible damage faults such as internal microcracks in the cells, and generates maintenance instructions containing BeiDou spatial three-dimensional coordinate parameters.

[0086] To verify the technical effectiveness of the photovoltaic fault precision decoupling diagnosis system in complex application scenarios, this embodiment constructs an experimental comparative test verification scheme. The test site is a mountainous photovoltaic test field with terrain undulations ranging from 20 meters to 150 meters. The test cycle is 120 operating days, covering cloudy conditions, local strong wind convergence, and high temperature and calm wind weather conditions. Deterministic defects are manually preset in the photovoltaic array, including microcracks in the cells, poor welding of metal busbars, shading by leaves and bird droppings, and potential-induced decay areas in the glass panel. In the data processing and visualization analysis stage, the data is fitted with DC voltage and current operating curves obtained from transient wide-span scanning, analyzed by fast Fourier transform of high-frequency AC impedance phase angle extraction, and generated in two-dimensional space initial temperature gradient field operator matrix. The experiment is set up with three parallel comparative test groups: the first group adopts a traditional single infrared thermal imaging inspection architecture, relying only on UAV hot spot edge detection. The second group adopts a traditional independent electrical microflow monitoring architecture, relying on the determination of steady-state extreme values ​​of inverter cascade voltage and current. The third group adopts the diagnostic system architecture based on air-to-ground coordinated triggering and thermoelectric closed-loop calibration provided by this invention.

[0087] Table 1. Comparison Test Data of Fault Identification Performance under Different Diagnostic Architectures for Photovoltaic Power Stations in Complex Micro-Topography According to the data in Table 1, the traditional single infrared thermal imaging inspection architecture exhibits a comprehensive fault detection rate of 83.74% due to the lag in surface heat conduction. Furthermore, the traditional single infrared thermal imaging inspection architecture generates 127 false alarms caused by micro-topographical wind-cooling differences. Because the accelerated local air convection caused by mountainous micro-topography carries away the accumulated heat on the photovoltaic module surface and forms pseudo-thermal gradient edges, the traditional single infrared thermal imaging inspection architecture lacks a thermodynamic background compensation mechanism, leading to the incorrect labeling of local wind-cooling abrupt change areas as heat-generating fault areas. Based on a space-ground collaborative and thermoelectric closed-loop diagnostic architecture, by extracting the electrical power residual from a benchmark string under the same micro-topographical environment, inversely calculating the local real convection heat dissipation compensation factor, and dynamically mapping and adjusting the edge detection activation threshold, the number of false alarms caused by micro-topographical wind-cooling differences is reduced to 8, confirming the technical effectiveness of the thermoelectric closed-loop calibration logic in filtering out spatial convection noise.

[0088] According to the data in Table 1, in terms of decoupling accuracy between shading and microcrack types, the traditional single infrared thermal imaging inspection architecture and the traditional independent electrical microfluidic monitoring architecture have decoupling accuracies of only 31.52% and 12.18%, respectively. Because leaf shading and microcracks in solar cells exhibit highly common-mode characteristics in terms of macroscopic DC power output reduction and local hot spot performance, traditional architectures cannot penetrate the thermodynamic appearance to distinguish internal physical failure modes. A new orthogonal evaluation dimension is proposed, using microsecond-level high-frequency AC impedance phase angle data extracted from a space-ground collaborative and thermoelectric closed-loop diagnostic architecture. Physical fractures in photovoltaic cells introduce abrupt changes in the parasitic capacitive reactance of the space charge depletion layer, leading to AC phase angle deflection, while surface shading only alters the DC parallel leakage path. Based on the space-ground collaborative and thermoelectric closed-loop diagnostic architecture, by performing joint matrix constraint judgment using AC impedance phase angle characteristics, normalized parallel impedance distortion characteristics, and infrared morphological tags, the decoupling accuracy between shading and microcrack types reaches 95.37%, achieving precise qualitative isolation between external recoverable shading interference and internal irreversible material damage.

