Reliability evaluation method of large photovoltaic medium-voltage direct-current power generation system

Abstract

The present application relates to photovoltaic power station reliability evaluation technical field, disclose a kind of reliability evaluation method of large photovoltaic medium voltage direct current power generation system, comprising: the equipment level reliability parameter of each core equipment is calculated;According to the preset topological structure, the reliability calculation formula of series-parallel system is combined, and the overall reliability of large photovoltaic medium voltage direct current power generation system is deduced in layers;And the availability and loss load probability are calculated;Determine whether photovoltaic module power degradation attenuation process conforms to brownian motion process;Brownian motion process is fused to build physical neural network model, and the output photovoltaic module power degradation amount prediction value;The reliability comprehensive evaluation report of large photovoltaic medium voltage direct current power generation system is generated.The present application effectively overcomes the defects that existing photovoltaic reliability evaluation mode is poor in adaptability, evaluation dimension is single, and component degradation prediction precision is low, to provide reliable basis for photovoltaic power station operation optimization and fault early warning.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic power plant reliability evaluation technology, and more specifically, to a reliability evaluation method for a large-scale photovoltaic medium-voltage DC power generation system. Background Technology

[0002] With the development of the photovoltaic industry, large-scale photovoltaic medium-voltage DC power generation systems have gradually become one of the mainstream architectures for photovoltaic power plants due to their advantages of low transmission loss and high power quality. This system adopts a series-parallel hybrid topology of "photovoltaic modules-combiner box-DC-DC converter-inverter". With multiple equipment layers and a complex topology, its operational reliability directly determines the power generation efficiency and power supply stability of the power plant.

[0003] Existing reliability evaluation methods for photovoltaic power plants are mostly designed for traditional AC photovoltaic power plants, and have the following technical shortcomings: First, they cannot adapt to the series-parallel hybrid topology of medium-voltage DC power generation systems, and only use a single fault statistics method to calculate system reliability, resulting in a large deviation between the evaluation results and the actual operating conditions; Second, the reliability evaluation of photovoltaic modules only focuses on post-fault statistics, without considering the power degradation characteristics during module operation, and cannot achieve early warning; Third, traditional data-driven degradation prediction models have a "black box problem," with weak generalization ability and low prediction accuracy under small sample data.

[0004] If a precise and comprehensive reliability assessment of a large-scale photovoltaic medium-voltage DC power generation system cannot be conducted, potential equipment failures and component degradation trends will not be detected in a timely manner, which may lead to unplanned power plant shutdowns, reduced power generation efficiency, or even power supply failures.

[0005] No effective solutions have yet been proposed to address the problems in the relevant technologies. Summary of the Invention

[0006] To address the problems in related technologies, this invention proposes a reliability evaluation method for large-scale photovoltaic medium-voltage DC power generation systems, in order to overcome the aforementioned technical problems existing in existing related technologies.

[0007] Therefore, the specific technical solution adopted by the present invention is as follows:

[0008] A reliability evaluation method for a large-scale photovoltaic medium-voltage DC power generation system includes:

[0009] Acquire operational data from multiple core devices in a large-scale photovoltaic medium-voltage DC power generation system; these core devices include photovoltaic modules, combiner boxes, DC-DC converters, and inverters.

[0010] Based on operational data, calculate the device-level reliability parameters for each core device; the device-level reliability parameters include failure rate, reliability, mean time between failures (MTBF), and mean time to repair (MTBT).

[0011] Based on the pre-defined topology of the large-scale photovoltaic medium-voltage DC power generation system, and combined with the reliability calculation formulas for series and parallel systems, the overall reliability of the large-scale photovoltaic medium-voltage DC power generation system is derived in layers; and the availability and failure probability of the large-scale photovoltaic medium-voltage DC power generation system are calculated to achieve a system-level macro-reliability evaluation of the large-scale photovoltaic medium-voltage DC power generation system.

[0012] Using photovoltaic module output power as the core degradation indicator and photovoltaic module conversion efficiency and electrical parameters as auxiliary degradation indicators, the determination of whether the photovoltaic module power degradation process conforms to the Brownian motion process is achieved through degradation increment independence verification, degradation increment normality verification, and degradation drift coefficient stability verification.

[0013] Based on the judgment results of the power degradation process of photovoltaic modules, a physical neural network model integrating the Brownian motion process is constructed. The physical neural network model is trained based on the preset multi-constraint total loss function and outputs the predicted value of photovoltaic module power degradation.

[0014] By integrating equipment-level reliability parameters, system-level macro-reliability evaluations, and predicted power degradation values ​​of photovoltaic modules, a comprehensive reliability evaluation report for large-scale photovoltaic medium-voltage DC power generation systems is generated.

[0015] Furthermore, based on operational data, the device-level reliability parameters of each core device are calculated, including:

[0016] Based on the time of equipment failure, the time of equipment failure repair, and the total operating time of the equipment, the number of failures of each core device is counted, and the original failure rate of each core device is calculated.

[0017] Using pre-acquired ambient temperature, humidity, and solar irradiance, the operating condition deviation of each core device is calculated, and the operating condition correction coefficient is determined based on the preset correction coefficient weights.

[0018] The original failure rate is corrected by the operating condition correction factor to obtain the actual failure rate; in order to eliminate the evaluation distortion caused by short-term start-stop and operating condition fluctuations.

[0019] Based on the actual failure rate, the reliability of each core device at any operating moment is calculated;

[0020] Based on the time of equipment failure and the time of equipment failure repair, calculate the average fault-free operating time and average repair time of each core device.

[0021] Summarize all calculation results to complete the calculation of equipment-level reliability parameters for each core device.

[0022] Furthermore, based on the pre-defined topology of the large-scale photovoltaic medium-voltage DC power generation system and combined with the reliability calculation formulas for series and parallel systems, the overall reliability of the large-scale photovoltaic medium-voltage DC power generation system is derived hierarchically; and the availability and failure probability of the large-scale photovoltaic medium-voltage DC power generation system are calculated to achieve a system-level macro-reliability evaluation of the large-scale photovoltaic medium-voltage DC power generation system, including:

[0023] Based on the topology of large-scale photovoltaic medium-voltage DC power generation system, photovoltaic modules are connected in series to form strings, strings are connected in parallel to form arrays, arrays are matched with combiner boxes and DC-DC converters in series, multiple series branches are connected in parallel to inverters, and multiple inverters are connected in parallel to the grid, the reliability of each level of large-scale photovoltaic medium-voltage DC power generation system is calculated in layers and level by level using the series and parallel reliability calculation formulas.

[0024] By combining the reliability calculation results of each level, the overall reliability of the entire large-scale photovoltaic medium-voltage DC power generation system is derived by summarizing and deducing layer by layer.

[0025] Based on the actual operating time and downtime of the large-scale photovoltaic medium-voltage DC power generation system, the availability of the large-scale photovoltaic medium-voltage DC power generation system is calculated to characterize its ability to supply power normally.

[0026] The available capacity, load demand, and probability of occurrence of each state of a large-scale photovoltaic medium-voltage DC power generation system under different operating conditions are statistically analyzed. The load failure probability of the large-scale photovoltaic medium-voltage DC power generation system is calculated, and the risk of power shortage of the large-scale photovoltaic medium-voltage DC power generation system is assessed.

[0027] The overall reliability, availability, and failure probability of a large-scale photovoltaic medium-voltage DC power generation system are integrated to complete the system-level macro-reliability evaluation of the large-scale photovoltaic medium-voltage DC power generation system.

[0028] Furthermore, using photovoltaic module output power as the core degradation indicator and photovoltaic module conversion efficiency and electrical parameters as auxiliary degradation indicators, the determination of whether the photovoltaic module power degradation process conforms to the Brownian motion process is achieved through verifications of degradation increment independence, degradation increment normality, and degradation drift coefficient stability, respectively:

[0029] The output power and conversion efficiency of photovoltaic modules and electrical parameters are respectively used as core degradation indicators and auxiliary degradation indicators, and the operating time, ambient temperature, ambient humidity and output power time series data of photovoltaic modules are extracted.

[0030] The output power time series data is processed at equal time intervals to generate a power time series sequence. The power difference between adjacent time points is calculated to obtain the power degradation increment sequence of the photovoltaic module. The autocorrelation coefficient of the power degradation increment sequence of the photovoltaic module is calculated one step behind. The independence of the degradation increment is verified based on the autocorrelation coefficient result.

[0031] The photovoltaic module power degradation increment sequence is sorted, and empirical cumulative distribution function and standard normal theoretical cumulative distribution function with the same mean and variance are constructed respectively. Kolmogorov-Smirnov test is used to calculate the maximum deviation between empirical cumulative distribution function and standard normal theoretical cumulative distribution function. Combined with the preset critical value, the normality of degradation increment is verified.

[0032] A fixed window length is set, and the photovoltaic module power degradation increment sequence is segmented and the drift degree is calculated segment by segment using a sliding window method. The stability of the degradation drift coefficient is verified based on the fluctuation of the drift degree.

[0033] Based on the combined verification results of the independence of degradation increment, the normality of degradation increment, and the stability of degradation drift coefficient, it is determined whether the power degradation process of photovoltaic modules conforms to the Brownian motion process. If all verifications are qualified, the process proceeds to the next step. If any verification fails, supplementary data is collected and all verifications are performed again.

[0034] Furthermore, based on the judgment results of the photovoltaic module power degradation process, a physical neural network model integrating the Brownian motion process is constructed. The physical neural network model is trained based on a preset multi-constraint total loss function, and the predicted value of photovoltaic module power degradation is output, including:

[0035] The inputs and outputs of the physical neural network model that integrates the Brownian motion process are determined; the inputs are time, ambient temperature, and ambient humidity, and the output is the power degradation of the photovoltaic module.

