Photovoltaic device stability prediction method based on multi-modal model and related device

Abstract

The embodiment of the invention provides a photovoltaic device stability prediction method based on a multi-modal model and a related device, and the method comprises the steps: inputting a multi-modal sample set into a trained multi-modal deep neural network model, and outputting a high-dimensional feature vector representing the state of a comprehensive device; and then, inputting the high-dimensional feature vector into a time sequence prediction sub-model to obtain a photoelectric conversion efficiency PCE prediction value of each time node in the full-life cycle of the photovoltaic device to be detected, and drawing a performance attenuation curve to realize accurate pre-judgment of the full-life cycle performance of the photovoltaic device. In addition, the stability rating of the photovoltaic device to be tested can be generated according to the performance attenuation curve, key factors influencing the stability of the device are positioned when the stability rating is detected to be in an unstable state, and device optimization and improvement suggestions are generated, so that the stability evaluation period is greatly shortened, the test cost is reduced, and efficient technical support is provided for device research and development and mass production optimization.

Description

Technical Field

[0001] This application belongs to the field of photovoltaic technology, and in particular relates to a method and related apparatus for predicting the stability of photovoltaic devices based on a multimodal model. Background Technology

[0002] Photovoltaic devices (such as perovskite solar cells) have high photoelectric conversion efficiency, but their efficiency is prone to degradation and their stability is poor. How to predict and delay the efficiency degradation of perovskite solar cells and improve the operational stability of the devices are key technical problems that urgently need to be solved in this field.

[0003] Most existing research on the reliability of photovoltaic devices uses single-mode data or methods to assess and predict device degradation. This approach fails to fully describe the entire chain of changes in perovskite solar cells from "material structure state - photoelectric performance - eventual failure" and also makes it difficult to accurately capture the complex laws governing efficiency decay caused by the interaction of multiple degradation factors. Summary of the Invention

[0004] This application provides a method and related apparatus for predicting the stability of photovoltaic devices based on a multimodal model. By using multimodal data fusion and time series prediction technology, it can accurately predict the performance and stability rating of photovoltaic devices throughout their entire life cycle. It can locate the core influencing factors of unstable states and generate targeted optimization suggestions, which can significantly shorten the stability assessment cycle and reduce testing costs, providing efficient technical support for device R&D and mass production optimization.

[0005] In a first aspect, embodiments of this application provide a method for predicting the stability of photovoltaic devices based on a multimodal model. The method includes: collecting multi-source data from the photovoltaic device under test, wherein the multi-source data includes at least two of image data, spectral data, initial photoelectric data, and environmental stress parameter sequences; performing modal adaptation preprocessing on the multi-source data to obtain a multimodal sample set; inputting a training sample set from the multimodal sample set into a trained multimodal deep neural network model, and outputting a high-dimensional feature vector representing the overall device state; the multimodal deep neural network model includes parallelly configured modal subnetworks and a feature fusion layer, wherein each modal subnetwork is used to extract features from the training sample set, and the feature fusion layer is used for... The features extracted from each modal sub-network are fused to obtain the high-dimensional feature vector; the high-dimensional feature vector is input into the time series prediction sub-model to obtain the predicted photoelectric conversion efficiency (PCE) at each time point throughout the entire life cycle of the photovoltaic device under test and to plot the performance degradation curve; a stability rating of the photovoltaic device under test is generated based on the performance degradation curve, the stability rating including stable and unstable states; if the stability rating is detected to be unstable, key factors affecting device stability are determined based on the performance degradation curve and the multi-source data, the key factors including at least one of defect regions, sensitive spectral bands, and environmental influencing factors; device optimization and improvement suggestions are generated based on the key factors.