[0089] According to the data in Table 1, for the potential-induced degradation effect caused by material-level global degradation, the traditional independent electrical microfluidic monitoring architecture lacks sensitivity to the slow degradation trend due to the lack of reference state alignment, resulting in a detection depth index error rate as high as 43.61%. Based on a space-ground collaborative and thermoelectric closed-loop diagnostic architecture, the large-area diffuse boundary features of the infrared thermal imaging spectrum are aligned with the component open-circuit voltage characteristic parameters and normalized series impedance distortion characteristics in the multi-dimensional electrical feature vector through cross-domain modal feature alignment. This eliminates the voltage drop interference caused by a single temperature increase, reducing the detection depth index error rate of potential-induced degradation to 3.24%. Combined with the spatial three-dimensional coordinate direct addressing mechanism established by the underlying trigger frame, the diagnostic system ensures closed-loop identification and location fixation of multi-source photovoltaic faults in complex physical environments.

[0090] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for accurate identification and root cause diagnosis of photovoltaic faults based on steady-state profile features, characterized in that, Includes the following steps: Step S100: Obtain the spatial initial thermal gradient field data and transient electrical characteristic data of the target photovoltaic string that are synchronized with time, and establish direct addressing from physical three-dimensional coordinates to the underlying network nodes through a pre-loaded offline spatial electrical topology hash table. Step S200: Extract the high-frequency electrical output data of the reference string to calculate the electrical power residual, and inversely calculate the adaptive closed-loop convection heat dissipation compensation factor under the current micro-topography environment based on the electrical power residual. Step S300: Apply the adaptive closed-loop convection heat dissipation compensation factor to the activation threshold of the edge detection algorithm, perform dynamic threshold anti-disturbance processing on the initial spatial thermal gradient field data, and extract standardized morphological topological attributes to construct a physical thermal feature vector. Step S400: Call the historical fault-free steady-state operation reference data to perform reference differential operation on the transient electrical characteristic data, calculate the impedance distortion characteristic parameters, and fuse the impedance distortion characteristic parameters with the AC impedance phase angle characteristics to construct a multi-dimensional electrical characteristic vector; Step S500: Input the physical thermal feature vector and the multidimensional electrical feature vector into a preset feature discrete judgment matrix to perform orthogonal logic joint judgment, and output the fault type label of the photovoltaic module and the corresponding spatial coordinate data.

2. The method for accurate identification and root cause diagnosis of photovoltaic faults based on steady-state profile features according to claim 1, characterized in that, The process of establishing direct addressing and acquiring data in step S100 includes: The local temperature gradient amplitude is calculated based on the pixel matrix of the image acquired by the infrared thermal imager. When the local temperature gradient amplitude is greater than the judgment benchmark value, the time capture channel is triggered to freeze the current timestamp and simultaneously latch the corresponding physical three-dimensional coordinates. Extract the latitude and longitude data from the physical three-dimensional coordinates, and use a spatial hashing algorithm to reduce its dimension and map it into discrete spatial hash key values; The offline spatial electrical topology hash table is retrieved using the spatial hash key as an index. The network protocol address of the group control gateway in the area where the abnormal target point is located and the identification code of the target inverter node are extracted. The underlying trigger frame carrying the freeze timestamp information is sent directly through the local area network link to control the underlying inverter to deviate from steady-state operation and perform transient DC voltage linear scanning and high-frequency AC ripple extraction.

3. The method for accurate identification and root cause diagnosis of photovoltaic faults based on steady-state profile features according to claim 1, characterized in that, The process of calculating the electrical power residual in step S200 includes: A local spatial search domain is established with the three-dimensional coordinates of the target photovoltaic string as the center. The absolute elevation difference between the neighboring candidate strings and the target photovoltaic string within the local spatial search domain is calculated. Strings with an absolute elevation difference less than the preset elevation difference tolerance threshold and in steady-state operation are selected as the benchmark strings. Using the frozen timestamp information as the time reference anchor point, the historical data slice of the reference group string is locked in the high-frequency ring buffer of the underlying group control gateway; Extract the instantaneous power corresponding to two adjacent valid sampling nodes of the frozen timestamp information, eliminate the discrete sampling frequency difference through linear interpolation operation of time domain approximation, and calculate the reference actual electrical output power aligned with the frozen timestamp information time. The transient environmental irradiance data and transient environmental temperature data, which are time-aligned with the frozen timestamp information, are substituted into the physical equivalent circuit model to calculate the theoretical output power of the reference string. The theoretical output power is then subtracted from the actual electrical output power of the reference string to obtain the electrical power residual characterizing the local wind cooling difference.