[0036] Based on the physical laws of photovoltaic module degradation, the drift coefficient and diffusion coefficient are estimated by using a sub-network fitting method or a preset function form fitting method.

[0037] The weight and bias parameters of the physical neural network are randomly initialized to set the initial parameter state of the physical neural network.

[0038] The photovoltaic module power degradation observation data and the physical constraint sampling points of the degradation process are input into the physical neural network. After unifying the data range, forward propagation calculation is performed to obtain the power degradation prediction result.

[0039] Based on the automatic differentiation technique preset by the physical neural network, the first and second derivatives of the power degradation prediction results are obtained and used for the calculation of various loss terms.

[0040] The data constraint loss, drift coefficient constraint loss, diffusion coefficient constraint loss, and stochastic differential equation residual term loss are calculated separately, and the total loss of the physical neural network model is obtained by weighting them according to the preset weight coefficients.

[0041] With minimizing the total loss as the optimization objective, the loss gradient is solved through backpropagation to expand the search range of the parameter space and select high-quality parameter combinations. In conjunction with the optimizer, the weights and bias parameters of the physical neural network are iteratively updated.

[0042] Repeatedly execute the data input, forward calculation, loss calculation and parameter update process until the total loss tends to converge or the preset maximum number of iterations is reached, and then end the training of the physical neural network model.

[0043] The time-series data and environmental data are input into the trained physical neural network model, which outputs a predicted value of the power degradation of the photovoltaic module to complete the reliability evaluation of the photovoltaic module.

[0044] Furthermore, the formula for calculating the data constraint loss is as follows:

[0045] ;

[0046] The formula for calculating the drift coefficient constraint loss is:

[0047] ;

[0048] The formula for calculating the diffusion coefficient constraint loss is:

[0049] ;

[0050] The formula for calculating the loss of the residual term in a stochastic differential equation is:

[0051] ;

[0052] In the formula, Indicates data constraint loss; This represents the drift coefficient constraint loss; This represents the diffusion coefficient constraint loss; This represents the loss of the residual term in a stochastic differential equation; This indicates the amount of degradation predicted by the physical neural network model; Indicates the measured degradation amount; N represents the number of data samples; i represents the i-th data sample; Indicates the theoretical drift coefficient; Represents the theoretical diffusion coefficient; E represents the mathematical expectation. Indicates variance.

[0053] Furthermore, the formula for calculating the total loss of the physical neural network model, obtained by weighting according to preset weight coefficients, is as follows:

[0054] ;

[0055] In the formula, L represents the total loss function; Indicates data constraint loss; This represents the drift coefficient constraint loss; This represents the diffusion coefficient constraint loss; This represents the loss of the residual term in a stochastic differential equation; Preset weighting coefficients representing data constraint loss; The preset weighting coefficient represents the drift coefficient constraint loss; The preset weighting coefficient represents the loss constrained by the diffusion coefficient; The preset weighting coefficients represent the loss of the residual term in a stochastic differential equation.

[0056] Furthermore, expanding the search range of the parameter space and filtering high-quality parameter combinations includes:

[0057] Construct a parameter candidate sample set, pre-set several initial parameter samples, and preset a judgment threshold for each set of parameter samples;

[0058] Multiple sets of weight parameters and bias parameters are randomly selected from the candidate parameter sample set and traversed and retrieved. The state information of the parameter samples is extracted, and the corresponding evaluation index is calculated and accumulated to obtain the comprehensive evaluation index of each set of parameter samples.

[0059] The comprehensive evaluation index of each group of parameter samples is compared with the preset judgment threshold. If the comprehensive evaluation index does not reach the preset judgment threshold, parameter samples are re-extracted and the search continues. Otherwise, the parameter sample type is distinguished by combining the parameter sample status information.

[0060] After completing the full sampling analysis of a single group of parameters to be retrieved, calculate the comprehensive evaluation value of each group of parameters, and judge the quality of the parameters based on the preset abnormal threshold, remove inferior parameters, and retain high-quality combinations of weight parameters and bias parameters.

[0061] The process of continuously extracting new parameter samples, repeating the retrieval, index calculation, and screening process, continues until the parameter candidate sample set is traversed, at which point the parameter space search and parameter screening end.

[0062] Furthermore, multiple sets of weight parameters and bias parameters are randomly selected from the parameter candidate sample set for traversal retrieval. The state information of the parameter samples is extracted, and the corresponding evaluation indicators are calculated and accumulated to obtain the comprehensive evaluation indicators for each set of parameter samples, including:

[0063] Multiple sets of weight parameters and bias parameters are extracted from the parameter candidate sample set using a random sampling method to form a parameter sample set to be traversed and retrieved.

[0064] For each group of weight parameters and bias parameters sample extracted, feature reduction is performed. Redundant or highly correlated dimensions are removed by analyzing the structural characteristics of the parameter samples, and the state information of the group of parameter samples is extracted from the reduced parameter features.

[0065] Based on the extracted state information, multiple individual evaluation indicators are calculated for each group of parameter samples; the individual evaluation indicators are used to characterize the performance of the group of parameter samples under different evaluation dimensions.

[0066] Iterate through all extracted parameter sample groups, repeatedly perform the operations of extracting state information and calculating individual evaluation indicators, and accumulate multiple individual evaluation indicators for each parameter sample group to obtain the comprehensive evaluation indicator corresponding to each parameter sample group.

[0067] Furthermore, feature reduction is performed on each extracted set of weight and bias parameter samples. Redundant or highly correlated dimensions are removed by analyzing the structural characteristics of the parameter samples, and the state information of the set of parameter samples is extracted from the reduced parameter features, including:

[0068] For each group of extracted weight parameters and bias parameters samples, the correlation between the values ​​of each parameter sample in the group of weight parameters and bias parameters samples is calculated to obtain the correlation analysis results of the group of weight parameters and bias parameters samples; the correlation analysis results are used to characterize the degree of association between the values ​​of different parameter samples in the group of weight parameters and bias parameters samples.

[0069] Based on the correlation analysis results, parameter sample pairs in the set of weight parameters and bias parameters that have a correlation degree exceeding a preset threshold are identified, and one parameter sample in each pair of highly correlated parameter samples is marked as a redundant parameter, forming a list of redundant parameter identifiers corresponding to the set of weight parameters and bias parameters.

[0070] From the original parameter sample set of the set of weight parameters and bias parameters, remove all redundant parameter samples listed in the redundant parameter identifier list, and retain the remaining parameter samples to obtain the parameter sample feature set after reduction of the set of weight parameters and bias parameters.

[0071] The values ​​of each parameter sample retained in the reduced parameter sample feature set are concatenated sequentially according to the order of their values ​​in the original parameter sample set to form a continuous feature sequence. This feature sequence is then used as the state information of the set of weight parameters and bias parameter samples.

[0072] The beneficial effects of this invention are as follows:

[0073] 1. This invention effectively overcomes the shortcomings of existing photovoltaic reliability evaluation methods, such as poor adaptability, single evaluation dimensions, and low accuracy of module degradation prediction. By constructing a three-layer reliability evaluation architecture of equipment level, system level, and module degradation level, it comprehensively completes the calculation of core equipment reliability parameters, and combines the system's series-parallel hybrid topology to solve the system reliability, availability, and failure probability, as well as the photovoltaic module degradation prediction, thus enriching the dimensions of reliability evaluation. At the same time, it embeds the physical laws of Wiener process degradation into the physical neural network loss function, taking into account both model fitting ability and interpretability, and can still guarantee the accuracy of degradation prediction under small sample conditions, thereby providing a reliable basis for photovoltaic power plant operation and maintenance optimization and fault early warning.

[0074] 2. This invention constructs a parameter candidate sample set and combines it with a multi-level threshold discrimination, random sampling traversal retrieval, feature reduction screening, and multi-dimensional index accumulation evaluation mechanism to achieve a large-scale parameter space search and refined selection of weight parameters and bias parameters of the physical neural network. By conducting correlation analysis and redundant dimension removal on the parameter samples, the redundancy of parameter features is effectively reduced, and the effective state information of the samples is accurately extracted. Based on the accumulation of multiple single evaluation indicators to obtain a comprehensive evaluation index, and combined with the judgment threshold, abnormal threshold, hierarchical screening, and classification to distinguish the quality of samples, iterative selection of massive parameter samples and elimination of inferior samples are achieved. This improves the training stability of the physical neural network model and the accuracy of photovoltaic module power degradation prediction, thereby ensuring the accuracy and credibility of the comprehensive evaluation results of the reliability of large-scale photovoltaic medium-voltage DC power generation systems. Attached Figure Description

[0075] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0076] Figure 1 This is a flowchart of a reliability evaluation method for a large-scale photovoltaic medium-voltage DC power generation system according to an embodiment of the present invention;

[0077] Figure 2 This is a schematic diagram of a reliability evaluation system for a large-scale photovoltaic medium-voltage DC power generation system according to an embodiment of the present invention;

[0078] Figure 3 This is a schematic diagram of the hardware operating environment according to an embodiment of the present invention;

[0079] Figure 4 This is a flowchart illustrating the system-level macroscopic reliability evaluation of a large-scale photovoltaic medium-voltage DC power generation system according to an embodiment of the present invention;

[0080] Figure 5 This is a flowchart of the dendritic cell algorithm according to an embodiment of the present invention.

