Method and system for monitoring reliability of photovoltaic converter in plateau special environment

Abstract

The invention discloses a method and a system for monitoring the reliability of a photovoltaic converter in a special plateau environment. The method comprises the following steps: firstly, acquiring electrical quantity, temperature quantity, environment quantity and operation quantity, filtering abnormal data, and realizing accurate alignment of multi-frequency signals in combination with a dynamic time warping algorithm; secondly, constructing a plateau sensitive feature set, and optimizing feature quality through physical consistency check and three-stage feature selection; then, an IGBT thermal fatigue equation and a capacitance aging equation are fused to establish a health index evolution model, a Bayesian physical information neural network is used for prediction, and a high-reliability confidence interval is output through Monte Carlo sampling. And finally, dynamically correcting the residual life based on the comprehensive environment factor, and triggering a hierarchical maintenance decision according to the health index state, the residual life and the confidence interval width. The service life prediction precision of the photovoltaic converter in the plateau environment is remarkably improved, the operation and maintenance cost is effectively reduced, and the equipment operation reliability is enhanced.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of photovoltaic power generation equipment health management, and particularly relates to a photovoltaic converter reliability monitoring method and system under a plateau special environment. BACKGROUND

[0002] The rapid development of photovoltaic power stations in plateau areas is facing a reliability management dilemma. The harsh environment above an altitude of 3000 meters constitutes a unique "triple strangulation chain": the thin air greatly weakens the efficiency of the heat dissipation system, making the converter power module continuously suffer in a high-temperature steamer; the strong ultraviolet rays act like an invisible carving knife, continuously eroding the capacitor insulation material and IGBT packaging shell day and night; the severe day and night temperature difference becomes a cold and hot blacksmith, quenching and tempering the metal weld layer every day. These coupled environmental stresses are not simply superimposed, but have a synergistic amplification effect - the heat dissipation lag under low air pressure exacerbates the thermal expansion stress caused by temperature difference, and ultraviolet light degradation reduces the toughness of materials resisting mechanical fatigue, resulting in a cliff-like decline in equipment life.

[0003] However, the current mainstream monitoring technology is still trapped in the "plain thinking" rut. The temperature warning system uses the heat transfer model based on standard atmospheric conditions, completely ignoring the fatal weakening of air pressure on the efficiency of air cooling / liquid cooling, like using a sea level barometer to predict plateau weather; the electrical characteristic diagnosis mechanically applies the harmonic distortion rate and other plain indicators, failing to capture the ripple anomaly caused by the boiling of capacitor electrolyte in a low air pressure environment, just like using a plain electrocardiogram to diagnose a plateau heart disease; the life prediction model is obsessed with data black boxes, regarding physical mechanisms such as thermal fatigue crack propagation and electrochemical aging as noise filtering, and its prediction results in the plateau scene with data scarcity are like building a tower on sand. More seriously, the operation and maintenance strategy is seriously out of touch with the actual environment: regular maintenance ignores the seasonal regularity of ultraviolet summer rampage and temperature difference winter rampant, and there is a monitoring gap in the critical equipment degradation period; fault repair is subject to the logistics dilemma on the plateau, often evolving into a vicious cycle of "small illness dragging into big illness".

[0004] This systemic failure is causing a chain crisis: power station operators are caught between "excessive maintenance burning money" and "fault downtime losing money", equipment manufacturers are plagued by high plateau failure frequency and are subject to quality claim responsibilities, and the green energy transformation of the entire industry chain is being hindered by the plateau reliability curse. The industry urgently needs a technological paradigm revolution - a monitoring method that can penetrate the "low pressure-radiation-temperature difference" coupled damage mechanism, like providing a special CT machine for the plateau converter; a prediction model that integrates physical nature and data intelligence, like an old Chinese doctor diagnosing plateau-specific diseases; and a dynamic operation and maintenance framework that converts environmental stress into decision parameters, enabling plateau power station operation to upgrade from "firefighting emergency" to "preventing disease type control". SUMMARY

[0005] The application proposes to use Bayesian inference, physical constraint fusion, data-driven and uncertainty quantification methods, aiming to accurately monitor the reliability of photovoltaic converters in plateau environments.

[0006] The specific scheme is as follows:

[0007] A photovoltaic converter reliability monitoring method in a plateau special environment, comprising the following steps:

[0008] S1, data acquisition and preprocessing: the electrical quantity, temperature quantity, environmental quantity and running quantity of the plateau converter are collected, and the data of different sampling frequencies are timestamp calibrated by abnormal value processing based on 3 sigma criterion, power change segment elimination, stable period selection and dynamic time warping (DTW) algorithm, to ensure high data quality and accurate correspondence;

[0009] S2, plateau sensitive feature engineering: a feature system including voltage ripple rate, high-frequency harmonic energy ratio, junction temperature fluctuation entropy, thermal response time, thermal shock coefficient and comprehensive stress index is constructed, and the optimal feature subset is selected through physical consistency and three-stage feature selection method;

[0010] S3, physical constraint modeling: IGBT thermal fatigue and capacitor aging models are established, the health state evolution equation is combined to quantitatively evaluate the health state of the equipment, and the physical residual is introduced to ensure that the data-driven model is consistent with the physical law;

[0011] S4, Bayesian-PINN fusion modeling: a network architecture including input layer, Bayesian hidden layer and output layer is constructed, the loss function of fusion data fitting term, physical constraint term and Bayesian regularization term is set, the data and physical model are deeply coupled, and the model generalization ability and prediction accuracy are improved;

[0012] S5, RUL prediction process: the prediction uncertainty is considered from the health index to the remaining useful life (RUL), the environment adaptive correction mechanism is introduced, and the precise maintenance decision support is realized based on the multi-index decision logic;

[0013] S6, plateau environment adaptive mechanism: according to the altitude and seasonal change, the model parameters and feature weights are automatically adjusted, so that the monitoring system maintains high performance monitoring effect under different environmental conditions.

[0014] Further, the step S1 data acquisition and preprocessing specifically includes:

[0015] (1) data source

[0016] The collected data includes the electrical quantity, temperature quantity, environmental quantity and running quantity of the plateau converter. The electrical quantity includes DC bus voltage (V dc ) and AC output current (I ac), with a sampling frequency of 10 kHz, measured by a Hall sensor. Temperature quantities include IGBT junction temperature (T j ) and heat sink temperature (T hs ), with a sampling frequency of 1 Hz, measured by a thermocouple. Environmental quantities involve ultraviolet intensity (UV), altitude (h), air pressure (P), with a sampling frequency of 0.1 Hz, provided by a weather station. Operational quantities are output power (P loss ) and switching frequency (f sw ), with a sampling frequency of 1 kHz, from the SCADA system.

[0017] (2) Pretreatment process

[0018] ① Outlier processing: based on the 3σ criterion, if the data x satisfies |x-μ|>3σ, it is determined as an outlier and is removed. Where μ represents the mean of the data, σ represents the standard deviation of the data. For the power variation segment, if the absolute value of the power change rate satisfies , the data segment is removed. Here, P loss represents the output power, and P rated is the rated power.

[0019] ② Selection of stable period: the light intensity needs to satisfy >800W / m 2 , the power fluctuation needs to satisfy <5% (in a 10min window), and the time window is limited to 10:00-14:00 (true solar time).

[0020] ③ Signal alignment: DTW (Dynamic Time Warping) algorithm is used, the goal is to minimize ∑d(a i ,b j ), where a i and b j are data points in two time series, and d(a i ,b j ) represents the distance between them. And the environmental data and electrical data are timestamped.

[0021] Further, the step S2 highland sensitive feature engineering specifically includes:

[0022] (1) Feature system

[0023] ① The calculation formula of voltage ripple rate is:

[0024]

[0025] In the formula, V dc is the DC bus voltage, For its average value, RMS represents the root mean square value, h is the altitude. This feature reflects the aging degree of the capacitor, the greater the ripple ratio, the worse the capacitor filtering effect, and the more serious the aging may be, considering the influence of altitude on the ripple ratio.

[0026] ②The calculation formula of high-frequency harmonic energy ratio is:

[0027]

[0028] In the formula, f sw is the switching frequency, and PSD(f) is the power spectral density. This feature reflects the degradation of IGBT switching characteristics, and the high-frequency harmonic energy ratio will increase as the performance of IGBT decreases.

[0029] ③The calculation formula of junction temperature fluctuation entropy is:

[0030]

[0031] In the formula, T j is the IGBT junction temperature, p(T j ) is its probability distribution, P loss0 is the reference power, and P loss is the actual output power. It measures the complexity of the junction temperature fluctuation, reflects the thermal fatigue damage, and the more complex the junction temperature fluctuation, the more serious the thermal fatigue damage may be.

[0032] ④The calculation formula of thermal response time is:

[0033]

[0034] In the formula, T j (t) is the junction temperature function with time, P loss (t) is the output power function with time, and corr represents the correlation. This feature reflects the performance of the heat dissipation system, and the longer the thermal response time, the more likely there is a problem or performance degradation in the heat dissipation system.

[0035] ⑤The calculation formula of thermal shock coefficient is:

[0036]

[0037] In the formula, ΔT day is the diurnal temperature difference, and N cycle is the number of thermal cycles. It measures the influence of diurnal temperature difference stress on the equipment, and the greater the temperature difference and the more the number of cycles, the greater the thermal shock coefficient, and the greater the environmental stress the equipment withstands.

[0038] ⑥The calculation formula of comprehensive stress index is:

[0039]

[0040] In the formula, UV is the intensity of ultraviolet light, and h is the altitude. The index comprehensively considers factors such as ultraviolet light, thermal shock, and altitude, and fully reflects the total damage degree of the plateau environment to the equipment.

[0041] (2) Feature fusion

[0042] ① Physical consistency test: verification Ensure that the ripple rate of the capacitor changes with temperature in accordance with the temperature characteristics of the capacitor. Verify that |corr(S T ,T a )| < 0.3 to exclude the interference of ambient temperature on the fluctuation of junction temperature.

[0043] ② Three-stage feature selection: first, remove features with small variance through variance thresholding, then perform physical screening to retain features that meet physical laws, and finally use the recursive feature elimination (RFE) algorithm to select the optimal feature subset.