[0006] Secondly, embodiments of this application provide a photovoltaic device stability prediction device based on a multimodal model. The device includes: a data acquisition unit for acquiring multi-source data of the photovoltaic device under test, the multi-source data including at least two of image data, spectral data, initial photoelectric data, and environmental stress parameter sequences; a processing unit for performing modal adaptation preprocessing on the multi-source data to obtain a multimodal sample set; inputting a training sample set from the multimodal sample set into a trained multimodal deep neural network model, and outputting a high-dimensional feature vector representing the overall device state; the multimodal deep neural network model includes parallelly configured modal subnetworks and a feature fusion layer, each modal subnetwork being used for feature extraction from the training sample set. The feature fusion layer is used to fuse the features extracted by each modal sub-network to obtain the high-dimensional feature vector; the high-dimensional feature vector is input into the time series prediction sub-model to obtain the predicted photoelectric conversion efficiency (PCE) values ​​at each time point in the entire life cycle of the photovoltaic device under test and to plot the performance degradation curve; a stability rating of the photovoltaic device under test is generated based on the performance degradation curve, the stability rating including stable and unstable states; if the stability rating is detected to be unstable, key factors affecting the stability of the device are determined based on the performance degradation curve and the multi-source data, the key factors including at least one of defect regions, sensitive spectral bands, and environmental factors; and device optimization and improvement suggestions are generated based on the key factors.

[0007] Thirdly, embodiments of this application provide a server, including a processor and a memory, wherein the memory stores a computer program, and when the processor invokes the computer program in the memory, it executes the step instructions of the method as described in any one of the first aspects.

[0008] Fourthly, embodiments of this application provide a computer-readable storage medium having a computer program / instructions stored thereon, which, when executed by a processor, implement the steps of any of the possible methods in the first aspect.

[0009] As can be seen, in this embodiment, multi-source data of the photovoltaic device under test is collected. The multi-source data includes at least two of the following: image data, spectral data, initial photoelectric data, and environmental stress parameter sequences. Modal adaptation preprocessing is performed on each of the multi-source data to obtain a multi-modal sample set. The training sample set from the multi-modal sample set is input into a trained multi-modal deep neural network model, which outputs a high-dimensional feature vector representing the overall device state. The multi-modal deep neural network model includes parallelly configured modal subnetworks and a feature fusion layer. Each modal subnetwork is used to extract features from the training sample set, and the feature fusion layer is used to extract features from each modal subnetwork. The extracted features are fused to obtain a high-dimensional feature vector. This high-dimensional feature vector is then input into a time-series prediction sub-model to obtain the predicted photoelectric conversion efficiency (PCE) values ​​for each time point throughout the entire lifespan of the photovoltaic device under test, and to plot performance degradation curves. Based on the performance degradation curves, a stability rating for the photovoltaic device under test is generated, including stable and unstable states. If an unstable state is detected, key factors affecting device stability are identified based on the performance degradation curves and multi-source data. These key factors include at least one of defect regions, sensitive spectral bands, and environmental factors. Based on these key factors, device optimization and improvement suggestions are generated. Therefore, compared to existing technologies, multi-modal data fusion and time-series prediction technologies enable accurate prediction and stability rating of photovoltaic device performance throughout its entire lifespan. This allows for the identification of core influencing factors of unstable states and the generation of targeted optimization suggestions, significantly shortening the stability assessment cycle, reducing testing costs, and providing efficient technical support for device R&D and mass production optimization. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 A flowchart illustrating a method for predicting the stability of photovoltaic devices based on a multimodal model, provided in an embodiment of this application; Figure 2 A schematic diagram of the residual efficiency prediction process for a comparative example provided in this application; Figure 3 A schematic flowchart of a perovskite device stability prediction method based on the fusion of EL images and electrical test data provided in an embodiment of this application; Figure 4 A schematic diagram of the multimodal Transformer prediction process based on EL images, PL spectra, and electrical parameters provided for embodiments of this application; Figure 5A schematic diagram of the closed-loop prediction process for multimodal and time series fusion provided in an embodiment of this application; Figure 6 A functional unit structure block diagram of a photovoltaic device stability prediction device based on a multimodal model provided in this application embodiment; Figure 7 This is a schematic diagram of the structure of a server provided in an embodiment of this application. Detailed Implementation

[0012] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present application.

[0013] The terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but in some embodiments includes steps or units not listed, or in some embodiments includes other steps or units inherent to these processes, methods, products, or apparatuses.

[0014] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0015] In the embodiments of this application, "and / or" describes the relationship between associated objects, indicating that three relationships can exist. For example, A and / or B can represent the following three situations: A exists alone; A and B exist simultaneously; B exists alone. Among them, A and B can be singular or plural.

[0016] In this embodiment, the symbol " / " can indicate that the preceding and following objects are in an "or" relationship. Alternatively, the symbol " / " can also represent a division sign, i.e., performing a division operation. For example, A / B can mean A divided by B.