[0081] In the picture:

[0082] 1. Data acquisition module; 2. Equipment reliability calculation module; 3. System reliability assessment module; 4. Degradation process verification module; 5. Degradation prediction modeling module; 6. Comprehensive evaluation generation module. Detailed Implementation

[0083] To further illustrate the various embodiments, the present invention provides accompanying drawings, which are part of the disclosure of the present invention. These drawings are mainly used to illustrate the embodiments and can be used in conjunction with the relevant descriptions in the specification to explain the operating principles of the embodiments. With reference to these drawings, those skilled in the art should be able to understand other possible implementation methods and the advantages of the present invention.

[0084] According to an embodiment of the present invention, a reliability evaluation method for a large-scale photovoltaic medium-voltage DC power generation system is provided.

[0085] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments, such as... Figure 1 As shown, the reliability evaluation method for a large-scale photovoltaic medium-voltage DC power generation system according to an embodiment of the present invention includes:

[0086] S1. Obtain operational data from multiple core devices in a large-scale photovoltaic medium-voltage DC power generation system; these core devices include photovoltaic modules, combiner boxes, DC-DC converters, and inverters.

[0087] S2. Based on operational data, calculate the device-level reliability parameters of each core device; device-level reliability parameters include failure rate, reliability, mean time between failures (MTBF), and mean time to repair (MTBT).

[0088] S3. Based on the preset topology of the large-scale photovoltaic medium-voltage DC power generation system, and combined with the reliability calculation formulas for series and parallel systems, the overall reliability of the large-scale photovoltaic medium-voltage DC power generation system is derived in layers; and the availability and failure probability of the large-scale photovoltaic medium-voltage DC power generation system are calculated to achieve a system-level macro-reliability evaluation of the large-scale photovoltaic medium-voltage DC power generation system.

[0089] S4. Using the output power of photovoltaic modules as the core degradation index and the conversion efficiency and electrical parameters of photovoltaic modules as auxiliary degradation indexes, the degradation increment independence verification, degradation increment normality verification, and degradation drift coefficient stability verification are used to determine whether the power degradation and attenuation process of photovoltaic modules conforms to the Brownian motion process.

[0090] S5. Based on the judgment results of the power degradation process of photovoltaic modules, a physical neural network model integrating the Brownian motion process is constructed. The physical neural network model is trained based on the preset multi-constraint total loss function and the predicted value of photovoltaic module power degradation is output.

[0091] S6 integrates equipment-level reliability parameters, system-level macro-reliability evaluation, and predicted photovoltaic module power degradation to generate a comprehensive reliability evaluation report for a large-scale photovoltaic medium-voltage DC power generation system.

[0092] In this optional embodiment, the calculation of device-level reliability parameters for each core device based on operational data includes:

[0093] Based on the time of equipment failure, the time of equipment failure repair, and the total operating time of the equipment, the number of failures of each core device is counted, and the original failure rate of each core device is calculated.

[0094] Using pre-acquired ambient temperature, humidity, and solar irradiance, the operating condition deviation of each core device is calculated, and the operating condition correction coefficient is determined based on the preset correction coefficient weights.

[0095] The original failure rate is corrected by the operating condition correction factor to obtain the actual failure rate; in order to eliminate the evaluation distortion caused by short-term start-stop and operating condition fluctuations.

[0096] Based on the actual failure rate, the reliability of each core device at any operating moment is calculated;

[0097] Based on the time of equipment failure and the time of equipment failure repair, calculate the average fault-free operating time and average repair time of each core device.

[0098] Summarize all calculation results to complete the calculation of equipment-level reliability parameters for each core device.

[0099] like Figure 4 As shown, in this optional embodiment, based on the preset topology of the large-scale photovoltaic medium-voltage DC power generation system and combined with the reliability calculation formulas for series and parallel systems, the overall reliability of the large-scale photovoltaic medium-voltage DC power generation system is derived hierarchically; and the availability and failure probability of the large-scale photovoltaic medium-voltage DC power generation system are calculated to achieve a system-level macroscopic reliability evaluation of the large-scale photovoltaic medium-voltage DC power generation system, including:

[0100] Based on the topology of large-scale photovoltaic medium-voltage DC power generation system, photovoltaic modules are connected in series to form strings, strings are connected in parallel to form arrays, arrays are matched with combiner boxes and DC-DC converters in series, multiple series branches are connected in parallel to inverters, and multiple inverters are connected in parallel to the grid, the reliability of each level of large-scale photovoltaic medium-voltage DC power generation system is calculated in layers and level by level using the series and parallel reliability calculation formulas.

[0101] By combining the reliability calculation results of each level, the overall reliability of the entire large-scale photovoltaic medium-voltage DC power generation system is derived by summarizing and deducing layer by layer.

[0102] Based on the actual operating time and downtime of the large-scale photovoltaic medium-voltage DC power generation system, the availability of the large-scale photovoltaic medium-voltage DC power generation system is calculated to characterize its ability to supply power normally.

[0103] The available capacity, load demand, and probability of occurrence of each state of a large-scale photovoltaic medium-voltage DC power generation system under different operating conditions are statistically analyzed. The load failure probability of the large-scale photovoltaic medium-voltage DC power generation system is calculated, and the risk of power shortage of the large-scale photovoltaic medium-voltage DC power generation system is assessed.

[0104] The overall reliability, availability, and failure probability of a large-scale photovoltaic medium-voltage DC power generation system are integrated to complete the system-level macro-reliability evaluation of the large-scale photovoltaic medium-voltage DC power generation system.

[0105] In this optional embodiment, using the output power of the photovoltaic module as the core degradation indicator and the conversion efficiency and electrical parameters of the photovoltaic module as auxiliary degradation indicators, the determination of whether the power degradation process of the photovoltaic module conforms to the Brownian motion process is carried out through verification of the independence of degradation increment, verification of the normality of degradation increment, and verification of the stability of degradation drift coefficient, respectively.

[0106] The output power and conversion efficiency of photovoltaic modules and electrical parameters are respectively used as core degradation indicators and auxiliary degradation indicators, and the operating time, ambient temperature, ambient humidity and output power time series data of photovoltaic modules are extracted.

[0107] The output power time series data is processed at equal time intervals to generate a power time series sequence. The power difference between adjacent time points is calculated to obtain the power degradation increment sequence of the photovoltaic module. The autocorrelation coefficient of the power degradation increment sequence of the photovoltaic module is calculated one step behind. The independence of the degradation increment is verified based on the autocorrelation coefficient result.

[0108] The photovoltaic module power degradation increment sequence is sorted, and empirical cumulative distribution function and standard normal theoretical cumulative distribution function with the same mean and variance are constructed respectively. Kolmogorov-Smirnov test is used to calculate the maximum deviation between empirical cumulative distribution function and standard normal theoretical cumulative distribution function. Combined with the preset critical value, the normality of degradation increment is verified.

[0109] A fixed window length is set, and the photovoltaic module power degradation increment sequence is segmented and the drift degree is calculated segment by segment using a sliding window method. The stability of the degradation drift coefficient is verified based on the fluctuation of the drift degree.

[0110] Based on the combined verification results of the independence of degradation increment, the normality of degradation increment, and the stability of degradation drift coefficient, it is determined whether the power degradation process of photovoltaic modules conforms to the Brownian motion process. If all verifications are qualified, the process proceeds to the next step. If any verification fails, supplementary data is collected and all verifications are performed again.

[0111] In this optional embodiment, based on the judgment result of the photovoltaic module power degradation process, a physical neural network model integrating the Brownian motion process is constructed. The physical neural network model is trained based on a preset multi-constraint total loss function, and the predicted value of photovoltaic module power degradation is output, including:

[0112] The inputs and outputs of the physical neural network model that integrates the Brownian motion process are determined; the inputs are time, ambient temperature, and ambient humidity, and the output is the power degradation of the photovoltaic module.

[0113] Based on the physical laws of photovoltaic module degradation, the drift coefficient and diffusion coefficient are estimated by using a sub-network fitting method or a preset function form fitting method.

[0114] The weight and bias parameters of the physical neural network are randomly initialized to set the initial parameter state of the physical neural network.

[0115] The photovoltaic module power degradation observation data and the physical constraint sampling points of the degradation process are input into the physical neural network. After unifying the data range, forward propagation calculation is performed to obtain the power degradation prediction result.

[0116] Based on the automatic differentiation technique preset by the physical neural network, the first and second derivatives of the power degradation prediction results are obtained and used for the calculation of various loss terms.

[0117] The data constraint loss, drift coefficient constraint loss, diffusion coefficient constraint loss, and stochastic differential equation residual term loss are calculated separately, and the total loss of the physical neural network model is obtained by weighting them according to the preset weight coefficients.

[0118] With minimizing the total loss as the optimization objective, the loss gradient is solved through backpropagation to expand the search range of the parameter space and select high-quality parameter combinations. In conjunction with the optimizer, the weights and bias parameters of the physical neural network are iteratively updated.

[0119] Repeatedly execute the data input, forward calculation, loss calculation and parameter update process until the total loss tends to converge or the preset maximum number of iterations is reached, and then end the training of the physical neural network model.

[0120] The time-series data and environmental data are input into the trained physical neural network model, which outputs a predicted value of the power degradation of the photovoltaic module to complete the reliability evaluation of the photovoltaic module.

[0121] like Figure 5 As shown, in this optional embodiment, expanding the parameter space search range and filtering high-quality parameter combinations includes:

[0122] Construct a parameter candidate sample set, pre-set several initial parameter samples, and preset a judgment threshold for each set of parameter samples;

[0123] Multiple sets of weight parameters and bias parameters are randomly selected from the candidate parameter sample set and traversed and retrieved. The state information of the parameter samples is extracted, and the corresponding evaluation index is calculated and accumulated to obtain the comprehensive evaluation index of each set of parameter samples.