[0044] ③ Final feature vector: x = [R r ,E h ,S T ,τ th ,Λ,K ts ] T .

[0045] Further, step S3 physical constraint modeling specifically includes:

[0046] (1) Core physical equation

[0047] ① IGBT thermal fatigue model:

[0048]

[0049] In the formula, N f represents the thermal fatigue life of IGBT, A is a constant, ΔT j is the junction temperature change, β is a material constant, E a is the activation energy, k B is the Boltzmann constant, and T m is the average temperature. This model describes the life characteristics of IGBT under thermal stress.

[0050] ② Capacitor aging model:

[0051]

[0052] In the formula, L is the capacitor life, L0 is the reference life, E a is the activation energy, T is the temperature, V is the voltage, V0 is the reference voltage, and n is the aging index. This model reflects the aging law of the capacitor under the combined action of electrical stress and thermal stress.

[0053] (2) Health state evolution equation

[0054] ① Health index definition: u(t) ∈ [0, 1], u(0) = 1 at initial time, u(t fail ) = 0.2 at failure time

[0055] The health index u(t) is a dimensionless parameter that quantifies the health state of the device. It is defined in the range [0, 1] to facilitate unified evaluation and monitoring of the device health state. At the initial time, u(0) = 1 indicates that the device is in a brand-new, undamaged state. At the failure time, u(t fail ) = 0.2 is determined according to the failure threshold of the device. When the health index drops to 0.2, it is considered that the device has reached the end of its life and needs to be maintained or replaced.

[0056] ② Degradation rate modeling:

[0057]

[0058] The degradation rate is the rate of change of the health index with time, reflecting the speed of deterioration of the device health state. It is composed of the negative sum of the degradation rates of IGBT and capacitor, indicating that the degradation of IGBT and capacitor is the main factor leading to the decline of the overall health state of the device. Specifically,

[0059]

[0060] In the formula, C1 and C2 are constants related to materials and processes; E a,IGBT and E a,Cap are the activation energies of IGBT and capacitor, respectively, reflecting the energy barrier that needs to be overcome by its internal failure mechanism; R is the gas constant; T m and T are the average temperatures of the environment where IGBT and capacitor are located; ΔT j is the change of IGBT junction temperature, reflecting the influence of thermal stress on IGBT degradation; β is a material constant related to the thermal mechanical properties of IGBT material; K ts is the thermal shock coefficient, quantifying the influence of thermal shock such as diurnal temperature difference on IGBT degradation; V and V0 are the actual voltage and reference voltage of the capacitor, respectively; n is the aging index, describing the influence law of voltage on capacitor aging rate; I uv is the ultraviolet intensity index, reflecting the influence degree of ultraviolet on capacitor aging. These parameters jointly determine the degradation rates of IGBT and capacitor, thereby affecting the evolution of the overall health state of the device.

[0061] ③ Physical residual:

[0062]

[0063] The physical residual is an important indicator for measuring the deviation between the health state evolution model and the physical law. According to the definition of the health index and the degradation rate modeling, in an ideal case, the change rate of the health index should be equal to -(Γ IGBT +Γ Cap ). Therefore, the physical residual evaluates the physical rationality of the model by calculating the difference between the actual model prediction of the health index change rate and this ideal value. If the physical residual is large, it means that the model prediction result deviates greatly from the physical law, and the model needs to be corrected and optimized. In the model training process, by minimizing the physical residual, the consistency between the data-driven model and the physical model can be ensured, and the reliability and accuracy of the prediction result can be improved.

[0064] Further, the step S4 of the Bayesian-PINN fusion modeling specifically includes:

[0065] (1) Network architecture

[0066] The input layer receives time t and feature vector x as input, and the output layer outputs health index u(t) after the processing of Bayesian hidden layer 1 and Bayesian hidden layer 2. The weights w of the Bayesian hidden layer are subject to normal distribution The output health index u(t) is input to the physical constraint module, and the prediction result is combined with the physical constraint to be used as a loss function for model training.

[0067] (2) Loss function

[0068]

[0069] In the formula,

[0070] ① Data fitting term: The data fitting term is used to measure the difference between the health index u pred predicted by the model and the real health index u true . By minimizing this error, the model can accurately fit the existing data. Wherein, N is the number of data samples, and the average is taken to eliminate the influence of data volume on the loss function, so that the size of the loss function is independent of the data volume, which is convenient for comparison and optimization between different scales of data sets.

[0071] ② Physical constraint term: The physical constraint term ensures that the health state evolution predicted by the model conforms to the physical law.‖PDE(u)‖ 2 is the square norm of the physical residual, which reflects the deviation between the model prediction result and the physical model. By minimizing this part of the loss, the model learns the health state evolution mode conforming to the physical law in the training process. And ReLU(u t+1 -u tThe physical constraint term is used to prevent non-physical increases in health indicators. Under normal circumstances, the health status of equipment should monotonically decrease or remain constant over time; an increase would violate physical principles. The Modified Linear Unit Function (ReLU) is used to penalize such unreasonable changes, ensuring that the trend of health indicator changes conforms to physical reality. Weighting coefficients β1 and β2 are used to balance the contributions of the two parts in the physical constraint term; their values ​​are adjusted according to the actual problem and model training to achieve the optimal physical constraint effect.

[0072] ③ Bayesian regularization term: The Bayesian regularization term is used to quantify the uncertainty of model parameters and prevent overfitting. The prior distribution p(w) is usually assumed to be a simple normal distribution, and its calculation formula is:

[0073]

[0074] In the formula, μ p It is the mean vector of the prior distribution, Σ p It is the covariance matrix of the prior distribution.

[0075] The posterior distribution q(w) is an update of the prior distribution, taking into account the distribution after observing the data. Its calculation formula is:

[0076]

[0077] In the formula, μ q It is the mean vector of the posterior distribution, Σ q It is the covariance matrix of the posterior distribution.

[0078] Bayesian regularization terms By measuring the difference between the posterior and prior distributions of model parameters, it is ensured that the model fully utilizes prior knowledge during training, avoiding overfitting of model parameters to training data. In equipment health status assessment, the prior distribution pre-determines a reasonable range for parameters based on physical models or engineering experience. The posterior distribution is adjusted during training based on actual data, reflecting the most likely values ​​of model parameters and their uncertainties under given data. By minimizing the Bayesian regularization term, the model can find a balance between data-driven approaches and physical constraints, ensuring that predictions not only conform to observed data but also align with the actual physical behavior of the equipment. This helps improve the model's generalization ability and reliability in few-shot learning and uncertainty quantification.

[0079] (3) Quantification of uncertainty

[0080] ① Monte Carlo sampling:

[0081]

[0082] In the formula, u(i) (t) represents the predicted value of the health indicator at the i-th sampling; x is the input feature vector; w (i) represents the model parameters at the i-th sampling; q(w) is the posterior distribution of the model parameters; M is the number of samplings, which is set to 100 here.

[0083] By sampling multiple sets of model parameters w (i) from the posterior distribution q(w), the input feature x is predicted multiple times using Bayesian neural networks (BNNs) to obtain multiple health indicator predictions u (i) (t). The physical meaning of this step is to quantify the impact of model parameter uncertainty on the prediction results, and to simulate the possible value range of the parameters by multiple samplings, thereby providing a basis for subsequent uncertainty analysis.

[0084] ②Statistical quantity calculation: mean standard deviation

[0085] In the formula, is the mean of the predicted value of the health indicator; σ u (t) is the standard deviation of the predicted value of the health indicator.

[0086] The mean and standard deviation of multiple predictions are calculated, and the mean reflects the central tendency of the prediction, and the standard deviation quantifies the uncertainty degree of the prediction. The larger the standard deviation, the more uncertain the model is in predicting the health indicator, which may be related to the uncertainty of the model parameters, data noise or limitations of the model structure. Through these two statistical quantities, the distribution characteristics of the prediction results can be described concisely, providing a basis for subsequent confidence interval estimation and decision support.

[0087] ③Confidence interval:

[0088]

[0089] In the formula, CI 95 (t) represents the 95% confidence interval; 1.96 is the critical value of the normal distribution corresponding to the 95% confidence level.

[0090] Based on the mean and standard deviation, a 95% confidence interval is constructed, which means that under multiple samplings and predictions, there is a 95% probability that the true health indicator u(t) falls within this interval. The confidence interval provides an uncertainty range for RUL prediction, helping to assess the reliability of the prediction results. A narrower confidence interval indicates that the prediction results are more reliable, while a wider confidence interval indicates that there is greater uncertainty, which requires more data or further model optimization.

[0091] The uncertainty quantification result is of great significance for subsequent RUL prediction and decision support. On the one hand, by providing the confidence interval of the health indicator, the remaining useful life range of the equipment can be more accurately estimated instead of a single point estimate, thereby providing more comprehensive information for operation and maintenance decision-making. For example, in the decision logic, when the confidence interval width exceeds a certain threshold, a manual inspection action is triggered to deal with greater uncertainty. On the other hand, uncertainty quantification helps to evaluate the reliability of the model, and when the model prediction uncertainty is large, the model can be adjusted or more data can be collected to improve the prediction accuracy, enhancing the robustness and credibility of the entire health state evaluation system.

[0092] Further, the step S5 RUL prediction process specifically includes:

[0093] (1) Conversion from health indicator u value to RUL

[0094] ① Failure time calculation:

[0095] In the formula, t fail represents the time when the equipment reaches the failure threshold; t is the time variable; is the mean of the health indicator prediction value σ u (t) is the standard deviation of the health indicator prediction value; 0.2 is the pre-set health indicator failure threshold.

[0096] The mean and standard deviation of the health indicator are used to determine the time range when the equipment may fail. Considering the prediction uncertainty, the earliest and latest failure times are given, reflecting the failure risk of the equipment at different confidence levels, which can provide a basis for subsequent RUL calculation and clearly indicate the remaining running time of the equipment under different conditions, helping to develop a reasonable maintenance plan.

[0097] ② Basic RUL calculation: RUL = t fail -t current ,

[0098] In the formula, RUL represents the remaining useful life; t fail is the predicted failure time; t current is the current time.

