A highland photovoltaic converter monitoring method and system based on physical information fusion

Abstract

The invention provides a kind of highland photovoltaic converter monitoring method and system based on physical information fusion, wherein the method comprises: modal feature extraction, interaction and multi-time sequence feature extraction and channel compression are carried out on multidimensional key data, and data value is excavated;Physical constraint fusion layer integrates power conservation, thermodynamic cycle and insulation degradation and other physical constraints into model training, ensures that modeling conforms to physical reality, improves monitoring reliability and accuracy;High-cold adaptability decoder calculates health index based on cumulative damage and thermal resistance change rate, warns potential failure, and optimizes converter performance to adapt to highland environment;Label-free training mechanism improves the problem of scarce labeled data, improves model generalization and robustness;Lightweight model design and efficient inference engine integration ensure that the system runs efficiently and stably on embedded platform.The invention significantly improves monitoring accuracy and fault warning timeliness, and provides a comprehensive, efficient and intelligent solution for highland photovoltaic converter reliability monitoring.

Description

Technical Field

[0001] This invention relates to the field of reliability monitoring technology for high-altitude photovoltaic converters, and in particular to a monitoring method and system for high-altitude photovoltaic converters based on physical information fusion. Background Technology

[0002] With the increasing global demand for clean energy, photovoltaic (PV) power generation projects in high-altitude regions have experienced rapid development due to their abundant solar energy resources. In these projects, high-altitude PV converters, as key equipment connecting PV power generation systems to the power grid, primarily function to convert DC power generated by PV cells into AC power and feed it into the grid, while ensuring power quality and stability. However, the unique characteristics of the high-altitude environment present numerous challenges to the reliable operation of energy storage converters. High-altitude regions typically exhibit unique environmental conditions such as low air pressure, low temperatures, and high altitudes. Low air pressure leads to decreased insulation performance, increasing the risk of partial discharge and thus affecting the insulation life of the converter; low temperatures worsen the converter's heat dissipation performance, causing increased junction temperatures of power devices and accelerating device aging; large diurnal temperature variations at high altitudes also adversely affect the performance and reliability of the converter. Furthermore, the relatively complex grid conditions and frequent grid disturbances in high-altitude regions further increase the uncertainty of converter operation. Traditional monitoring methods for high-altitude PV converters mainly rely on data-driven models. While these models can handle large amounts of monitoring data, they are prone to learning patterns that do not conform to physical laws when lacking physical constraints, resulting in low reliability of monitoring results. Moreover, these methods typically require a large amount of labeled data for training. However, in high-altitude areas, labeled data is often scarce due to the harsh environment and difficulties in data collection, limiting the model's generalization ability and accuracy. Furthermore, existing monitoring systems lack real-time performance and stability on embedded platforms, making it difficult to meet the stringent requirements of real-time monitoring of high-altitude photovoltaic converters. More importantly, traditional methods lack compensation mechanisms tailored to the unique characteristics of the high-altitude environment, failing to accurately assess the converter's operating status and thus hindering the timely detection of performance degradation and potential faults caused by environmental factors. Summary of the Invention

[0003] To address the aforementioned issues, this invention discloses a monitoring method and system for plateau photovoltaic converters that can adapt to the special environment of plateaus, improve monitoring accuracy and reliability, reduce dependence on labeled data, and ensure real-time performance and embedded compatibility. This is of great significance for ensuring the stable operation of plateau photovoltaic power generation projects, reducing operation and maintenance costs, and improving energy utilization efficiency.

[0004] The specific plan is as follows:

[0005] A monitoring method for high-altitude photovoltaic converters based on physical information fusion includes the following steps:

[0006] S1. Multi-dimensional feature extraction: Modal feature extraction, interaction, multi-temporal feature extraction, and channel compression are performed on the collected multi-dimensional key data to mine data value;

[0007] S2. Physical constraint fusion: Physical constraints such as power conservation, thermodynamic cycles, and insulation degradation are incorporated into model training to ensure that the modeling conforms to physical reality and improve the credibility and accuracy of monitoring.

[0008] S3, High Altitude Adaptability Decoding: Calculates the health index based on cumulative damage and thermal resistance change rate, provides early warning of potential faults, and optimizes converter performance to adapt to the high-altitude environment;

[0009] S4. Improved Unlabeled Training Mechanism: By reconstructing the loss function and using plateau data augmentation strategies, the dependence on labeled data is reduced, and the model's generalization ability is improved.

[0010] Furthermore, step S1, multivariate feature extraction, includes:

[0011] (1) Data preprocessing

[0012] When monitoring photovoltaic converters in high-altitude areas, the first step is to collect key data, including the three-phase voltage V. abc Three-phase current I abc IGBT junction temperature T j Air pressure P, heat dissipation temperature difference ΔT, and switching frequency f sw It has a total of 10 dimensions. These data form the input matrix. Where T is the time series length, representing the number of data points in each sampling period; This represents the real number field, used to define the range of values ​​for the data. This process aims to organize the raw data into a uniform format, providing a foundation for subsequent feature extraction, ensuring data quality and consistency, and thus improving the accuracy and reliability of the entire monitoring method.

[0013] (2) Modal feature extraction

[0014] Each modality undergoes a one-dimensional convolution operation independently:

[0015] F i =Conv1D(X i W i ,b i ), i = 1, 2, ..., 10

[0016] In the formula F i X is the feature map of the i-th modality, i.e., the extracted feature vector, used to represent the key characteristics of this modality; Conv1D is a one-dimensional convolution operation, which, for time-series data, can extract local features and capture short-term correlations and patterns in the data; iIt is the input data of the i-th mode, such as V in three-phase voltage. a W represents the representation of the original signal in the corresponding dimension. i These are the convolution kernel and weight matrix, used to perform convolution operations on the input data, and through training, learn the representation of different features; b i It is a bias term used to adjust the offset of the convolution result, increasing the flexibility and fitting ability of the model.

[0017] The kernel size is set to 3, and the number of output channels is 64. This process uses convolution operations to automatically extract local features from each modality of data, mapping the original high-dimensional data onto multiple feature channels. Each channel corresponds to a specific feature expression, thereby reducing the data dimensionality while retaining key information. This lays the foundation for subsequent feature interaction and fusion, enabling the model to process and analyze data more efficiently.

[0018] (3) Intermodal feature interaction

[0019] Extracting only single-mode features is insufficient to fully characterize the complex operating state of high-altitude photovoltaic converters; therefore, it is necessary to fuse modal features through a channel attention mechanism.

[0020] F fuse =F⊙α

[0021] In the formula F fuse The first part is the fused feature map, which integrates information from various modes and can more comprehensively reflect the operating status of the converter; the second part is the element-wise multiplication operation, used to apply attention weights to the feature map to achieve weighting of different features; the third part is the original feature map, which contains features from each mode and is the set of features extracted independently from each mode. The fourth part is the attention weight, used to emphasize important features and suppress unimportant features, guiding the model to focus on key information.

[0022] The attention weight α is dynamically calculated based on global information, integrating features from different modes. This allows the model to simultaneously consider the combined impact of electrical characteristics (such as voltage and current), thermal characteristics (such as IGBT junction temperature and heat dissipation temperature difference), and environmental factors (such as air pressure) on the converter's operating state, thus enhancing the expressive power of features. Its function is to enable the model to automatically allocate attention to features of different modes, highlighting features that have a greater impact on the converter's state. For example, the importance of current features will significantly increase under overload conditions, thereby enhancing the model's sensitivity and ability to identify abnormal converter states, providing a more discriminative feature foundation for subsequent feature extraction and analysis.

[0023] (4) Multi-scale temporal feature extraction

[0024] To capture features at different time scales, dilated convolutions with dilation rates of 1, 2, 4, and 8 were used for extraction.

[0025]

[0026] In the formula This is the feature map at the d-th dilation rate, used to capture features at different time scales, such as short-term fluctuations and long-term trends. DilatedConv is a dilated convolution operation that expands the receptive field by inserting holes into the convolution kernel, enabling the kernel to extract features over a wider time range. d is the dilation rate, which controls the size of the holes and thus affects the sampling interval of the convolution kernel. The larger the dilation rate, the wider the receptive field, and the more time-scale features can be captured.

[0027] Finally, the feature maps at different inflation rates are concatenated to obtain a multi-scale feature map:

[0028]

[0029] In the formula F multi-scale It is a comprehensive feature map that integrates features from different time scales; Concat is a concatenation operation used to merge feature maps of different scales into a complete feature map.

[0030] This process expands the model's ability to perceive the time dimension, enabling it to simultaneously capture both instantaneous changes in the converter's operating state (such as instantaneous voltage and current fluctuations caused by short-circuit faults) and evolutionary trends over longer time scales (such as the gradual rise in IGBT junction temperature). This is crucial for accurately assessing the converter's health status, as different faults and aging phenomena may manifest at different time scales. Its role is to enhance the model's ability to model the dynamic operation of the converter, allowing the model to analyze operating data more comprehensively and meticulously, thereby more accurately identifying fault modes and aging signs, and providing richer and more comprehensive feature information for subsequent health status assessments and fault early warnings.