[0124] The comprehensive evaluation index of each group of parameter samples is compared with the preset judgment threshold. If the comprehensive evaluation index does not reach the preset judgment threshold, parameter samples are re-extracted and the search continues. Otherwise, the parameter sample type is distinguished by combining the parameter sample status information.

[0125] After completing the full sampling analysis of a single group of parameters to be retrieved, calculate the comprehensive evaluation value of each group of parameters, and judge the quality of the parameters based on the preset abnormal threshold, remove inferior parameters, and retain high-quality combinations of weight parameters and bias parameters.

[0126] The process of continuously extracting new parameter samples, repeating the retrieval, index calculation, and screening process, continues until the parameter candidate sample set is traversed, at which point the parameter space search and parameter screening end.

[0127] It should be explained that this invention utilizes the dendritic cell algorithm to expand the search range of the parameter space and select high-quality parameter combinations. The dendritic cell algorithm is an intelligent optimization algorithm that simulates the working mechanism of dendritic cells in the biological immune system. Relying on the operational logic of sample state perception, multi-dimensional index calculation, threshold judgment and classification, and iterative screening, it completes the identification and selection of massive candidate samples. This algorithm can be adapted to scenarios involving the full-domain retrieval and selection of superior and inferior parameters of neural network weights and bias parameters by configuring parameters such as sample size, judgment threshold, anomaly threshold, and index calculation rules, balancing the comprehensiveness of parameter search and the efficiency of screening. In this invention, the combination of weight parameters and bias parameters is used as the algorithm's sample. Through the algorithm's state extraction, index accumulation, multi-level threshold comparison, and sample classification and elimination mechanisms, a large-scale parameter space traversal is achieved, and high-quality parameter combinations suitable for the operation requirements of the physical neural network are selected.

[0128] Specifically, a parameter candidate sample set is constructed for the weight and bias parameters of the physical neural network. Several initial parameter samples are pre-set within this set, and a separate judgment threshold is preset for each set (the threshold ranges from [0.2, 0.8], preferably 0.5; this threshold serves as the comprehensive evaluation criterion for determining whether a sample enters the classification stage). The number of initial parameter sample sets ranges from 30 to 200, preferably 100. Each set consists of a complete combination of weight and bias parameters of the physical neural network, ensuring that the parameter structure of a single sample is consistent with the actual operating parameter structure of the network. The parameter candidate sample set is constructed based on the discretized and dimension-reduced parameter features, including only valid parameter samples after redundant features have been removed.

[0129] Multiple sets of weight and bias parameter samples are randomly selected from the constructed parameter candidate sample set for traversal retrieval. Based on the effective parameter features retained after dimensionality reduction, the operational status information corresponding to each set of parameter samples is extracted. Following a unified index calculation rule (which calculates individual evaluation indicators item by item based on the sample operational status information, including three basic indicators: parameter operational stability, parameter matching, and parameter adaptability), the individual evaluation indicators are solved. The index values ​​are then successively accumulated into the index record of the corresponding parameter sample. Finally, all individual evaluation indicators are integrated to generate a comprehensive evaluation index corresponding to each set of parameter samples. The number of samples randomly selected in a single run ranges from 5 to 20, with 10 sets being preferred. The calculation precision of individual evaluation indicators is uniformly retained to four decimal places, and the index accumulation is a linear superposition of multiple retrieval results for the same sample.

[0130] The comprehensive evaluation index corresponding to each set of parameter samples is compared one by one with the preset judgment threshold mentioned above. If the value of the comprehensive evaluation index is less than the judgment threshold, it means that the current sample does not meet the basic usage requirements. New parameter samples are then extracted from the parameter candidate sample set to continue the retrieval and index calculation process. If the value of the comprehensive evaluation index is greater than or equal to the judgment threshold, the basic index of the sample is deemed to meet the standard. Based on the extracted parameter sample operation status information, the parameter sample type is distinguished according to the sample classification rules (the sample classification rules are based on the characteristic bias of the operation status information, dividing the samples into three categories: stable samples, matching samples, and adaptive samples; the classification judgment criteria are based on the numerical proportion of the three types of single evaluation indicators. The parameter stable index with the highest proportion is classified as a stable sample, the parameter matching index with the highest proportion is classified as a matching sample, and the parameter adaptive index with the highest proportion is classified as an adaptive sample).

[0131] After completing the sampling, analysis, and classification of all parameter samples to be retrieved in the current batch, a comprehensive evaluation value is calculated based on the comprehensive evaluation index and operating status information of each group of parameter samples (the comprehensive evaluation value is calculated by weighting the comprehensive evaluation index and the status characteristic coefficient, with the status characteristic coefficient ranging from [0.3, 0.7], preferably 0.5). A preset abnormal threshold is called (the abnormal threshold ranges from [0.1, 0.4], preferably 0.2, which is the standard for separating the quality of samples) to judge the quality of samples. Samples with a comprehensive evaluation value lower than the abnormal threshold are judged as inferior parameter samples and are directly removed. Samples with a comprehensive evaluation value greater than or equal to the abnormal threshold are judged as qualified samples, and the combination of weight parameters and bias parameters of this group is retained and included in the high-quality parameter candidate library.

[0132] The algorithm continuously extracts new parameter samples from the parameter candidate sample set, and repeatedly executes the complete process of sample retrieval, state information extraction, index calculation and accumulation, threshold comparison, sample classification, quality judgment and sample elimination, traversing and processing all samples in the parameter candidate sample set one by one. When all parameter samples in the parameter candidate sample set have completed the retrieval and screening operation, the current parameter space search and parameter screening work is terminated. The final combination of weight parameters and bias parameters remaining in the high-quality parameter candidate library is the optimal parameter combination screened by the dendritic cell algorithm, which can be directly used for iterative optimization of physical neural networks.

[0133] In this optional embodiment, multiple sets of weight parameters and bias parameters are randomly selected from the parameter candidate sample set for traversal retrieval. The state information of the parameter samples is extracted, and the corresponding evaluation indicators are calculated and accumulated to obtain the comprehensive evaluation indicators for each set of parameter samples, including:

[0134] Multiple sets of weight parameters and bias parameters are extracted from the parameter candidate sample set using a random sampling method to form a parameter sample set to be traversed and retrieved.

[0135] For each group of weight parameters and bias parameters sample extracted, feature reduction is performed. Redundant or highly correlated dimensions are removed by analyzing the structural characteristics of the parameter samples, and the state information of the group of parameter samples is extracted from the reduced parameter features.

[0136] Based on the extracted state information, multiple individual evaluation indicators are calculated for each group of parameter samples; the individual evaluation indicators are used to characterize the performance of the group of parameter samples under different evaluation dimensions.

[0137] Iterate through all extracted parameter sample groups, repeatedly perform the operations of extracting state information and calculating individual evaluation indicators, and accumulate multiple individual evaluation indicators for each parameter sample group to obtain the comprehensive evaluation indicator corresponding to each parameter sample group.

[0138] Specifically, a random sampling method is used to extract multiple sets of weight parameters and bias parameters from the parameter candidate sample set, forming the parameter sample set to be traversed and retrieved. The random sampling rules are as follows: all samples in the parameter candidate sample set are uniformly numbered, and samples are selected unbiasedly using random number generation. Each sampling does not deliberately limit the sample location or parameter type, ensuring sampling randomness. The number of sample groups extracted in a single sampling ranges from 5 to 20 groups, with 10 groups being the preferred value. After sampling, all selected samples are uniformly classified into the set to be traversed and retrieved, and subsequent processing is carried out sequentially.

[0139] For each set of extracted weight and bias parameter samples, feature reduction is performed. Redundant or highly correlated dimensions are removed by analyzing the structural characteristics of the parameter samples. The state information of the parameter samples is then extracted from the reduced parameter features. The feature reduction criteria are as follows: (Dimension Removal Standard: The threshold for the dimension correlation coefficient is set to [0.7, 0.9], with a preferred value of 0.8. When the correlation coefficient of two parameter feature dimensions is greater than or equal to this threshold, they are considered highly correlated, and one is removed. Feature dimensions without practical representation or impact on parameter performance are directly identified as redundant and removed). During dimension screening, core feature dimensions that directly reflect the parameter operating characteristics and are related to the network's operating effect are prioritized for retention. Only useless and highly repetitive dimension content is removed to avoid loss of effective information due to dimension reduction. After feature reduction, sample state information is extracted based on the retained effective parameter features. State information is defined as various feature data that reflect the operating performance of weight and bias parameters, parameter matching degree, and iteration stability, serving as the basis for index calculation.

[0140] Based on the extracted state information, multiple individual evaluation indicators are calculated for each group of parameter samples. These individual evaluation indicators characterize the performance of the parameter samples under different evaluation dimensions. This invention pre-defines three types of individual evaluation indicators: parameter stability indicators, parameter matching indicators, and parameter adaptation indicators (indicator definitions: parameter stability indicators characterize the fluctuation range of parameters during iteration; parameter matching indicators characterize the degree of fit between weight parameters and bias parameters; parameter adaptation indicators characterize the overall fit between parameters and network operation requirements). The calculation of individual indicators uniformly follows pre-defined operation rules (operation rules: using sample state information as input, the calculation is completed according to the quantification logic of the corresponding evaluation dimension; all individual indicator calculation results are retained to four decimal places to ensure consistent indicator numerical precision), thus sequentially solving all individual evaluation indicators for a single group of samples.