[0099] The basic RUL is the expected length of time that the equipment can still operate normally under the current health state and operating conditions. The confidence interval CI RUL quantifies the uncertainty range of RUL, reflecting the reliability of the prediction result. It provides key information for equipment maintenance decision-making, helping to determine the optimal maintenance time, avoiding both the waste of resources caused by premature maintenance and the equipment failure caused by late maintenance.

[0100] (2) Environment adaptive correction

[0101] ① Environment factor calculation:

[0102] Where, Λ env is the comprehensive environment factor; UV is the ultraviolet intensity; ΔT day is the day-night temperature difference; h is the altitude. The formula considers the effects of ultraviolet intensity, day-night temperature difference and altitude on equipment life, and quantifies the effects of different environmental factors into a comprehensive factor. Among them, ultraviolet light accelerates material aging, day-night temperature difference causes thermal fatigue, and altitude affects heat dissipation and insulation performance. It can be used to correct the basic RUL, so that it is more in line with the equipment life under actual environmental conditions, and improve the prediction accuracy.

[0103] Note: ① Ultraviolet intensity term: UV: ultraviolet radiation intensity (unit: W / m 2 ). Denominator 100: high-altitude strong ultraviolet reference value, 100 W / m 2 is the high-altitude photovoltaic system strong ultraviolet threshold defined by IEC 62446-3. High-altitude characteristics: The ultraviolet intensity on Qinghai-Tibet Plateau in summer can reach 120-150 W / m 2 . Coefficient 0.4: The weight of ultraviolet in total environmental stress (40%), which is based on the analysis of ultraviolet contribution rate in the study of high-altitude photovoltaic backboard aging.

[0104] ② Daily average temperature difference term: ΔT day : The difference between the highest and lowest temperature in 24 hours (unit: ℃), denominator 30: The typical day-night temperature difference on high altitude, 30℃ is the threshold of high-altitude thermal cycle stress defined by GB / T 36545, and the daily average temperature difference on Qinghai-Tibet Plateau is generally in the range of 20-35℃. Coefficient 0.3: The weight of thermal shock in total stress (30%), which is based on the fact that the temperature difference contributes to 32% of the main causes of failure in IGBT module thermal fatigue test.

[0105] ③ Altitude term: h: Altitude (unit: meters), denominator 3000: High-altitude environment starting altitude reference, 3000m is the demarcation line of high-altitude electrical equipment defined by IEC 62804 standard. Coefficient 0.2: The weight of altitude in total stress (20%), which is based on the fact that the direct contribution rate of altitude in high-altitude photovoltaic power station failure statistics is 18-22%.

[0106] ② RUL correction:

[0107] Where, RUL 校正 is the corrected remaining useful life; RUL is the basic remaining useful life; is the corrected RUL confidence interval; CI RUL is the basic RUL confidence interval.

[0108] The basic RUL is corrected by introducing a comprehensive environmental factor. Considering that the plateau environment can accelerate equipment aging, the corrected RUL can more truly reflect the remaining useful life of the equipment in the actual environment. The prediction result better adapts to the special environment such as plateau, avoids the prediction deviation caused by environmental factors, and provides a more accurate basis for maintenance decision of equipment in complex environment.

[0109] ③Plateau extension:

[0110] wherein, σ u′ (t) is the corrected health index standard deviation; σ u (t) is the original health index standard deviation; h is the altitude.

[0111] In the plateau environment, due to the influence of low air pressure and other factors, the uncertainty of the health index is increased. This method corrects the standard deviation by considering the influence of altitude on uncertainty, which reflects the uncertainty change of the prediction result in the plateau environment. It provides more accurate uncertainty quantification results for subsequent uncertainty analysis and decision support, and helps to more comprehensively evaluate the equipment health status and remaining useful life.

[0112] (3) Decision logic

[0113] The decision logic is as follows:

[0114] ① Normal state: when u>0.6 and RUL>1000h and σ total <0.05, the decision action is routine monitoring.

[0115] ② Warning state: when 0.3 total <0.15, the decision action is preparation for maintenance.

[0116] ③ Emergency state: when u≤0.3 and RUL≤500h, the decision action is immediate shutdown.

[0117] ④ Confidence anomaly: when CI width>0.3×RUL, the decision action is manual inspection.

[0118] wherein, u is the health index, RUL is the remaining useful life, σ total is the total standard deviation, and CI width is the confidence interval width.

[0119] Further, the step S6 plateau environment self-adapting mechanism specifically comprises:

[0120] (1) Dynamic adjustment strategy

[0121] The dynamic adjustment strategy is as follows:

[0122] ① Physical constraint weight adjustment: Wherein, a is the physical constraint weight, a0 is the initial physical constraint weight, and h is the altitude. As the altitude increases, the physical constraint weight is increased, the role of physical constraint in the model is strengthened, and the model pays more attention to physical laws to cope with the complexity of the influence of highland environment on equipment.

[0123] ② KL regularization weight adjustment: Wherein, γ is the KL regularization weight, γ0 is the initial KL regularization weight, and h is the altitude. The KL regularization weight is appropriately reduced to reduce the degree of restriction of regularization on the model, avoid the model being too conservative due to the particularity of the plateau environment, and improve the flexibility and adaptability of the model.

[0124] ③ Feature set correction: Wherein, σ T′ is the corrected feature standard deviation, σ T is the original feature standard deviation, and h is the altitude. The influence of factors such as low air pressure on features is compensated to improve the representation ability of features in the plateau environment.

[0125] ④ Degradation rate enhancement: Wherein, Γ′ is the enhanced degradation rate, Γ is the original degradation rate, and h is the altitude. Considering the characteristics of the plateau environment accelerating the degradation of equipment, the degradation rate is enhanced to make the model more accurately reflect the life consumption of equipment in the plateau environment.

[0126] (2) Seasonal strategy

[0127] Summer: UV weight 50%, thermal shock weight 30%.

[0128] Winter: UV weight 30%, thermal shock weight 50%.

[0129] Spring and autumn: UV weight 40%, thermal shock weight 40%.

[0130] A photovoltaic converter reliability monitoring system in a plateau special environment comprises:

[0131] A data acquisition module is used to acquire electrical quantity, temperature quantity, environmental quantity and running quantity data of the plateau converter.

[0132] A data preprocessing module is used to perform outlier processing, stable period selection and signal alignment on the collected data.

[0133] A feature engineering module is used to extract plateau sensitive features and perform physical consistency verification and feature selection.

[0134] Physical constraint modeling module: used to establish physical model and health state evolution equation, calculate physical residual error;

[0135] Bayesian-PINN fusion modeling module: used to build fusion model, set loss function, and quantify uncertainty;

[0136] RUL prediction module: used to predict remaining useful life and make environmental correction, and make maintenance decision;

[0137] Highland environment adaptive module: used to dynamically adjust model parameters and feature weights according to altitude and season.

[0138] Further, the data acquisition module includes Hall sensor, thermocouple and weather station interface, respectively used to collect electrical quantity, temperature quantity and environmental quantity data. The Bayesian-PINN fusion modeling module is realized by using deep learning framework, and has the function of variational inference to estimate the posterior distribution parameters of Bayesian hidden layer weights.

[0139] The specific engineering implementation guide is as follows:

[0140] (1) System deployment

[0141] After the field sensor collects data, it is transmitted to the edge computing node for preliminary processing, then the features are extracted, and then the feature data is sent to the cloud model service for RUL prediction, and finally the prediction result is fed back to the operation and maintenance system.

[0142] (2) Parameter calibration

[0143] The parameter calibration information is as follows:

[0144] E a,IGBT The calibration range is 90-110kJ / mol, and the calibration method is accelerated aging test. Among them, E a,IGBT is the activation energy of IGBT, which is a physical quantity representing the energy barrier that atoms need to overcome in the process of diffusion or migration.

[0145] The calibration range of β is 5.2-6.3, and the calibration method is thermal cycle test. Among them, β is a material constant in the IGBT thermal fatigue model, reflecting the thermal mechanical properties of the material.

[0146] The calibration range of n is 3.0-3.5, and the calibration method is voltage stress test. Among them, n is the aging index in the capacitance aging model, which describes the influence law of voltage on the aging rate of capacitance.

[0147] u threshold The calibration range is 0.15-0.25, and the calibration method is historical failure data analysis. Among them, u threshold is the failure threshold of the health index, which is used to judge whether the equipment has reached the failure state.

[0148] (3) Maintenance strategy

[0149] The maintenance strategy is as follows:

[0150] When RUL>2000h, annual routine inspection is carried out. Among them, RUL is the remaining useful life, indicating the expected time for the device to work normally in the current state

[0151] When 1000h

[0152] When 500h

[0153] When RUL≤500h, weekly inspection is carried out and a shutdown plan is made.

[0154] The beneficial effects of the present application are:

[0155] The present application significantly improves the life prediction accuracy of photovoltaic converters in plateau environment, effectively reduces the operation and maintenance cost, and enhances the equipment operation reliability.

[0156] Specifically,

[0157] 1. Through multi-dimensional data acquisition and fine pretreatment, the data quality and accuracy are significantly improved, providing a high signal-to-noise ratio and high synchronization data basis for state monitoring, effectively reducing the interference of data noise on the monitoring results.

[0158] 2. A system containing multiple sensitive features is constructed and the feature quality is optimized, accurately capturing key information of the device health state in plateau environment, significantly enhancing the feature's representation ability for the device degradation process.

[0159] 3. The health index evolution model is established by fusing IGBT thermal fatigue equation and capacitance aging equation, strictly following the physical law constraint model prediction, significantly improving the model's explanation ability and prediction accuracy for the device health state.

[0160] 4. A network architecture is constructed by fusing data fitting, physical constraints and Bayesian regularization, realizing the deep coupling of data-driven and physical model, significantly improving the model's generalization ability and prediction accuracy under small sample and high noise conditions.

[0161] 5. Considering the prediction uncertainty and introducing an environment adaptive correction mechanism, the prediction results are dynamically adjusted to adapt to complex environmental changes, significantly improving the reliability and accuracy of the remaining useful life prediction, and providing a scientific basis for maintenance decision-making.