[0017] In the embodiments of this application, "at least one item" or its similar expression refers to any combination of these items, including any combination of a single item or a plurality of items. "One or more" means one or more, while "multiple" means two or more. For example, "at least one item" of a, b, or c can represent the following seven cases: a, b, c; a and b; a and c; b and c; a, b, and c. Each of a, b, and c can be an element or a set containing one or more elements.

[0018] In the embodiments of this application, "equal to" can be used with "greater than" and is applicable to technical solutions used when "greater than" is used; it can also be used with "less than" and is applicable to technical solutions used when "less than" is used. When "equal to" is used with "greater than", it is not used with "less than"; when "equal to" is used with "less than", it is not used with "greater than".

[0019] To address the aforementioned technical challenges, this application provides a method and apparatus for predicting the stability of photovoltaic devices based on a multimodal model. By integrating multimodal data fusion and time series prediction techniques, it enables accurate prediction and stability rating of the photovoltaic device's performance throughout its entire lifecycle. It can pinpoint the core influencing factors of unstable states and generate targeted optimization suggestions, significantly shortening the stability assessment cycle and reducing testing costs, thus providing efficient technical support for device R&D and mass production optimization.

[0020] The technical solution of this application and how it solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will be described below with reference to the accompanying drawings. Please refer to... Figure 1 , Figure 1 A flowchart illustrating a photovoltaic device stability prediction method based on a multimodal model, as provided in this application embodiment, is shown below. Figure 1 As shown, the method includes the following steps S101-S107: Step S101: Collect multi-source data of the photovoltaic device under test.

[0021] The multi-source data includes at least two of the following: image data, spectral data, initial photoelectric data, and environmental stress parameter sequences. However, the multi-source data described in this application is not limited to the above four types; other types of data can be added as needed. For example, X-ray diffraction (XRD) information on material phase transitions can be added, or infrared thermography can be used to monitor the battery's heat distribution in real time to further enrich the characterization of the failure mechanism. If new detection methods (such as terahertz spectroscopy) become sensitive to PSC degradation in the future, they can also be incorporated into this system as a new modal extension.

[0022] In this embodiment, a photovoltaic device refers to a semiconductor device that directly converts solar energy into electrical energy using the photovoltaic effect. Its core structure typically includes functional layers such as a light absorption layer, an electrode layer, and a charge transport layer. Photovoltaic conversion can be achieved based on different material systems and device structures. Common types include crystalline silicon photovoltaic devices (monocrystalline silicon, polycrystalline silicon), thin-film photovoltaic devices (perovskite, copper indium gallium selenide, cadmium telluride, etc.), and tandem photovoltaic devices (such as perovskite-crystalline silicon tandem).

[0023] The core performance indicator of these devices is photoelectric conversion efficiency (PCE), and its long-term stability directly determines the device's service life and application value. Environmental stress (temperature, humidity, and light), intrinsic material defects, and residual problems from the manufacturing process can all cause irreversible degradation of device performance. The method described in this application is applicable to the full life cycle stability prediction of various photovoltaic devices. Through multimodal data fusion and time-series prediction models, it enables the prediction of performance degradation trends, stability rating, and failure attribution analysis of photovoltaic devices with different material systems.

[0024] In some embodiments, the image data is used to reflect the grain morphology and defect distribution of the photovoltaic device under test, the spectral data is used to reflect the material structure and stress information of the device, the initial photoelectric data includes open-circuit voltage Voc, short-circuit current density Jsc, fill factor FF and initial photoelectric conversion efficiency PCE, and the environmental stress parameter sequence includes temperature T, relative humidity RH and light intensity during testing or use.