[0031] (5) Feature channel compression

[0032] After completing multi-scale feature extraction, to reduce feature dimensionality, computational complexity, and retain key features, 1×1 convolution is used for channel compression:

[0033] F final =Conv1D(F multi-scale ,1,64)

[0034] In the formula F final The final feature map has fewer channels but retains key features, which can reduce the computational burden while retaining information that is crucial for converter state assessment; 1 is the size of the convolution kernel, which is 1×1, so that the convolution only operates on the channel dimension, realizing feature fusion and compression between channels; 64 is the number of output channels, representing the number of channels in the compressed feature map.

[0035] This process removes redundant features, retaining the most valuable information for converter condition monitoring, making subsequent processing more efficient. Its purpose is to reduce the model's computational resource consumption, improve its operational efficiency, avoid overfitting, enhance its generalization ability, and ensure stable and rapid operation of the model in the actual monitoring scenario of high-altitude photovoltaic converters. This provides a concise and effective feature representation for subsequent health status assessment and fault early warning.

[0036] Furthermore, step S2, physical constraint fusion, includes:

[0037] (1) Power conservation constraint

[0038] In high-altitude environments, the power conversion of high-altitude photovoltaic converters must adhere to the power conservation principle, meaning that input power and output power should remain balanced, but the impact of efficiency degradation at high altitudes must be considered. Therefore, a power conservation constraint loss function is constructed:

[0039]

[0040] In the formula It is the power conservation constraint loss, which measures the error in power balance and is used to constrain the power conversion process of the converter during model training; P in Input power, i.e., the power that the converter obtains from the power grid, is the energy source that drives the converter's operation; P out It is the output power, that is, the power provided by the converter to the load, which is the actual output energy of the converter; η(P,T) j The efficiency is determined by considering the gas pressure P and the IGBT junction temperature T. j The impact on efficiency reflects the combined effect of high-altitude environmental factors on the energy conversion efficiency of the converter; V dc It is the DC-side voltage, the voltage value of the DC bus of the converter, which is the intermediate energy transmission link connecting the AC power grid and the load; I dc This is the DC-side current, the current value of the converter's DC bus, which, together with the DC voltage, determines the power transmission on the DC side; k h f is the switching harmonic loss coefficient of the h-th order, representing the loss weight at different harmonic orders, used to calculate the energy loss generated by the switching device during high-frequency switching; sw It refers to the switching frequency, the switching frequency of the power devices in the converter. The higher the switching frequency, the greater the harmonic loss.

[0041] The formula for calculating efficiency is:

[0042]

[0043] In the formula, η0 is the reference efficiency, the efficiency of the converter under standard conditions, representing the energy conversion efficiency of the converter in an ideal environment. P is the air pressure, the air pressure value in a high-altitude environment. Lower air pressure leads to decreased heat dissipation efficiency and changes in insulation performance, thus affecting the efficiency of the converter. T j It refers to the junction temperature of the IGBT, a power device. An increase in junction temperature will increase the on-resistance and switching losses of the device, thereby reducing the efficiency of the converter.

[0044] This process integrates the principle of power conservation from physical laws into the training of the monitoring model, ensuring that the model's depiction of the converter's power conversion process conforms to physical reality. This improves the accuracy of the model's assessment of the converter's operating status and avoids unreasonable predictions that violate physical laws. Simultaneously, physical constraints enhance the model's efficiency in utilizing data, especially when labeled data is limited. By guiding the model's learning through physical constraints, the model's generalization ability and adaptability to high-altitude environments are improved, enabling the model to more reliably assess the converter's health status and performance.

[0045] (2) Thermodynamic Cycle Constraints

[0046] Junction temperature variation in power devices is one of the key factors affecting converter reliability. To accurately assess the cumulative impact of IGBT junction temperature variation on device lifetime, an IGBT junction temperature cumulative damage model under thermodynamic cycling constraints is adopted:

[0047]

[0048] In the formula ΔD acc Cumulative damage represents the cumulative impact of IGBT junction temperature variations on device lifespan, used to assess the long-term reliability of converters. N is the number of sampling points, the number of samples taken within a cycle or time period, reflecting the temporal resolution of the data; ΔT j (k) represents the junction temperature change at the k-th sampling point, the temperature difference between adjacent sampling points, and is a key indicator for measuring junction temperature fluctuations; m is the junction temperature change exponent, describing the degree of influence of junction temperature changes on damage, usually determined through experimental or theoretical analysis; A is the normalization factor, used to normalize damage values ​​to a reasonable range for easier model processing and comparison; Q is the activation energy, a constant related to materials and processes, representing the thermal energy threshold that causes damage and reflecting the thermal stability of the material; R is the gas constant, a universal constant, a constant related to the amount of substance, used to convert temperature into units of thermodynamic energy; T j (k) The junction temperature of the IGBT at the kth sampling point is the junction temperature of the power device at that moment, which is a direct factor affecting damage accumulation.

[0049] The constraints are:

[0050] D acc ≥δ·σ(ΔTj ·t on )

[0051] In the formula D acc δ is the cumulative damage threshold, which is the maximum allowable cumulative damage value. Exceeding this threshold may lead to a significant degradation in device performance or failure. δ is the safety factor, a coefficient greater than 1, used to ensure that the converter operates within a safe range, taking into account the uncertainty of model predictions and risk factors in actual operation. σ is the standard deviation, which measures the dispersion of junction temperature changes and reflects the statistical characteristics of junction temperature fluctuations. on It is the conduction time, the time during which a power device is in the conduction state. During the conduction period, the junction temperature change has a direct impact on damage accumulation.

[0052] σ(ΔT j ·t on The term "cumulative damage" measures the dispersion of the product of junction temperature change and conduction time, reflecting the statistical characteristics of junction temperature change under different operating conditions. Multiplying this by a safety factor δ yields a conservative cumulative damage reference value to ensure device safety during operation. Separate constraints are established to dynamically adjust and ensure the cumulative damage threshold D. acc The constraint is large enough to cover the cumulative damage that may be encountered in actual operation. This constraint has several advantages: First, it dynamically adjusts the safety limit, adjusting the cumulative damage threshold based on junction temperature changes and conduction time during actual operation to ensure device safety under different operating conditions. Second, it improves the model's adaptability by introducing a safety factor and statistical indicators, enhancing the model's adaptability to different operating conditions and avoiding safety risks caused by uncertainties in model predictions. Third, it ensures reliability, guaranteeing that IGBT devices will not fail prematurely due to cumulative damage exceeding the threshold, thereby extending device lifespan and improving converter reliability. In summary, constraint D... acc ≥δ·σ(ΔT j ·t on This constraint is designed to ensure that the cumulative damage threshold is large enough to cover the cumulative damage in actual operation, thereby guaranteeing the safe operation of IGBT devices in high-altitude photovoltaic converters. This constraint helps to dynamically adjust the safety limits, improving the model's adaptability and reliability.

[0053] The above process establishes a quantitative relationship between IGBT junction temperature changes and cumulative damage based on thermodynamic principles, incorporating the physical degradation process of the device into the monitoring model. Its functions are twofold: first, to enable the model to dynamically assess the health status of the IGBT, promptly identifying potential damage accumulation caused by junction temperature fluctuations and providing a scientific basis for preventative maintenance; second, to ensure that the converter operates within a safe range through constraints, avoiding premature device failure due to excessive junction temperature changes, thereby extending the converter's lifespan, improving its reliability and stability in harsh high-altitude environments, and ensuring the long-term reliable operation of the energy storage system.

[0054] (3) Insulation degradation constraint

[0055] In high-altitude environments, reduced air pressure leads to decreased insulation performance, thus requiring constraints on the inverter's insulation state. Firstly, the partial discharge initiation voltage needs to be corrected:

[0056]

[0057] In the formula V PDIV This is the corrected partial discharge initiation voltage, a value that takes into account the effects of high-altitude environments, used to evaluate the insulation performance of the converter under actual operating conditions; V PDIV0 This is the partial discharge initiation voltage under standard conditions, a measurement taken in a laboratory or standard environment, used as the basis for correction; P is the air pressure, specifically the air pressure value in a high-altitude environment. Lower air pressure reduces the breakdown voltage of insulating materials, increasing the risk of partial discharge; c p is the air pressure correction factor, ranging from 0.8 to 1.2, used to correct the initiation voltage under different air pressures, reflecting the nonlinear effect of air pressure on the partial discharge initiation voltage; T is the ambient temperature, the temperature value in a high-altitude environment. Temperature changes affect the physical properties of materials, and thus affect the partial discharge initiation voltage.

[0058] The constraints are:

[0059] |V phase | peak <0.7V PDIV

[0060] In the formula |V phase | peak The peak phase voltage is the maximum amplitude of the output phase voltage of the converter, and it is a key indicator for measuring voltage stress; V PDIV It is the partial discharge initiation voltage, a value corrected for the high-altitude environment. It represents the maximum voltage that the insulating material can withstand under the current environmental conditions without partial discharge.