[0162] 6. The model parameters and feature weights are automatically adjusted according to the altitude and seasonal changes, significantly enhancing the adaptability of the monitoring system to the special environment of the plateau, and ensuring that the system can maintain high-performance monitoring effects under different environmental conditions. BRIEF DESCRIPTION OF DRAWINGS

[0163] Figure 1 Overall method implementation flowchart.

[0164] Figure 2 Overall method implementation technology roadmap.

[0165] Figure 3 Data acquisition and preprocessing flowchart.

[0166] Figure 4 Plateau sensitive feature engineering flowchart.

[0167] Figure 5 Bayesian-PINN fusion modeling flowchart.

[0168] Figure 6 RUL prediction and decision flowchart. DETAILED DESCRIPTION

[0169] The present application will be further illustrated below in conjunction with the drawings and specific embodiments, and it should be understood that the following specific embodiments are only used to illustrate the present application and not to limit the scope of the present application.

[0170] As shown in the figure, the present application provides a photovoltaic converter reliability monitoring method under special plateau environment, which comprises the following steps:

[0171] 1. Data acquisition and preprocessing implementation

[0172] (1) Sensor selection and deployment

[0173] According to the different characteristics of electrical quantity, temperature quantity, environmental quantity and operation quantity, appropriate sensors are selected. For the collection of DC bus voltage (V dc ) and AC output current (I ac ), high-precision Hall sensors are used, which should have high sampling frequency (10 kHz) response capability, can accurately capture the rapid change of electrical quantity, and have good linearity and stability, to ensure the reliability of the collected data.

[0174] For the measurement of IGBT junction temperature (T j ) and heat sink temperature (T hs ), thermocouples are used, which have fast response time and high precision, can adapt to the real-time monitoring demand of temperature change, the sampling frequency is 1 Hz, which meets the dynamic tracking requirement of temperature, and its temperature measurement range covers the normal operation and possible high temperature working condition of the equipment, to ensure the effectiveness of the data.

[0175] The environmental quantity data such as ultraviolet intensity (UV), altitude (h), air pressure (P) and the like are provided by a weather station, which is equipped with professional environmental monitoring sensors and has high-precision and high-reliability data acquisition capability, with a sampling frequency of 0.1 Hz, and can stably obtain environmental information within a certain time.

[0176] The output power (P loss ) and switching frequency (f sw ) in the running quantity data are from a SCADA system connected to the control system of the device, which can accurately collect these running parameters through the built-in power monitoring module and switching frequency counter, with a sampling frequency of 1 kHz, to ensure real-time control of the running state of the device.

[0177] (2) Data acquisition system construction

[0178] A distributed data acquisition system is constructed, connecting various sensors to data acquisition cards, which have multi-channel high-precision A / D conversion function, can simultaneously collect multiple electrical signals, and ensure the stability and synchronization of the sampling frequency; at the same time, the acquisition card is optimized for the characteristics of temperature signals, with cold-end compensation and filtering function, to improve the collection accuracy of temperature data; environmental quantity and running quantity data are connected to the system through the corresponding interface module, and all acquisition cards are connected to the host computer through the industrial bus, to realize centralized collection and transmission of data.

[0179] (3) Implementation of abnormal value processing algorithm

[0180] An abnormal value processing program based on the 3σ criterion is written, which first calculates the mean (μ) and standard deviation (σ) of each group of data, and for each data point x, judges whether it satisfies |x-μ|>3σ, if it satisfies, it is determined as an abnormal value and is removed.

[0181] For the processing of power dramatic change section, the power change rate is calculated, that is, the power change rate is obtained by numerically differentiating the output power (P loss ) data, and then compared with 0.1 times the rated power (0.1P rated ), if it exceeds the threshold, the data in this section is removed. In the implementation process, the sliding window method can be used to calculate the power change rate, the window length can be selected according to the sampling frequency of the actual data and the power change characteristics, for example, for power data with a sampling frequency of 1 kHz, a sliding window of 10 sampling points can be selected to calculate the power change rate, which can smooth the influence of noise and accurately detect the power dramatic change section.

[0182] After removing outliers, the data is interpolated to ensure the continuity of the data over time. Methods such as linear interpolation or spline interpolation can be used. The appropriate interpolation algorithm is selected according to the characteristics of the data and the extent of missing data. For example, linear interpolation can quickly fill in data gaps for occasional single outliers, while spline interpolation can provide a smoother transition for data gaps over a longer period of time, ensuring that the interpolated data can reflect the actual operating status of the equipment.

[0183] (4) Stable period selection strategy

[0184] Write a program to determine light intensity, extract light intensity information from environmental data provided by a weather station, and set a threshold of 800 W / m². 2 When the light intensity remains above the threshold for 10 consecutive minutes (based on true solar time), it is marked as the starting point of a possible stable period, and the corresponding start timestamp is recorded.

[0185] Calculate the power fluctuation rate within a 10-minute time window, for the output power (P) loss The data is statistically analyzed to calculate the ratio of its standard deviation to its mean, i.e., the power fluctuation rate. When the power fluctuation rate is less than 5%, combined with the judgment result of the light intensity, the period is determined to be a stable period, and the data of the period is extracted for subsequent feature engineering analysis. If the power fluctuation rate requirement is not met, the sliding time window is continued to make the next judgment until a stable period that meets the conditions is found.

[0186] The time range of the data for the stable period is limited to 10:00-14:00 (true solar time). This is based on the operating patterns and environmental characteristics of solar energy equipment. During this period, the equipment is usually in a relatively stable operating state and is less affected by external interference, so the collected data is more representative. By comparing with the timestamp information, the stable period data that meets the time range requirements is selected as the basic dataset for subsequent feature extraction and analysis.

[0187] (5) Application of signal alignment algorithm

[0188] Dynamic Time Warping (DTW) algorithm is used to timestamp and calibrate environmental and electrical data. First, the two sets of data are represented as two time series, for example, the environmental data series is [a1, a2, ..., a...]. n The electrical data sequence is [b1, b2, ..., b n ], where n and m are the number of sampling points for the two sets of data, and n ≠ m due to the different sampling frequencies.

[0189] A distance matrix d(i,j) is defined, where d(i,j) represents the distance between the i-th sample point of the environmental data and the j-th sample point of the electrical data, which can be calculated using the Euclidean distance or other suitable distance measurement methods, for example, d(i,j) = |a i -b j |, and then according to the recursive formula of the DTW algorithm, the cumulative distance matrix is calculated, and the path that minimizes the cumulative distance is found, which is the optimal time alignment path between the two sets of data.

[0190] According to the optimal time alignment path, the environmental data and the electrical data are interpolated or resampled to have the same timestamp sequence, thereby realizing the alignment of the signals. The interpolation method can be linear interpolation, polynomial interpolation, etc. according to the data characteristics, to ensure that the aligned data can accurately reflect the running state of the equipment under different environmental conditions, and provide accurate time correspondence for subsequent feature extraction and analysis, improving the quality and reliability of feature engineering.

[0191] 2. High-sensitivity feature engineering implementation

[0192] (1) Voltage ripple rate calculation

[0193] From the preprocessed DC bus voltage data (V dc ), the root mean square value (RMS) is calculated, which is calculated according to the definition formula of the root mean square value, that is: Where N is the number of voltage data sampling points, is the average value of the voltage data. The root mean square value is calculated to quantify the degree of voltage fluctuation and reflect the filtering effect of the capacitor.

[0194] Get the altitude (h) information of the current device location, which can be obtained from the environmental data provided by the weather station or real-time obtained through the GPS positioning module, and substitute it into the voltage ripple rate calculation formula: Calculate the voltage ripple rate R r , which can comprehensively reflect the aging degree of the capacitor and the influence of the plateau low-pressure environment on the capacitor performance, providing key information for subsequent health state evaluation.

[0195] (2) High-frequency harmonic energy ratio calculation

[0196] The current or voltage signal during IGBT switching is subjected to Fourier transform, and the fast Fourier transform (FFT) algorithm can be used to convert the time domain signal to the frequency domain signal to obtain its power spectral density (PSD(f)).

[0197] Determine the switching frequency (f sw) and its frequency range, the switching frequency is determined according to the specification parameters of the device or through spectrum analysis, and the switching frequency is usually between several hundred Hz to several kHz, for example, for a certain photovoltaic inverter, the switching frequency is 20 kHz, then the high frequency band is defined as 2 times the switching frequency to 5 times the switching frequency, that is, 40 kHz to 100 kHz, and the low frequency band is defined as 0 to the switching frequency, that is, 0 to 20 kHz.

[0198] The harmonic energy of the high frequency band and the low frequency band is calculated, and the power spectrum density in the high frequency band and the low frequency band is integrated, respectively, which can be realized by numerical integration method, such as trapezoidal rule or Simpson rule, to calculate the high frequency harmonic energy and the low frequency harmonic energy Then according to the formula: The high frequency harmonic energy ratio E h This feature can effectively reflect the degradation of IGBT switching characteristics, and provides an important basis for evaluating the health state of IGBT.

[0199] (3) Junction temperature fluctuation entropy calculation

[0200] From the IGBT junction temperature data (T j ) collected by the temperature sensor, the probability distribution p(T j ) of the junction temperature is calculated, which can be calculated by histogram method or kernel density estimation method, and the junction temperature data is divided into several intervals, and the frequency of data in each interval is calculated to obtain the probability distribution function of the junction temperature.

[0201] The junction temperature fluctuation entropy is calculated according to the formula: , wherein P loss0 is the reference power, and P loss is the actual output power. The reference power can be determined according to the rated power of the device, for example, for a device with a rated power of 10 kW, P loss0 is 10 kW, and the actual output power is obtained from the SCADA system. By calculating the junction temperature fluctuation entropy, the complexity of the junction temperature fluctuation can be quantified, and the thermal fatigue damage can be reflected, which provides a key indicator for the health state evaluation of the device.

