A method for predicting the lifetime of a photovoltaic device based on artificial intelligence

Abstract

This application discloses an artificial intelligence-based method for predicting the lifetime of photovoltaic devices, belonging to the field of electronic digital data processing technology. The method includes: calculating the deviation of the current degradation rate from the historical average degradation rate based on the current operating condition feature vector of the photovoltaic device; comparing the deviation with a preset threshold to generate a gating signal for updating model parameters; updating the parameters of the degradation prediction model based on a loss function including a parameter drift penalty term when the gating signal allows updates; outputting the predicted remaining lifetime and its uncertainty using the updated degradation prediction model, and performing Bayesian weighted fusion on the predicted remaining lifetime based on the uncertainty to obtain an estimated remaining lifetime; determining the maximum single-step lifetime reduction based on the current environmental stress intensity, and constraining and correcting the estimated remaining lifetime to obtain a constrained and corrected estimated remaining lifetime.

Description

Technical Field

[0001] This application relates to the field of digital data processing technology, and more specifically, to an artificial intelligence-based method for predicting the lifetime of photovoltaic devices. Background Technology

[0002] As the scale of photovoltaic power generation systems continues to expand, photovoltaic devices, as the core power generation unit, are subject to the combined effects of multiple factors such as fluctuations in light intensity, changes in ambient temperature, damp heat aging, and mechanical stress during long-term operation, exhibiting complex nonlinear degradation characteristics. To ensure the stable operation of photovoltaic systems and improve operation and maintenance management, it is urgent to accurately predict and dynamically assess the remaining service life of photovoltaic devices.

[0003] In existing technologies, traditional methods based on physical degradation models usually rely on preset mechanistic assumptions and empirical parameters, which are difficult to accurately characterize the actual degradation process in complex operating environments and have weak adaptability. While data-driven methods based on artificial intelligence can build predictive models using historical operating data, they still have problems in the process of processing digital electrical data, such as online updates of model parameters being susceptible to noise interference, unstable extraction of time-series features, and difficulty in effectively quantifying and integrating the uncertainty of prediction results, thus affecting the reliability and consistency of prediction results.

[0004] Furthermore, existing methods lack constraints on the model update process during dynamic processing of continuously running data, which can easily lead to parameter drift accumulation and thus reduce long-term prediction performance. At the same time, in the process of multi-source operation monitoring data fusion and time series modeling, the insufficient utilization of correlation and uncertainty information between data limits the further improvement of prediction accuracy.

[0005] In summary, how to achieve stable online updates of photovoltaic device lifetime prediction models, effective constraint of parameter drift, and fusion utilization of uncertainty information within an artificial intelligence framework based on electrical digital data processing, thereby improving the accuracy and robustness of remaining lifetime prediction, has become an urgent technical problem to be solved. Summary of the Invention

[0006] To overcome a series of shortcomings in existing technologies, the purpose of this application is to provide an artificial intelligence-based method for predicting the lifetime of photovoltaic devices, comprising the following steps: Based on the current operating condition feature vector of the photovoltaic device, calculate the deviation of the current degradation rate from the historical average degradation rate; The deviation is compared with a preset threshold to generate a gating signal for updating the model parameters; When the gating signal allows updates, the parameters of the degradation prediction model are updated based on the loss function that includes a parameter drift penalty term; The updated degradation prediction model is used to output the remaining useful life prediction value and its uncertainty. Based on the uncertainty, the remaining useful life prediction value is fused by Bayesian weighted fusion to obtain the remaining useful life estimate. The maximum single-step life reduction is determined based on the current environmental stress intensity, and the remaining service life estimate is constrained and corrected to obtain the constrained and corrected remaining service life estimate.

[0007] In some embodiments, the gating signal generation method is as follows: when the deviation is lower than the lower limit of the preset information threshold, a first gating signal that allows full parameter updates is generated; when the deviation is between the upper and lower limits of the preset information threshold, a second gating signal that allows parameter updates under a limited learning rate is generated; when the deviation exceeds the upper limit of the preset information threshold, a third gating signal that prohibits parameter updates is generated.

[0008] In some embodiments, the loss function includes a prediction residual loss term, a parameter drift penalty term, and a prediction monotonicity soft constraint loss term, wherein: the prediction residual loss term is calculated using the symmetric Huber loss function to characterize the error between the current predicted value and the actual degraded observation value; the parameter drift penalty term is obtained by calculating the weighted L2 norm between the updated model parameter vector and the corresponding model parameter vector of the previous update period, and is used to constrain the magnitude of the model parameter update; the prediction monotonicity soft constraint loss term is used to constrain the changing trend of the degraded prediction sequence.

[0009] In some embodiments, the degradation prediction model adopts a hybrid prediction architecture that combines a one-way causal long short-term memory network and a Bayesian neural network. The one-way causal long short-term memory network is used to extract time-series features from a historical operating condition feature sequence of length T time steps and output the final hidden state vector. The Bayesian neural network is used to receive the final hidden state vector and perform probability mapping, outputting the degradation prediction result and the corresponding uncertainty.

[0010] In some embodiments, the method for outputting the predicted remaining useful life and its uncertainty using the updated degradation prediction model is as follows: Based on the updated degradation prediction model, multiple sets of remaining useful life prediction results and corresponding prediction variance components are obtained. The sample variance is calculated based on multiple sets of remaining useful life prediction results, and the sample variance is determined as cognitive uncertainty; at the same time, the prediction variance component of the model's explicit output is extracted, and the prediction variance component is determined as random uncertainty. Based on the comparison data between the historical prediction results of the most recent preset quantity and the actual degradation quantity, the cognitive uncertainty and random uncertainty are calibrated by temperature to obtain the corresponding correction coefficients, and the cognitive uncertainty and random uncertainty are corrected based on the correction coefficients. The overall uncertainty is calculated based on the corrected cognitive uncertainty and random uncertainty. Multiple sets of remaining useful life prediction results are constructed into multiple Gaussian mixture components of a Gaussian mixture model, and the mixture weights are determined based on the reciprocal of the overall uncertainty corresponding to each Gaussian mixture component. Bayesian weighted fusion is performed on each Gaussian mixture component based on the mixture weight to obtain the mean and variance of the mixture distribution, and the mean is determined as the predicted value of the remaining useful life. Based on the overall uncertainty, the exponential smoothing coefficient is determined, and a first-order exponential smoothing filter is applied to the remaining useful life prediction value to obtain the smoothed remaining useful life prediction value and the corresponding uncertainty.

[0011] In some embodiments, the method for determining the maximum single-step lifetime reduction based on the current environmental stress intensity is as follows: The current environmental stress intensity is obtained by weighted summing of the thermal stress component, light-induced attenuation stress component, and mechanical stress component of the photovoltaic device at the current moment. The environmental stress intensity is input into a preset piecewise mapping function. In the low stress range, the maximum single-step lifetime reduction is determined according to the step amount corresponding to the nominal decay rate. In the high stress range, the maximum single-step lifetime reduction is determined according to the power function of the environmental stress intensity. Based on the predicted remaining lifetime value and the maximum single-step lifetime decrease output at the previous moment, determine the upper bound of the constraint on the remaining lifetime at the current moment. If the predicted remaining useful life at the current moment is lower than the upper limit of the constraint, the predicted remaining useful life at the current moment is directly output; otherwise, the predicted remaining useful life at the current moment is corrected to the lower limit of the constraint, and the corrected predicted remaining useful life is output.

[0012] In some embodiments, the method further includes the following steps: In the process of obtaining the estimated remaining useful life, the magnitude of change of the predicted remaining useful life is calculated, and it is determined whether the magnitude of change exceeds the preset abrupt change judgment threshold. If the limit is not exceeded, the estimated remaining useful life after constraint correction will be output directly. If the result exceeds the limit, an attribution judgment is performed, and the attribution result is obtained.

[0013] In some embodiments, the method further includes the following steps: Perform the following actions based on the attribution results: If it is determined to be a true acceleration of degradation, the constraint coefficient of the parameter drift penalty term is reduced, and the estimated remaining useful life after constraint correction is output. If the error is determined to be a fluctuation in operating conditions or an incorrect update, the parameter update of the degradation prediction model is rolled back, and the current abrupt prediction value is identified as an outlier and masked so that it does not participate in the generation and output of the remaining useful life estimate. The output is rolled back to the remaining useful life estimate corresponding to the state before the update.

[0014] In some embodiments, the method for determining whether the change magnitude exceeds a preset mutation determination threshold is as follows: Obtain the standard deviation of the predicted remaining service life sequence within the historical preset period, and determine the first-level warning threshold and the second-level confirmation threshold based on the standard deviation; When the change exceeds the first-level warning threshold but does not exceed the second-level confirmation threshold, the current state will be marked as suspended. Determine whether the cumulative change amplitude of multiple subsequent sampling cycles meets the condition of continuous increase: if it does, determine that the change amplitude exceeds the preset sudden change judgment threshold; otherwise, release the suspended state and restore normal output. When the change exceeds the secondary confirmation threshold, it is directly determined that the change exceeds the preset mutation judgment threshold.

[0015] In some embodiments, the method further includes: when the photovoltaic device belongs to a photovoltaic array containing two or more devices of the same type, using degradation rate observation information of other devices in the array to assist in updating the degradation state of the current device, specifically including the following steps: Based on the historical sequence of degradation rate of photovoltaic array within a preset time window, the Pearson correlation coefficient between each photovoltaic device is calculated, and a spatial correlation coefficient matrix of degradation rate is constructed. A significance test was performed on the spatial correlation coefficient matrix to screen and retain device associations with Pearson correlation coefficients greater than a preset correlation threshold and corresponding statistical significance levels that meet the preset significance level. Based on the mean degradation rate within the array, the residual sequence of degradation rate of each photovoltaic device is calculated, and the continuity anomaly of the residual sequence is determined. When the absolute value of the residual of any photovoltaic device exceeds the standard deviation of a preset multiple within a preset number of consecutive update cycles, it is marked as an abnormal device and its data is removed. When the current photovoltaic device performs model parameter updates, the degradation rates of other photovoltaic devices that have been screened for correlation and not marked as anomalies are weighted and fused to obtain auxiliary monitoring signals; The auxiliary monitoring signal is constructed as an auxiliary loss term and introduced into the loss function of the degradation prediction model. The contribution weight of the auxiliary loss term is adaptively adjusted based on the current photovoltaic device observation data quality.