[0025] In this embodiment, at least two of the following multi-source data are selected: image data, spectral data, initial photoelectric data, and environmental stress parameter sequences. All data are collected in the initial stage after the photovoltaic device under test is manufactured (e.g., within 24 hours after production). Real-time monitoring throughout the entire life cycle is not required. The specific data collection method is as follows: Image data is obtained by capturing images of the device surface and / or cross-section using high-resolution microscopic imaging equipment. These images reflect the device's grain morphology (such as grain size and density) and defect distribution (such as the location and morphology of pinholes, cracks, and impurity agglomerations). If a scanning electron microscope (SEM) is used, the magnification can be adjusted according to the device type, and the image resolution can be adapted to the subsequent model input requirements. Specifically, a Zeiss Sigma 300 SEM with a magnification of 5000x and an image resolution of 2048×2048 pixels can be selected. Spectral data: Characteristic spectra reflecting the material structure of the device are acquired using spectroscopic testing equipment, which are used to reflect the material's crystal structure, chemical bond state, and stress information; such as Raman spectroscopy and / or photoluminescence (PL) spectroscopy. The Raman spectroscopy scanning range is adapted to the characteristic peak range of the device material, and the PL spectroscopy excitation wavelength is selected according to the material's band gap; specifically, a Horiba LabRAM HR Evolution Raman spectrometer (scanning range 100-800 cm⁻¹) can be used. - ¹, spectral resolution 0.5 cm - ¹) Edinburgh FLS1000 fluorescence spectrometer (excitation wavelength 532 nm, scanning range 600-800 nm). Initial photoelectric data: Photovoltaic device performance testing equipment is used to test under standard test conditions to obtain core parameters characterizing the initial photoelectric conversion capability of the device. These parameters include at least the photoelectric conversion efficiency (PCE), and may also include open-circuit voltage (Voc), short-circuit current density (Jsc), and fill factor (FF). For example, using a Newport Oriel Sol3A photovoltaic device tester under AM 1.5G spectrum, standard illuminance, and room temperature conditions, the measured initial data could be: Voc = 1.02V, Jsc = 23.5mA / cm², FF = 0.78, PCE = 18.9%. Environmental stress parameter sequence: Based on the target usage scenario of the device under test, environmental stress time series data for the entire life cycle are obtained, including at least one or more of temperature (T), relative humidity (RH) and light intensity. The time step of the time series data is matched with the prediction accuracy requirements. For example, if the target usage scenario is an outdoor scenario in East China, the environmental historical data of the region for the past 10 years are retrieved from the meteorological database and organized into a time series (time step of 1 day, total length of 10 years) according to the daily average value, which is used to simulate the environmental service conditions of the device throughout its entire life cycle.

[0026] Step S102: Perform modal adaptation preprocessing on the multi-source data to obtain a multimodal sample set.

[0027] Specifically, appropriate preprocessing operations are performed for different modal data characteristics. The core objectives are to eliminate noise interference, unify data format and dimensions, and ensure that the data can be adapted to subsequent model inputs. The preprocessing logic and examples for various types of data are as follows: Image data preprocessing: This includes at least noise reduction, normalization, and size standardization. Additional steps such as grayscale conversion and enhancement can be added as needed to ensure that the image data meets the input requirements of the modal subnetwork. Spectral data preprocessing: includes at least baseline correction, smoothing and noise reduction, and normalization. Interpolation alignment steps can be added as needed to standardize the length of spectral data. Initial photoelectric data preprocessing: Standardization or normalization methods are used to eliminate the influence of dimensions and ensure the comparability between different parameters; Environmental stress parameter sequence preprocessing: includes at least missing value imputation, outlier removal and normalization / standardization operations, and must be aligned with the time series according to the preset time step to ensure that it matches the input time series of the subsequent time series prediction sub-model; The preprocessed modal data are combined according to the rule of "one device corresponds to one set of multimodal data" to construct a multimodal sample set. The size of the sample set is determined according to the model training requirements and is divided into a training sample set and a validation sample set according to a preset ratio. For example, a sample set is constructed by collecting multi-source data from 100 photovoltaic devices of the same model, of which 80 sets are used as the training sample set and 20 sets are used as the validation sample set.

[0028] Step S103: Input the training sample set from the multimodal sample set into the trained multimodal deep neural network model, and output a high-dimensional feature vector representing the state of the integrated device.

[0029] The multimodal deep neural network model includes parallel modal subnetworks and a feature fusion layer. Each modal subnetwork is used to extract features from the training sample set, and the feature fusion layer is used to fuse the features extracted by each modal subnetwork to obtain the high-dimensional feature vector.