[0202] (4) Thermal response time calculation

[0203] The IGBT junction temperature function (T j (t)) and the output power function (P loss (t)) are obtained, and the corresponding data sequences are collected from the temperature sensor and the SCADA system, respectively, and are converted into time sequence function form. In the data analysis software, the data can be preprocessed by using time series analysis tools, such as noise removal and smoothing processing, to improve the data quality.

[0204] The correlation between junction temperature and output power is calculated, and Pearson correlation coefficient or other correlation indicators are used to calculate T j (t) and P loss (t) at different time lags, find the time point t with the maximum correlation rate, that is: This time point is the thermal response time. By analyzing the thermal response time, the response speed of the heat dissipation system to power changes can be evaluated, and the performance state of the heat dissipation system can be judged.

[0205] Considering the influence of altitude on thermal response time, multiply the calculated thermal response time by the altitude correction coefficient e 0.0001h , to get the final thermal response time τ th , the altitude data is obtained from the weather station, through this correction method, the influence of high altitude environment on the performance of heat dissipation system can be more accurately reflected, and more reliable data support for equipment maintenance decision can be provided.

[0206] (5) Thermal shock coefficient calculation

[0207] Obtain the day-night temperature difference (ΔT day ) data from the weather station, the weather station will provide the highest temperature and the lowest temperature every day, and the day-night temperature difference can be obtained by calculating the difference between the two, and record the number of thermal cycles (N cycle ), the number of thermal cycles can be counted according to the running time of the equipment and the change rule of local day-night temperature difference, for example, for the equipment running in the wild for a long time, every time it experiences a day-night alternation, it can be counted as a thermal cycle, through the temperature recorder installed on the equipment site to record the temperature change curve, and automatically count the number of thermal cycles. According to the formula:

[0208] Calculate the thermal shock coefficient K ts , which considers the influence of day-night temperature difference and number of thermal cycles on equipment, and can effectively quantify the damage degree of thermal shock to equipment, providing important reference basis for equipment life prediction and maintenance decision.

[0209] (6) Comprehensive stress index calculation

[0210] Collect ultraviolet intensity (UV), thermal shock coefficient (K ts ) and altitude (h) data, among which ultraviolet intensity is obtained from the weather station, thermal shock coefficient is calculated by the above steps, and altitude data comes from the weather station or GPS positioning module. According to the formula: The comprehensive stress index Lambda is calculated by weighting and summing the ultraviolet intensity, thermal shock coefficient and altitude, so as to obtain a comprehensive stress index capable of comprehensively reflecting the total damage degree of the equipment in the plateau environment. The index provides comprehensive environmental factor consideration for subsequent environmental self-adaptive correction and equipment health state evaluation, and helps to improve the accuracy and reliability of the prediction results.

[0211] (7) Physical consistency verification and feature selection

[0212] Verify whether the change of the ripple ratio of the capacitor with temperature meets the temperature characteristics of the capacitor, that is, verify where T c is the shell temperature of the capacitor. By collecting voltage ripple ratio data at different temperatures, the R r vs. T c scatter plot is established, and linear regression analysis is performed to observe the sign of the regression coefficient. If the regression coefficient is negative, it indicates that the ripple ratio decreases with the increase of temperature, which meets the temperature characteristics of the capacitor. If it does not meet, the feature extraction process needs to be checked and corrected to ensure the physical rationality and reliability of the feature.

[0213] Verify the degree of interference of the ambient temperature on the junction temperature fluctuation entropy, that is, verify |corr(S T ,T a )|<0.3, where T a is the ambient temperature, and the correlation between the junction temperature fluctuation entropy S T and the ambient temperature T a is calculated. If the absolute value of the correlation is greater than 0.3, it indicates that the ambient temperature has a greater impact on the junction temperature fluctuation entropy, and further analysis and corresponding correction measures need to be taken, for example, an ambient temperature compensation term is introduced in the feature extraction process to improve the accuracy and representativeness of the feature.

[0214] Three-stage feature selection is performed. First, the variance of each feature is calculated, and features with a variance less than a threshold value are removed. For example, if the value of a certain feature changes little in all samples, and its variance is lower than the threshold value, it is considered that the feature has weak discrimination ability for the health state of the equipment, and can be removed from the feature set. Then, the remaining features are screened according to physical laws and professional knowledge, and features that meet the physical laws are retained, for example, features that are irrelevant to the aging mechanism of the equipment or contradict the physical model are removed. Finally, the recursive feature elimination (RFE) algorithm is used to sort and select the features based on the prediction performance of the Bayesian-PINN model, and the features with the smallest contribution to the model prediction are gradually removed until the optimal feature subset is obtained. The optimal feature number and feature combination are determined through cross-validation, and finally the selected feature vector x is obtained: x = [R r , E h , S T , τ thΛ, K ts ] T , provide simplified and effective feature inputs for subsequent modeling and prediction.

[0215] 3. Physical constraint modeling implementation

[0216] (1) Core physical equation parameter determination

[0217] IGBT thermal fatigue model: Determine the value range of each parameter through experiment calibration. Determine constant A: refer to the reliability data and thermal fatigue life test results provided by IGBT manufacturers, combine with accelerated life test method, conduct life test on IGBT under different junction temperature changes (ΔT j ) and average temperature (T m ) conditions, record its failure time (N f ), determine the value range of constant A through nonlinear regression fitting. Calibration of material constant β: use thermal cycle test method to conduct junction temperature change cycle test of different amplitudes on IGBT samples, record the failure cycle number, and according to the test data, fit to obtain the value of β, usually the value range of β is between 5.2-6.3, which is closely related to the thermal mechanical properties of the material. Calculation of activation energy E a : through analyzing the failure mechanism and physical and chemical process of IGBT, combining with Arrhenius equation, according to the experimental data, determine the value range of activation energy E a is 90-110kJ / mol.

[0218] For capacitor aging model: Determine each parameter through experiment and data analysis. Determine reference life L0: according to the data provided by capacitor manufacturers and related standard test methods, conduct long-term aging test on capacitors under standard temperature (such as 25℃) and rated voltage conditions, record the failure time of capacitors, and determine the reference life L0. Calibration of aging index n: use voltage stress experiment, apply different multiples of rated voltage (such as 1.1 times, 1.2 times, etc.) to capacitors, record the failure time, and according to the model fitting, obtain the value of aging index n, usually the value range of n is between 3.0-3.5, for example, through experimental data analysis, the value of n is 3.2, the aging index reflects the influence degree of voltage on the aging rate of capacitor, that is, the higher the voltage, the faster the capacitor ages, and the relationship is exponential. Determination of capacitor activation energy E a : combine the material properties and aging mechanism of capacitors, conduct accelerated aging test at different temperatures, record the failure time of capacitors, and use Arrhenius equation to fit to obtain the value of E a .

[0219] (2) Health state evolution equation construction

[0220] The health indicator u(t) is defined to have a range of values ​​[0,1]. The initial time u(0) = 1 indicates that the equipment is brand new and undamaged, and the failure time u(t) = 1 indicates that the equipment is undamaged. fail A degradation rate of 0.2 indicates that the equipment has reached the end of its lifespan and requires maintenance or replacement. This definition is based on the equipment's failure criteria and engineering experience. Establish a degradation rate model:

[0221] ① Calculate the IGBT degradation rate: The specific values ​​of each parameter are calculated based on the calibration results. For example, C1 is determined according to the IGBT manufacturing process and material properties, and E... a,IGBT Take 100 kJ / mol, R as the gas constant 8.314 J / (mol·K), T m ΔT represents the average operating temperature of the IGBT (e.g., 80℃, which needs to be converted to Kelvin 353.15K). j For junction temperature changes (e.g., 30K), β is taken as 5.8, K. ts Let Γ be the thermal shock coefficient (e.g., 0.5). Substituting this value into the formula, we can obtain Γ. IGBT The value reflects the degradation rate of the IGBT under current operating conditions.

[0222] ② Calculate the capacitance degradation rate: For example, C2 is determined according to the capacitor's manufacturing process, V is the actual voltage the capacitor withstands (e.g., 400V), V0 is the reference voltage (e.g., 380V), n is taken as 3.2, and E... a,Cap Take 80 kJ / mol, T as the capacitor's operating temperature (e.g., 60℃, converted to 333.15 K), I uv Substituting the ultraviolet intensity index (e.g., 0.8) into the formula, we obtain Γ. Cap The value quantifies the aging rate of the capacitor under electrical stress, thermal stress, and ultraviolet radiation.

[0223] Γ IGBT and Γ Cap Add them together and take the negative value to get the rate of change of the health indicators. The evolution of health indicators over time is calculated using numerical integration methods (such as the Euler method or the Runge-Kutta method) to obtain the time series of u(t). This series can intuitively reflect the dynamic changes in the health status of the equipment, providing basic data for subsequent RUL prediction.

[0224] ③ Calculate the physical residual: By comparing the rate of change of health indicators predicted by the actual model Compared with the degradation rate (Γ) calculated based on the physical model IGBT +Γ Capbetween the difference, get the physical residual, if the physical residual is large, the model prediction result and the physical law exist deviation, need to adjust and optimize the model, for example, check the accuracy of parameter value, the rationality of model structure, etc., by constantly reducing the physical residual, ensure that the Bayesian-PINN model is driven by data at the same time, meet the physical constraints, improve the credibility and reliability of the prediction result.4. Bayesian-PINN fusion modeling implementation

[0225] (1) Network architecture building

[0226] The Bayesian-PINN model is built using a deep learning framework. The input layer of the model receives time t and feature vector x (including the above selected six features: R r ,E h ,S T ,τ th ,Λ env ,K ts ) as input. The number of neurons in the input layer is determined according to the dimension of the feature vector.

[0227] The Bayesian hidden layer is built. The Bayesian neural network is realized by using the variational inference method. The probability layer is used instead of the traditional fully connected layer. Bayesian hidden layer 1 and Bayesian hidden layer 2 contain 128 neurons. The ReLU function is selected as the activation function to increase the expression ability of the model. The weights w of the Bayesian hidden layer are subject to normal distribution In the model training process, the posterior distribution parameters (mean μ w and variance ) of the weights are estimated by the variational inference method.