[0016] Compared with the prior art, this application has the following beneficial effects: This application constructs an adaptive parameter update strategy based on a degradation bias gating mechanism and integrates parameter drift constraints and Bayesian uncertainty weighted estimation to achieve high-precision dynamic prediction and constraint correction of the remaining service life of photovoltaic devices while ensuring model stability. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the overall process of an artificial intelligence-based photovoltaic device lifetime prediction method in one embodiment of this application.

[0018] Figure 2 This is a flowchart illustrating a gate signal generation method in one embodiment of this application.

[0019] Figure 3 This is a schematic diagram of a hybrid architecture for a degradation prediction model in one embodiment of this application.

[0020] Figure 4 This is a flowchart illustrating the attribution determination method in one embodiment of this application. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of the embodiments of this invention will be described in more detail below with reference to the accompanying drawings. In the drawings, the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The described embodiments are some embodiments of this invention, but not all embodiments.

[0022] Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0023] The embodiments and directional terms described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0024] like Figure 1 As shown, an artificial intelligence-based method for predicting the lifetime of photovoltaic devices includes the following steps: Acquire current operational monitoring data of photovoltaic devices and the historical average degradation rate of photovoltaic devices; wherein, the monitoring data includes at least one of output power, photovoltaic device temperature, ambient irradiance and wind speed; The current operation monitoring data is processed by outlier removal and normalization to obtain the current operating condition feature vector, and the deviation of the current degradation rate from the historical average degradation rate is calculated based on the operating condition feature vector. The deviation is compared with a preset threshold to generate a gating signal for updating the model parameters; When the gating signal indicates that an update is allowed, the parameters of the degradation prediction model are updated based on the loss function that includes a parameter drift penalty term; The updated degradation prediction model is used to output the remaining useful life prediction value and its uncertainty. Based on the uncertainty, the remaining useful life prediction value is subjected to Bayesian weighted fusion processing to obtain a smooth remaining useful life estimate. The maximum single-step life reduction is determined based on the current environmental stress intensity, and the remaining life estimate is constrained and corrected to ensure that the remaining life estimate satisfies the monotonically decreasing constraint. In the process of obtaining the estimated remaining useful life, the magnitude of change of the predicted remaining useful life is calculated, and it is determined whether the magnitude of change exceeds the preset abrupt change judgment threshold. If the limit is not exceeded, the estimated remaining useful life after constraint correction will be output directly. If the number exceeds the limit, an attribution judgment is performed, and the following operations are performed based on the attribution results: If it is determined to be a true acceleration of degradation, the constraint coefficient of the parameter drift penalty term is reduced, and the estimated remaining useful life after constraint correction is output. If the error is determined to be a fluctuation in operating conditions or an incorrect update, the parameter update of the degradation prediction model is rolled back, and the current abrupt prediction value is identified as an outlier and masked so that it does not participate in the generation and output of the remaining useful life estimate. The output is rolled back to the remaining useful life estimate corresponding to the state before the update.

[0025] The aforementioned AI-based photovoltaic device lifetime prediction method achieves real-time perception of device degradation status by performing outlier removal, normalization, and degradation deviation analysis on the current operation monitoring data of photovoltaic devices. It reduces the impact of operating condition fluctuations on the stability of the lifetime prediction model by introducing a model update gating mechanism based on confidence thresholds and an adaptive parameter update strategy including parameter drift penalties. Furthermore, it achieves smooth output and physical consistency constraints for remaining lifetime prediction results by combining Bayesian weighted fusion and monotonically decreasing constraint correction. Simultaneously, it distinguishes between actual accelerated degradation and erroneous updates through mutation detection, attribution judgment, and model rollback mechanisms, suppressing the propagation of abnormal prediction results. This significantly improves the accuracy, stability, and robustness of photovoltaic device lifetime prediction, meeting the needs of long-term online lifetime assessment of photovoltaic devices under complex operating environments.

[0026] In some embodiments, the method for obtaining current operational monitoring data of photovoltaic devices is as follows: Construct a multi-source sensing data acquisition system for obtaining the operating status of photovoltaic devices, and perceive the multi-dimensional physical parameters of photovoltaic devices during operation; The sampling parameters of each data source in the data acquisition system are configured uniformly, and a data synchronization mechanism based on time identifiers is established; Based on uniformly configured sampling parameters, various monitoring data during the operation of photovoltaic devices are collected synchronously to form raw operating data; The raw operational data is aggregated and processed, and data integrity is checked and repaired to obtain a continuous data sequence. Based on preset rules, abnormal or missing data is processed, and valid operational data that has undergone quality control is output. Based on effective operational data and combined with historical operational information, characteristic parameters are calculated and updated to obtain monitoring data characterizing the current operating status of photovoltaic devices.

[0027] For example, in a distributed photovoltaic power station, a single photovoltaic module is selected as the monitoring object. A multi-source acquisition system is constructed by deploying sensors on its surface and support structure. The output power is acquired in real time by the module-level inverter, the module temperature is obtained by the attached temperature sensor, the ambient irradiance is measured by the horizontal irradiance sensor, and the wind speed is provided by the meteorological station equipment.

[0028] During actual operation, with 10:00:00 on April 8, 2026 as the time reference point, all sensors synchronously sampled at a frequency of 1Hz. For example, at this time point, the collected data showed an output power of 320W, a component temperature of 48℃, and an irradiance of 850W / m². 2 The wind speed is 3.2 m / s, and the data is bound using a unified timestamp to form the original operational data record for that moment.

[0029] During the data aggregation process, data from 10:00:00 to 10:05:00 is continuously spliced ​​in 5-minute time windows. If irradiance data is missing at 10:02:15, it is completed by linear interpolation based on the irradiance data at 10:02:14 and 10:02:16, thus forming a continuous data sequence.

[0030] During the anomaly detection process, for example, if the component temperature suddenly increases to 85℃ at 10:03:30, which is significantly deviated from the historical normal operating range (such as 30℃~60℃), it is identified as an anomaly and is removed or smoothed out to avoid interfering with subsequent feature calculations.

[0031] When performing feature calculations based on valid operating data, for example, if the output power within the window decreases by about 3.5% compared to the average power under the same historical irradiation conditions, and the temperature sensitivity coefficient increases by 0.02 / W·℃, a monitoring feature vector reflecting the current operating status of the photovoltaic module is updated.

[0032] The photovoltaic device operation monitoring data acquisition method described in this application constructs a multi-source sensor data acquisition system and uniformly configures and synchronizes the sampling parameters of each data source, enabling collaborative sensing and synchronous acquisition of multi-dimensional physical parameters during photovoltaic device operation. By aggregating, processing, detecting, and repairing missing data from the raw operation data, the continuity and consistency of the monitoring data are ensured. Furthermore, by incorporating anomaly data processing and quality control mechanisms, the impact of abnormal sampling, data loss, and environmental interference on the monitoring results is reduced. Simultaneously, by dynamically calculating and updating characteristic parameters based on historical operation information, an accurate representation of the current operating status of the photovoltaic device is achieved, thereby significantly improving the reliability, completeness, and timeliness of the operation monitoring data and meeting the requirements for long-term online status sensing and lifespan assessment of photovoltaic devices.

[0033] In some embodiments, current operation monitoring data is collected through a distributed multi-point sensor network, which includes a thermistor array mounted on the back panel, a Hall effect current sensor mounted on the DC side of the inverter, an irradiance meter mounted on the front panel, and an ultrasonic anemometer mounted at a weather station.

[0034] In some embodiments, the method for outlier removal and normalization of current operational monitoring data is as follows: The acquired current operational monitoring data is classified according to the monitoring channels to form corresponding multi-channel time series data sequences. Based on the statistical distribution characteristics of the data from each monitoring channel, preliminary anomaly detection is performed to identify candidate abnormal data points. The candidate abnormal data points are subjected to temporal neighborhood consistency verification and physical constraint condition determination. Data points determined to be abnormal are replaced or corrected, and data points not determined to be abnormal are retained to obtain the data sequence after anomaly processing. Trend continuity analysis is performed on continuously occurring candidate abnormal data points to identify data fluctuations caused by changes in operating conditions. When multiple channels show abnormalities at the same time, joint verification is performed based on the physical correlation between monitoring parameters to correct the abnormality judgment results. The normalized parameters for each monitoring channel are determined based on historical operational data. The normalized parameters are used to characterize the statistical distribution characteristics of the data from the corresponding monitoring channel. The normalization parameters are used to perform a dimensionless transformation on the anomaly-handled data, and the transformation results are subjected to range constraints to obtain standardized operating data.

[0035] For example, in the actual operation of a distributed photovoltaic power station, the current operation monitoring data of a photovoltaic module is selected as the processing object, and its monitoring channels include output power, module temperature, ambient irradiance and wind speed.

[0036] During the outlier removal and normalization process, the data from each monitoring channel are first classified according to time sequence to form a multi-channel time series. For example, in the sampling interval from 10:00:00 to 10:00:10, the output power sequence is generally stable between 310W and 340W, but an abnormal sudden change of 900W occurs at 10:00:05. Based on the statistical distribution characteristics of this channel (such as a mean of approximately 325W and a standard deviation of approximately 15W), it is identified as a candidate outlier data point. Similarly, the temperature channel fluctuates within the normal range of 40℃ to 55℃, but an outlier of 10℃ occurs at 10:00:06, which is also marked as a candidate outlier point.