[0141] The process iterates through all extracted parameter sample groups, repeatedly performing state information extraction and single-item evaluation index calculation. Multiple single-item evaluation indices for each parameter sample group are then summed to obtain the comprehensive evaluation index for each group. For the remaining sample groups in the search set to be iterated through, the entire process of feature reduction, state information extraction, and single-item index calculation is executed iteratively for each group, ensuring that each sample group undergoes standardization. The index summation rule (summation method: linear summation mode, directly adding the values ​​of all single-item evaluation indices corresponding to the same sample without setting additional weight coefficients) is used. The comprehensive evaluation index is defined as the sum of all single-item evaluation indices for a single sample group, used to comprehensively evaluate the combined performance of the weighted and biased parameter samples in that group. Once all samples in the search set to be iterated through have completed the summation operation, the comprehensive evaluation index for all samples is obtained, providing complete data support for subsequent threshold comparison and sample selection.

[0142] In this optional embodiment, feature reduction processing is performed on each set of extracted weight parameters and bias parameters samples. Redundant or highly correlated dimensions are eliminated by analyzing the structural characteristics of the parameter samples, and the state information of the set of parameter samples is extracted from the reduced parameter features, including:

[0143] For each group of extracted weight parameters and bias parameters samples, the correlation between the values ​​of each parameter sample in the group of weight parameters and bias parameters samples is calculated to obtain the correlation analysis results of the group of weight parameters and bias parameters samples; the correlation analysis results are used to characterize the degree of association between the values ​​of different parameter samples in the group of weight parameters and bias parameters samples.

[0144] Based on the correlation analysis results, parameter sample pairs in the set of weight parameters and bias parameters that have a correlation degree exceeding a preset threshold are identified, and one parameter sample in each pair of highly correlated parameter samples is marked as a redundant parameter, forming a list of redundant parameter identifiers corresponding to the set of weight parameters and bias parameters.

[0145] From the original parameter sample set of the set of weight parameters and bias parameters, remove all redundant parameter samples listed in the redundant parameter identifier list, and retain the remaining parameter samples to obtain the parameter sample feature set after reduction of the set of weight parameters and bias parameters.

[0146] The values ​​of each parameter sample retained in the reduced parameter sample feature set are concatenated sequentially according to the order of their values ​​in the original parameter sample set to form a continuous feature sequence. This feature sequence is then used as the state information of the set of weight parameters and bias parameter samples.

[0147] Specifically, for each extracted set of weight parameters and bias parameters, the correlation between the values ​​of each parameter within that set is calculated to obtain the correlation analysis results between the weight parameters and bias parameters. The Pearson correlation coefficient is used to quantify the correlation (correlation calculation rule: for each pair of independent parameter values ​​within a single set of samples, the correlation degree is calculated for each parameter combination, and the corresponding correlation coefficient is output as the quantification result). The correlation analysis results are used to visually represent the degree of linkage between different parameter values ​​in the set of samples. All correlation coefficient calculation results are uniformly retained to four decimal places to ensure consistent calculation precision.

[0148] Based on the correlation analysis results, parameter pairs in the set of weight parameters and bias parameters whose correlation exceeds a preset threshold are identified. One parameter in each highly correlated parameter pair is marked as a redundant parameter, forming a redundant parameter identification list for that set of samples. The correlation threshold is set to [0.7, 0.9], with a preferred value of 0.8 (this threshold is the standard for determining high correlation; when the correlation coefficient between two sets of parameters is greater than or equal to this threshold, the correlation is considered too high, indicating feature redundancy). For each pair of parameters determined to be highly correlated, redundant parameters are marked according to a unified screening rule (redundancy marking rule: parameters with greater data fluctuation and weaker impact on overall performance are preferentially removed; if the performance of two parameters is basically the same, one is randomly selected and marked as a redundant parameter). After completing the traversal and identification of all parameters, all marked parameters are summarized and a standardized redundant parameter identification list is generated, clearly indicating the redundant parameter number, location, and corresponding category.

[0149] From the original parameter set of the weight and bias parameter samples, all redundant parameters listed in the redundant parameter identification list are removed, and the remaining parameters are retained to obtain the parameter feature set after sample reduction. During the removal operation, the corresponding parameters are removed item by item according to the redundant parameter identification list; no parameters outside the list are additionally deleted or retained. The parameters remaining after removal are the valid feature parameters. All valid parameters together constitute the parameter feature set after sample reduction. This set only retains parameters with low correlation and independent characterization.

[0150] The parameter values ​​retained in the reduced parameter feature set are sequentially concatenated according to their order in the original parameter set to form a continuous feature sequence. This feature sequence serves as the state information for the weight and bias parameter samples. The sequence concatenation rule (using the inherent order of the original parameter set without changing the parameter positions, sequentially concatenates the values ​​of all retained parameters to form a single, complete, and ordered feature sequence) defines the sample state information as the continuous feature sequence obtained after feature reduction and ordered concatenation. This sequence fully carries the core operational characteristics of this set of parameters and serves as the fundamental data source for subsequent indicator calculations and sample classification.

[0151] According to another embodiment of the invention, such as Figure 2 As shown, a reliability evaluation system for a large-scale photovoltaic medium-voltage DC power generation system is also provided, the system comprising:

[0152] Data acquisition module 1 is used to acquire the operating data of multiple core devices in a large-scale photovoltaic medium-voltage DC power generation system; the multiple core devices include photovoltaic modules, combiner boxes, DC-DC converters and inverters;

[0153] Equipment reliability calculation module 2 is used to calculate the equipment-level reliability parameters of each core device based on operational data; the equipment-level reliability parameters include failure rate, reliability, mean time between failures (MTBF), and mean time to repair (MTBF).

[0154] System reliability assessment module 3 is used to derive the overall reliability of the large-scale photovoltaic medium-voltage DC power generation system in layers based on the preset topology of the large-scale photovoltaic medium-voltage DC power generation system and the reliability calculation formula of series and parallel systems; and to calculate the availability and failure probability of the large-scale photovoltaic medium-voltage DC power generation system in order to achieve system-level macro reliability evaluation of the large-scale photovoltaic medium-voltage DC power generation system.

[0155] The degradation process verification module 4 is used to determine whether the degradation process of photovoltaic module power conforms to the Brownian motion process by using the output power of photovoltaic module as the core degradation index and the conversion efficiency and electrical parameters of photovoltaic module as auxiliary degradation indexes. It uses the degradation increment independence verification, degradation increment normality verification, and degradation drift coefficient stability verification respectively.

[0156] The degradation prediction modeling module 5 is used to construct a physical neural network model that integrates the Brownian motion process based on the judgment results of the photovoltaic module power degradation process. The physical neural network model is trained based on the preset multi-constraint total loss function and outputs the predicted value of photovoltaic module power degradation.

[0157] The comprehensive evaluation generation module 6 is used to integrate equipment-level reliability parameters, system-level macro reliability evaluation, and photovoltaic module power degradation prediction values ​​to generate a comprehensive reliability evaluation report for a large-scale photovoltaic medium-voltage DC power generation system.

[0158] To facilitate understanding of the above technical solutions of the present invention, the following provides a detailed description of the reliability evaluation of the large-scale photovoltaic medium-voltage DC power generation system in actual practice.

[0159] Step 1: Collect basic equipment operation data: Through the online monitoring system for primary equipment in the photovoltaic power station, collect core operating data of photovoltaic modules, combiner boxes, DC-DC converters, and inverters, including: total equipment operating time T, and the time of fault occurrence. Fault repair time The system monitors equipment input and output power, and simultaneously collects auxiliary environmental data such as temperature, humidity, and solar irradiance using power plant environmental monitoring sensors, with a sampling period of 1 minute.

[0160] Step 2: Calculate device-level reliability parameters: Based on the collected operational data, calculate the failure rate of each core device. Reliability Mean Time Between Failures (MTBF) Average repair time Equipment-level reliability parameters are used in calculating the failure rate. When introducing operating condition disturbance corrections, the operating condition correction coefficients are determined based on time-series data of temperature, humidity, and irradiance. Regarding the original failure rate Make corrections to obtain Based on ambient temperature ,humidity Solar irradiance Time-series data is used to identify abnormal operating conditions such as short-term equipment start-up and shutdown, sudden load changes, and abrupt environmental changes, and to construct operating condition correction coefficients. , This represents the original statistical failure rate. , To adjust the weighting coefficients, a value of 0.1-0.5 is used. For deviation from operating conditions, The reference temperature is 25℃. The reference humidity is 60%. With a reference irradiance of 1000 W / m², The specific calculation formula to eliminate the evaluation distortion caused by short-term start-stop and operating condition fluctuations is as follows:

[0161] failure rate : .

[0162] Reliability : .

[0163] Mean time between failures (MTBF) ): .

[0164] Mean time to repair ( ): .

[0165] Step 3: Based on the topology of the photovoltaic medium-voltage DC power generation system, which consists of "modules connected in series to form strings - strings connected in parallel to form arrays - array + combiner box + DC-DC converter connected in series - multiple series branches connected in parallel to inverters - multiple inverters connected in parallel to the grid", and combined with the reliability calculation formulas for series and parallel systems, the overall reliability of the large-scale photovoltaic medium-voltage DC power generation system is derived hierarchically; at the same time, the availability A and the loss of load probability (LOLP) of the large-scale photovoltaic medium-voltage DC power generation system are calculated to achieve a system-level macro-reliability evaluation of the large-scale photovoltaic medium-voltage DC power generation system.

[0166] The reliability of a series system is calculated as follows: .

[0167] The reliability calculation method for parallel systems is as follows: .