[0061] The above process modifies the partial discharge initiation voltage based on the characteristics of the plateau environment and establishes a constraint relationship between the peak phase voltage and the modified initiation voltage, thereby incorporating insulation degradation constraints into the monitoring model. Its functions are twofold: First, it ensures that the converter's operating voltage is always lower than the partial discharge initiation voltage of the insulation material under actual conditions, preventing faults caused by insulation breakdown and ensuring the safe operation of the converter. Second, by dynamically adjusting the insulation constraints to adapt to changes in the plateau environment, it improves the model's sensitivity and adaptability to environmental changes, enabling the monitoring system to provide timely warnings of insulation deterioration. This provides strong support for the insulation maintenance of plateau photovoltaic converters and ensures the long-term stable operation and safe and reliable power supply of energy storage systems in plateau regions.

[0062] Furthermore, step S3, high-altitude cold adaptation decoding, includes:

[0063] (1) Health Index (HI) Calculation

[0064] To comprehensively assess the overall health status of high-altitude photovoltaic converters, a Health Index (HI) is introduced as a key indicator, and its calculation formula is as follows:

[0065] HI=1-ω1D acc -ω2ΔR th

[0066] In the formula, HI is the health index, ranging from 0 to 1, used to measure the overall health status of the converter. A higher HI value indicates a better health status for the converter; conversely, a lower HI value indicates more damage or aging problems. ω1 and ω2 are weighting coefficients, representing the degree of influence of cumulative damage and thermal resistance changes on the health index, respectively, satisfying ω1 + ω2 = 1. These two weighting coefficients can be adjusted according to the actual operating characteristics and failure modes of the converter to reflect the relative importance of different factors in the health status assessment. acc This is cumulative damage, calculated using thermodynamic cycling constraints, which reflects the degree of damage accumulation caused by junction temperature changes during power device operation. ΔR th It is the rate of change of thermal resistance, which is the proportion of change of the thermal resistance of the heat dissipation system relative to its initial value, and is used to evaluate the performance degradation of the heat dissipation system.

[0067] The above process integrates key performance degradation indicators of the converter (i.e., cumulative damage to power devices and changes in the thermal resistance of the cooling system) into a unified health index. Its purpose is twofold: first, to provide a concise and intuitive indicator to characterize the overall health status of the converter, facilitating operators to quickly understand the equipment's condition and make maintenance decisions; second, by weighting and fusing different degradation indicators, it fully considers the multi-physics coupling characteristics of the converter, enabling the health index to more comprehensively reflect the actual operating state of the converter, avoiding the bias caused by single-indicator assessments, thereby improving the accuracy and reliability of health status assessment and providing a strong basis for subsequent fault warning and maintenance strategy formulation.

[0068] (2) Fault warning signs

[0069] ①Insulation risk

[0070] The calculation formula for the insulation risk of the converter is as follows:

[0071]

[0072] In the formula R ins R_ins is the insulation risk value, ranging from 0 to 1, used to measure the degree of risk in the converter's insulation condition. A higher R_ins value indicates a greater insulation risk, requiring more attention. σ is an activation function, such as the sigmoid function, used to map the linear combination result to a probability range of 0 to 1, facilitating interpretation and setting warning thresholds. W is a weight vector containing the weights of each feature, used to highlight features that have a greater impact on insulation risk. dV / dt max THD represents the maximum voltage change rate, the maximum rate of change of the voltage waveform per unit time. It reflects the steepness of the voltage change; a steep voltage change may exert greater stress on the insulating material. P represents air pressure, the air pressure value in high-altitude environments. Lower air pressure reduces insulation performance and increases the risk of insulation breakdown. V It is the total harmonic distortion rate of voltage, which measures the harmonic content of the voltage waveform. Harmonic distortion increases the possibility of partial discharge, thereby affecting the insulation life.

[0073] When R ins An early warning is triggered when the pressure is greater than 0.6 and P < 70 kPa. This condition is set based on the characteristics of the high-altitude environment and the assessment logic of insulation risk. That is, in a low-pressure environment, a higher insulation risk value indicates that the insulation condition of the converter is close to a dangerous level and timely measures need to be taken.

[0074] ② Heat dissipation failure

[0075] The calculation method for determining heat dissipation failure is as follows:

[0076]

[0077] In the formula R cool : Heat dissipation failure flag, a binary variable, with a value of 0 or 1. 1 indicates heat dissipation failure and is used to clearly indicate whether the heat dissipation system has malfunctioned. This indicator function outputs 1 when the condition within the parentheses is met, and 0 otherwise, enabling a binary judgment of heat dissipation failure. ΔT: Heat dissipation temperature difference, the difference between the radiator surface temperature and the ambient temperature, is a key indicator of radiator performance. A larger temperature difference indicates that the radiator is effectively dissipating heat. The rate of temperature change, or the rate at which the radiator temperature changes over time, reflects the speed at which the cooling system responds to temperature changes.

[0078] When R cool An early warning is triggered when ΔT = 1 and ΔT > 15℃. This condition takes into account both the judgment of heat dissipation failure and the situation of excessively high heat dissipation temperature difference. That is, when the heat dissipation system fails and the radiator temperature is too high, the converter may face the risk of overheating, and timely warning and measures need to be taken.

[0079] The above process establishes early warning indicators for insulation risk and heat dissipation failure based on health indices and key risk indicators, providing real-time assurance for the safe operation of the converter. Its functions are twofold: First, it promptly detects potential insulation and heat dissipation problems in the converter under high-altitude conditions, issuing early warning signals so that operators can respond quickly and take preventative maintenance measures to prevent further deterioration of the fault, reducing downtime and maintenance costs. Second, by setting clear early warning conditions and thresholds, it improves the accuracy and reliability of early warnings, reduces the probability of false alarms and missed alarms, ensures the safe and stable operation of the converter, guarantees the reliability and availability of the high-altitude photovoltaic system, thereby improving the overall energy storage system's operational efficiency and economic benefits, and enhancing its adaptability and survivability in high-altitude areas.

[0080] (3) Plateau compensation mechanism

[0081] To ensure the converter maintains good performance in high-altitude environments, air pressure compensation is applied to the switching frequency.

[0082]

[0083] In the formula The compensated switching frequency, which takes into account the influence of air pressure, is used to adjust the converter's operating parameters to adapt to the high-altitude environment. sw : Initial switching frequency, the initial switching frequency of the converter's power devices, the optimal operating frequency determined under standard conditions. P: Air pressure, the air pressure value in a high-altitude environment. Changes in air pressure directly affect heat dissipation and insulation performance, and thus affect the optimal switching frequency of the converter.

[0084] The above process dynamically adjusts the switching frequency of the converter based on changes in air pressure in the high-altitude environment to optimize its performance and reliability. Its purpose is twofold: first, by increasing the switching frequency, it compensates for the reduced heat dissipation efficiency caused by low air pressure, enhancing the converter's heat dissipation capacity in high-altitude environments, ensuring that power devices operate within safe temperature ranges, and reducing the risk of failure due to overheating; second, adjusting the switching frequency helps improve the converter's insulation performance, reducing the probability of partial discharge and improving the converter's insulation reliability in low-pressure environments. This, in turn, enhances the overall performance and adaptability of the converter, ensuring stable operation in high-altitude areas, extending its service life, reducing maintenance costs, and guaranteeing the efficient operation and long-term reliable power supply of the energy storage system in high-altitude environments.

[0085] Furthermore, improvements to the label-free training mechanism in step S4 include:

[0086] (1) Loss Function Reconstruction

[0087] The total loss function takes into account multiple constraints, and its calculation formula is as follows:

[0088]

[0089] In the formula This is the total loss function, which integrates loss terms from multiple constraints to guide the model training process, ensuring it simultaneously meets multiple physical constraints and performance requirements. λ1, λ2, λ3, and λ4 are weight coefficients, controlling the proportion of each loss term in the total loss. By adjusting these weights, the impact of different constraints on model training can be balanced, making it suitable for practical applications. This is the reconstructed three-phase voltage, the voltage value reconstructed by the model based on the extracted features, used to evaluate the model's ability to reproduce voltage characteristics. V abc 1 represents the original three-phase voltage, the actual measured voltage value, which serves as the reference standard for the reconstructed voltage. 2 is the L2 norm, which measures the error between the reconstructed voltage and the original voltage. It is used in the voltage reconstruction loss term to ensure that the model can accurately reproduce the voltage characteristics. It is a power conservation constraint loss that ensures the balance of input and output power, making the power conversion process of the model conform to physical laws. is the gradient of accumulated damage over time, measuring the rate of damage change. It is used to constrain the aging gradient loss term, ensuring the model accurately reflects the damage accumulation process. ||·||1 is the L1 norm, used to constrain the sparsity of the gradient, making the damage accumulation model simpler and avoiding overfitting. k is the pressure sensitivity coefficient, measuring the influence of air pressure on the perceived insulation degradation loss. It is used to adjust the weights of the perceived insulation degradation loss term, ensuring it has an appropriate impact under different air pressures. This is the predicted change in partial discharge initiation voltage. The model predicts the change in initiation voltage, which is used for the insulation degradation sensing loss term, enabling the model to accurately capture changes in insulation performance. F It is the Frobenius norm, used to measure the error of the matrix, used for the insulation degradation sensing loss term, and to evaluate the model's prediction accuracy for changes in the partial discharge initiation voltage.