[0037] Subsequently, temporal neighborhood consistency checks were performed on candidate outlier data points. For example, regarding the power outlier at 10:00:05, the power values ​​at 10:00:04 and 10:00:06 were 322W and 328W respectively, showing a smooth trend without abrupt changes. Therefore, this point was determined to be an outlier, and the neighborhood mean was used. Replace and correct the data points; while retain the data points that do not show any abnormal characteristics.

[0038] Furthermore, in trend continuity analysis, for example, if the power continuously decreases from 10:00:07 to 10:00:09, and the irradiance decreases synchronously, then the change is determined to be a normal operating condition fluctuation caused by changes in lighting conditions, rather than abnormal data, thus avoiding the erroneous rejection of real operating changes.

[0039] During the multi-channel joint verification process, for example, if an abnormal power drop and a sudden change in irradiance occur simultaneously at 10:00:06, but the changes in wind speed and temperature are within the normal range, then based on the physical positive correlation between power and irradiance, the irradiance drop at that moment is re-determined as a real environmental change, thereby correcting the original abnormal determination result.

[0040] In the normalization stage, for example, if the power channel mean is determined to be 330W and the standard deviation is 20W based on historical operating data, the corrected power data is standardized and transformed into dimensionless values. Similarly, temperature, irradiance and wind speed are normalized based on their respective historical statistical parameters so that data with different dimensions are mapped to a unified scale space.

[0041] Finally, by imposing range constraints on the normalization results (e.g., limiting them to a certain range) The range is used to suppress the residual effects of extreme anomalies, thereby obtaining standardized operating data for subsequent photovoltaic device condition assessment and lifetime prediction.

[0042] The method for outlier removal and normalization of operational monitoring data described in this application achieves accurate identification and correction of candidate outlier data points by performing statistical distribution analysis, temporal neighborhood consistency verification, and physical constraint determination on multi-channel time-series data. It reduces the risk of data fluctuations caused by changes in operating conditions being misjudged as anomalies by performing trend continuity analysis on continuous outlier data and combining this with joint verification of the physical correlation between multi-channel monitoring parameters. Furthermore, it determines the normalized parameters corresponding to each monitoring channel based on historical operational data and performs dimensionless transformation and range constraint processing on the anomaly-processed data to achieve data unification and standardization for different monitoring parameters. This significantly improves the accuracy, consistency, and comparability of operational monitoring data, providing a stable and reliable data foundation for photovoltaic device degradation status analysis and lifetime prediction, and meeting the long-term online monitoring needs under complex operating environments.

[0043] In some embodiments, the method for obtaining the operating condition feature vector is as follows: based on standardized operating data, the data is combined according to a preset feature organization method to obtain the operating condition feature vector at the current moment.

[0044] In some embodiments, the deviation is calculated as follows: Based on the operating condition feature vector, relevant information on the degradation rate used to characterize the current degradation state is extracted; The average historical degradation rate corresponding to the current photovoltaic device is retrieved as a comparison benchmark. The deviation of the current degradation rate from the historical average degradation rate is calculated by comparing the information related to the current degradation rate with the historical average degradation rate.

[0045] For example, in a real-world photovoltaic module monitoring scenario, degradation rate-related information is first extracted based on the current operating condition feature vector. This information is comprehensively characterized by parameters such as the output power attenuation trend, the rate of change in power conversion efficiency per unit irradiance, and the rate of change in temperature sensitivity. For instance, the current degradation rate is calculated to be 0.42% / month based on 30 consecutive days of operating data.

[0046] Subsequently, the average historical degradation rate of the photovoltaic module during the stable operation phase is retrieved from the historical database. For example, based on long-term statistical results of the past 12 months, its average historical degradation rate is 0.30% / month. This value is used as a benchmark to reflect the average degradation level of the equipment under normal aging conditions.

[0047] In the process of calculating the deviation, the current degradation rate is compared with the historical average degradation rate. For example, the deviation is calculated using the difference method. This deviation represents the degree to which the current rate of degradation of photovoltaic modules accelerates relative to historical normal levels.

[0048] In another example, if the current degradation rate of a photovoltaic module is 0.28% / month, while the historical average degradation rate is 0.30% / month, the calculated deviation is -0.02% / month, indicating that the current degradation status is slightly better than the historical average and falls within the normal fluctuation range.

[0049] The method for obtaining the operating condition feature vector and calculating the deviation described in this application combines standardized operating data according to a preset feature organization method to achieve a unified feature expression of the current operating condition of photovoltaic devices. It extracts degradation rate-related information representing the current degradation state based on the operating condition feature vector and compares it with the historical average degradation rate to quantify the difference between the current degradation behavior and historical degradation patterns. Furthermore, by calculating the deviation of the current degradation rate relative to the historical average degradation rate, it enhances the sensitivity to abnormal degradation changes and performance degradation trends, thereby significantly improving the accuracy and stability of photovoltaic device degradation state identification. This provides a reliable basis for subsequent model parameter updates and remaining service life prediction, meeting the needs of long-term online degradation assessment of photovoltaic devices under complex operating conditions.

[0050] In some embodiments, the method for determining the preset threshold is as follows: Obtain current season information and retrieve historical deviation data for the same period and deviation data for the most recent thirty days; Statistical analysis was performed on the historical deviation data and the deviation data of the most recent 30 days to obtain the 95th percentile value of the historical deviation distribution and the 95th percentile value of the deviation distribution of the most recent 30 days. Obtain the sample size for the same period in history and the sample size for the most recent 30 days. Based on the ratio between the sample size for the same period in history and the sample size for the most recent 30 days, adaptively determine the weight parameters corresponding to the distribution of deviations in the same period in history and the distribution of deviations in the most recent 30 days. The sum of each weight parameter is 1. Based on the weighting parameters, the 95th percentile of the historical same-period deviation distribution and the 95th percentile of the deviation distribution in the most recent 30 days are weighted and fused to obtain the basic confidence threshold. At the beginning of each calendar month, the basic confidence threshold is updated based on the statistical results of the deviation distribution corresponding to the previous calendar month to obtain the preset confidence threshold.

[0051] For example, in the life assessment of a photovoltaic power station, the analysis is performed using the summer operation phase as the current season. The system first retrieves the deviation data of the photovoltaic module for the same period in history (e.g., the past three summer months), and at the same time obtains the deviation data of the most recent 30 days as a short-term operation reference.

[0052] In the statistical analysis process, for example, the historical deviation data for the same period is generally distributed between 0.05% / month and 0.18% / month, and its 95th percentile value is calculated to be 0.17% / month; while the deviation data for the most recent 30 days has increased slightly due to the recent high temperature, with a distribution range of 0.06% / month to 0.22% / month, and its 95th percentile value is 0.20% / month.

[0053] In determining the sample size and weights, for example, if the historical data set is 900 sets and the recent 30 days data set is 300 sets, then the ratio is 3:1. Based on this ratio, the weight parameters are adaptively determined as follows: historical data set weight 0.75, recent 30 days weight 0.25.

[0054] In the calculation of the basic confidence threshold, the two 95th percentile values ​​are weighted, for example: 0.17×0.75+0.20×0.25=0.1275+0.05=0.1775% / month, thus obtaining a basic confidence threshold of approximately 0.178% / month.

[0055] Furthermore, in the monthly update mechanism, for example, if the deviation statistics for August are obtained in early September and it is found that the overall deviation level in the last 30 days has decreased compared to July, with its 95th percentile value dropping to 0.16% / month, then the basic confidence threshold is corrected and updated based on the latest statistical results, so that the preset confidence threshold is adjusted to, for example, around 0.17% / month, thereby better reflecting the current trend of operational status changes.

[0056] The method for determining the pre-set confidence threshold described in this application achieves synergistic analysis of long-term periodic variation characteristics and short-term operational fluctuation characteristics by combining current seasonal information with historical deviation data from the same period and deviation data from the most recent 30 days. It enhances sensitivity to abnormal degradation deviation changes by separately statistically analyzing the 95th percentile values ​​corresponding to the historical and recent deviation distributions. Based on this, it adaptively determines the weight parameters corresponding to different deviation distributions according to the ratio of historical to recent sample sizes, and weights and fuses the statistical results at different time scales to dynamically generate the basic confidence threshold. Simultaneously, by updating the basic confidence threshold at the beginning of each calendar month based on the statistical results of the deviation distribution from the previous calendar month, the pre-set confidence threshold can adaptively adjust with seasonal and operational changes, thereby significantly improving the accuracy and stability of model parameter update judgments and reducing the interference of abnormal operating conditions or short-term fluctuations on the life prediction model update process.

[0057] In some embodiments, the gating signal generation method is as follows: when the deviation is lower than the lower limit of the preset information threshold, a first gating signal that allows full parameter updates is generated; when the deviation is between the upper and lower limits of the preset information threshold, a second gating signal that allows parameter updates under a limited learning rate is generated; when the deviation exceeds the upper limit of the preset information threshold, a third gating signal that prohibits parameter updates is generated.

[0058] For example, during the online update of a photovoltaic module lifetime prediction model, the currently calculated deviation is used as the input, and a gating signal is determined in combination with a preset threshold range (e.g., lower limit 0.12% / month, upper limit 0.20% / month).

[0059] In the first case, for example, if the calculated deviation is 0.08% / month at a certain moment, which is less than the preset lower limit of the confidence threshold of 0.12% / month, it indicates that the current degradation state is better than the historical normal fluctuation range. The model is determined to be in a stable or conservative update stage. Therefore, the first gating signal is generated to allow the model to perform a full parameter update in order to fully absorb the long-term trend changes brought about by the new data.

[0060] In the second case, for example, if the deviation is 0.15% / month, which is in the range of 0.12% / month to 0.20% / month, it indicates that the current degradation level is fluctuating normally but has a slight trend of change. At this time, a second gating signal is generated to restrict the model to update parameters only under a small learning rate, such as reducing the learning rate to 30% to 50% of the original value, in order to avoid the model overfitting short-term fluctuations.