[0168] Therefore, the reliability of the entire photovoltaic power station system can be calculated based on the specific form of the power station topology using the reliability calculation methods for series systems and parallel systems described above.

[0169] From a broader perspective, the reliability of a photovoltaic power plant is reflected in its ability to provide the amount of electricity expected by the power grid, i.e., the system's availability. The calculation logic is the ratio of the time the photovoltaic power station can operate normally to the total operating time of the photovoltaic power station:

[0170] ;

[0171] in, For system uptime, This refers to the system downtime.

[0172] Loss of Load Probability (LOLP) is defined as the probability that a power plant's available generating capacity is less than its load demand within a specific time period. It quantifies the level of risk a system faces regarding power shortages and is a probability value between 0 and 1.

[0173]

[0174] The available capacity for each state i is The probability of occurrence is The load is D.

[0175] Step 4: Verify the Wiener process characteristics of photovoltaic module degradation: Using the output power of the photovoltaic module as the core degradation indicator, collect data on the time during the module's operation. Ambient temperature ,humidity And output power data, and verify whether its degradation process conforms to the Wiener process using a three-step method:

[0176] The incremental independence verification process involves: first, collecting time-series data of photovoltaic module output power at equal time intervals to construct a power time-series sequence, and then calculating the power difference between adjacent time points to obtain the module power degradation increment sequence. Let the photovoltaic module power degradation increment sequence be... The first step is to calculate the series mean: Formula for calculating the autocorrelation coefficient with a 1-step lag: The time series autocorrelation algorithm is used to calculate the autocorrelation coefficient of the degraded increment sequence lagged by one step. If the value of the autocorrelation coefficient lagged by one step is close to 0, the increment is determined to be independent.

[0177] The normality verification specifically involves: arranging the power degradation increment sequence of the modules (i.e., the photovoltaic modules mentioned above) in ascending order, constructing both the empirical cumulative distribution function and the standard normal theoretical cumulative distribution function with the same mean and variance, and using the Kolmogorov-Smirnov test to calculate the maximum deviation between the two distributions. Let the sample cumulative distribution function be... Theoretical standard distribution function The KS test maximum bias statistic: The maximum deviation value is compared with the critical value corresponding to the preset significance level. If the maximum deviation value is less than the critical value, the degradation increment is determined to follow a normal distribution.

[0178] The stability verification of the drift coefficient is as follows: A fixed window length is set, and a sliding window method is used to segment the subsequence on the power degradation increment sequence. The drift value is calculated for each window. If the drift value calculated for each window has a small fluctuation range, no obvious upward or downward trend, and remains stable overall, then the drift coefficient is determined. It is a constant.

[0179] Step 5: Construct a PINN-Wiener fusion model to evaluate component reliability:

[0180] A physical neural network (PINN) model incorporating the Wiener process is constructed to achieve accurate prediction of photovoltaic module power degradation. The specific steps are as follows:

[0181] Determine the model input and output: time Ambient temperature ,humidity As input to the model, the power degradation of photovoltaic modules As model output.

[0182] It conforms to the Wiener process, and the stress it experiences during operation includes ambient temperature. ,humidity The physical law governing the degradation of photovoltaic modules is expressed as follows:

[0183] ;

[0184] in, This represents the power degradation of the component. To be affected by ambient temperature ,humidity The drift coefficient affected To be affected by ambient temperature ,humidity The diffusion coefficient of the influence This is the increment of Brownian motion.

[0185] Network input set to time Ambient temperature ,humidity The network output is the power degradation of the photovoltaic modules. ; Regarding the relevant characteristic parameters in the degradation process (including drift coefficient) diffusion coefficient Curve fitting and parameter estimation can be accomplished using two methods: sub-network fitting and pre-defined function fitting. The specific implementation steps are as follows:

[0186] 1) Using sub-networks for fitting and parameter estimation: A nested network architecture is constructed by building a main prediction network paired with multiple lightweight sub-networks. The main network uses time... Ambient temperature ,humidity As input, output the predicted power degradation value of photovoltaic modules. Sub-networks, relying on the same input data, output environmental impact coefficients respectively. Volatility coefficient Two types of feature parameters to be estimated. Time-dependent. Ambient temperature ,humidity and corresponding measured values ​​of component power degradation The network is trained using measured sample data. Subnetworks and the main network are trained iteratively in sync. Network parameters are automatically updated via backpropagation, learning the influence of temperature and humidity on the drift and diffusion coefficients. Subnetworks directly output dynamic values ​​of the drift and diffusion coefficients through forward propagation, without requiring manually preset analytical formulas. During the synchronous output of the degradation fitting curve by the main network, various influencing parameters are dynamically fitted and calculated. Ultimately, degradation curve fitting and feature parameter estimation are completed synchronously, achieving collaborative solution.

[0187] 2) A direct assumption function approach is used to achieve fitting and parameter estimation, combining the actual power degradation law of photovoltaic modules. A degradation baseline function expression containing time, temperature, and humidity variables is pre-defined (this baseline function is the analytical form corresponding to the Wiener stochastic differential equation of the physical law of photovoltaic module degradation mentioned above), clarifying the basic structure of the function and the unknown parameters to be determined. The sequentially collected runtime data and the synchronously monitored on-site temperature and humidity environmental data are substituted into the preset function one by one to build the corresponding relationship. Relying on the model iterative training, the values ​​of unknown parameters inside the function are continuously corrected. By continuously reducing the fitting error, the entire power degradation curve is fitted, and finally the values ​​of all characteristic parameters within the preset function are determined, completing the parameter estimation work.

[0188] The network structure needs to be determined and adjusted based on the actual data scale. The loss function is roughly as follows:

[0189] The formula for calculating data constraint loss is:

[0190] ;

[0191] The formula for calculating the drift coefficient constraint loss is:

[0192] ;

[0193] The formula for calculating the diffusion coefficient constraint loss is:

[0194] ;

[0195] The formula for calculating the loss of the residual term in a stochastic differential equation is:

[0196] ;

[0197] In the formula, Indicates data constraint loss; This represents the drift coefficient constraint loss; This represents the diffusion coefficient constraint loss; This represents the loss of the residual term in a stochastic differential equation; This indicates the amount of degradation predicted by the physical neural network model; Indicates the measured degradation amount; N represents the number of data samples; E is the expected value. Let be the variance; i represents the i-th data sample; Indicates the theoretical drift coefficient; Represents the theoretical diffusion coefficient, in the formula Drift coefficient Time corresponding to the i-th sample The product of represents the Wiener degradation process at time t. The theoretically expected degradation amount is as follows.

[0198] The formula for calculating the total loss is:

[0199] ;

[0200] In the formula, L represents the total loss function; Indicates data constraint loss; This represents the drift coefficient constraint loss; This represents the diffusion coefficient constraint loss; This represents the loss of the residual term in a stochastic differential equation; Preset weighting coefficients representing data constraint loss; The preset weighting coefficient represents the drift coefficient constraint loss; The preset weighting coefficient represents the loss constrained by the diffusion coefficient; These represent the preset weighting coefficients for the loss of the residual term in the stochastic differential equation. The preset weighting coefficients use optimally selected values: =1, =0.5, =0.5, =0.2.

[0201] Training Method: The weights and biases of the neural network are randomly initialized. Observed data, physical constraint points, and boundary / initial condition points are input into the network to obtain predicted values. The derivative of the network output with respect to input variables (such as time and spatial coordinates) is calculated using automatic differentiation, and this is used for the physical constraint loss. Data loss, physical loss, and boundary / initial condition loss are calculated separately to obtain the total loss. Based on the total loss, the gradient is calculated through backpropagation. The network and physical parameters are updated using an optimizer (such as Adam or L-BFGS) to minimize the total loss. Forward propagation, loss calculation, and parameter updates are repeated until the loss converges or the maximum number of iterations is reached. Output Degradation Prediction: New time, temperature, and humidity data are input into the trained model to output a predicted value for the photovoltaic module's power degradation. This degradation value is then used to determine the current reliability status of the module.

[0202] Utilizing the automatic differentiation function built into the physical neural network framework, in time... As the independent variable, the predicted power degradation value output by the model. Automatic differentiation directly yields the degradation amount. Regarding time First derivative:

[0203] ;

[0204] And the second derivative of the rate of change of the instantaneous degradation rate with time:

[0205] ;

[0206] The obtained derivative results are directly used in the subsequent calculation and solution of drift coefficient constraint loss, diffusion coefficient constraint loss, and stochastic differential equation residual loss; the data constraint loss, drift coefficient constraint loss, diffusion coefficient constraint loss, and stochastic differential equation residual loss are calculated separately in sequence, and then combined with the optimized preset weight coefficients. =1, =0.5, =0.5, =0.2 is used to perform a weighted summation operation to obtain the total loss of the model; with minimizing the total loss as the optimization objective, the total loss deviation is propagated layer by layer using the backpropagation algorithm, and the loss gradient corresponding to the parameters of each layer of the network is solved layer by layer; this application prioritizes the use of the L-BFGS optimizer to iteratively update the network weights and bias parameters. During the training process, the Adam optimizer can be used to complete the early fast parameter convergence. After the parameter update is completed, data is input again for iterative training; the above training process is repeated until the total loss of the model tends to converge stably, or the number of training iterations reaches the preset maximum number of iterations, and then the network training is terminated, completing the physical neural network training.