[0090] The weights are set as follows:

[0091] λ1=1.0, λ2=0.8, λ3=0.5, λ4=0.3 (P<70kPa)

[0092] In high-altitude environments, when the air pressure is below 70 kPa, the weight of perceived insulation degradation loss increases to enhance attention to changes in insulation performance.

[0093] The above process, by reconstructing the loss function, integrates various physical constraints and performance requirements into the model's training process, enabling it to learn effectively even without a large amount of labeled data. Its advantages are twofold: first, it fully utilizes the physical information and inherent patterns in the data, reducing reliance on labeled data and lowering the workload and cost of data preparation, making it particularly suitable for scenarios where data acquisition is difficult, such as high-altitude environments; second, by balancing different loss terms, it ensures that the model is well-trained in multiple aspects (voltage reconstruction, power conservation, damage accumulation modeling, and insulation degradation perception), improving the model's overall performance and generalization ability. This allows it to more accurately assess the health status of the converter and predict potential faults, providing strong support for the reliable operation of high-altitude photovoltaic converters and enhancing the intelligence and practicality of the entire monitoring system.

[0094] (2) Plateau data augmentation strategy

[0095] ① Enhanced pressure-temperature coupling

[0096] To enhance the model's adaptability to changes in the plateau environment, a pressure-temperature coupling enhancement strategy is adopted, as shown in the formula:

[0097] P aug =P×np.random.uniform(0.85,1.15)

[0098]

[0099] In the formula P aug This is the enhanced air pressure value, generated by introducing random fluctuations on top of the original air pressure, used to simulate the natural changes in air pressure in plateau regions. `np.random.uniform(0.85, 1.15)` generates uniformly distributed random numbers between 0.85 and 1.15 to achieve random fluctuations in the air pressure value, thus expanding the model's ability to recognize air pressure changes. Tj_aug It is the enhanced IGBT junction temperature, a junction temperature value adjusted according to changes in gas pressure, reflecting the physical coupling relationship between gas pressure and junction temperature.

[0100] By simulating pressure and temperature variations in a high-altitude environment, more diverse training data is generated to expand the model's adaptability to environmental changes. Its benefits include: first, improving the model's robustness to uncertainties in the high-altitude environment, enabling it to maintain stable performance under conditions of significant pressure and temperature fluctuations, and avoiding a decrease in model prediction accuracy due to environmental changes; second, enhancing the model's understanding of the coupling relationship between pressure and temperature, allowing it to more accurately capture the combined impact of these two key environmental factors on the converter's operating status, thereby improving the model's reliability and effectiveness in practical applications and ensuring the stable operation and accurate monitoring of the monitoring system in the complex high-altitude environment.

[0101] ② Power grid disturbance injection

[0102] To simulate actual disturbances in the power grid, disturbances are injected into the three-phase voltage data, using the following formula:

[0103]

[0104] In the formula, np.random.rayleigh(scale=0.05, size=V) abc `.shape` generates random noise with a Rayleigh distribution to simulate disturbances in the power grid. The Rayleigh distribution is commonly used to describe random noise and interference signals, and can reflect the actual fluctuations in the power grid relatively well. `scale = 0.05` is the scaling parameter of the Rayleigh distribution, which controls the amplitude of the noise. An appropriate scaling parameter is set according to the characteristics of the actual power grid disturbances to generate reasonable disturbance noise. `size = V` abc The shape parameter ensures that the generated noise has the same shape as the original voltage data, guaranteeing that the disturbance can be correctly added to the three-phase voltage data.

[0105] The above process enhances the model's adaptability to actual operating conditions by adding simulated grid disturbances to the training data. Its benefits are twofold: first, it improves the model's resilience to grid disturbances, enabling it to accurately assess the converter's operating status under conditions of significant grid fluctuations, avoiding misjudgments and omissions; second, it enriches the diversity of training data, allowing the model to learn in more varied scenarios, thereby improving its generalization ability and reliability in practical applications. This ensures the stable operation and accurate monitoring of the monitoring system in complex grid environments, guaranteeing the safe and reliable power supply to the high-altitude photovoltaic converter.

[0106] A monitoring system for high-altitude photovoltaic converters based on physical information fusion includes: a data acquisition and preprocessing module for collecting and preprocessing key data; a multimodal temporal fusion network for extracting temporal features of multimodal data; a physical constraint fusion layer for integrating physical laws into the feature extraction process; a high-altitude adaptability decoder for calculating health indices and fault warning indicators, and performing high-altitude compensation; a labelless training module for optimizing model parameters and improving model generalization ability; and an embedded deployment optimization module for ensuring real-time monitoring and stable operation of the monitoring system on an embedded platform.

[0107] Furthermore, the multimodal temporal fusion network includes: a modal feature extraction unit: performing one-dimensional convolution operations on each modal data to extract local features; an intermodal feature interaction unit: fusing modal features through a channel attention mechanism; a multi-scale temporal feature extraction unit: capturing features at different time scales using dilated convolutions with different dilation rates; and a feature channel compression unit: using 1×1 convolutions to compress channels and remove redundant features.

[0108] Furthermore, the physical constraint fusion layer includes: a power conservation constraint unit: constructing a power conservation constraint loss function to ensure that the model modeling of the converter power conversion process conforms to physical reality; a thermodynamic cycle constraint unit: using an IGBT junction temperature cumulative damage model to constrain the evaluation of the cumulative impact of the model on device lifetime; and an insulation degradation constraint unit: correcting the partial discharge initiation voltage and establishing a constraint relationship between the phase voltage peak and the corrected initiation voltage.

[0109] Furthermore, the high-altitude adaptability decoder includes: a Health Index (HI) calculation unit: calculating the health index by comprehensively considering cumulative damage and thermal resistance change rate; a fault warning sign unit: establishing warning signs for insulation risk and heat dissipation failure respectively; and a high-altitude compensation mechanism unit: performing air pressure compensation on the switching frequency to optimize converter performance.

[0110] Furthermore, the label-free training module includes: a loss function reconstruction unit, which reconstructs the total loss function by comprehensively considering multiple constraints; and a plateau data augmentation strategy unit, which enhances the model's adaptability to changes in the plateau environment by employing pressure-temperature coupling enhancement and power grid disturbance injection strategies.

[0111] Furthermore, the embedded deployment optimization module includes:

[0112] Lightweight model design unit: Employs techniques such as grouped convolution, depthwise separable convolution, model pruning, and quantization to optimize model computation efficiency and resource consumption; High-efficiency inference engine integration unit: Integrates a high-efficiency inference engine to improve the inference speed and efficiency of the model on embedded platforms; Embedded platform adaptation unit: Performs adaptation and verification on STM32H7 series chips to ensure the effective application of the monitoring system in plateau energy storage converters.

[0113] The embedded deployment optimization module achieves real-time performance and embedded adaptation, as detailed below:

[0114] (1) Lightweight Model Design

[0115] To ensure the real-time performance of the monitoring method on the embedded platform, the following lightweight model design strategy is adopted:

[0116] ① Grouped Convolution Replaces Dilated Convolution: The computationally intensive dilated convolution is replaced with grouped convolution. Grouped convolution divides the input channels into multiple groups, each undergoing an independent convolution operation, and the results are then concatenated. This method significantly reduces computational cost and the number of model parameters while maintaining feature extraction capabilities. The aim is to reduce computational complexity, enabling the model to run quickly on resource-constrained embedded devices, while preserving as much of the multi-scale feature extraction capability as possible from the original dilated convolution. This ensures the model's accuracy in perceiving the converter's operating status remains unaffected, thereby improving model efficiency while maintaining monitoring effectiveness and meeting the real-time monitoring needs of high-altitude photovoltaic converters.

[0117] ② Application of Depthwise Separable Convolution: Depthwise separable convolution decomposes standard convolution into two steps: depthwise convolution and pointwise convolution. Depthwise convolution performs convolution operations on each input channel individually, while pointwise convolution fuses the outputs of the depthwise convolution across channels using 1×1 convolutions. This decomposition further reduces computational cost and model size while maintaining high feature extraction efficiency. The reason for using depthwise separable convolution is that it can reduce computational resource consumption while maintaining the model's ability to express features, enabling the model to run efficiently on embedded platforms, quickly process real-time acquired data, and promptly output the health status and fault warning information of the converter, ensuring the real-time performance and reliability of the monitoring system, and improving the operational safety and management efficiency of the entire energy storage system.