[0061] In the third case, for example, if the deviation reaches 0.25% / month, exceeding the preset threshold limit of 0.20% / month, it indicates that the photovoltaic module may have an abnormal accelerated degradation or sudden performance deterioration risk. At this time, a third gating signal is generated to prohibit the model parameter update, thereby preventing abnormal data from polluting the model structure and triggering further fault diagnosis or alarm mechanisms.

[0062] The gating signal generation method described in this application achieves adaptive control of the parameter update behavior of the lifetime prediction model by comparing the deviation corresponding to the current degradation rate with a preset confidence threshold in a hierarchical manner. When the deviation is below the lower limit of the preset confidence threshold, a first gating signal is generated, allowing full parameter updates to enhance the model's learning ability to stable degradation patterns. When the deviation is between the upper and lower limits of the preset confidence threshold, a second gating signal is generated, allowing parameter updates under a limited learning rate, to reduce the impact of moderate fluctuations on the stability of model parameters. When the deviation exceeds the upper limit of the preset confidence threshold, a third gating signal is generated, prohibiting parameter updates to prevent abnormal data, sudden operating conditions, or error disturbances from causing erroneous learning in the model. Based on this, dynamic gating adjustment of the model update process is achieved, thereby significantly improving the stability, robustness, and prediction reliability of the lifetime prediction model in complex operating environments. A flowchart of the gating signal generation method of this application is provided below. Figure 2 .

[0063] In some embodiments, the loss function including the parameter drift penalty term includes a prediction residual loss term, a parameter drift penalty term, and a prediction monotonicity soft constraint loss term, wherein: The prediction residual loss term is used to characterize the error between the current predicted value and the actual degraded observation value. It is calculated using the symmetric Huber loss function, and its breakpoint parameter is set to 0.5 times the standard deviation of the historical degradation rate to balance the ability to suppress outliers and the stability of fitting normal samples. The parameter drift penalty term is used to constrain the magnitude of model parameter updates. It is obtained by calculating the weighted L2 norm between the updated model parameter vector and the corresponding model parameter vector in the previous update cycle. The weight matrix of the weighted L2 norm is composed of the diagonal elements of the Fisher information matrix corresponding to each parameter. The diagonal elements of the Fisher information matrix are estimated based on the exponential moving average of the squared mean of the gradients over the past thirty update cycles to improve the numerical stability of parameter sensitivity estimation under limited data conditions. The predictive monotonic soft constraint loss term is used to constrain the changing trend of the degradation prediction sequence. It applies a second penalty to non-monotonic decreasing changes in the degradation prediction sequence. The penalty amount is proportional to the square of the corresponding violation magnitude to ensure the continuity and reasonableness of the degradation prediction results.

[0064] For example, during the training of a photovoltaic module lifetime prediction model, degradation data from multiple consecutive time windows are selected for loss function calculation, with the output power attenuation rate used as the degradation observation value.

[0065] In calculating the residual loss term, for example, if the model predicts a degradation rate of 0.38% / month at a certain point in time, while the actual observed rate is 0.32% / month, the residual is 0.06% / month. Based on historical statistics, the standard deviation of the degradation rate is 0.10% / month, so the breakpoint parameter is set to 0.5 times this, i.e., 0.05% / month. In this case, the symmetric Huber loss function is used to handle this residual: since 0.06% / month is slightly larger than the breakpoint range of 0.05% / month, the loss function weakens this bias, thereby suppressing the impact of abnormal errors while maintaining the model's stability in fitting normal samples.

[0066] In the calculation of the parameter drift penalty term, for example, after the model's parameter vector is... After the (k-1)th update, the parameter vector is When there is a significant difference in L2 values ​​between the two, a penalty constraint is introduced. The weight matrix is ​​composed of the diagonal elements of the Fisher information matrix. For example, if a key parameter has a high Fisher information value (e.g., 0.8), then that parameter has a larger weight in the drift penalty, thus limiting its rapid changes. The diagonal elements of the Fisher information matrix are estimated using the exponential moving average of the squared gradients over the past 30 update cycles. For example, parameters with large recent gradient fluctuations have higher EMA values, thereby enhancing their stability constraints and preventing the model from over-updating under small data perturbations.

[0067] In predicting the monotonic soft-constraint loss term, for example, the degradation sequence output by the model is... (Unit: % / month), where a non-monotonic decreasing trend appears between 0.33 and 0.32, indicating a reverse fluctuation. The magnitude of this violation of monotonicity, 0.01, is used as a penalty, and the penalty value (0.01) is calculated in squared form. 2 =0.0001), and a loss function is added to suppress non-physically reasonable fluctuations in the predicted sequence.

[0068] The loss function described in this application, which includes a parameter drift penalty term, achieves joint constraints on the update process of the photovoltaic device degradation prediction model by introducing a prediction residual loss term, a parameter drift penalty term, and a prediction monotonicity soft constraint loss term. Specifically, the error between the current predicted value and the actual degradation observation value is calculated using the symmetric Huber loss function, and the inflection point parameter is set to 0.5 times the standard deviation of the historical degradation rate. This ensures the stability of fitting normal samples while improving the ability to suppress abnormal biases. A parameter drift penalty term is constructed based on the Fisher information matrix, and the parameter sensitivity is estimated using the exponential moving average of the gradient squared mean over the past thirty update cycles. This achieves adaptive constraints on the update magnitude of key model parameters, thereby reducing the risk of model overfitting and parameter drift under limited data conditions. Furthermore, a secondary penalty is applied to the non-monotonic decreasing changes in the degradation prediction sequence to enhance the continuity and trend rationality of the degradation prediction results. This significantly improves the stability, robustness, and prediction accuracy of the lifetime prediction model under complex operating conditions, meeting the requirements for long-term online degradation assessment of photovoltaic devices.

[0069] In some embodiments, the weight coefficients corresponding to the predicted residual loss term, the parameter drift penalty term, and the predicted monotonicity soft constraint loss term are set independently and are all real numbers greater than zero; wherein, the weight coefficient corresponding to the parameter drift penalty term is the constraint coefficient of the parameter drift penalty term, and the weight coefficient corresponding to the predicted monotonicity soft constraint loss term is fixedly set to one-tenth of the weight coefficient of the predicted residual loss term.

[0070] For example, in the training process of a photovoltaic module lifetime prediction model, the total loss function, which includes a parameter drift penalty term, consists of three parts: prediction residual loss term, parameter drift penalty term, and prediction monotonicity soft constraint loss term. The weight coefficients of each term are set independently and are all positive real numbers.

[0071] In specific settings, for example, the weighting coefficient of the predicted residual loss term is set to... , as the basic fitting weights; the constraint coefficient corresponding to the parameter drift penalty term is set to It is used to control the sensitivity of the model parameter update magnitude, so that the model avoids excessive drift when learning new data.

[0072] Meanwhile, the weighting coefficient for the predicted monotonic soft constraint loss term is fixed at one-tenth of the weighting coefficient for the predicted residual loss term, i.e. This is used to apply weak constraints to the non-monotonic fluctuations of the predicted sequence while ensuring the physical rationality of the degradation trend, so as to avoid excessively affecting the model's ability to express real short-term fluctuations.

[0073] For example, in a single training iteration, if the prediction residual loss is 0.05, the parameter drift penalty is 0.02, and the monotonicity constraint loss is 0.01, then the total loss L is calculated as follows: L=1.0×0.05+0.3×0.02+0.1×0.01=0.05+0.006+0.001=0.057.

[0074] In some embodiments, the degradation prediction model employs a hybrid prediction architecture combining a unidirectional causal long short-term memory network and a Bayesian neural network, wherein: A one-way causal long short-term memory network is used to extract temporal features from a historical operating condition feature sequence of length T time steps and output the final hidden state vector. It performs forward inference based only on the current time and historical time data. A Bayesian neural network is used to receive the final hidden state vector and perform probability mapping to output the degradation prediction result and the corresponding uncertainty. The network weight parameters are modeled using a Gaussian distribution, and the mean and log-variance parameters are both trained parameters for optimization. During the inference phase, the Bayesian neural network uses the Monte Carlo Dropout strategy for posterior approximation. By performing no less than 50 random forward propagations and statistically fusing multiple output results, stable prediction results and uncertainties are obtained. During the online update process, only the mean parameters of the Bayesian neural network and the output layer parameters of the unidirectional causal long short-term memory network are updated, while the recurrent kernel parameters of the unidirectional causal long short-term memory network are frozen. The time step length T is adaptively set according to the degradation time constant of the photovoltaic device, and its initial value corresponds to the number of time steps corresponding to one percent of the rated life. When the historical data length is insufficient, all available historical data is used as input, and proportional weight compensation is applied to the corresponding sample loss.

[0075] In this embodiment, the hybrid architecture of the degradation prediction model is as follows: Figure 3 As shown in the diagram, the input is a historical operating condition feature sequence with a length of T time steps; a unidirectional causal long short-term memory network extracts the temporal features from this sequence, with its recurrent kernel parameters remaining frozen during online updates, and only the output layer parameters participating in the update; a Bayesian neural network receives the final hidden state vector and, through performing at least 50 Monte Carlo Dropout random forward propagations, jointly outputs the predicted remaining useful life and the corresponding uncertainty.

[0076] The degradation prediction model described in this application combines a unidirectional causal long short-term memory network (LSTM) with a Bayesian neural network to extract degradation time-series features and model prediction uncertainties for photovoltaic devices. The unidirectional causal LSTM network performs causal time-series modeling based on current and historical data, improving the stability and real-time performance of degradation state feature extraction. The Bayesian neural network combines Gaussian prior modeling with Monte Carlo Dropout posterior approximation to achieve a joint output of degradation prediction results and corresponding uncertainties, enhancing the reliability and robustness of lifetime prediction results. Furthermore, by freezing the cyclic kernel parameters and updating only a portion of the network parameters, the risk of parameter drift and catastrophic forgetting during online updates is reduced. Simultaneously, by combining adaptive time step length setting and sample weight compensation mechanisms, the model's adaptability and prediction stability under different operating stages are improved, thus meeting the long-term online lifetime prediction requirements of photovoltaic devices under complex operating conditions.