[0207] Step Six: Generate a Comprehensive Reliability Evaluation Report: Integrating equipment-level reliability parameters, system-level reliability indicators, and photovoltaic module degradation prediction results, the report analyzes the system from three dimensions: equipment health status, system operational reliability, and module degradation trends. This generates a comprehensive reliability evaluation report for the large-scale photovoltaic medium-voltage DC power generation system, identifying potentially faulty equipment, system reliability bottlenecks, and module degradation early warning times, providing specific suggestions for power plant operation and maintenance optimization.

[0208] It should be explained that the one-year operating data of the power station's combiner box, DC-DC converter, and inverter was collected: the operating condition correction coefficient was determined based on the on-site temperature, humidity, and irradiance data. =1.0, for the original failure rate Corrections were made. The combiner box had a total operating time of 8760 hours, with 2 failures and repair times of 5 hours and 4 hours respectively; the DC-DC converter had a total operating time of 8760 hours, with 1 failure and a repair time of 3 hours; and the inverter had a total operating time of 8760 hours and 0 failures.

[0209] According to the calculation formula in step two of this invention, the reliability parameters of each device are calculated:

[0210] Combiner box: =1.0×2 / 8760≈2.28×10-4 / h; = ≈0.81; =(8760-5-4) / 2=4375.5h; =(5+4) / 2=4.5h.

[0211] DC-DC converter: =1.0×1 / 8760≈1.14×10-4 / h; = ≈0.90; =8760-3=8757h; =3h.

[0212] Inverter: =1.0×0 / 8760=0; = =1; =8760h; =0h.

[0213] The power station topology is as follows: 20 photovoltaic modules are connected in series to form 1 string to 60 strings are connected in parallel to form 1 array to 1 array + 1 combiner box + 1 DC-DC converter are connected in series to form 1 branch to 30 branches and connected in parallel to 10 inverters for grid connection.

[0214] According to the series-parallel hierarchical calculation method in step three of this invention:

[0215] String reliability (20 components in series): =[ Components ] 20 .

[0216] Array reliability (60 sets of series and parallel): =[ Components ] 20 .

[0217] Branch reliability (array + combiner box + DC-DC series): = × junction box × DC-DC .

[0218] Single inverter reliability (30 branches in parallel): =1-[1- ] 30 .

[0219] Overall system reliability (10 inverters in parallel): system =1-[1- × junction box ] 10 .

[0220] Substituting the equipment reliability parameters, the overall reliability of the power station system is calculated. system ≈0.95, System Availability =8700 / 8760≈0.993, probability of load failure =0.002 indicates that the overall system reliability is high and the risk of load failure is low.

[0221] Collect 6 months of operating data (time) of photovoltaic modules from this power station Ambient temperature ,humidity Output power), verified by the three-step method in step four of this invention, determines that the component degradation process conforms to the Wiener process.

[0222] Construct a PINN-Wiener fusion model, setting the model as an input layer (3 neurons) - hidden layers (2 layers, each with 64 neurons) - output layer (1 neuron), with loss term weight coefficients... =1, =0.5, =0.5, =0.2, using the Adam optimizer, learning rate 0.001, and 1000 iterations.

[0223] After training, input environmental data for the next 3 months (temperature 25±5℃, humidity 60±10%). The model outputs a predicted component power degradation value of 5%, indicating that the current reliability of the component is good. It is estimated that the degradation will reach 10% after 12 months, and a component maintenance plan needs to be developed in advance.

[0224] Based on the evaluation results of the aforementioned equipment, systems, and components, a reliability evaluation report for the 100MW photovoltaic medium-voltage DC power generation system was generated. The core conclusions are as follows:

[0225] Equipment health status: Inverters are fault-free and have optimal reliability; DC-DC converters have good reliability; combiner boxes have experienced a high number of failures and are therefore key monitoring equipment.

[0226] System reliability: The overall system reliability is 0.95, availability is 0.993, the probability of load failure is low, and the overall operation is stable.

[0227] Component degradation trend: The current degradation rate is 5%, and the reliability is good. It is expected that the degradation rate will reach 10% in 12 months. It is recommended to carry out component cleaning and testing in advance.

[0228] In addition, this application also provides an electronic device. For example... Figure 3 The diagram illustrates the hardware operating environment of an electronic device, which may include: a processor (e.g., CPU), memory, a user interface, a network interface, and a communication bus. The communication bus is used to enable communication between components. The user interface may include a display screen and an input unit such as a keyboard; optionally, the user interface may also include a standard wired interface or a wireless interface. The network interface may optionally include a standard wired interface or a wireless interface. The memory may be high-speed RAM or stable non-volatile memory, such as disk storage. Alternatively, the memory may be a storage device independent of the aforementioned processor.

[0229] Those skilled in the art will understand that Figure 3 The electronic devices shown do not constitute a limitation on electronic devices and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0230] like Figure 3 As shown, a memory, as a type of computer storage medium, may include an operating system, a network communication module, a user interface module, and device management programs. The operating system is a program that manages and controls the hardware and software resources of electronic devices, supporting the operation of electronic devices and other software or programs. Figure 3 In the electronic device shown, the user interface is mainly used to connect to the terminal and communicate with the terminal, such as receiving user signaling data sent by the terminal; the network interface is mainly used to communicate with the backend server; the processor can be used to call the program stored in the memory and execute the steps of the method or system described above.

[0231] Furthermore, embodiments of this application also propose a computer-readable storage medium storing a device management program, which, when executed by a processor, implements the steps of the method or system described above.

[0232] The specific implementation of the computer-readable storage medium in this application is basically the same as the embodiments of the methods or systems described above, and will not be repeated here. Furthermore, to achieve the above objectives, this application also provides a computer program product, comprising: a computer program, which, when executed by a processor, implements the steps of the methods or systems described above.

[0233] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A reliability evaluation method for a large-scale photovoltaic medium-voltage DC power generation system, characterized in that, include: The system acquires operational data from multiple core devices in a large-scale photovoltaic medium-voltage DC power generation system, including photovoltaic modules, combiner boxes, DC-DC converters, and inverters. Based on operational data, device-level reliability parameters for each core device are calculated; these device-level reliability parameters include failure rate, reliability, mean time between failures (MTBF), and mean time to repair (MTBF). Based on the pre-defined topology of the large-scale photovoltaic medium-voltage DC power generation system, and combined with the reliability calculation formulas for series and parallel systems, the overall reliability of the large-scale photovoltaic medium-voltage DC power generation system is derived in layers; and the availability and failure probability of the large-scale photovoltaic medium-voltage DC power generation system are calculated to achieve a system-level macro-reliability evaluation of the large-scale photovoltaic medium-voltage DC power generation system. Using photovoltaic module output power as the core degradation indicator and photovoltaic module conversion efficiency and electrical parameters as auxiliary degradation indicators, the determination of whether the photovoltaic module power degradation process conforms to the Brownian motion process is achieved through degradation increment independence verification, degradation increment normality verification, and degradation drift coefficient stability verification. Based on the judgment results of the power degradation process of photovoltaic modules, a physical neural network model integrating the Brownian motion process is constructed to output the predicted value of the power degradation of photovoltaic modules. By integrating equipment-level reliability parameters, system-level macro-reliability evaluations, and predicted power degradation values ​​of photovoltaic modules, a comprehensive reliability evaluation report for large-scale photovoltaic medium-voltage DC power generation systems is generated.

2. The reliability evaluation method for a large-scale photovoltaic medium-voltage DC power generation system according to claim 1, characterized in that, The calculation of device-level reliability parameters for each core device based on operational data includes: Based on the time of equipment failure, the time of equipment failure repair, and the total operating time of the equipment, the number of failures of each core device is counted, and the original failure rate of each core device is calculated. Using pre-acquired ambient temperature, humidity, and solar irradiance, the operating condition deviation of each core device is calculated, and the operating condition correction coefficient is determined based on the preset correction coefficient weights. The original failure rate is corrected by the operating condition correction factor to obtain the actual failure rate; in order to eliminate the evaluation distortion caused by short-term start-stop and operating condition fluctuations. Based on the actual failure rate, the reliability of each core device at any operating moment is calculated; Based on the time of equipment failure and the time of equipment failure repair, calculate the average fault-free operating time and average repair time of each core device. Summarize all calculation results to complete the calculation of equipment-level reliability parameters for each core device.

3. The reliability evaluation method for a large-scale photovoltaic medium-voltage DC power generation system according to claim 1, characterized in that, The process involves, based on a pre-defined topology of a large-scale photovoltaic medium-voltage DC power generation system and combining reliability calculation formulas for series and parallel systems, deriving the overall reliability of the system in layers; and calculating the availability and failure probability of the system to achieve a system-level macro-reliability evaluation of the large-scale photovoltaic medium-voltage DC power generation system, including: Based on the topology of large-scale photovoltaic medium-voltage DC power generation system, photovoltaic modules are connected in series to form strings, strings are connected in parallel to form arrays, arrays are matched with combiner boxes and DC-DC converters in series, multiple series branches are connected in parallel to inverters, and multiple inverters are connected in parallel to the grid, the reliability of each level of large-scale photovoltaic medium-voltage DC power generation system is calculated in layers and level by level using the series and parallel reliability calculation formulas. By combining the reliability calculation results of each level, the overall reliability of the entire large-scale photovoltaic medium-voltage DC power generation system is derived by summarizing and deducing layer by layer. Based on the actual operating time and downtime of the large-scale photovoltaic medium-voltage DC power generation system, the availability of the large-scale photovoltaic medium-voltage DC power generation system is calculated to characterize its ability to supply power normally. The available capacity, load demand, and probability of occurrence of each state of a large-scale photovoltaic medium-voltage DC power generation system under different operating conditions are statistically analyzed. The load failure probability of the large-scale photovoltaic medium-voltage DC power generation system is calculated, and the risk of power shortage of the large-scale photovoltaic medium-voltage DC power generation system is assessed. The overall reliability, availability, and failure probability of a large-scale photovoltaic medium-voltage DC power generation system are integrated to complete the system-level macro-reliability evaluation of the large-scale photovoltaic medium-voltage DC power generation system.