[0118] ③ Model Pruning and Quantization: The trained model is pruned to remove unimportant connection weights, reducing sparsity. Then, the model is quantized, converting 32-bit floating-point numbers (FP32) to 8-bit integers (INT8), reducing storage requirements and computational complexity. Dynamic range calibration is used during quantization, with special handling for key features such as high-amplitude currents to ensure the quantized model maintains accuracy in high-altitude environments. The purpose of model pruning and quantization is to further optimize the model's storage and computational performance, enabling it to run on embedded devices with limited storage space and computing power. Pruning removes redundant parameters, reducing computational operations; quantization lowers numerical accuracy requirements, speeding up computation and reducing storage usage. Simultaneously, dynamic range calibration ensures the accuracy of key features, guaranteeing the model's monitoring accuracy in harsh high-altitude environments. This allows the monitoring system to accurately assess the converter's health status while maintaining efficient operation, providing a reliable basis for equipment maintenance and management, and ensuring the stable operation and long-term reliable power supply of the high-altitude photovoltaic system.

[0119] (2) Integration of high-efficiency inference engine

[0120] To improve the inference speed and efficiency of the model on embedded platforms, a high-efficiency inference engine is integrated. Specific measures include:

[0121] ①TVM Compilation Optimization: The model is compiled and optimized using the TVM (TensorVirtual Machine) framework. TVM is an open-source machine learning compiler framework that can generate highly optimized executable code for different hardware platforms (such as embedded chips based on the ARM architecture). TVM fully leverages the parallel computing capabilities of the hardware through techniques such as loop optimization and memory access optimization, thereby improving model inference speed. The purpose of using TVM compilation optimization is to utilize its powerful compilation capabilities to generate efficient executable code tailored to the characteristics of embedded hardware platforms, maximizing the computational potential of the hardware, and thus significantly improving the model's inference speed. This meets the stringent time requirements for real-time monitoring of high-altitude photovoltaic converters, ensuring that monitoring data can be processed and analyzed promptly, identifying potential faults and issuing early warnings in a timely manner, and guaranteeing the safe and stable operation of the energy storage system.

[0122] ② Instruction-Level Parallelism and SIMD Instruction Set Utilization: Deeply explore the instruction set characteristics of embedded processors and fully utilize the Single Instruction Multiple Data (SIMD) instruction set for parallel computing. For example, on chips such as the STM32H7 with the ARM architecture, the NEON instruction set is used to accelerate operations such as convolution and matrix multiplication. SIMD instructions allow the same operation on multiple data elements within a single instruction cycle, significantly improving computational efficiency. The aim of this measure is to accelerate computationally intensive operations in the model, such as convolution and matrix operations, by leveraging the parallel computing capabilities of modern processors. The SIMD instruction set can process multiple data points simultaneously, greatly improving computational speed, thereby reducing the time latency of model inference. This enables the monitoring system to respond more quickly to changes in converter state, promptly detect fault precursors, enhance the system's real-time performance and reliability, and ensure efficient monitoring and safe management of high-altitude photovoltaic converters.

[0123] ③ Memory Management Optimization: Static memory allocation and memory reuse techniques are employed to reduce the overhead and fragmentation issues associated with dynamic memory allocation. During model inference, the memory layout is rationally planned, compactly arranging input data, intermediate features, and output results in memory to improve cache hit rate and reduce memory access latency. Simultaneously, memory reuse reduces memory consumption, ensuring the model can run smoothly on resource-constrained embedded devices. The goal of memory management optimization is to improve memory utilization efficiency, reduce memory access latency, and ensure the smoothness of the model inference process. Static memory allocation and memory reuse techniques avoid fragmentation and latency issues that may occur during dynamic memory allocation, while a compact memory layout helps improve cache utilization and accelerate data read and write speeds. This is particularly important for real-time monitoring applications on embedded platforms, as it ensures efficient model operation with limited hardware resources, timely processing of monitoring data, and guarantees the safe operation and reliable power supply of the energy storage system.

[0124] (3) Embedded platform adaptation and verification

[0125] To ensure the effective application of the monitoring method in actual high-altitude photovoltaic converters, embedded platform adaptation and verification are carried out. Specific measures are as follows:

[0126] ① STM32H7 Series Chip Adaptation: The STM32H7 series chips are high-performance 32-bit ARM Cortex-M7 core microcontrollers from STMicroelectronics, featuring a main frequency of up to 480MHz and rich peripheral interfaces, making them suitable for running complex monitoring algorithms. The aforementioned lightweight model and optimized inference engine were ported and adapted to the STM32H7 platform. The floating-point unit (FPU) and DSP instruction set of the chip were utilized to further improve model inference speed. Simultaneously, the FreeRTOS operating system was used to schedule tasks, ensuring the real-time performance and synchronization of model inference with other monitoring tasks (such as data acquisition and communication). The STM32H7 series chips were chosen because their powerful computing capabilities and rich peripheral interfaces meet the operational requirements of the monitoring algorithms, while their low power consumption is suitable for applications with high energy efficiency requirements, such as high-altitude photovoltaic converters. By leveraging the chip's FPU and DSP instruction sets, its computational advantages are fully utilized to accelerate the model inference process; the FreeRTOS operating system can effectively manage multiple monitoring tasks, ensuring their coordinated operation, achieving efficient real-time monitoring, and guaranteeing the stable operation and reliable power supply of the energy storage system in high-altitude environments.

[0127] ② Real-time performance assurance measures: On the embedded platform, the following measures are taken to ensure the real-time performance of model inference:

[0128] Interrupt Priority Configuration: Properly setting the interrupt priority for model inference tasks ensures timely response and initiation of inference tasks at critical moments such as data acquisition completion, preventing data loss or processing delays. The purpose of interrupt priority configuration is to ensure that model inference tasks can immediately obtain processor resources after data acquisition to begin data processing and analysis, thereby guaranteeing the timeliness and effectiveness of monitoring data, avoiding untimely fault diagnosis due to data delays, and ensuring the safe operation of the energy storage system.

[0129] Task scheduling optimization: A priority-based preemptive scheduling algorithm is adopted to allocate processor resources according to the urgency and importance of tasks. For model inference tasks, a higher priority is set to ensure that they can acquire processor resources for computation in a timely manner. The purpose of task scheduling optimization is to improve the utilization efficiency of processor resources and ensure that critical tasks (such as model inference) are executed first. By setting reasonable priorities, the processor can quickly switch between multiple tasks, meeting the strict time requirements of real-time monitoring, and promptly outputting the health status and fault warning information of the converter, thus ensuring the safe and stable operation of the energy storage system.

[0130] Pipeline processing organizes data acquisition, preprocessing, model inference, and result output into a pipeline structure. During the inference calculation of the previous data batch, the acquisition and preprocessing of the next data batch are performed simultaneously, improving overall processing efficiency and reducing end-to-end latency. The purpose of pipelined processing is to fully utilize processor resources and reduce waiting time by overlapping the execution processes of different tasks, thereby improving the overall throughput and response speed of the system. This processing method effectively shortens the total time from data acquisition to result output, enabling the monitoring system to respond more quickly to changes in converter status, promptly detect potential faults, enhance the real-time performance and reliability of the system, and ensure the safe operation and efficient management of high-altitude photovoltaic converters.

[0131] ③ Power consumption optimization strategy: To address the power consumption limitations of embedded devices, the following power consumption optimization measures are adopted:

[0132] Dynamic Voltage and Frequency Scaling (DVFS): This feature dynamically adjusts the processor's voltage and frequency based on the workload of the model inference task. It reduces frequency and voltage during lighter inference tasks to decrease power consumption, and increases frequency and voltage during heavier tasks to ensure real-time performance. The purpose of DVFS is to dynamically optimize processor power consumption while maintaining system real-time performance. By adjusting voltage and frequency according to task load, a balance between power consumption and performance can be achieved under different operating conditions, extending the battery life of embedded devices, reducing heat dissipation requirements, improving device reliability and stability, and ensuring long-term stable operation of the monitoring system in harsh high-altitude environments.

[0133] Peripheral clock gating: Peripheral clocks are only activated when needed (e.g., ADCs, communication interfaces), avoiding unnecessary power consumption. During model inference, peripheral clocks unrelated to the current task are turned off. The purpose of peripheral clock gating is to reduce the static power consumption of embedded devices. By precisely controlling the peripheral clock signals, it is ensured that peripherals consume power only when needed, thereby significantly reducing the overall power consumption of the system. This is particularly important in high-altitude areas where power supply may be limited, improving the energy efficiency of energy storage systems and ensuring the continuous operation of monitoring systems.

[0134] Low-power mode utilization: During data acquisition intervals or other idle periods, the processor is placed in a low-power mode (such as sleep mode) to reduce power consumption. The processor is then woken up by an interrupt or timer to perform task processing. The purpose of utilizing low-power mode is to minimize energy consumption when the device is idle. The processor in sleep mode consumes extremely low power, conserving energy when not performing tasks. When a new monitoring task arrives, the processor is quickly woken up by an interrupt or timer to ensure the continuity of monitoring operations. This strategy can effectively extend the lifespan of the equipment, reduce maintenance costs, improve its adaptability and survivability in high-altitude and remote areas, and ensure the long-term reliable operation of the energy storage system while maintaining the real-time performance of the monitoring system.