[0077] In some embodiments, the method for outputting the predicted remaining useful life and its uncertainty using the updated degradation prediction model is as follows: Input the historical operating condition feature sequence corresponding to the current moment into the updated degradation prediction model, perform multiple Monte Carlo random forward propagations, and obtain multiple sets of remaining useful life prediction results and corresponding prediction variance components. The sample variance is calculated based on multiple sets of remaining useful life prediction results, and the sample variance is determined as cognitive uncertainty; at the same time, the prediction variance component of the model's explicit output is extracted, and the prediction variance component is determined as random uncertainty. Based on historical comparison data between the most recent preset quantity of historical prediction results and the actual degradation quantity, the cognitive uncertainty and random uncertainty are calibrated by temperature to obtain the corresponding correction coefficients, and the cognitive uncertainty and random uncertainty are corrected by the correction coefficients. The overall uncertainty is calculated based on the corrected cognitive uncertainty and random uncertainty. Multiple sets of remaining useful life prediction results obtained from multiple random forward propagations are constructed as multiple Gaussian mixture components of a Gaussian mixture model. The mixture weights are determined based on the reciprocal of the overall uncertainty corresponding to each Gaussian mixture component, and the mixture weights exceeding the preset upper limit are constrained and allocated. Based on the constrained mixing weights, Bayesian weighted fusion is performed on each Gaussian mixture component to calculate the mean and variance of the mixture distribution, and the mean is determined as the predicted remaining useful life. Based on the current overall uncertainty, the exponential smoothing coefficient is determined, and the remaining useful life prediction value is subjected to first-order exponential smoothing filtering to obtain the smoothed remaining useful life prediction value and the corresponding uncertainty.

[0078] For example, in the lifespan prediction process of a certain photovoltaic module, at the current moment... The historical operating condition characteristic sequence (including output power decay sequence, component temperature sequence, irradiance sequence and ambient wind speed sequence) is used as input, and the remaining useful life (RUL) is estimated using the updated degradation prediction model.

[0079] First, by performing multiple Monte Carlo stochastic forward propagations, such as 50 stochastic forward calculations, different random deactivation or weight perturbations are introduced into the model each time, resulting in 50 sets of remaining lifetime prediction results, such as [820 days, 790 days, 805 days, 810 days, ...]. At the same time, the model outputs the corresponding prediction variance components, such as a local prediction variance value accompanying each forward propagation, to characterize the impact of data noise.

[0080] Based on the above 50 sets of prediction results, the sample variance is calculated. For example, the calculated sample variance is 400 (days). 2 This is used as cognitive uncertainty to characterize the prediction fluctuations caused by parameter uncertainty in the model; simultaneously, the prediction variance component of the model's explicit output is extracted, for example, the average prediction variance is 250 (days). 2 This is used as random uncertainty to reflect the sources of uncertainty caused by data noise or measurement errors.

[0081] Subsequently, temperature calibration was performed based on the comparison error between the most recent 30 historical prediction results and the actual degradation observations. For example, if the model is found to have a systematic overestimation trend in the near future, the cognitive uncertainty correction coefficient of 0.9 and the random uncertainty correction coefficient of 1.1 are obtained respectively, and the two types of uncertainty are corrected to improve the reliability of uncertainty estimation.

[0082] In the calculation of overall uncertainty, for example, the corrected cognitive uncertainty is 360 (days). 2 The random uncertainty is 275 (days). 2 The overall uncertainty can then be expressed as a weighted combination of the two (e.g., a sum of squares or a linear combination), yielding a combined uncertainty of approximately 635 (days). 2 ), used to uniformly characterize the reliability of predictions.

[0083] Furthermore, the 50 sets of RUL prediction results are constructed as multiple mixture components of a Gaussian mixture model. For example, each Monte Carlo prediction result corresponds to a Gaussian distribution component, with its mean being the single prediction value and its variance being the corresponding uncertainty. Then, the mixture weights are determined based on the reciprocal of the overall uncertainty of each component; for example, components with lower uncertainty (more stable predictions) receive higher weights. If the weight of a component exceeds a preset upper limit (e.g., 0.2), normalization constraints are applied to redistribute the weights within a reasonable range.

[0084] Subsequently, Bayesian weighted fusion is performed on each Gaussian component based on the constrained mixture weights to calculate the overall mean and variance of the mixture distribution. For example, the mean after fusion is 805 days and the variance is 390 days. 2 And this mean will be used as the final remaining useful life prediction.

[0085] Finally, the exponential smoothing coefficient is determined based on the current overall uncertainty. For example, a smaller smoothing coefficient of 0.6 is used when the uncertainty is large, and 0.8 is used when the uncertainty is small. The RUL prediction value is then subjected to first-order exponential smoothing. For example, the final smoothed remaining useful life is 798 days. The corresponding uncertainty is output simultaneously for subsequent operation and maintenance decisions and risk warnings.

[0086] The method described in this application for outputting the predicted remaining useful life and its uncertainty involves inputting historical operating condition characteristic sequences into an updated degradation prediction model and performing multiple Monte Carlo random forward propagations to jointly obtain the remaining useful life prediction results and corresponding uncertainties. By separately calculating the sample variance and model output variance components, cognitive uncertainty and random uncertainty are distinguished, and temperature calibration is performed in conjunction with historical prediction errors to improve the accuracy and reliability of uncertainty estimation. Furthermore, a Gaussian mixture model is constructed, and the mixture weights are adaptively allocated based on the reciprocal of the overall uncertainty to achieve Bayesian weighted fusion of multiple prediction results, thereby enhancing the stability and robustness of the lifetime prediction results. Simultaneously, exponential smoothing filtering based on the overall uncertainty is used to reduce the impact of short-term fluctuations on the prediction results, achieving a smooth and continuous remaining useful life estimation output, meeting the long-term online lifetime prediction requirements of photovoltaic devices under complex operating environments.

[0087] In some embodiments, the method for determining the maximum single-step lifetime reduction based on the current environmental stress intensity is as follows: Obtain the photovoltaic device temperature, cumulative irradiance dose, and wind speed parameters at the current moment; The thermal stress component is calculated using the Arrhenius model based on the difference between the photovoltaic device temperature and the reference temperature; the light-induced attenuation stress component is calculated based on the ratio between the cumulative irradiance dose and the rated annual irradiance dose; and the mechanical stress component is calculated based on the positive part of the difference between the wind speed parameter and the safe wind speed threshold, combined with the wind speed corresponding to the rated pressure limit of the photovoltaic device. The weighting coefficients corresponding to the thermal stress component, the photo-induced attenuation stress component, and the mechanical stress component are obtained. The current environmental stress intensity is obtained by weighted summation of each stress component based on the weighting coefficients. The weighting coefficients are obtained by partial least squares regression fitting based on the historical accelerated aging test data of the corresponding device model. The environmental stress intensity is input into a preset piecewise mapping function. In the low stress range, the maximum single-step lifetime reduction is determined according to the step size corresponding to the nominal decay rate. In the high stress range, the maximum single-step lifetime reduction is determined according to the power function of the environmental stress intensity. The power exponent of the power function is not less than 1. Based on the predicted remaining lifetime value and the maximum single-step lifetime decrease output at the previous moment, determine the upper bound of the constraint on the remaining lifetime at the current moment. If the predicted remaining useful life at the current moment is lower than the upper limit of the constraint, the predicted remaining useful life at the current moment is directly output; otherwise, the predicted remaining useful life at the current moment is corrected to the lower limit of the constraint, and the corrected predicted remaining useful life is output.

[0088] For example, in a real-world photovoltaic power station, a single photovoltaic module at a given time... Environmental monitoring data was used as input for calculations. At that moment, the module temperature was measured at 65℃, and the cumulative irradiation dose was 5.2 kWh / m². 2 (Accounting for 0.012% of the annual rated irradiation dose), wind speed is 14 m / s, reference temperature is set at 25℃, safe wind speed threshold is 12 m / s, and the wind speed corresponding to the rated pressure limit is 25 m / s.

[0089] First, in the calculation of thermal stress components, based on the Arrhenius model, the temperature difference is... Substituting 65-25=40℃ into the exponential model, we obtain the thermal stress component, which corresponds to an acceleration factor of 1.85, characterizing the accelerated effect of high temperature on lifespan decay.

[0090] In the calculation of photo-induced attenuation stress components, for example, if the ratio of cumulative irradiation dose to annual rated value is 0.012, then the photo-induced attenuation stress component is 0.012, which characterizes the contribution of cumulative irradiation to material aging.

[0091] In the calculation of mechanical stress components, the portion of wind speed exceeding the safety threshold is 14-12=2m / s, which is then normalized by combining it with the rated pressure wind speed of 25m / s. For example, the mechanical stress component is obtained as 0.08, which characterizes the intensity of the influence of wind load on structural fatigue.

[0092] Subsequently, the three stress components are weighted and fused. For example, the weight coefficients obtained by partial least squares regression are: thermal stress 0.6, light-induced attenuation 0.25, and mechanical stress 0.15. Then the environmental stress intensity S is calculated as: S=0.6×1.85+0.25×0.012+0.15×0.08≈1.11.

[0093] In determining the maximum single-step lifetime reduction, the environmental stress intensity is input into a preset piecewise mapping function. For example, in the low-stress range (S < 0.5), the calculation is based on a nominal decay rate of 0.5 days / step, while in the high-stress range (S ≥ 0.5), a power function is used. Perform mapping. Since the current S=1.11 is in the high stress range, the maximum single-step life reduction is calculated as follows: Where k=10, then sky.

[0094] In lifetime constraint calculations, for example, if the predicted remaining lifetime at the previous moment was 800 days, then the current upper limit of the constraint is 800 - 11.5 = 788.5 days. If the current model predicts 790 days, it exceeds the upper limit of the constraint and is corrected to 788.5 days as the final output; if the predicted value is 780 days, it is within the constraint range and the original prediction result is output directly.