4. The method for accurate identification and root cause diagnosis of photovoltaic faults based on steady-state profile features according to claim 3, characterized in that, The process of reversely calculating the adaptive closed-loop convection heat dissipation compensation factor in step S200 is as follows: When the theoretical output power is greater than the set lower limit of grid-connected power generation, based on the relative ratio of the electrical power residual to the theoretical output power, and combined with the pre-calibrated empirical correlation coefficient of the photovoltaic module thermoelectric response, the macroscopic initial convection heat dissipation compensation factor introduced by the system is proportionally superimposed and corrected, and the local real convection heat dissipation compensation factor is calculated and obtained as the adaptive closed-loop convection heat dissipation compensation factor.

5. The method for accurate identification and root cause diagnosis of photovoltaic faults based on steady-state profile features according to claim 1, characterized in that, The dynamic threshold anti-interference processing in step S300 includes: Spatial Gaussian smoothing filtering is performed on the obtained initial two-dimensional temperature distribution matrix. The local temperature partial derivatives in the horizontal and vertical directions are calculated using discrete difference operators, and vector magnitude synthesis is performed to construct the initial spatial thermal gradient field data. A bidirectional limiter is used to clamp the boundary range of the maximum and minimum values ​​of the adaptive closed-loop convection heat dissipation compensation factor. The adaptive activation threshold is obtained by dividing the standard activation threshold of the standard morphological edge detection algorithm by the adaptive closed-loop convection heat dissipation compensation factor with amplitude limiting constraint. The initial spatial thermal gradient field data is binarized using the adaptive activation threshold, and pixel coordinates greater than or equal to the adaptive activation threshold are extracted as effective edge response points to generate a binarized thermal gradient edge image.

6. The method for accurate identification and root cause diagnosis of photovoltaic faults based on steady-state profile features according to claim 5, characterized in that, The process of extracting standardized morphological topological attributes and constructing physical thermal feature vectors in step S300 includes: Connectivity analysis is performed on the binarized thermal gradient edge image to calculate the actual equivalent physical heating area, boundary perimeter, and aspect ratio of the minimum bounding rectangle for each independent thermal anomaly connected region. The morphological compactness parameter is calculated using the ratio of pixel area to the square of boundary perimeter. When the aspect ratio of a physical quantity is greater than a set threshold, the physical thermal topology attribute is classified as a linear radial boundary label. When the aspect ratio physical quantity is not greater than the set threshold, the actual equivalent physical heating area is less than the set first area constraint threshold, and the morphological compactness parameter is greater than the compactness limit threshold, it is classified as a closed sharp boundary label. When the actual equivalent physical heating area is greater than the second area constraint threshold and the morphological compactness parameter is less than the compactness limit threshold, it is classified as a large area diffuse boundary label. Each category label is encapsulated into a multidimensional physical thermal feature vector using one-hot encoding rules.