[0030] In this embodiment, the core structure of the multimodal deep neural network model is "parallel configuration of each modal sub-network + feature fusion layer". Each modal sub-network is adapted to the feature extraction requirements of the corresponding modal data, and the feature fusion layer realizes the correlation modeling and fusion of cross-modal features, and outputs a high-dimensional feature vector representing the state of the integrated device.

[0031] In some embodiments, the modal subnetwork includes an image subnetwork, a spectral subnetwork, and a numerical subnetwork. The image subnetwork uses a convolutional neural network (CNN) or a lightweight CNN to extract image features, which include texture and structural features. The spectral subnetwork uses a spectral analysis network or a coding network to extract spectral features, which include peak positions and intensity features. The numerical subnetwork uses a fully connected network or a convolutional network to extract numerical features of the initial photoelectric data and the environmental stress parameter sequence.

[0032] The image sub-network employs a convolutional neural network (CNN) or a lightweight CNN capable of extracting image texture and structural features. It takes preprocessed image data as input and outputs a fixed-dimensional image feature vector. Image features include texture and structural features; texture features reflect the arrangement of grains and defect edges on the device surface, while structural features reflect grain size and defect distribution. The spectral sub-network employs a spectral analysis or encoding network capable of extracting peak positions and intensities. It takes standardized spectral data as input and outputs a fixed-dimensional spectral feature vector. Spectral features include peak position and peak intensity features; peak position features reflect the crystal structure type and chemical bond type of the device material, while peak intensity features reflect the content and distribution density of the corresponding structure or chemical bond. The numerical sub-network employs a fully connected or convolutional network capable of extracting numerical features. It takes standardized initial photoelectric data and environmental stress parameters as input and outputs a fixed-dimensional numerical feature vector. Numerical features reflect the magnitude relationship, trend of change, and correlation between parameters. Each sub-network outputs a fixed-dimensional feature vector for subsequent feature fusion.

[0033] In some embodiments, the feature fusion layer employs a multi-head self-attention mechanism of a Transformer network to achieve cross-modal feature association modeling through self-attention weight calculation; or, a multimodal variational autoencoder (VAE) is employed to encode each modal feature into jointly distributed latent variables, and then generate a fused high-dimensional feature vector through a decoding process.

[0034] The feature fusion layer employs a fusion mechanism capable of modeling cross-modal correlations, such as the multi-head self-attention mechanism of the Transformer network or the multimodal variational autoencoder (VAE). For example, if the Transformer multi-head self-attention mechanism is used (the number of heads can be set to 8, and the hidden layer dimension can be set to 512), the feature vectors output by each modal sub-network are input, and cross-modal correlations (such as the correlation between "defect region-spectral feature peak attenuation") are mined through self-attention weight calculation, and finally a 1024-dimensional high-dimensional feature vector is output.

[0035] Input the training sample set of the multimodal sample set into the trained model (the optimizer, learning rate, number of iterations and other parameters of the model training can be adjusted according to the training effect), and output the high-dimensional feature vector corresponding to each training sample.

[0036] It should be noted that, besides Transformer and VAE, other architectures can be used to implement the network module for multimodal feature fusion. For example, convolutional neural networks with attention mechanisms can be combined, or graph neural networks can be used to treat each modality's features as nodes for fusion. These alternatives are still based on the idea of ​​multimodal deep learning and are equivalent replacements for the core technology.

[0037] Step S104: Input the high-dimensional feature vector into the time series prediction sub-model to obtain the predicted photoelectric conversion efficiency (PCE) values ​​at each time point in the entire life cycle of the photovoltaic device under test and plot the performance degradation curve.

[0038] Among them, the performance degradation curve, with time as the horizontal axis and photoelectric conversion efficiency (PCE) as the vertical axis, clearly presents the complete PCE change trajectory of the device from the initial service state to the failure state. It can directly reflect the degradation characteristics of different time stages, such as the division of the initial slow degradation stage, the intermediate accelerated degradation stage, and the final rapid degradation stage, as well as the difference in degradation rate of each stage.