[0095] The method for determining the maximum single-step lifetime decline described in this application acquires parameters such as photovoltaic device temperature, cumulative irradiance dose, and wind speed, and constructs thermal stress, light-induced attenuation stress, and mechanical stress components to achieve a comprehensive characterization of the current environmental stress state. Specifically, based on the Arrhenius model, the irradiance dose ratio, and the wind speed safety threshold difference, the environmental impact corresponding to different degradation mechanisms is quantitatively analyzed, improving the accuracy of environmental stress assessment. Furthermore, by combining weighting coefficients obtained from fitting historical accelerated aging experimental data, each stress component is weighted and fused to obtain the current environmental stress intensity. The maximum single-step lifetime decline is adaptively determined through a preset piecewise mapping function, achieving dynamic adjustment of the lifetime decay rate under different stress ranges. Simultaneously, the current prediction result is constrained and corrected by combining the remaining lifetime prediction value from the previous moment, ensuring that the remaining lifetime prediction result meets the monotonically decreasing constraint, thereby improving the physical rationality, stability, and reliability of the lifetime prediction result.

[0096] In some embodiments, the method for determining whether the change magnitude exceeds a preset mutation determination threshold is as follows: Obtain the current remaining useful life prediction value and the corresponding remaining useful life prediction sequence for the past thirty days; Linear regression analysis is performed based on the remaining useful life prediction sequence to obtain the corresponding linear regression slope, and the expected single-step decline is estimated based on the linear regression slope. Calculate the absolute value of the difference between the current predicted remaining useful life and the expected decrease in one step, and determine the absolute value as the magnitude of change; Obtain the standard deviation of the predicted remaining useful life series over the past year, and determine the first-level warning threshold and the second-level confirmation threshold based on the standard deviation. The first-level warning threshold is twice the standard deviation, and the second-level confirmation threshold is four times the standard deviation. The change amplitude is compared with the first-level warning threshold and the second-level confirmation threshold respectively. When the change amplitude exceeds the first-level warning threshold but does not exceed the second-level confirmation threshold, the system enters a suspended state and acquires the change amplitude corresponding to the next three consecutive sampling periods. Determine whether the cumulative change amplitude corresponding to the last three sampling periods shows a continuous increasing trend. If the continuous increasing condition is met, it is determined that the change amplitude exceeds the preset sudden change judgment threshold; otherwise, the suspended state is released and normal output is restored. When the change exceeds the secondary confirmation threshold, it is directly determined that the change exceeds the preset mutation judgment threshold; The first-level warning threshold and the second-level confirmation threshold are periodically updated based on the latest statistical data. During the recovery period corresponding to a major maintenance event, the automatic update of the preset mutation judgment threshold is frozen. The recovery period is determined based on the standard recovery time constant of the corresponding event type.

[0097] For example, in the online prediction process of the remaining useful life of a photovoltaic module, the system at the current moment... The predicted remaining useful life is 720 days. Simultaneously, the predicted remaining useful life sequence for the past 30 days is retrieved, for example: [750 days, 748 days, 747 days, 745 days, 742 days, ..., 723 days].

[0098] First, based on the above 30-day prediction sequence, a linear regression analysis was performed, and the linear regression slope was found to be approximately -1.0 day / day, indicating that the life expectancy prediction value decreased by an average of 1 day per day. Based on this, the single-step expected decrease was estimated to be 1 day.

[0099] Then, the deviation between the current predicted value and the theoretical decreasing trend is calculated. For example, if the predicted value at the previous time was 723 days, based on the expected decrease, the theoretical predicted value at the current time should be 722 days, while the actual predicted value is 720 days. The change is then calculated as follows: .

[0100] Next, obtain the statistical distribution of the predicted remaining useful life series over the past year. For example, if its standard deviation is 1.5 days, then: Level 1 warning threshold = 2 × 1.5 = 3 days; The secondary confirmation threshold is 4 × 1.5 = 6 days.

[0101] In this example, since the current change over 2 days has not exceeded the first-level warning threshold of 3 days, the system determines that the current change is within the normal fluctuation range and continues to output the prediction result normally.

[0102] In another embodiment, for example, if a calculation shows a change of 4 days, this exceeds the first-level warning threshold of 3 days but does not reach the second-level confirmation threshold of 6 days. Therefore, the system enters a suspended state and continues to monitor the next three sampling cycles.

[0103] For example, the variation ranges for the last three sampling periods are as follows: Period 1: 4.2 days; Period 2: 4.8 days; Period 3: 5.5 days.

[0104] Since the cumulative change shows a continuous increasing trend, the change is determined to be a real lifespan mutation event. This confirms that the change exceeds the preset mutation judgment threshold and triggers an abnormal degradation alarm or model protection mechanism.

[0105] Conversely, if the changes in the last three cycles are 4.2 days, 3.6 days, and 2.8 days respectively, it indicates that the abnormal fluctuations have not continued to increase, but have gradually returned to stability. Therefore, the suspension status should be lifted and the normal prediction output should be restored.

[0106] Furthermore, in extreme cases, such as when the current change rate directly reaches 7 days, exceeding the secondary confirmation threshold of 6 days, there is no need to enter the suspension observation phase. Instead, a lifespan mutation is directly determined, and an abnormal response is immediately triggered.

[0107] In addition, during the dynamic updating of thresholds, for example, the standard deviation is recalculated monthly based on the latest one-year forecast data, and the first-level warning threshold and the second-level confirmation threshold are automatically updated; however, if the photovoltaic module has just completed major maintenance (such as module replacement or inverter repair), the threshold is frozen and automatically updated within the corresponding recovery period (e.g., 30 days) to avoid short-term fluctuations after maintenance interfering with the mutation judgment mechanism.

[0108] The method for determining abrupt changes in the aforementioned variation range in this application estimates the expected single-step decline by performing linear regression analysis combining the current predicted remaining useful life with the predicted sequence of the past thirty days. It then determines the variation range based on the deviation between the current predicted value and the expected single-step decline, thus achieving quantitative detection of abnormal fluctuations in lifespan prediction. Furthermore, it adaptively determines the primary warning threshold and the secondary confirmation threshold based on the statistical characteristics of the predicted value sequence of the past year, improving the adaptability of the abrupt change determination threshold to long-term operational status changes. On this basis, it reduces the risk of misjudgment caused by short-term noise fluctuations by setting a suspension state and a trend confirmation mechanism for multiple consecutive sampling periods, and performs rapid confirmation when the variation range exceeds the secondary confirmation threshold, improving the responsiveness to real abnormal degradation events. Simultaneously, it incorporates a threshold freezing mechanism during the recovery period of major maintenance events to avoid maintenance interventions from disturbing the abrupt change determination results, thereby significantly improving the accuracy, stability, and reliability of lifespan prediction anomaly detection.

[0109] In some embodiments, the attribution determination method is as follows: When the magnitude of the change exceeds the preset mutation judgment threshold, the corresponding abnormal change is marked as a mutation event; Obtain the operating condition feature vectors within one sampling window before and after the occurrence of the mutation event, calculate the mean of the corresponding operating condition feature vectors, and construct a regularized covariance matrix based on the sample covariance matrix; Based on the regularized covariance matrix, the Mahalanobis distance between the mean values ​​of the characteristic vectors of the operating conditions before and after the mutation event is calculated, and the Mahalanobis distance is compared with a preset threshold obtained based on the statistical analysis of the historical operating condition change magnitude to determine whether the current mutation is related to the change in operating conditions. If the current mutation is determined to be related to the change in operating conditions, the operation and maintenance event records within a preset time range before and after the occurrence of the mutation event are retrieved, and if there are operation and maintenance events of a preset type, the determination results of the correlation between the change in operating conditions are weighted and enhanced. The degradation rate time series corresponding to the mutation event is obtained. The degradation rate time series consists of degradation rate data at the time of mutation occurrence and the time before and after sampling. Seasonal difference processing is performed on it to obtain the degradation rate series after removing seasonal components. A trend significance test was performed on the degradation rate sequence after removing seasonal components to determine whether there was an accelerating degradation trend. Based on the results of the correlation determination of operating condition changes, the results of weight enhancement processing, and the results of the determination of degradation acceleration trend, the system uses preset attribution decision rules to perform fusion determination and output the corresponding attribution category.

[0110] For example, during the prediction of the remaining useful life of a photovoltaic module, it was detected that the predicted remaining useful life value at the current moment suddenly dropped from 720 days at the previous moment to 690 days, with a change of 30 days, exceeding the preset mutation judgment threshold of 20 days. Therefore, this abnormal change was marked as a mutation event.

[0111] Subsequently, the operating condition feature vectors are obtained within one sampling window before and after the mutation event. For example, the average operating condition feature vector obtained statistically within the window before the mutation is: [Component temperature 45℃, irradiance 820W / m] 2 [Wind speed 3m / s, output power 325W]; The mean value corresponding to the window after the mutation is: [Component temperature 63℃, irradiance 910W / m] 2 Wind speed 11m / s, output power 295W.

[0112] Furthermore, a sample covariance matrix is ​​constructed based on historical operating condition sample data, and a regularized covariance matrix is ​​generated by adding a diagonal regularization term to reduce numerical instability caused by feature correlation. Subsequently, the Mahalanobis distance between the operating condition means before and after the abrupt change is calculated using this regularized covariance matrix; for example, the calculated result is 5.8.

[0113] The Mahalanobis distance is compared with a threshold (e.g., 4.0) obtained based on historical operating condition changes. Since 5.8 is greater than 4.0, it is determined that the current mutation is strongly correlated with the operating condition change, that is, the mutation may be affected by changes in the environment or operating conditions.

[0114] Based on this, further retrieval of maintenance event records within 24 hours before and after the sudden change event is performed. For example, if preset maintenance events such as "inverter parameter adjustment" and "component cleaning and maintenance" are found during this period, the determination results of the correlation between operating conditions and changes are weighted to enhance the credibility of "changes in operating conditions leading to sudden changes".