7. The method for accurate identification and root cause diagnosis of photovoltaic faults based on steady-state profile features according to claim 1, characterized in that, The process of obtaining impedance distortion characteristic parameters and AC impedance phase angle characteristics in step S400 includes: The voltage and current sampling points in the transient electrical characteristic data are substituted into the main equation of the standard photovoltaic single diode physical equivalent circuit model. The initial iterative estimate is calculated using the extracted short-circuit current, open-circuit voltage and maximum power characteristic points. The equivalent series resistance and equivalent parallel resistance of the photovoltaic module are iteratively approximated by using the nonlinear least squares objective function and the Levenberg-Marquardt numerical optimization algorithm. Using real-time environmental irradiance parameters and real-time backplane temperature parameters as the primary keys for retrieval and matching, the bilinear interpolation algorithm is used to extract the historical benchmark equivalent series resistance and historical benchmark equivalent parallel resistance under the same meteorological conditions from the historical fault-free operation status feature library. The degradation drift rate of the equivalent series resistance and equivalent parallel resistance relative to their respective historical reference values ​​is calculated, and a sensitivity amplification factor is introduced for linear stretching to obtain the impedance distortion characteristic parameters of the series differential amplifier and the parallel differential amplifier. The high-frequency AC impedance phase angle data obtained by traversing the injected high-frequency AC disturbance signal is used to extract the AC impedance phase angle characteristic value corresponding to the characteristic resonant frequency point where the imaginary reactance zero-crossing point or the extreme point of the first derivative of the phase angle abrupt change is located.

8. The method for accurate identification and root cause diagnosis of photovoltaic faults based on steady-state profile features according to claim 7, characterized in that, The process of constructing the multidimensional electrical feature vector in step S400 includes: Retrieve the historical benchmark mean and standard deviation sets calculated statistically during historical fault-free operation cycles; The standard normal distribution drift transformation logic is adopted by extracting physical feature values, subtracting the baseline statistical mean, and dividing by the baseline statistical standard deviation to eliminate the heterogeneous dimensional differences between the series differential amplifier impedance distortion feature parameters, the parallel differential amplifier impedance distortion feature parameters, and the AC impedance phase angle feature values, thereby generating normalized series impedance distortion features, normalized parallel impedance distortion features, and normalized AC phase angle features, respectively. The three normalized features mentioned above are mapped and concatenated in a fixed channel arrangement order in a continuous memory address space to assemble a high-dimensional structured multidimensional electrical feature vector.

9. The method for accurate identification and root cause diagnosis of photovoltaic faults based on steady-state profile features according to claim 8, characterized in that, The method for constructing the preset feature discrete decision matrix in step S500 is as follows: The Gaussian mixture distribution model was used to fit the probability density of the healthy baseline sample set in the early stage of production, and the baseline mean and standard deviation of series impedance, parallel impedance and AC phase angle under healthy operating conditions were obtained. Based on the Laida criterion and standard deviation, the basic confidence boundary is calculated and superimposed with the product of the total historical operating time of the photovoltaic module and the annualized degradation sensitivity coefficient. The series resistance positive drift threshold, parallel resistance negative drift threshold and AC impedance phase angle deflection threshold with time dimension degradation compensation are established. Each normalized feature in the multidimensional electrical feature vector is compared with its corresponding threshold after dimensional alignment. Based on the exceedance of the comparison, an electrical state judgment vector composed of Boolean discrete variables is generated and constructed as a feature discrete judgment matrix.

10. The method for accurate identification and root cause diagnosis of photovoltaic faults based on steady-state profile features according to claim 9, characterized in that, The process of performing orthogonal logic joint determination in step S500 includes: When the morphological label of the physical thermal feature vector is a closed sharp boundary, and the feature discrete judgment matrix indicates that the parallel impedance has a significant negative drift and the series impedance and AC phase angle have not been distorted, it is judged as an external shielding recoverable fault. When the morphological label of the physical thermal feature vector is a linear radial boundary, and the feature discrete judgment matrix indicates that the series impedance has a significant positive drift or the AC phase angle has a significant deflection, it is judged as an internal irreversible damage fault caused by microcracks in the cell. When the morphological label of the physical thermal feature vector is a large-area diffuse boundary, and the feature discrete judgment matrix indicates that the series impedance has a significant positive drift and the open-circuit voltage attenuation in the transient electrical feature data exceeds the voltage tolerance threshold, it is determined to be a potential-induced attenuation fault caused by material-level global degradation, and an automated maintenance instruction containing the target physical coordinates and high-voltage reverse DC bias parameters is generated and sent to the operation and maintenance management platform.

Images