[0039] In some embodiments, the time series prediction sub-model employs a Long Short-Term Memory (LSTM) network or an improved structure thereof, the improved structure including a bidirectional LSTM and a gated recurrent unit (GRU). The specific process by which the time series prediction sub-model generates the predicted power conversion efficiency (PCE) values ​​for each time point throughout the entire lifespan of the photovoltaic device under test and plots the performance degradation curve includes: inputting the high-dimensional feature vector at time t and the corresponding environmental stress parameters into the LSTM or its improved structure; processing the input data at time t through multiple memory gate units to output the predicted device performance value at time t+1; and processing the data at time t+1... The environmental stress parameters at time t+1 are integrated with the performance prediction value at time t+1 to form the input data at time t+1. t is updated to t+1, and the steps "inputting the high-dimensional feature vector at time t and the corresponding environmental stress parameters into the LSTM or its improved structure" and subsequent steps are repeated iteratively to obtain the PCE prediction values ​​at times t+2, t+3...t+n, until the predicted PCE value decreases to a preset device failure threshold, covering the entire lifespan. The PCE prediction values ​​at each time point within the entire lifespan are correlated sequentially along the time axis to form a continuous performance degradation curve.

[0040] In this embodiment, the time series prediction sub-model adopts a time series prediction model capable of capturing temporal dependencies. Optional models include Long Short-Term Memory (LSTM) networks and their improved structures (such as bidirectional LSTM, Gated Recurrent Units (GRUs)). The core process involves iterative prediction to achieve full-lifecycle PCE prediction and plotting the decay curve. The full-lifecycle PCE prediction and decay curve plotting process is based on a closed-loop iterative logic of "initial input - single-step prediction - iterative update - curve plotting." By integrating the prediction result of the previous moment with the corresponding environmental stress parameters as the input for the next moment, continuous prediction throughout the entire lifecycle is achieved until the predicted PCE value drops to a preset failure threshold, ensuring that the prediction covers the complete lifecycle of the device from service to failure.

[0041] The logical steps for the time series prediction sub-model to generate full-lifecycle PCE prediction values ​​and plot performance degradation curves are as follows: Steps S401-S405: Step S401, Initial Input: Integrate the high-dimensional feature vector with the environmental stress parameters at the initial time step (time t) to form the initial input data, and input it into the time series prediction sub-model; Step S402, Single-step prediction: The time series prediction model processes the input data at time t through its internal memory mechanism and outputs the PCE prediction value at time t+1. Step S403, Iterative update: Integrate the environmental stress parameters at time t+1 with the PCE prediction value at time t+1 to form the input data at time t+1; Step S404, iterative loop: update time step t to t+1, repeat the above single-step prediction step and subsequent steps, and obtain the PCE prediction value for each subsequent time node in sequence; when the predicted PCE value drops to the preset device failure threshold (such as 50% of the initial PCE, which can be preset according to device type, application scenario and industry standard), stop the iteration to achieve full life cycle coverage. Step S405, plot the degradation curve: associate the PCE prediction values ​​of each time node (t, t+1, t+2...t+n) in the whole life cycle in the order of time axis, and use conventional data visualization tools to plot a continuous performance degradation curve. The curve can reflect the degradation trend of device performance (such as initial slow degradation, mid-term accelerated degradation, and final rapid degradation).

[0042] It should be noted that although LSTM performs well in this invention, in some cases, simplified structures such as Transformer-based time series models (e.g., self-attention-based time series prediction models) or Gate RecurrentUnit (GRU) can be used to replace LSTM to predict performance degradation curves. These variations can achieve similar prediction results while keeping the input-output relationship unchanged.

[0043] As can be seen, in this embodiment, the LSTM and improved gating mechanism can accurately capture the long-term time-series dependence of PCE degradation in photovoltaic devices. Compared with traditional time-series prediction models (such as ARIMA and traditional RNN), the mean absolute error (MAE) of the PCE prediction value is reduced, ensuring the accuracy of the degradation trend prediction throughout the entire life cycle. Through the logic of "iterative update-threshold stop", the entire cycle from the start of service to failure of the device can be automatically covered without manual intervention to determine the prediction duration, avoiding the problem of incomplete stability assessment caused by insufficient test cycle in traditional testing. The plotted continuous performance degradation curve can intuitively show the difference in degradation rate at different stages of the device, providing accurate time-series data support for subsequent stability rating and key factor positioning, and helping to quickly identify the core stage of performance degradation.

[0044] In some embodiments, before generating the stability rating of the photovoltaic device under test based on the performance degradation curve, the method further includes: calculating the coefficient of determination R², mean absolute error MAE, and root mean square error RMSE on the validation set of the multimodal sample set to verify the model prediction accuracy.