[0228] The output layer outputs the health index u(t). The output layer uses a linear activation function to ensure that the output value is within the range [0, 1]. The output health index u(t) is input into the physical constraint module, combined with the physical constraint condition, and used as part of the loss function for model training. The physical constraint module calculates the time derivative of the health index and the degradation rate (Γ IGBT + Γ Cap ) calculated according to the physical model to obtain the physical residual, which is used to build the physical constraint term loss function.

[0229] (2) Loss function setting and optimization

[0230] Data fitting term loss function: Collect labeled historical data, including the actual health index values (u true ) of the device at different time points, which can be obtained through actual device maintenance records, laboratory tests or simulation data generation, etc. The health index predicted by the model (u pred) with the actual values, and take the average to get

[0231] Physical constraint term loss function: The construction of the square norm of the physical residual ‖PDE(u)‖ 2 , where: The ReLU(u t+1 -u t ) term is calculated, and the adjacent time points t and t+1 in the time series data are traversed to calculate the change amount u(t+1)-u(t) of the health indicator. If the change amount is positive (i.e. the health indicator rises), it is converted to a non-zero value (such as taking the positive value itself) by the ReLU function, otherwise it is taken as zero. The sum of these non-zero values is taken to get this part of the loss. This constraint term is used to prevent the health indicator from increasing non-physically and ensure that the model's predicted health state evolution conforms to the physical law. The weight coefficients β1 and β2 can be determined by experience or through hyperparameter search, such as initially setting β1=0.1 and β2=0.01, and adjusting them according to the model training effect later.

[0232] Bayesian regularization term loss function: The calculation of the square norm of the physical residual ‖PDE(u)‖ The prior distribution p(w) is assumed to be a normal distribution, and the posterior distribution q(w) is determined by the variational inference parameters of the Bayesian hidden layer. In the deep learning framework, the KL divergence calculation function provided by is used to calculate This regularization term is used to quantify the uncertainty of the model parameters and prevent overfitting of the model, which is particularly important in small sample learning.

[0233] The above three parts of the loss function are added to get the total loss function: The model is trained by selecting a suitable optimization algorithm (such as the Adam optimizer) to adjust the model parameters (including the weight posterior distribution parameters of the Bayesian hidden layer and other parameters of the network) to minimize the total loss function. During the training process, the downward trend of each loss is monitored, as well as the prediction performance on the validation set, such as prediction accuracy and uncertainty quantification effect. The hyperparameters (such as α, γ, β1, β2, etc.) are adjusted as needed to obtain the best model performance.

[0234] (3) Uncertainty quantification implementation

[0235] Monte Carlo sampling: sample multiple sets of model parameters w i (i=1,2,...,M, M=100) from the posterior distribution q(w) i , use the Bayesian neural network (BNN) to perform forward propagation on the input feature vector x to get the corresponding health indicator prediction value ui (t) was sampled 100 times to obtain 100 predicted values. These predicted values ​​reflect the impact of the uncertainty of the model parameters on the prediction results. By simulating the possible range of parameter values ​​through multiple samplings, basic data is provided for subsequent uncertainty analysis.

[0236] Statistical calculation: Based on the predicted value u obtained from sampling i (t), calculate the mean: Sum of standard deviation: The mean reflects the central trend of the prediction, that is, the average predicted value of the health indicator after considering the uncertainty of the parameters. The standard deviation quantifies the degree of uncertainty of the prediction. The larger the standard deviation, the more uncertain the model's prediction of the health indicator is, which may be affected by data noise, model parameter uncertainty or model structure limitations. These two statistics can concisely describe the distribution characteristics of the prediction results and provide a basis for subsequent confidence interval estimation and decision support.

[0237] Confidence interval construction: Based on the mean and standard deviation, according to the formula: A 95% confidence interval is constructed, where 1.96 is the critical value of the normal distribution corresponding to the 95% confidence level. In practical applications, different confidence levels (such as 90% or 99%) can be selected as needed, and the critical value can be adjusted accordingly. This confidence interval provides an uncertainty range for RUL prediction, helping decision-makers assess the reliability of the prediction results. For example, when the confidence interval is narrow, the prediction results are more reliable, and a more accurate maintenance plan can be formulated accordingly; while when the confidence interval is wide, it indicates that there is greater uncertainty, which may require more data or further model optimization, or even trigger manual inspection to deal with potential equipment failure risks.

[0238] 5. Implementation of RUL Prediction Process

[0239] (1) Calculation of conversion of health indicator u-value to RUL

[0240] Failure time calculation: based on the average of health indicators and standard deviation σ u (t), calculate the estimated time for the equipment to reach the failure threshold: By iterating through the mean of health indicators in the time series, find the first one that satisfies... The time point t is the expected failure time. For example, in the predicted time series of health indicator mean values, it is found that when t = 1000 hours, Then t fail = 1000 hours. Similarly, calculate the earliest failure time considering uncertainties: And the latest expiration time: For example, if at t = 800 hours, Then At t = 1200 hours, Then At t = 1200 hours, these failure times take into account the uncertainty of the prediction, giving a range of time in which the device is likely to fail, providing more comprehensive information for subsequent maintenance decisions.

[0241] Basic RUL calculation: Obtain the current time t current For example, assume the current device has been running for 500 hours, then t current = 500 hours. Calculate the basic remaining useful life: RUL = t fail - t current = 1000 - 500 = 500 hours. This indicates that without considering environmental factor correction, the device is expected to operate normally for another 500 hours.

[0242] Calculate the confidence interval of the basic RUL: hours. This confidence interval reflects the uncertainty range of the basic RUL, i.e. there is a 95% probability that the remaining useful life of the device is between 300 and 700 hours, providing a risk assessment basis for maintenance decisions, for example, maintenance personnel can reasonably arrange maintenance plans according to this interval to ensure that maintenance is completed before the device fails, while avoiding the waste of resources caused by premature maintenance.

[0243] (2) Environment adaptive correction implementation

[0244] Environmental factor calculation: Collect ultraviolet intensity (UV), diurnal temperature difference (ΔT day ) and altitude (h) data, for example, assume that the ultraviolet intensity of the location where the device is located is 120 (the unit is determined according to the specific measuring instrument), the diurnal temperature difference is 35℃, and the altitude is 3600 meters. Substituting into the formula: Calculate to get:

[0245] This comprehensive environmental factor reflects the comprehensive influence of ultraviolet, diurnal temperature difference and altitude on the device life under the current environment, the larger the value, the more serious the adverse effects of the environment on the device life, and the greater the correction of the basic RUL.

[0246] RUL correction: According to the basic RUL (such as 500 hours) and the comprehensive environmental factor (0.653), calculate the corrected remaining useful life: hours, which indicates that after considering the highland environmental factors, the actual remaining useful life of the device is expected to be 467.3 hours, which is due to the complex effects of low air pressure and other factors in the highland environment on the aging process of the device.

[0247] Similarly, calculate the corrected RUL confidence interval: The corrected confidence interval also reflects the impact of environmental factors on RUL uncertainty, providing decision-makers with a more realistic range of equipment life prediction under actual environmental conditions, helping to develop more reasonable and reliable maintenance strategies.

[0248] High altitude extension: Obtain the altitude of the device's location (e.g., 2000 meters), and calculate the corrected health index standard deviation based on the original health index standard deviation (assumed to be 0.05): This step considers the impact of low air pressure and other factors in high-altitude environments on the uncertainty of health index prediction, quantifies the additional uncertainty brought by high-altitude environments, and provides a more accurate data basis for subsequent uncertainty analysis and decision support. For example, when evaluating equipment health status and developing maintenance plans, this increased uncertainty needs to be considered to ensure the robustness and reliability of decisions.

[0249] (3) Decision logic implementation

[0250] Write a decision logic program based on the current values of health index (u), remaining useful life (RUL), total standard deviation (σ total ), and confidence interval width (CI width) according to the following rules:

[0251] ① Normal state: When u > 0.6 and RUL > 1000h and σ total < 0.05, the decision action is routine monitoring. At this time, the equipment health status is good, the remaining useful life is long, and the uncertainty is low. Only routine monitoring (such as monthly or quarterly) is required for equipment inspection and data collection, recording equipment operating status, and no additional maintenance operations are required.

[0252] ② Warning state: When 0.3 < u ≤ 0.6 and 500h < RUL ≤ 1000h and σ total < 0.15, the decision action is to prepare for maintenance, indicating that the equipment health status is starting to decline, the remaining useful life is gradually decreasing, and the uncertainty is within an acceptable range. At this time, you should start preparing for the necessary supplies and personnel, develop a maintenance plan, and increase the monitoring frequency of the equipment (such as data collection and analysis every two weeks) to timely grasp the changes in equipment status and ensure that maintenance is completed before the equipment fails.

[0253] ③ Emergency state: when u ≤ 0.3 and RUL ≤ 500h, the decision action is to shut down immediately, which means that the device health status is seriously deteriorated and is about to face the risk of failure, and the device must be stopped running immediately to avoid more serious consequences such as device damage, safety accidents or power supply interruption caused by device failure, etc. Before executing the shutdown operation, necessary safety measures should be taken and professional personnel should be arranged to handle the scene as soon as possible.

[0254] ④ Confidence anomaly: when CI width > 0.3 × RUL, the decision action is manual inspection, which indicates that the uncertainty of the prediction result is too large and may affect the reliability of the decision. At this time, the operation and maintenance personnel should be dispatched to the scene to manually inspect the device and verify the actual running state of the device, find out the reasons for the increase of uncertainty, such as sensor failure, data transmission error, environmental mutation, etc., and take corresponding measures according to the scene, such as repairing the sensor, re-collecting data, adjusting model parameters, etc., to improve the prediction accuracy and the credibility of the decision.

[0255] In practical application, the above decision logic is integrated into the device health status monitoring system, by real-time acquisition of prediction results (including u, RUL, σ total , CI width, etc.), automatically triggering the corresponding decision action, and generating maintenance work order or alarm information, sending to the operation and maintenance personnel, guiding them to carry out corresponding operation, realizing the automation and intelligentization of device maintenance decision, improving the operation and maintenance efficiency and device reliability.