[0115] Subsequently, the corresponding degradation rate time series was obtained. For example: the degradation rate before the mutation was 0.32% / month; the degradation rate at the time of the mutation was 0.47% / month; and the degradation rate after the mutation was 0.56% / month.

[0116] Seasonal differencing was performed on the degradation rate sequence to remove the periodic effects of diurnal temperature variation and seasonal radiation fluctuations, resulting in a deseasonalized degradation rate sequence. Further trend significance tests (e.g., the Mann-Kendall trend test) were used to analyze its changing trend. If the test results showed a significant and continuous upward trend in the degradation rate, it was determined that accelerated degradation was present.

[0117] Finally, combining: Results of correlation determination for changes in operating conditions; Enhanced results for operational events; Results of assessment of accelerated degradation trend; The analysis was performed according to the pre-set attribution decision rules.

[0118] For example: If the operating condition changes are highly correlated, maintenance events occur, and there is no obvious trend of accelerated degradation, the output attribution category is "operating condition disturbance type sudden change"; If the correlation between the operating condition changes is weak but there is a significant trend of accelerated degradation, the output attribution category is "accelerated device aging mutation". If both strong operating condition changes and significant accelerated degradation exist simultaneously, the output attribution category will be "compound factor mutation".

[0119] The attribution judgment method described in this application marks abnormal changes as mutation events when the change magnitude exceeds a preset mutation judgment threshold. It then calculates the Mahalanobis distance by combining the mean of the characteristic vectors of operating conditions before and after the mutation event with the regularized covariance matrix, thereby achieving a quantitative analysis of the correlation between mutation events and operating condition changes. Furthermore, it enhances the judgment results of the correlation between operating condition changes by combining historical operating condition change statistical thresholds and maintenance event records, improving the ability to identify the impact of maintenance operations and environmental disturbances. Based on this, it identifies the true acceleration trend of degradation by performing seasonal differencing and trend significance testing on the degradation rate time series. Simultaneously, it integrates the results of operating condition change correlation, maintenance event impact, and degradation trend analysis using preset attribution decision rules for fusion judgment, thereby accurately distinguishing between true degradation acceleration and abnormal situations such as operating condition fluctuations and erroneous updates, improving the accuracy, stability, and reliability of life prediction anomaly handling. A general schematic diagram of the attribution judgment process of this application can be found in [link to application]. Figure 4 .

[0120] In some embodiments, the method for reducing the constraint coefficient of the parameter drift penalty term is as follows: Obtain the current degradation rate and the historical average degradation rate, and calculate the multiple by which the current degradation rate exceeds the historical average degradation rate; The degree of degradation acceleration is determined based on the multiple exceeded, and the degree of degradation acceleration includes mild acceleration, moderate acceleration and severe acceleration. The corresponding reduction ratio of the constraint coefficient is determined according to the degree of degradation acceleration. Among them, mild acceleration corresponds to a reduction ratio of 20% of the baseline value, moderate acceleration corresponds to a reduction ratio of 50% of the baseline value, and severe acceleration corresponds to a reduction ratio of 80% of the baseline value. The constraint coefficients corresponding to the parameter drift penalty terms are adjusted based on the reduction ratio of the constraint coefficients, and the adjusted constraint coefficients are maintained during the period of degradation acceleration. During the period when the constraint coefficient is decreasing, the learning rate for updating model parameters is adaptively scaled based on the inverse of the degradation acceleration to adjust the step size of model parameter updates in a synchronous manner. Continuously monitor whether the subsequent degradation rate is within the upper bound corresponding to the historical degradation rate mean plus one standard deviation for five consecutive update cycles; When the degradation rate meets the conditions for five consecutive update cycles, the constraint coefficient is restored cycle by cycle according to the preset fixed step size, and the constraint coefficient is restored to the baseline value within ten update cycles.

[0121] The parameter drift penalty term constraint coefficient adjustment method described in this application classifies the degree of degradation acceleration by calculating the multiple by which the current degradation rate exceeds the historical average degradation rate. By setting different constraint coefficient reduction ratios for mild, moderate, and severe acceleration, the model can adaptively release parameter update constraints according to the degree of degradation acceleration, thereby enhancing its responsiveness to the actual degradation acceleration state. Furthermore, the model parameter update learning rate is synchronously scaled in conjunction with the degree of degradation acceleration, achieving coordinated adjustment of parameter constraint strength and update step size, reducing the risk of unstable updates during severe degradation. Simultaneously, by continuously monitoring subsequent degradation rate changes and restoring constraint coefficients periodically at a preset step size after the degradation state stabilizes, the model update mechanism smoothly transitions from rapid adaptation to stable convergence, thereby improving the stability, robustness, and adaptability of the lifetime prediction model in complex degradation scenarios.

[0122] In some embodiments, the parameter update step of the rollback degradation prediction model is implemented through a parameter snapshot buffer; the parameter snapshot buffer adopts a first-in-first-out circular queue structure to store the most recent... A complete snapshot of the model parameters after the first successful parameter update, in which... The value is set based on the system's storage resource limits, with a default value of 10; After each successful parameter update, the current model parameter state is serialized and stored in the parameter snapshot buffer. At the same time, the corresponding timestamp, the deviation at the time of update trigger, and the loss function values ​​before and after the update are recorded to form a parameter update history. When a rollback operation is triggered, the most recently stored parameter snapshot before the current mutation determination is retrieved from the parameter snapshot buffer, and the parameters of the degradation prediction model are restored based on the parameter snapshot, so that the model parameters are rolled back to the corresponding historical state. After the parameter rollback is completed, the input feature vectors and real observations corresponding to the rolled-back parameter update process are added to the calibration monitoring set. In subsequent model updates, the prediction error is calculated in real time for the samples in the calibration monitoring set. When the prediction error is lower than the threshold corresponding to the historical prediction error mean plus one standard deviation for three consecutive times, the corresponding sample is removed from the calibration monitoring set. Otherwise, the conservative update mode is triggered, and the learning rate corresponding to the current update step is reduced to one-fifth of the normal learning rate to reduce the risk of parameter oscillation. When suppressing the output of mutation prediction values, linear interpolation is used to smooth the output results at the current time. The starting point of the linear interpolation is the mean of the prediction values ​​corresponding to the three sampling periods before the mutation occurs, and the ending point is the prediction value after the next update, so as to ensure the continuity and stability of the remaining lifetime prediction sequence.

[0123] In some embodiments, the method further includes: when the photovoltaic device belongs to a photovoltaic array containing two or more devices of the same type, using degradation rate observation information of other devices in the array to assist in updating the degradation state of the current device, specifically including the following steps: Obtain the historical sequence of degradation rates for each photovoltaic device of the same model within the photovoltaic array over the past ninety days. The Pearson correlation coefficient between each photovoltaic device was calculated based on the historical degradation rate sequence, and a spatial correlation coefficient matrix of degradation rate was constructed. Significance tests were performed on the spatial correlation coefficient matrix of degradation rates to screen and retain device associations with Pearson correlation coefficients greater than 0.7 and corresponding two-sided p-values ​​of t-test less than 0.05. Continuously monitor the residual of the degradation rate of each photovoltaic device in the array relative to the mean degradation rate of the array, and determine whether the absolute value of the corresponding residual exceeds three times the standard deviation of the degradation rate in the array for seven consecutive update cycles. When any photovoltaic device meets the above conditions, it is marked as an abnormal device and removed from the subsequent spatial correlation coefficient matrix calculation. When updating the parameters of the current photovoltaic device, the degradation rates of other photovoltaic devices that meet the screening criteria are weighted and fused, and the corresponding Pearson correlation coefficient is used as the weight to obtain the auxiliary supervision signal. The auxiliary monitoring signal is introduced as an auxiliary loss term into the loss function of the degradation prediction model, and the contribution weight of the auxiliary loss term is adaptively adjusted based on the current photovoltaic device observation data quality. The observation data quality is determined in real time based on the outlier ratio in the outlier removal step.

[0124] For example, in a distributed photovoltaic power station, a photovoltaic array contains 20 photovoltaic modules of the same model, and it is necessary to update the degradation status of the photovoltaic device numbered PV-08.

[0125] First, the historical degradation rate sequence of all 20 photovoltaic modules in the array over the past 90 days was obtained. For example, the average degradation rate of PV-08 over the past 90 days was approximately 0.34% / month, while the degradation rates of the other modules ranged from 0.30% / month to 0.36% / month.

[0126] Subsequently, Pearson correlation coefficients were calculated between each pair of components based on their historical degradation rate sequences. For example, the correlation coefficient between PV-08 and PV-03 was 0.86; the correlation coefficient between PV-08 and PV-11 was 0.79; and the correlation coefficient between PV-08 and PV-15 was 0.42.

[0127] Furthermore, significance tests were performed on the above correlations. For example, the two-sided p-value for PV-03 was 0.003; for PV-11, the p-value was 0.012; and for PV-15, the p-value was 0.18.

[0128] Therefore, only the component associations that satisfy "correlation coefficient > 0.7 and p value < 0.05" are retained, that is, PV-03 and PV-11 are retained, while the correspondence of PV-15 is removed, thereby constructing an effective degradation rate spatial correlation coefficient matrix.

[0129] During operation, the residuals of the degradation rate of each component relative to the average degradation rate of the array are continuously monitored. For example, the current average degradation rate of the array is 0.33% / month, while component PV-17 has shown a high degradation rate of over 0.50% / month for seven consecutive update cycles. The absolute value of its residuals has consistently exceeded three times the standard deviation of the array degradation rate (e.g., 0.04% / month), i.e., over 0.12% / month.

[0130] Since PV-17 meets the continuous anomaly condition, it is marked as an abnormal device and removed in the subsequent spatial correlation coefficient matrix calculation to avoid local faulty components interfering with the overall collaborative update.