[0045] Among them, R², MAE, and RMSE complement each other in representing prediction accuracy: R² is used to measure the degree of fit between the predicted value and the true value (the value ranges from [0,1], and the closer it is to 1, the better the fit); MAE is used to reflect the average absolute deviation between the predicted value and the true value (the smaller the value, the smaller the deviation); and RMSE is used to amplify the impact of larger deviations (more sensitive to extreme errors, and the smaller the value, the higher the accuracy). Through multi-indicator collaborative verification, the misjudgment of accuracy by a single indicator can be completely avoided, ensuring that the model's prediction ability meets the requirements of stability rating.

[0046] The specific process for verifying the model prediction accuracy is as follows: Steps S501-S504: Step S501: Retrieve the multimodal sample set constructed in step S102 and separate the pre-divided validation set (the validation set consists of samples in the multimodal sample set that did not participate in model training, including multi-source preprocessed data of the validation samples and the corresponding real PCE data. The real PCE data can be obtained through a small amount of accelerated aging tests or short-term field tests and is used for comparison with the predicted values).

[0047] Step S502: Input the multi-source preprocessed data of the validation set into the trained multimodal deep neural network model to obtain the high-dimensional feature vector of each sample in the validation set; then input the high-dimensional feature vector into the time series prediction sub-model to obtain the PCE prediction value sequence of each sample in the validation set (corresponding to the prediction period of the preset duration).

[0048] Step S503: Based on the PCE predicted value sequence of each sample in the validation set and the corresponding real PCE data sequence, calculate the three types of indicators: coefficient of determination R², mean absolute error MAE, and root mean square error RMSE.

[0049] Step S504: Preset the accuracy standards for each indicator (determined according to the type of photovoltaic device, application scenario, and industry accuracy requirements): If all three types of indicators meet the preset standards, it indicates that the model prediction accuracy is qualified, and the subsequent step of "generating the stability rating of the photovoltaic device under test based on the performance degradation curve" is executed; if the preset standards are not met, model optimization operations are performed (such as adjusting network parameters, supplementing training samples, etc.), and after optimization, steps S502-S503 are repeated until the accuracy meets the standards before entering the stability rating stage.

[0050] The preset accuracy standard can be flexibly adjusted according to the device type (such as crystalline silicon, perovskite) and application scenario (such as outdoor, indoor), which has strong adaptability and ensures that the technical solution can output reliable prediction results in different scenarios.

[0051] As can be seen, in this embodiment, by using multi-indicator collaborative verification, models with substandard prediction accuracy are screened out in advance, avoiding erroneous stability ratings based on performance degradation curves with large deviations, and ensuring that the rating results can truly reflect the stability level of the device. If the model accuracy is substandard, targeted model optimization can improve the prediction capability, avoiding the transmission of accuracy defects to subsequent key factor localization and optimization suggestion generation stages, and reducing the cost of blind optimization caused by incorrect attribution.

[0052] Step S105: Generate a stability rating for the photovoltaic device under test based on the performance degradation curve.

[0053] The stability rating includes both stable and unstable states.

[0054] A stable state refers to the condition in which the photovoltaic device under test exhibits gradual performance degradation without significant accelerated deterioration throughout its entire lifespan, meeting the long-term service requirements of the preset application scenario. Specifically, this is characterized by a low average degradation rate throughout the entire lifespan, a controllable maximum degradation rate during the accelerated degradation phase, a service life that reaches or exceeds the preset target, and gradual degradation during the initial service phase with no risk of early failure.

[0055] An unstable state refers to a situation where the photovoltaic device under test exhibits a significant accelerated performance degradation phase throughout its entire life cycle, with the degradation rate exceeding the preset acceptable range, the service life failing to meet the application scenario requirements, or the degradation occurring too rapidly in the initial service phase, posing a risk of early failure.

[0056] The stability of photovoltaic devices is essentially reflected in their ability to resist performance degradation throughout their entire lifespan, and the performance degradation curve directly reflects the changing trend of PCE over time (such as degradation rate, total degradation, and degradation stage characteristics). Therefore, the core method for stability rating is to extract key characteristic parameters from the performance degradation curve, establish a mapping relationship between key characteristic parameters and stability status (i.e., rating criteria), and finally output the stability rating result by comparing the characteristic parameters of the degradation curve of the device under test with the rating criteria.