[0135] ④ System Integration and Verification: The monitoring algorithm, optimization model, and embedded software are integrated into the control system of the high-altitude photovoltaic converter for system-level testing and verification. The system is tested under different altitudes, air pressures, and temperatures to evaluate its performance indicators, such as health status assessment accuracy, fault warning accuracy, and real-time performance. The effectiveness and reliability of the system are verified by comparing data with actual converter operating data. The purpose of system integration and verification is to ensure the effectiveness and reliability of the monitoring method in practical applications. By testing under various actual operating conditions, the performance of the monitoring system can be comprehensively evaluated, potential problems can be identified and resolved, and system parameters and configurations can be optimized, thereby ensuring the stable operation and accurate monitoring of the monitoring system on the high-altitude photovoltaic converter. This is crucial for improving the availability and security of the energy storage system, effectively reducing downtime due to failures, lowering operation and maintenance costs, and enhancing the overall operational efficiency and competitive advantage of the energy storage system.

[0136] Through the aforementioned real-time and embedded adaptation optimizations, this monitoring method can operate efficiently on embedded platforms, meeting the real-time monitoring needs of high-altitude photovoltaic converters and providing strong support for the safe and reliable operation of the equipment.

[0137] The beneficial effects of this invention are as follows:

[0138] This invention addresses the unique operating environment of high-altitude photovoltaic converters and the shortcomings of existing monitoring technologies. It proposes a monitoring method and system for high-altitude photovoltaic converters based on physical constraint fusion. Through a series of innovative improvements, it significantly enhances monitoring accuracy, environmental adaptability, data utilization efficiency, system real-time performance, and stability, providing strong support for the reliable operation of high-altitude photovoltaic converters. Specifically:

[0139] (1) A multi-feature extraction and interaction model was constructed: through modal feature extraction, inter-modal feature interaction, scale multi-time series feature extraction and feature channel compression, the key information in the data was fully explored, so that the model could reflect the operating status of the converter more comprehensively and accurately, and significantly improve the monitoring accuracy.

[0140] (2) A physical constraint fusion layer model was introduced: physical constraints such as power conservation, thermodynamic cycle and insulation degradation were integrated into the model training, so that the model could follow physical laws while learning data characteristics, thereby improving the credibility and accuracy of monitoring results.

[0141] (3) A high-altitude cold-adaptive decoder model was designed, which includes functional modules such as health index calculation, fault warning signs and plateau compensation mechanism, so that the monitoring system can accurately assess the operating status of the converter in the plateau environment and enhance the environmental adaptability of the system.

[0142] (4) The unlabeled training mechanism was optimized: by reconstructing the loss function and adopting the plateau data augmentation strategy, the dependence on labeled data was reduced, and the generalization ability and data utilization efficiency of the model were improved.

[0143] (5) Real-time performance and embedded adaptation optimization were achieved: The use of lightweight model design, high-efficiency inference engine integration and deep optimization for embedded platforms ensured the real-time performance and stability of the system and met the strict requirements of plateau photovoltaic converter for real-time monitoring. Attached Figure Description

[0144] Figure 1 This is the overall flowchart of the method.

[0145] Figure 2 This is a flowchart of the data acquisition and preprocessing process.

[0146] Figure 3 This is a diagram of a multimodal temporal fusion network architecture.

[0147] Figure 4 This is a diagram of the physical constraint fusion layer architecture.

[0148] Figure 5 This is a flowchart of a high-altitude cold-adaptive decoder.

[0149] Figure 6 An improved architecture diagram for the label-free training mechanism.

[0150] Figure 7 Optimize the architecture diagram for embedded deployment. Detailed Implementation

[0151] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are for illustrative purposes only and are not intended to limit the scope of the present invention.

[0152] As shown in the figure, this invention provides a monitoring method for high-altitude photovoltaic converters based on physical information fusion, and the specific steps are as follows:

[0153] 1. Data Acquisition and Preprocessing

[0154] Obtaining high-quality, accurate data reflecting the operating status of photovoltaic converters in high-altitude areas provides a reliable foundation for subsequent feature extraction and health assessment.

[0155] In high-altitude environments, industrial-grade high-precision sensors are used to collect various data from the converter. Specifically, this includes:

[0156] -Three-phase voltage V abc It employs a Hall voltage sensor with an accuracy of ±0.2% and a frequency response range of DC-10kHz to acquire the three-phase voltage output of the converter and reflect its electrical performance.

[0157] -Three-phase current I abc It adopts a Hall current sensor with an accuracy of ±0.5% and an overload capacity of 10 times the rated current to collect the three-phase current output of the converter and evaluate its electrical performance.

[0158] -IGBT junction temperature T j Thermocouples (such as K-type thermocouples) are embedded inside the IGBT module, with a temperature measurement range of -40℃ to +250℃ and an accuracy of ±1℃, to monitor the junction temperature of power devices in real time and provide data support for thermal management.

[0159] - Air pressure P: A high-precision air pressure sensor (such as BMP388) is selected, with a measurement range of 30kPa to 110kPa, an accuracy of ±0.12hPa, and a resolution of 0.001hPa. The air pressure value of the plateau environment is collected and used to correct the calculation of insulation performance and heat dissipation efficiency.

[0160] - Heat dissipation temperature difference ΔT: Thermocouples are placed on the surface of the heat sink and in the environment to calculate the heat dissipation temperature difference and evaluate the working efficiency of the heat dissipation system.

[0161] -Switching frequency f sw The switching frequency of power devices is measured using a frequency meter with an accuracy of ±0.01%, which is used to analyze the dynamic performance of the converter.

[0162] These sensors are capable of operating in a wide temperature range (-40℃ to +85℃), adapting to extreme temperature changes at high altitudes, and have good electromagnetic compatibility (EMC), enabling them to operate stably in converter environments with strong electromagnetic interference.

[0163] The acquired data is transmitted to the embedded controller for the following preprocessing operations:

[0164] ① Normalization: Normalize the data to the range [0,1]. The formula is as follows: Where X norm For the normalized data, X min and X max These are the minimum and maximum values ​​of the data, respectively.

[0165] ② Filtering and noise reduction: A Butterworth filter with a cutoff frequency of 500Hz is used to remove high-frequency noise and ensure the clarity and accuracy of the signal.

[0166] ③ Feature window segmentation: The preprocessed data is divided into sliding windows with a window length of 100ms and an overlap rate of 50%, forming time-series data segments for subsequent feature extraction.

[0167] The preprocessed data is stored in the controller's memory, preparing it for subsequent feature extraction and analysis. Simultaneously, the controller is also responsible for transmitting the data to a remote monitoring center via a CAN bus or Ethernet interface, enabling remote monitoring and management of the data.

[0168] 2. Multimodal temporal fusion network

[0169] By integrating multimodal data, extracting time-series features, and capturing the correlation and complementarity between different modalities, a comprehensive feature representation is formed, providing high-quality feature input for subsequent health status assessment and fault early warning.

[0170] ① Modal feature extraction: For each modal data (e.g., V) a Perform a one-dimensional convolution operation to extract preliminary features. The convolution kernel size is 3, the number of output channels is 64, and the ReLU activation function is used, with the formula F. i =Conv1D(X i W i ,b i ), where F i Let X be the feature map of the i-th mode. i For the input data of the i-th mode, W and b i These are the convolution kernel and the bias, respectively. Through this process, discriminative local features can be extracted from each modality, laying the foundation for subsequent feature fusion.

[0171] ② Intermodal Feature Interaction: Features from different modalities are fused through a channel attention mechanism. First, Global Average Pooling (GAP) is performed, compressing each channel of the feature map into a single value, with the formula gap = GlobalAveragePooling1D(F). Then, a fully connected layer (FC) is used to adjust the channel weights and calculate the attention weights, with the formulas fc1 = Dense(64, activation = 'relu')(gap) and fc2 = Dense(64, activation = 'sigmoid')(fc1). Finally, the attention weights are applied to the original feature map, with the formula F. fuse =F⊙α, where α is the attention weight. This process can highlight important features, suppress unimportant features, and enhance the expressive power of features.

[0172] ③ Multi-scale temporal feature extraction: Multi-scale features are extracted using dilated convolutions with dilation rates of 1, 2, 4, and 8 to capture feature information at different time scales. The formula is as follows: Dilated convolution expands the receptive field by inserting holes into the convolution kernel, enabling it to capture both short-term and long-term dependencies simultaneously. These features at different scales are then concatenated and fused using the following formula: A multi-scale feature map containing rich temporal information is obtained.

[0173] ④ Feature Channel Compression: The number of feature channels is compressed using 1×1 convolution, reducing computational complexity while maintaining the discriminative power of the features. The formula is F. final =Conv1D(F multi-scale (1,64). A 1×1 convolution can change the number of channels in the feature map, achieving feature fusion and dimensionality reduction, resulting in the final feature map F. final This feature map contains comprehensive feature information of multimodal data, providing a high-quality feature representation for subsequent physical constraint fusion and health status assessment.