[0256] 6. High altitude environment adaptive mechanism implementation

[0257] (1) Dynamic adjustment strategy implementation

[0258] Physical constraint weight adjustment: according to the altitude (h) of the device location, according to the formula: Adjust the physical constraint weight α, for example, the initial physical constraint weight α0 is set to 1.0, when the device is at an altitude of 3000 meters, α = 1.0 × (1 + 0.1 × 3000 / 3000) = 1.1 is calculated, which shows that with the increase of altitude, the physical constraint weight increases, which strengthens the role of physical constraint in model training and prediction, makes the Bayesian-PINN model pay more attention to physical law, in order to cope with the complexity of the influence of high altitude environment on the device, improve the adaptability and prediction accuracy of the model under high altitude conditions.

[0259] KL regularization weight adjustment: according to the altitude h, use the formula: Adjust the KL regularization weight γ, for example, the initial KL regularization weight γ0 is 0.1, when the altitude is 3000 meters, γ = 0.1 / (1+0.05×3000 / 3000) = 0.1 / 1.05≈0.095 is calculated, appropriately reduce the KL regularization weight, reduce the degree of restriction of the model by the Bayesian regularization, avoid the model too conservative due to the particularity of the plateau environment, improve the flexibility and adaptability of the model, so that it can better fit the device running data under the plateau environment, while maintaining certain generalization ability, prevent overfitting phenomenon.

[0260] Feature set correction: according to the original feature standard deviation (σ T ) and altitude (h), according to the formula: Correct the feature standard deviation, for example, the original feature standard deviation σ T is 0.1, when the altitude is 2000 meters, the corrected feature standard deviation σ T′ = 0.1×(1+0.05×2000 / 1000) = 0.1×1.1 = 0.11 is calculated, in this way, compensate for the influence of low air pressure and other factors on the feature, improve the representation ability of the feature in the plateau environment, ensure that the feature can accurately reflect the running state of the equipment under the plateau condition, so as to provide more reliable data input for model training and prediction, improve the model performance.

[0261] Degradation rate enhancement: according to the original degradation rate (Γ) and altitude (h), using the formula: Enhance the degradation rate, for example, the original degradation rate is 0.01 / hour, when the altitude is 3000 meters, the enhanced degradation rate Γ' = 0.01×(1+0.005×3000 / 1000) = 0.01×1.015 = 0.01015 / hour is calculated, considering the characteristics of the plateau environment accelerating the degradation of the equipment, by enhancing the degradation rate, the model can more accurately reflect the life consumption of the equipment under the plateau environment, so as to improve the accuracy and reliability of RUL prediction, provide more actual environmental conditions for the maintenance decision of the equipment.

[0262] (2) Seasonal strategy implementation

[0263] According to the seasonal change of the region where the equipment is located, adjust the weight distribution of ultraviolet and thermal shock in the comprehensive stress index.

[0264] Summer strategy: Set the UV weight to 50% and the thermal shock weight to 30%. This is because the UV intensity is higher in summer, and the impact of equipment material aging is more significant. At the same time, the day-night temperature difference is relatively small, and the impact of thermal shock is relatively weakened. By increasing the weight of UV, the main influencing factor of UV on equipment life in summer is highlighted, so that the model can more accurately assess the health status and remaining useful life of the equipment in summer, providing targeted guidance for equipment maintenance decisions in summer, such as strengthening sun protection measures and material aging detection for equipment before summer arrives, and reasonably arranging summer maintenance plans.

[0265] Winter strategy: Set the UV weight to 30% and the thermal shock weight to 50%. In winter, due to the large day-night temperature difference, thermal shock causes more serious fatigue damage to the equipment, while the UV intensity is relatively low. Therefore, by increasing the weight of thermal shock, the key influence of day-night temperature difference on equipment life in winter is emphasized. The model will focus more on considering the thermal shock factor in winter, timely discovering equipment damage caused by thermal fatigue, and guiding winter maintenance work, such as strengthening equipment insulation measures and thermal cycle monitoring to prevent equipment failure caused by thermal shock.

[0266] Spring and autumn strategy: Set the UV weight and thermal shock weight to 40%. The environmental conditions in spring and autumn are relatively moderate, with moderate UV intensity and day-night temperature difference, so both factors have a balanced impact on equipment life. Therefore, give the same weight to both factors to ensure that the model can consider the impact of both factors on equipment, and ensure balanced equipment health status assessment in spring and autumn. For example, conduct regular equipment inspection and maintenance in spring and autumn, and reasonably arrange maintenance time and content based on model prediction results to ensure stable operation of equipment during seasonal transition.

[0267] 7. Engineering implementation guide implementation

[0268] (1) System deployment implementation

[0269] Build on-site sensor network: Install various sensors (such as Hall sensors, thermocouples, UV sensors, etc.) on devices such as photovoltaic inverters to ensure that the installation location of the sensors is reasonable and can accurately collect the required electrical, temperature, environmental and operating data. The installation of sensors should comply with relevant standards and the requirements of equipment manufacturers, for example, temperature sensors should be installed in a location that can directly reflect the IGBT junction temperature and heat sink temperature, avoiding external interference, and at the same time, good protection measures should be taken for the sensors to ensure their long-term stable operation in harsh outdoor environments.

[0270] Deploying edge computing nodes: Deploy edge computing devices such as industrial-grade computers or edge computing servers near the device site, which have sufficient computing power and data storage capacity, can receive data collected by on-site sensors, and perform preliminary data processing, including data cleaning, outlier detection, and feature extraction, etc. Edge computing nodes are connected to sensors through wired (such as industrial Ethernet) or wireless (such as 4G / 5G) communication methods to ensure real-time and reliable data transmission. At the same time, the edge computing node is also responsible for sending the processed feature data to the cloud model service for further RUL prediction analysis.

[0271] Cloud model service deployment: Deploy Bayesian-PINN model services on cloud servers, which have high-performance computing resources and large-capacity data storage to meet the needs of model training and large-scale data prediction. Model services interact with edge computing nodes through RESTful API or other communication protocols, receive feature data and return RUL prediction results. At the same time, the cloud server is also responsible for storing and managing historical data, model parameters, and maintaining decision records, etc. to provide data support for the operation and optimization of the entire system.

[0272] Operation system integration: Feedback the RUL prediction results to the operation and management system (such as the enterprise's device management system or a dedicated photovoltaic power station operation software). The operation and management system automatically generates maintenance work orders, alarm information and maintenance suggestions based on the prediction results and decision logic, and pushes them to the operation personnel. Operation personnel can view the health status, remaining useful life, maintenance priority and other information of the device through the operation system, and make corresponding maintenance plans and work arrangements. At the same time, the operation system can also record the execution of maintenance work and the historical maintenance records of the device, providing data support for the continuous optimization of the model and the whole life cycle management of the device, and realizing the closed-loop management of device operation.

[0273] (2) Maintenance strategy implementation

[0274] According to the predicted remaining useful life (RUL), formulate the corresponding maintenance strategy:

[0275] When RUL>2000 hours, perform annual routine inspection, including appearance inspection, electrical connection inspection, cooling system inspection, control system calibration, etc. to ensure that the device can maintain good performance after a long period of operation. At the same time, record the operation data and inspection results of the device, update the health file of the device, and provide data support for subsequent maintenance decisions. Annual routine inspection is usually arranged during the low period of device operation or during planned power outage to reduce the impact on power supply.

[0276] When 1000 hours < RUL < 2000 hours, quarterly inspection is carried out, and the quarterly inspection is based on the annual inspection, and the performance test of key components and the replacement of part components (such as consumables) are added, for example, the function test of the driving circuit of the IGBT, the capacitance value test of the capacitor, and the replacement of the filter of the cooling system, etc., through the quarterly inspection, the potential fault hidden danger of the equipment can be found in time, the measures are taken in advance to repair or replace, the service life of the equipment is prolonged, the reliable operation of the equipment is ensured, and the quarterly inspection can be flexibly selected at a suitable time within the quarter according to the actual operation of the equipment and the power grid scheduling, and interference on the operation of the equipment is minimized.

[0277] When 500 hours < RUL < 1000 hours, monthly inspection and preparation of spare parts are carried out, the monthly inspection is more detailed and in-depth, including comprehensive inspection, cleaning and fastening of internal components of the equipment, and according to the health state prediction result and the fault mode analysis of the equipment, corresponding spare parts (such as IGBT modules, capacitors, circuit boards, etc.) are prepared to ensure that the equipment can be repaired and replaced in time when a fault occurs, and downtime is reduced, and the monthly inspection requires that the operation and maintenance personnel have high technical level and maintenance ability, can accurately judge the fault position and reason of the equipment, and take effective maintenance measures, and the preparation of spare parts should be based on factors such as fault probability and maintenance cycle to reasonably plan and store the spare parts, so as to balance the maintenance cost and equipment availability.

[0278] When RUL < 500 hours, weekly inspection and stoppage plan are carried out, the weekly inspection focuses on the real-time operation state and key performance indicators of the equipment, the monitoring frequency is increased, abnormal changes of the equipment are captured in time, and a detailed stoppage maintenance plan is made, including stoppage time, maintenance content, personnel arrangement, spare part list, etc., so that the stoppage maintenance work can be completed in an orderly manner before the equipment reaches the failure time, the stoppage maintenance should be carried out in a period with low power grid load as much as possible to reduce the influence on power supply, and relevant users and departments are notified in advance to make good power-off preparation and emergency measures, during the stoppage, comprehensive equipment maintenance, component replacement and system debugging are carried out to restore the good operation state of the equipment and ensure the stable operation of the power system.

[0279] The above is only the preferred embodiment of the present application, and it should be noted that for those skilled in the art, without departing from the core concept of the present application, a number of adjustments and improvements can be made, and these adjustments and improvements should also be considered as the protection scope of the present application.