4. The reliability evaluation method for a large-scale photovoltaic medium-voltage DC power generation system according to claim 1, characterized in that, The method of determining whether the photovoltaic module power degradation process conforms to Brownian motion includes using photovoltaic module output power as the core degradation indicator and photovoltaic module conversion efficiency and electrical parameters as auxiliary degradation indicators, respectively, through verification of the independence of degradation increment, verification of the normality of degradation increment, and verification of the stability of degradation drift coefficient. The output power and conversion efficiency of photovoltaic modules and electrical parameters are respectively used as core degradation indicators and auxiliary degradation indicators, and the operating time, ambient temperature, ambient humidity and output power time series data of photovoltaic modules are extracted. The output power time series data is processed at equal time intervals to generate a power time series sequence. The power difference between adjacent time points is calculated to obtain the power degradation increment sequence of the photovoltaic module. The autocorrelation coefficient of the power degradation increment sequence of the photovoltaic module is calculated one step behind. The independence of the degradation increment is verified based on the autocorrelation coefficient result. The photovoltaic module power degradation increment sequence is sorted, and empirical cumulative distribution function and standard normal theoretical cumulative distribution function with the same mean and variance are constructed respectively. Kolmogorov-Smirnov test is used to calculate the maximum deviation between empirical cumulative distribution function and standard normal theoretical cumulative distribution function. Combined with the preset critical value, the normality of degradation increment is verified. A fixed window length is set, and the photovoltaic module power degradation increment sequence is segmented and the drift degree is calculated segment by segment using a sliding window method. The stability of the degradation drift coefficient is verified based on the fluctuation of the drift degree. Based on the combined verification results of the independence of degradation increment, the normality of degradation increment, and the stability of degradation drift coefficient, it is determined whether the power degradation process of photovoltaic modules conforms to the Brownian motion process. If all verifications are qualified, the process proceeds to the next step. If any verification fails, supplementary data is collected and all verifications are performed again.

5. The reliability evaluation method for a large-scale photovoltaic medium-voltage DC power generation system according to claim 1, characterized in that, The process of constructing a physical neural network model that integrates the Brownian motion process based on the judgment results of the photovoltaic module power degradation process, training the physical neural network model based on a preset multi-constraint total loss function, and outputting the predicted value of photovoltaic module power degradation includes: The inputs and outputs of the physical neural network model that integrates the Brownian motion process are determined; the inputs are time, ambient temperature, and ambient humidity, and the output is the power degradation of the photovoltaic module. Based on the physical laws of photovoltaic module degradation, the drift coefficient and diffusion coefficient are estimated by using a sub-network fitting method or a preset function form fitting method. The weight and bias parameters of the physical neural network are randomly initialized to set the initial parameter state of the physical neural network. The photovoltaic module power degradation observation data and the physical constraint sampling points of the degradation process are input into the physical neural network. After unifying the data range, forward propagation calculation is performed to obtain the power degradation prediction result. Based on the automatic differentiation technique preset by the physical neural network, the first and second derivatives of the power degradation prediction results are obtained and used for the calculation of various loss terms. The data constraint loss, drift coefficient constraint loss, diffusion coefficient constraint loss, and stochastic differential equation residual term loss are calculated separately, and the total loss of the physical neural network model is obtained by weighting them according to the preset weight coefficients. With minimizing the total loss as the optimization objective, the loss gradient is solved through backpropagation to expand the search range of the parameter space and select high-quality parameter combinations. In conjunction with the optimizer, the weights and bias parameters of the physical neural network are iteratively updated. Repeatedly execute the data input, forward calculation, loss calculation and parameter update process until the total loss tends to converge or the preset maximum number of iterations is reached, and then end the training of the physical neural network model. The time-series data and environmental data are input into the trained physical neural network model, which outputs a predicted value of the power degradation of the photovoltaic module to complete the reliability evaluation of the photovoltaic module.

6. The reliability evaluation method for a large-scale photovoltaic medium-voltage DC power generation system according to claim 5, characterized in that, The formula for calculating the data constraint loss is as follows: ； The formula for calculating the drift coefficient constraint loss is as follows: ； The formula for calculating the diffusion coefficient constraint loss is as follows: ； The formula for calculating the loss of the residual term in the stochastic differential equation is as follows: ； In the formula, Indicates data constraint loss; This represents the drift coefficient constraint loss; This represents the diffusion coefficient constraint loss; This represents the loss of the residual term in a stochastic differential equation; This indicates the amount of degradation predicted by the physical neural network model; Indicates the measured degradation amount; N represents the number of data samples; i represents the i-th data sample; Indicates the theoretical drift coefficient; Represents the theoretical diffusion coefficient; E represents the mathematical expectation. Indicates variance.

7. The reliability evaluation method for a large-scale photovoltaic medium-voltage DC power generation system according to claim 5, characterized in that, The formula for calculating the total loss of the physical neural network model based on preset weight coefficients is as follows: ； In the formula, L represents the total loss function; Indicates data constraint loss; This represents the drift coefficient constraint loss; This represents the diffusion coefficient constraint loss; This represents the loss of the residual term in a stochastic differential equation; Preset weighting coefficients representing data constraint loss; The preset weighting coefficient represents the drift coefficient constraint loss; The preset weighting coefficient represents the loss constrained by the diffusion coefficient; The preset weighting coefficients represent the loss of the residual term in a stochastic differential equation.

8. The reliability evaluation method for a large-scale photovoltaic medium-voltage DC power generation system according to claim 5, characterized in that, The process of expanding the search range of the parameter space and filtering high-quality parameter combinations includes: Construct a parameter candidate sample set, pre-set several initial parameter samples, and preset a judgment threshold for each set of parameter samples; Multiple sets of weight parameters and bias parameters are randomly selected from the candidate parameter sample set and traversed and retrieved. The state information of the parameter samples is extracted, and the corresponding evaluation index is calculated and accumulated to obtain the comprehensive evaluation index of each set of parameter samples. The comprehensive evaluation index of each group of parameter samples is compared with the preset judgment threshold. If the comprehensive evaluation index does not reach the preset judgment threshold, parameter samples are re-extracted and the search continues. Otherwise, the parameter sample type is distinguished by combining the parameter sample status information. After completing the full sampling analysis of a single group of parameters to be retrieved, calculate the comprehensive evaluation value of each group of parameters, and judge the quality of the parameters based on the preset abnormal threshold, remove inferior parameters, and retain high-quality combinations of weight parameters and bias parameters. The process of continuously extracting new parameter samples, repeating the retrieval, index calculation, and screening process, continues until the parameter candidate sample set is traversed, at which point the parameter space search and parameter screening end.

9. The reliability evaluation method for a large-scale photovoltaic medium-voltage DC power generation system according to claim 8, characterized in that, The process involves randomly selecting multiple sets of weight parameters and bias parameters from the candidate parameter sample set, traversing and retrieving them, extracting the state information of the parameter samples, calculating and accumulating the corresponding evaluation indicators, and obtaining the comprehensive evaluation indicators for each set of parameter samples, including: Multiple sets of weight parameters and bias parameters are extracted from the parameter candidate sample set using a random sampling method to form a parameter sample set to be traversed and retrieved. For each group of weight parameters and bias parameters sample extracted, feature reduction is performed. Redundant or highly correlated dimensions are removed by analyzing the structural characteristics of the parameter samples, and the state information of the group of parameter samples is extracted from the reduced parameter features. Based on the extracted state information, multiple individual evaluation indicators are calculated for each group of parameter samples; the individual evaluation indicators are used to characterize the performance of the group of parameter samples under different evaluation dimensions. Iterate through all extracted parameter sample groups, repeatedly perform the operations of extracting state information and calculating individual evaluation indicators, and accumulate multiple individual evaluation indicators for each parameter sample group to obtain the comprehensive evaluation indicator corresponding to each parameter sample group.

10. The reliability evaluation method for a large-scale photovoltaic medium-voltage DC power generation system according to claim 9, characterized in that, The step of performing feature reduction processing on each group of extracted weight parameters and bias parameters samples, eliminating redundant or highly correlated dimensions by analyzing the structural characteristics of the parameter samples, and extracting the state information of the group of parameter samples from the reduced parameter features includes: For each extracted set of weight parameters and bias parameters, the correlation between the values ​​of each parameter sample in the set of weight parameters and bias parameters is calculated to obtain the correlation analysis results of the set of weight parameters and bias parameters. The correlation analysis results are used to characterize the degree of association between the values ​​of different parameter samples in the set of weight parameters and bias parameters. Based on the correlation analysis results, parameter sample pairs in the set of weight parameters and bias parameters that have a correlation degree exceeding a preset threshold are identified, and one parameter sample in each pair of highly correlated parameter samples is marked as a redundant parameter, forming a list of redundant parameter identifiers corresponding to the set of weight parameters and bias parameters. From the original parameter sample set of the set of weight parameters and bias parameters, remove all redundant parameter samples listed in the redundant parameter identifier list, and retain the remaining parameter samples to obtain the parameter sample feature set after reduction of the set of weight parameters and bias parameters. The values ​​of each parameter sample retained in the reduced parameter sample feature set are concatenated sequentially according to the order of their values ​​in the original parameter sample set to form a continuous feature sequence. This feature sequence is then used as the state information of the set of weight parameters and bias parameter samples.

Images