[0174] 3. Physical Constraint Fusion Layer

[0175] By incorporating physical laws into the feature extraction process, the physical rationality of the model and the monitoring accuracy are ensured. Through power conservation, thermodynamic cycles and insulation degradation constraints, the model's ability to model the physical state of the converter is improved, providing a physical basis for health status assessment and fault early warning.

[0176] ① Power conservation constraint: Calculate the input-output power balance, considering the effects of air pressure and temperature on efficiency in high-altitude environments. The formula is as follows: Where P in and P out These represent the input power and output power, η(P,T) j To take into account air pressure P and junction temperature T j efficiency, V dc and I dc These represent the DC side voltage and current, k h f is the switching harmonic loss coefficient. sw The switching frequency is used. This constraint ensures that the model output conforms to the physical law of power conservation, improving the model's accuracy in depicting power flow.

[0177] ② Thermodynamic Cyclic Constraints: Constraints are based on the IGBT junction temperature cumulative damage model. The formula is as follows: Where ΔD acc For cumulative damage, ΔT j (k) represents the junction temperature change at the k-th sampling point, m is the junction temperature change exponent, A is the normalization factor, Q is the activation energy, R is the gas constant, and T j(k) represents the junction temperature at the k-th sampling point. The constraint condition is D. acc ≥δ·σ(ΔT j ·t on ), where D acc δ is the cumulative damage threshold, σ is the safety factor, and t is the standard deviation. on This is the conduction time. This constraint can effectively simulate the thermal damage process of IGBTs, providing a basis for device lifetime assessment and thermal management.

[0178] ③ Insulation degradation constraint: Modify the partial discharge initiation voltage to adapt to the high-altitude environment. The formula is as follows: Where V PDIV V is the corrected partial discharge initiation voltage. PDIV0 The partial discharge initiation voltage under standard conditions is given by P, where P is the gas pressure, T is the ambient temperature, and c is the temperature. p This is the pressure correction factor. The constraint condition is |V phase | peak <0.7V PDIV , where |V phase | peak This represents the peak value of the phase voltage. This constraint allows for accurate assessment of the converter's insulation condition and prevention of insulation faults.

[0179] By combining physical laws with data-driven methods through these three physical constraints, the model can more accurately reflect the actual operating status of the converter, thereby improving the reliability and accuracy of monitoring.

[0180] 4. Cold-Adaptability Decoder

[0181] Based on the extracted features, a health index and fault warning indicators are calculated to assess the health status of the converter and provide early warning of potential faults. At the same time, a plateau compensation mechanism is used to adapt to the special characteristics of the plateau environment, providing decision support for the stable operation of the converter.

[0182] (1) Health Index (HI) Calculation: Comprehensive Cumulative Damage D acc and the rate of change of thermal resistance ΔR th Calculate the health index. The formula is HI = 1 - ω1D acc -ω2ΔR th The HI value is a weighting coefficient that satisfies ω1 + ω2 = 1. The HI value reflects the overall health status of the converter; the closer the HI value is to 1, the healthier the converter, providing maintenance personnel with an intuitive basis for equipment health assessment.

[0183] (2) Fault warning signs:

[0184] ① Insulation risk: through formula THD VCalculate the insulation risk value. Here, σ is the activation function (e.g., Sigmoid), which maps the linear combination result to a probability range of 0 to 1; W is the weight vector, containing the weights of each feature; dV / dt max Maximum voltage change rate; P is air pressure; THD V Let R be the total harmonic distortion of the voltage. ins When the pressure is >0.6 and the air pressure P <70kPa, an insulation fault warning is triggered to promptly remind maintenance personnel to pay attention to the insulation status and prevent equipment damage and safety accidents caused by insulation faults.

[0185] ② Heat dissipation failure indicator: indicated by the formula Calculate the heat dissipation failure indicators. Among them, This is an indicator function that outputs 1 when the condition within the parentheses is met, and 0 otherwise; ΔT is the temperature difference due to heat dissipation. Let R be the rate of temperature change. cool When ΔT = 1 and ΔT > 20℃, a heat dissipation failure warning is triggered, prompting maintenance personnel to check the heat dissipation system in a timely manner to avoid converter overheating failure due to poor heat dissipation.

[0186] (3) High-altitude compensation mechanism: Considering the impact of the high-altitude environment on the switching frequency, a high-altitude compensation mechanism is adopted to compensate the switching frequency for air pressure. The formula is as follows: in For the compensated switching frequency, f sw Where is the original switching frequency, and P is the air pressure. By adjusting the switching frequency, the performance of the converter is optimized, making it better adaptable to the low-pressure environment of high altitudes, thereby improving the converter's operating efficiency and stability.

[0187] 5. Improvements to the label-free training mechanism

[0188] By optimizing model parameters through unlabeled training, the model's generalization ability and robustness are improved, and its dependence on labeled data is reduced. At the same time, data augmentation is used to improve the model's adaptability to the plateau environment, thereby enhancing the model's performance and reliability in practical applications.

[0189] (1) Loss Function Reconstruction: Construct a total loss function that includes voltage reconstruction loss, power conservation loss, aging gradient constraint loss, and insulation degradation sensing loss. The formula is as follows: Wherein, λ1=1.0, λ2=0.8, λ3=0.5, λ4=0.3 (when the air pressure P<70kPa). By comprehensively considering multiple physical constraints and loss terms for feature reconstruction, the model can still effectively learn the characteristics and patterns of the converter even without a large amount of labeled data, thus improving the model's generalization ability.

[0190] (2) Plateau data augmentation:

[0191] ① Enhanced pressure-temperature coupling: through formula P aug =P×np.random.uniform(0.85,1.15) and Implementation, where P aug T represents the enhanced air pressure value. j_aug This refers to the enhanced IGBT junction temperature. This method simulates fluctuations in air pressure and temperature at high altitudes, enabling the model to learn data characteristics under different environmental conditions and enhancing its adaptability to changes in the high-altitude environment.

[0192] ② Power grid disturbance injection: via formula This method simulates the unique power grid noise found in high-altitude areas, improving the model's adaptability to real-world operating conditions. It also exposes the model to more diverse data during training, enhancing its robustness and resistance to interference in actual operation.

[0193] 6. Embedded Deployment Optimization

[0194] The model and inference process were optimized to ensure real-time monitoring and stable operation of the entire monitoring system on the embedded platform, meeting the stringent resource constraints and real-time requirements of the high-altitude photovoltaic converter.

[0195] (1) Lightweight Model Design: Grouped convolution is used instead of dilated convolution, with the number of groups set to 4 to reduce computation. Depthwise separable convolution is also applied, decomposing standard convolution into two steps: depthwise convolution and pointwise convolution, further reducing computational complexity. The model is pruned at a rate of 30% to remove unimportant connection weights. Finally, INT8 quantization is used to convert 32-bit floating-point numbers (FP32) into 8-bit integers (INT8), reducing the model's storage requirements and computational complexity, enabling the model to run efficiently on resource-constrained embedded devices.

[0196] (2) High-efficiency inference engine integration: TVM compilation optimization is adopted to generate high-efficiency code for the ARM Cortex-M7 architecture. Parallel computing is performed using the NEON instruction set to improve inference speed. Static memory allocation and memory reuse techniques are adopted to reduce the overhead and fragmentation problems caused by dynamic memory allocation, optimize memory management, and improve system stability and reliability.

[0197] (3) Embedded Platform Adaptation and Verification: Deployed on an STM32H743VI microcontroller with a main frequency of 480MHz. The FreeRTOS operating system is used for task scheduling to ensure the real-time performance and synchronization of model inference with other monitoring tasks (such as data acquisition and communication). Power consumption is reduced through Dynamic Voltage and Frequency Scaling (DVFS), dynamically adjusting the processor's voltage and frequency according to the load of the model inference task. During data acquisition intervals or other idle periods, the processor is placed in a low-power mode (such as sleep mode), and woken up by interrupts or timers to perform task processing, optimizing power management and extending the system's operating time.

[0198] Through the above embedded deployment optimization measures, the high-altitude photovoltaic converter integrated monitoring system is ensured to operate efficiently, in real time, and stably on the embedded platform, enabling it to work reliably for a long time in the harsh high-altitude environment and providing strong protection for the safe operation of the energy storage converter.

[0199] The above description is only a preferred embodiment of the present invention. It should be noted that those skilled in the art can make several adjustments and improvements without departing from the core concept of the present invention, and these adjustments and improvements should also be considered within the scope of protection of the present invention.