Claims

1. A method for monitoring reliability of a photovoltaic inverter in a high-altitude special environment, characterized in that, The method comprises the following steps: S1, data acquisition and preprocessing: the electrical quantity, temperature quantity, environmental quantity and operation quantity of the highland inverter are collected, the abnormal value processing based on 3σ criterion, the power dramatic change section elimination, the stable time period selection and the dynamic time warping (DTW) algorithm are used to calibrate the time stamp of the data with different sampling frequencies, so that the data quality is high and the correspondence is accurate; S2, highland sensitive feature engineering: a feature system including voltage ripple rate, high-frequency harmonic energy ratio, junction temperature fluctuation entropy, thermal response time, thermal shock coefficient and comprehensive stress index is constructed, and the optimal feature subset is selected through physical consistency and three-stage feature selection method; S3, physical constraint modeling: the IGBT thermal fatigue and capacitor aging models are established, the health state evolution equation is combined to quantitatively evaluate the equipment health state, and the physical residual is introduced to ensure that the data-driven model is consistent with the physical law; S4, Bayesian-PINN fusion modeling: a network architecture including an input layer, a Bayesian hidden layer and an output layer is constructed, a loss function including a fusion data fitting term, a physical constraint term and a Bayesian regularization term is set, the data and the physical model are deeply coupled, and the model generalization ability and prediction accuracy are improved; S5, RUL prediction process: the conversion method from the health index to the remaining useful life (RUL) considers the prediction uncertainty, an environment adaptive correction mechanism is introduced, and accurate maintenance decision support is realized based on the multi-index decision logic; S6, highland environment adaptive mechanism: according to the altitude and seasonal changes, the model parameters and feature weights are automatically adjusted, so that the monitoring system maintains high performance monitoring effect under different environmental conditions.

2. The method for monitoring reliability of photovoltaic inverter in high altitude special environment according to claim 1, characterized in that, In the data acquisition and preprocessing of step S1, the 3σ criterion is used to eliminate abnormal data in the electrical quantity, and the power dramatic change section data is eliminated; the stable time period is selected according to the light intensity and power fluctuation rate, and the time window is limited to 10:00-14:00; the DTW algorithm is used to calibrate the time stamp of the environmental data and the electrical data.

3. The method for monitoring reliability of photovoltaic inverter in high altitude special environment according to claim 1, characterized in that, In the highland sensitive feature engineering of step S2, the voltage ripple rate considers the influence of altitude on the ripple rate and reflects the capacitor aging degree, the formula is: wherein V dc is the DC bus voltage, is the average value, RMS indicates the root mean square value, and h is the altitude; The high-frequency harmonic energy ratio reflects the IGBT switching characteristic degradation, the formula is: where f sw is the switching frequency and PSD(f) is the power spectral density; The junction temperature fluctuation entropy measures the complexity of the junction temperature fluctuation and is related to the thermal fatigue damage, the formula is: In the formula, T j is the IGBT junction temperature, p(T j ) is its probability distribution, P loss0 is the reference power, P loss is the actual output power; the thermal response time reflects the performance of the heat dissipation system, and the formula is: where T j (t) is a function of the junction temperature over time, P loss (t) is a function of the output power over time, corr represents the correlation; The thermal shock coefficient quantifies the influence of the diurnal temperature difference stress on the equipment, the formula is: where ΔT is the diurnal temperature difference, N is the number of thermal cycles, and T is the temperature. day where ΔT is the diurnal temperature difference, N is the number of thermal cycles, and T is the temperature. cycle where ΔT is the di The comprehensive stress index comprehensively reflects the total damage degree of the highland environment to the equipment, the formula is: In the formula, UV is the ultraviolet intensity, and h is the altitude; The physical consistency verifies the relationship between the capacitor ripple rate and the temperature, and the relationship between the junction temperature fluctuation entropy and the environmental temperature; The three-stage feature selection removes the features with variance threshold, selects the optimal feature subset through physical screening and recursive feature elimination algorithm.

4. The method for monitoring reliability of photovoltaic inverter in high altitude special environment according to claim 1, characterized in that, In the physical constraint modeling of step S3, the IGBT thermal fatigue model describes the life characteristics of IGBT under thermal stress, the formula is: where N f represents the thermal fatigue life of the IGBT, A is a constant, ΔT j is the junction temperature change, β is a material constant, E a is the activation energy, k B is the Boltzmann constant, T m is the average temperature; The capacitor aging model describes the life characteristics of the capacitor under the combined action of thermal stress and electrical stress, the formula is: where L is the capacitance lifetime, L0is the reference lifetime, E a is the activation energy, T is the temperature, V is the voltage, V0is the reference voltage, and n is the aging exponent. The health state evolution equation defines the health index range as [0, 1], the degradation rate is composed of the negative value of the sum of the IGBT and capacitor degradation rates, and the physical residual is used to evaluate the deviation of the model from the physical law.

5. The method for monitoring reliability of photovoltaic inverter in high altitude special environment according to claim 1, characterized in that, In the step S4, the network architecture input layer receives the time t and the feature vector x, the Bayesian hidden layer weight obeys the normal distribution, and the output layer outputs the health index; in the loss function, the data fitting term measures the difference between the prediction and the true health index, the physical constraint term ensures that the health state evolution conforms to the physical law, and the Bayesian regularization term quantifies the uncertainty of the model parameters, and the loss function is specifically: where the data fitting term: u pred is the model predicted health indicator, u true is the true health indicator, N is the number of data samples; the physical constraint term: ‖PDE(u)‖ 2 is the squared norm of the physical residual, ReLU(u t+1 -u t ) term is used to prevent non-physical increase of the health indicator, β1 and β2 are weight coefficients; the Bayesian regularization term: p(w) is the prior distribution, q(w) is the posterior distribution; The uncertainty quantification is realized through Monte Carlo sampling, calculation of the mean and standard deviation, and construction of the confidence interval, and specifically: the formula of Monte Carlo sampling is: In the formula, u (i) (t) represents the predicted value of the health index corresponding to the i-th sampling; x is the input feature vector; w (i) represents the model parameters of the i-th sampling; q(w) is the posterior distribution of the model parameters; M is the number of samplings; The mean of the statistical quantity calculation is The standard deviation is In the formula, u(t) is the mean of the health index prediction value; σ u (t) is the standard deviation of the health index prediction value; The confidence interval is: wherein CI 95 (t) indicates the 95% confidence interval; 1.96 is the normal distribution critical value corresponding to a 95% confidence level.

6. The method for monitoring reliability of photovoltaic inverter in high altitude special environment according to claim 1, characterized in that, In the RUL prediction process of the step S5, the failure time, the basic RUL and the confidence interval thereof are calculated from the health index u value to the RUL, specifically: In the formula, t fail denotes the time when the equipment is expected to reach the failure threshold. t is a time variable; σ is the mean value of the health index prediction value u (t) is the standard deviation of the health index prediction value; 0.2 is a pre-set health index failure threshold, specifically: RUL=t fail -t current , In the formula, RUL represents the remaining useful life; t fail is the predicted failure time; t current is the current time; The environmental self-adaptive correction corrects the base RUL and the health index standard deviation by integrating environmental factors, specifically: In the formula, RUL 校正 is the corrected remaining useful life; RUL is the base remaining useful life; is the corrected RUL confidence interval; CI RUL is the base RUL confidence interval; the highland extension: In the formula, σ u′ (t) is the corrected health index standard deviation; σ u (t) is the original health index standard deviation; h is the altitude; The decision logic formulates the maintenance decision under different states according to the health index, the remaining useful life, the total standard deviation and the confidence interval width, and specifically:

1. Normal state: when u > 0.6 and RUL > 1000h and σ total <0.05, the decision action is routine monitoring; ② Early warning state: when 0.3 < u < 0.6 and 500h < RUL < 1000h and σ < 0.15, the decision action is to prepare maintenance; total < 0.15, the decision action is to prepare maintenance; ③ Emergency state: when u≤0.3 and RUL≤500h, the decision action is to shut down immediately; ④ Confidence anomaly: when the CI width>0.3×RUL, the decision action is manual inspection; where u is a health indicator, RUL is the remaining useful life, σ total is the total standard deviation and CI width is the confidence interval width.

7. The method for monitoring reliability of photovoltaic inverter in high altitude special environment according to claim 1, characterized in that, In the highland environment self-adapting mechanism of the step S6, the dynamic adjustment strategy adjusts the physical constraint weight, the KL regular weight, the feature set and the degradation rate according to the altitude, wherein the physical constraint weight is adjusted as follows: In the formula, α is the physical constraint weight, α0 is the initial physical constraint weight, and h is the altitude; the KL regular weight is adjusted as follows: In the formula, γ is the KL regular weight, and γ0 is the initial KL regular weight; the feature set is corrected as follows: In the formula, σ T ′ is the corrected feature standard deviation, and σ T is the original feature standard deviation; The seasonal strategy adjusts the weights of ultraviolet and thermal shock in the comprehensive stress index according to different seasons, and specifically: Summer: ultraviolet weight 50%, thermal shock weight 30%; Winter: ultraviolet weight 30%, thermal shock weight 50%; Spring and autumn: ultraviolet weight 40%, thermal shock weight 40%.

8. A high-altitude special environment photovoltaic converter reliability monitoring system, characterized in that, The method for implementing any one of steps 1-7 comprises: A data acquisition module for acquiring electrical quantity, temperature quantity, environmental quantity and operation quantity data of the highland converter; A data preprocessing module for performing outlier processing, stable period selection and signal alignment on the collected data; A feature engineering module for extracting highland sensitive features and performing physical consistency verification and feature selection; A physical constraint modeling module for establishing a physical model and a health state evolution equation and calculating a physical residual; A Bayesian-PINN fusion modeling module for constructing a fusion model, setting a loss function and quantifying uncertainty; a RUL prediction module for predicting the remaining useful life and performing environmental correction and formulating a maintenance decision; A highland environment adaptive module for dynamically adjusting model parameters and feature weights according to altitude and season.

9. The high-altitude special environment photovoltaic inverter reliability monitoring system according to claim 8, characterized in that, The data acquisition module includes a Hall sensor, a thermocouple and a weather station interface for acquiring electrical quantity, temperature quantity and environmental quantity data, respectively.

10. The high-altitude special environment photovoltaic inverter reliability monitoring system according to claim 8, characterized in that, The Bayesian-PINN fusion modeling module is implemented by using a deep learning framework and has a variational inference function to estimate the posterior distribution parameters of the Bayesian hidden layer weight.

Images