[0131] When updating model parameters in PV-08, the degradation rates of associated components that meet the screening criteria are weighted and fused. For example: The current degradation rate of PV-03 is 0.35% / month, with a corresponding correlation coefficient of 0.86; The current degradation rate of PV-11 is 0.33% / month, with a corresponding correlation coefficient of 0.79.

[0132] The auxiliary monitoring signal is calculated as follows: (0.35×0.86+0.33×0.79)÷(0.86+0.79)≈0.340% / month.

[0133] Subsequently, this auxiliary supervision signal is introduced as an auxiliary loss term into the loss function of the degradation prediction model corresponding to PV-08, in order to constrain its prediction results to be consistent with the overall degradation trend of the array.

[0134] Furthermore, the contribution weight of the auxiliary loss term is dynamically adjusted based on the current quality of the PV-08 observation data. For example, if the outlier rate of the current PV-08 data is only 2% during the outlier removal process, indicating high data quality, the weight of the auxiliary loss term is set low, such as 0.15. If the outlier rate rises to 20%, indicating a decrease in the reliability of the current observations, the weight of the auxiliary loss term is automatically increased to 0.40, thereby enhancing the auxiliary role of array collaborative information in model updates.

[0135] The degradation state-assisted update method based on array spatial correlation described in this application obtains the historical degradation rate sequence of devices of the same model within the photovoltaic array and constructs a degradation rate spatial correlation coefficient matrix to achieve quantitative analysis of the correlation of device degradation behavior within the array. By screening the Pearson correlation coefficient and the corresponding significance test results, stable correlations between devices are retained, improving the reliability of auxiliary supervision information. On this basis, by continuously monitoring the degradation rate residuals of each device and removing devices that deviate from the overall degradation law of the array over a long period, the interference of abnormal devices on the array collaborative analysis results is reduced. At the same time, when the current device performs parameter updates, auxiliary supervision signals are constructed by combining the degradation rate information of other highly correlated devices, and the contribution weight of the auxiliary loss term is adaptively adjusted according to the current observation data quality, thereby enhancing the stability and accuracy of degradation state updates under limited observation conditions and improving the robustness and collaborative perception capability of the lifetime prediction model in the photovoltaic array scenario.

[0136] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting the lifetime of photovoltaic devices based on artificial intelligence, characterized in that, include: Based on the current operating condition feature vector of the photovoltaic device, calculate the deviation of the current degradation rate from the historical average degradation rate; The deviation is compared with a preset threshold to generate a gating signal for updating the model parameters; When the gating signal allows updates, the parameters of the degradation prediction model are updated based on the loss function that includes a parameter drift penalty term; The updated degradation prediction model is used to output the remaining useful life prediction value and its uncertainty. Based on the uncertainty, the remaining useful life prediction value is fused by Bayesian weighted fusion to obtain the remaining useful life estimate. The maximum single-step life reduction is determined based on the current environmental stress intensity, and the remaining service life estimate is constrained and corrected to obtain the constrained and corrected remaining service life estimate.

2. The method according to claim 1, characterized in that, The method for generating the gating signal is as follows: when the deviation is lower than the lower limit of the preset threshold, a first gating signal that allows full parameter updates is generated; when the deviation is between the upper and lower limits of the preset threshold, a second gating signal that allows parameter updates under the condition of a limited learning rate is generated. When the deviation exceeds the preset upper limit of the threshold, a third gating signal is generated to prohibit parameter updates.

3. The method according to claim 1, characterized in that, The loss function includes a prediction residual loss term, a parameter drift penalty term, and a prediction monotonicity soft constraint loss term, wherein: the prediction residual loss term is calculated using the symmetric Huber loss function to characterize the error between the current predicted value and the actual degraded observation value; the parameter drift penalty term is obtained by calculating the weighted L2 norm between the updated model parameter vector and the corresponding model parameter vector of the previous update period, and is used to constrain the magnitude of model parameter updates; the prediction monotonicity soft constraint loss term is used to constrain the changing trend of the degraded prediction sequence.

4. The method according to claim 1, characterized in that, The degradation prediction model adopts a hybrid prediction architecture that combines a one-way causal long short-term memory network and a Bayesian neural network. The one-way causal long short-term memory network is used to extract time-series features from the historical operating condition feature sequence with a length of T time steps and output the final hidden state vector. The Bayesian neural network is used to receive the final hidden state vector and perform probability mapping, outputting the degradation prediction result and the corresponding uncertainty.

5. The method according to claim 1, characterized in that, The method for outputting the remaining useful life prediction value and its uncertainty using the updated degradation prediction model is as follows: Based on the updated degradation prediction model, multiple sets of remaining useful life prediction results and corresponding prediction variance components are obtained. The sample variance is calculated based on multiple sets of remaining useful life prediction results, and the sample variance is determined as cognitive uncertainty; at the same time, the prediction variance component of the model's explicit output is extracted, and the prediction variance component is determined as random uncertainty. Based on the comparison data between the historical prediction results of the most recent preset quantity and the actual degradation quantity, the cognitive uncertainty and random uncertainty are calibrated by temperature to obtain the corresponding correction coefficients, and the cognitive uncertainty and random uncertainty are corrected based on the correction coefficients. The overall uncertainty is calculated based on the corrected cognitive uncertainty and random uncertainty. Multiple sets of remaining useful life prediction results are constructed into multiple Gaussian mixture components of a Gaussian mixture model, and the mixture weights are determined based on the reciprocal of the overall uncertainty corresponding to each Gaussian mixture component. Bayesian weighted fusion is performed on each Gaussian mixture component based on the mixture weight to obtain the mean and variance of the mixture distribution, and the mean is determined as the predicted value of the remaining useful life. Based on the overall uncertainty, the exponential smoothing coefficient is determined, and a first-order exponential smoothing filter is applied to the remaining useful life prediction value to obtain the smoothed remaining useful life prediction value and the corresponding uncertainty.

6. The method according to claim 1, characterized in that, The method for determining the maximum single-step lifetime reduction based on the current environmental stress intensity is as follows: The current environmental stress intensity is obtained by weighted summing of the thermal stress component, light-induced attenuation stress component, and mechanical stress component of the photovoltaic device at the current moment. The environmental stress intensity is input into a preset piecewise mapping function. In the low stress range, the maximum single-step lifetime reduction is determined according to the step amount corresponding to the nominal decay rate. In the high stress range, the maximum single-step lifetime reduction is determined according to the power function of the environmental stress intensity. Based on the predicted remaining lifetime value and the maximum single-step lifetime decrease output at the previous moment, determine the upper bound of the constraint on the remaining lifetime at the current moment. If the predicted remaining useful life at the current moment is lower than the upper limit of the constraint, the predicted remaining useful life at the current moment is directly output; otherwise, the predicted remaining useful life at the current moment is corrected to the lower limit of the constraint, and the corrected predicted remaining useful life is output.

7. The method according to claim 1, characterized in that, The method further includes the following steps: In the process of obtaining the estimated remaining useful life, the magnitude of change of the predicted remaining useful life is calculated, and it is determined whether the magnitude of change exceeds the preset abrupt change judgment threshold. If the limit is not exceeded, the estimated remaining useful life after constraint correction will be output directly. If the result exceeds the limit, an attribution judgment is performed, and the attribution result is obtained.

8. The method according to claim 7, characterized in that, The method for determining whether the change exceeds the preset mutation threshold is as follows: Obtain the standard deviation of the predicted remaining service life sequence within the historical preset period, and determine the first-level warning threshold and the second-level confirmation threshold based on the standard deviation; When the change exceeds the first-level warning threshold but does not exceed the second-level confirmation threshold, the current state will be marked as suspended. Determine whether the cumulative change amplitude of multiple subsequent sampling periods meets the condition of continuous increase: if it does, determine that the change amplitude exceeds the preset abrupt change judgment threshold; Otherwise, the suspended state will be lifted and normal output will resume; When the change exceeds the secondary confirmation threshold, it is directly determined that the change exceeds the preset mutation judgment threshold.

9. The method according to claim 7, characterized in that, The method further includes the following steps: Perform the following actions based on the attribution results: If it is determined to be a true acceleration of degradation, the constraint coefficient of the parameter drift penalty term is reduced, and the estimated remaining useful life after constraint correction is output. If the error is determined to be a fluctuation in operating conditions or an incorrect update, the parameter update of the degradation prediction model is rolled back, and the current abrupt prediction value is identified as an outlier and masked so that it does not participate in the generation and output of the remaining useful life estimate. The output is rolled back to the remaining useful life estimate corresponding to the state before the update.

10. The method according to any one of claims 1-9, characterized in that, The method further includes: when the photovoltaic device belongs to a photovoltaic array containing two or more devices of the same type, using the degradation rate observation information of other devices in the array to assist in updating the degradation state of the current device, specifically including the following steps: Based on the historical sequence of degradation rate of photovoltaic array within a preset time window, the Pearson correlation coefficient between each photovoltaic device is calculated, and a spatial correlation coefficient matrix of degradation rate is constructed. A significance test was performed on the spatial correlation coefficient matrix to screen and retain device associations with Pearson correlation coefficients greater than a preset correlation threshold and corresponding statistical significance levels that meet the preset significance level. Based on the mean degradation rate within the array, the residual sequence of degradation rate of each photovoltaic device is calculated, and the continuity anomaly of the residual sequence is determined. When the absolute value of the residual of any photovoltaic device exceeds the standard deviation of a preset multiple within a preset number of consecutive update cycles, it is marked as an abnormal device and its data is removed. When the current photovoltaic device performs model parameter updates, the degradation rates of other photovoltaic devices that have been screened for correlation and not marked as anomalies are weighted and fused to obtain auxiliary monitoring signals; The auxiliary monitoring signal is constructed as an auxiliary loss term and introduced into the loss function of the degradation prediction model. The contribution weight of the auxiliary loss term is adaptively adjusted based on the current photovoltaic device observation data quality.

Images