Claims

1. A monitoring method for high-altitude photovoltaic converters based on physical information fusion, characterized in that, Includes the following steps: S1. Multidimensional feature extraction: Modal feature extraction, interaction, multi-temporal feature extraction, and channel compression are performed on the collected multidimensional key data. S2, Physical Constraint Fusion: Integrating physical constraints such as power conservation, thermodynamic cycling, and insulation degradation into model training; S3, High Altitude Adaptability Decoding: Calculates the health index based on cumulative damage and thermal resistance change rate, provides early warning of potential faults, and optimizes converter performance to adapt to the high-altitude environment; S4. Improved unlabeled training mechanism: By reconstructing the loss function and using plateau data augmentation strategies, the dependence on labeled data is reduced and the model's generalization ability is improved; The multivariate feature extraction in step S1 includes: S11. Data preprocessing: Collect key data and organize it into a unified format; S12. Modal Feature Extraction: Perform a one-dimensional convolution operation on each modal data to extract local features. The specific formula is as follows: ， In the formula F i The feature map of the i-th mode; This is a one-dimensional convolution operation; X i W is the input data for the i-th mode; i It is a convolution kernel; b i It is a bias term; S13. Intermodal Feature Interaction: Modal features are fused through a channel attention mechanism, with the specific formula as follows: In the formula It is the fused feature map; It is an element-wise multiplication operation; F is the original feature map; It is attention weight; S14. Multi-scale temporal feature extraction: Features at different time scales are captured using dilated convolutions with varying dilation rates, specifically dilated convolutions with dilation rates of 1, 2, 4, and 8. In the formula This is the feature map under the d-th expansion rate; For dilated convolution, d is the dilation rate; By concatenating feature maps at different expansion rates, a multi-scale feature map is obtained: In the formula It is a comprehensive feature map that integrates features from different time scales; Concat is a concatenation operation; S15. Feature Channel Compression: Channel compression is performed using 1×1 convolution to remove redundant features. The specific formula is as follows: In the formula This is the final feature map; 1 indicates the size of the convolution kernel is 1×1; 64 is the number of output channels.

2. The monitoring method according to claim 1, characterized in that, The physical constraint fusion in step S2 includes: S21. Power Conservation Constraint: Construct a power conservation constraint loss function to ensure that the model's depiction of the converter's power conversion process conforms to physical reality. The specific formula for the power conservation constraint loss function is as follows: In the formula It is a power conservation constraint loss; It is the input power; It is the output power; It refers to efficiency, which takes into account the gas pressure P and the IGBT junction temperature T. j Impact on efficiency; It is the DC side voltage; It is the DC side current; f is the switching harmonic loss coefficient of the h-th order; sw It is the switching frequency; The formula for calculating efficiency is: In the formula It is the baseline efficiency; P is the air pressure; T is the reference efficiency. j It refers to the junction temperature of the IGBT; S22. Thermodynamic Cyclic Constraints: An IGBT junction temperature cumulative damage model is used to evaluate the cumulative impact of the constraint model on device lifetime. The IGBT junction temperature cumulative damage model is as follows: In the formula It represents cumulative damage; N is the number of sampling points; is the junction temperature change at the k-th sampling point; m is the junction temperature change exponent; A is the normalization factor; Q is the activation energy; R is the gas constant; The junction temperature of the IGBT at the kth sampling point; The constraints are: In the formula It is the cumulative damage threshold; It is the safety factor; It is the standard deviation; It is the conduction time; S23. Insulation degradation constraint: Correct the partial discharge initiation voltage and establish a constraint relationship between the phase voltage peak and the corrected initiation voltage. Specifically, the partial discharge initiation voltage is corrected as follows: In the formula This is the corrected partial discharge initiation voltage; It is the partial discharge initiation voltage under standard conditions; P is the gas pressure; c p This is the air pressure correction factor; T is the ambient temperature. The constraints are: In the formula It is the peak value of the phase voltage; It is the partial discharge initiation voltage.

3. The monitoring method according to claim 1, characterized in that, The high-altitude cold adaptability decoding in step S3 includes: S31. Health Index (HI) Calculation: The health index is calculated by combining cumulative damage and the rate of change of thermal resistance. The specific formula is as follows: In the formula, HI is the health index, with a value ranging from 0 to 1; , These are weighting coefficients, representing the degree of influence of cumulative damage and changes in thermal resistance on the health index, respectively, satisfying... + = 1; It is cumulative damage; It is the rate of change of thermal resistance; S32. Fault Warning Signs: Establish warning signs for insulation risk and heat dissipation failure respectively. The formula for the insulation risk of the converter is: In the formula This is the insulation risk value, ranging from 0 to 1; W is the activation function; W is the weight vector. It is the maximum rate of change of voltage; P is the air pressure; It is the total harmonic distortion of voltage; when And an early warning is triggered when P < 70 kPa; The calculation method for heat dissipation failure is as follows: In the formula It is a heat dissipation failure indicator, a binary variable that takes the value 0 or 1, where 1 indicates heat dissipation failure; It is an indicator function that outputs 1 when the condition inside the parentheses is met, and 0 otherwise; It is the temperature difference for heat dissipation; It is the rate of temperature change; when =1 and Timely triggering of warnings; S33. High-altitude compensation mechanism: Air pressure compensation is applied to the switching frequency to optimize converter performance. The specific formula is as follows: In the formula It is the compensated switching frequency; is the original switching frequency; P is the air pressure.

4. The monitoring method according to claim 1, characterized in that, The improvements to the label-free training mechanism in step S4 include: S41. Loss Function Reconstruction: Taking into account various constraints, the total loss function is reconstructed, and the formula is as follows: In the formula It is the total loss function; , , , These are weighting coefficients; It is the reconstructed three-phase voltage; It is the original three-phase voltage; It is the L2 norm; It is a power conservation constraint loss; It is the gradient of cumulative damage with respect to time; It is the L1 norm; k is the barometric pressure sensitivity coefficient; This is the predicted change in the partial discharge initiation voltage; It is the Frobenius norm; S42. Plateau Data Augmentation Strategy: A pressure-temperature coupling augmentation and power grid disturbance injection strategy are employed to enhance the model's adaptability to changes in the plateau environment. The formula for the pressure-temperature coupling augmentation is as follows: In the formula This is the enhanced air pressure value; It generates uniformly distributed random numbers between 0.85 and 1.15; This refers to the enhanced IGBT junction temperature. The formula for power grid disturbance injection is: In the formula is the random noise that generates the Rayleigh distribution; scale=0.05 is the scale parameter of the Rayleigh distribution; The generated noise has the same shape as the original voltage data.

5. A monitoring system for high-altitude photovoltaic converters based on physical information fusion, characterized in that, The method for monitoring high-altitude photovoltaic converters according to any one of claims 1-4 includes: Data acquisition and preprocessing module: used to collect key data and perform preprocessing; Multimodal temporal fusion network: used to extract temporal features from multimodal data; Physical constraint fusion layer: used to incorporate physical laws into the feature extraction process; High-altitude adaptability decoder: used to calculate health index and fault warning signs, and to perform high-altitude compensation; Unlabeled training module: used to optimize model parameters and improve model generalization ability; Embedded Deployment Optimization Module: Used to ensure real-time monitoring and stable operation of the monitoring system on embedded platforms.

6. The monitoring system according to claim 5, characterized in that, The multimodal temporal fusion network includes: Modal feature extraction unit: Performs one-dimensional convolution operation on each modal data to extract local features; Intermodal feature interaction unit: fuses modal features through a channel attention mechanism; Multi-scale temporal feature extraction unit: Captures features at different time scales using dilated convolutions with varying dilation rates; Feature channel compression unit: 1×1 convolution is used to compress channels and remove redundant features.

7. The monitoring system according to claim 5, characterized in that, The physical constraint fusion layer includes: Power conservation constraint unit: Construct a power conservation constraint loss function to ensure that the model's modeling of the converter's power conversion process conforms to physical reality; Thermodynamic cycle constraint unit: The cumulative damage model of IGBT junction temperature is used to evaluate the cumulative impact of the constraint model on device lifetime; Insulation degradation constraint unit: corrects the partial discharge initiation voltage and establishes a constraint relationship between the phase voltage peak and the corrected initiation voltage.

8. The monitoring system according to claim 5, characterized in that, The cold-adaptive decoder includes: Health Index (HI) Calculation Unit: Calculates the health index by combining cumulative damage and thermal resistance change rate; Fault warning sign unit: Establish warning signs for insulation risk and heat dissipation failure respectively; High-altitude compensation mechanism unit: Performs air pressure compensation on the switching frequency to optimize converter performance.

9. The monitoring system according to claim 5, characterized in that, The unlabeled training module includes: Loss function reconstruction unit: Reconstructs the total loss function by comprehensively considering multiple constraints; Plateau data augmentation strategy unit: Employs pressure-temperature coupling enhancement and power grid disturbance injection strategies to enhance the model's adaptability to changes in the plateau environment; The embedded deployment optimization module includes: Lightweight model design unit: Employs grouped convolution, depthwise separable convolution, model pruning and quantization techniques to optimize model computation efficiency and resource consumption; High-efficiency inference engine integration unit: Integrates a high-efficiency inference engine to improve the inference speed and efficiency of the model on the embedded platform; Embedded platform adapter unit: Adaptation and verification are performed on the STM32H7 series chips to ensure the effective application of the monitoring system in the plateau energy storage converter.

Images