[0057] The rating criteria must be set in conjunction with the material system of photovoltaic devices, application scenarios, industry standards, and actual usage requirements to ensure the scientific validity and usability of the rating criteria. In this embodiment, the stability rating criteria are set based on, but are not limited to, industry standard references (referring to the attenuation limit requirements for stability testing in international standards such as IEC 61215 (crystalline silicon photovoltaic devices) and IEC 61646 (thin-film photovoltaic devices), application scenario requirements (devices used in outdoor photovoltaic power stations must meet long-term stability requirements (e.g., attenuation not exceeding 20% ​​within a 25-year service life), while requirements can be appropriately relaxed for indoor photovoltaic devices), differences in material characteristics (the current mainstream stability level of perovskite photovoltaic devices is lower than that of crystalline silicon devices, requiring targeted adjustments to the rating threshold), and quantifiable operability (selecting key characteristic parameters that can be directly extracted and calculated from the attenuation curve, avoiding indicators that are difficult to quantify), etc.

[0058] The key characteristic parameters include the average daily decay rate over the entire life cycle, the maximum stage decay rate, and the time to reach the failure threshold. The average daily decay rate over the entire life cycle refers to the ratio of the total PCE decay to the number of days of service over the entire life cycle (total decay = initial PCE - PCE at failure). The maximum stage decay rate refers to the maximum daily decay rate of each stage (slow decay, accelerated decay, etc.) in the decay curve. The time to reach the failure threshold refers to the number of days required for the PCE to decrease from its initial value to the failure threshold (usually 50% of the initial PCE).

[0059] Step S106: If the stability rating is detected as the unstable state, the key factors affecting the stability of the device are determined based on the performance degradation curve and the multi-source data.

[0060] The key factors include at least one of the following: defect region, sensitive spectral band, and environmental influencing factors.

[0061] In some embodiments, determining the key factors affecting device stability based on the performance degradation curve and the multi-source dataset includes: identifying abnormal degradation intervals in the performance degradation curve, wherein the abnormal degradation intervals include time points of abrupt changes in degradation rate, accelerated degradation intervals, and final failure threshold time points; acquiring multi-source data corresponding to the abnormal intervals of the degradation curve, and visually locating the target image defect region and target sensitive spectral band that contribute the most to degradation through the attention weights output by the feature fusion layer; analyzing the correlation between the environmental stress parameters and the abnormal intervals of the degradation curve using the controlled variable method to determine the target environmental influencing factors that dominate the degradation; and fusing the target image defect region, the target sensitive spectral band, and the target environmental influencing factors by combining a preset multimodal feature-failure mechanism mapping library to output the priority ranking of the key factors.

[0062] Among them, the decay rate abrupt change time point refers to the time node where the performance decay rhythm changes significantly, the accelerated decay interval refers to the continuous period when the decay rate is significantly higher than other stages, and the final failure threshold time point refers to the time node when the PCE drops to the failure threshold. By performing slope analysis on the performance decay curve and abrupt change point detection algorithms (such as the sliding window method and the Paineman test), the decay rate abrupt change time point, the accelerated decay interval, and the final failure threshold time point can be identified, and the time range of the decay anomaly interval can be defined.

[0063] The multimodal feature-failure mechanism mapping library is a structured database built upon common knowledge, experimental data, and literature conclusions in the photovoltaic device field. It stores the correspondence between image features, spectral features, and numerical features of different photovoltaic devices (crystalline silicon, perovskite, copper indium gallium selenide, etc.) and device failure phenomena and mechanisms, as well as the influence weight of each feature on failure. Its core function is to correlate superficial features extracted from multi-source data, such as "target defect region, target sensitive spectral band, and target environmental factors," with the underlying physical / chemical mechanisms of photovoltaic device degradation, thereby achieving accurate attribution and avoiding one-sided judgments based solely on data statistics.

[0064] The multimodal feature-failure mechanism mapping library adopts a hierarchical structured design, covering four core levels: "feature layer - failure phenomenon layer - mechanism layer - weight layer". Each level corresponds one-to-one through association rules, as shown in Table 1 below: Table 1. Four-level content of the multimodal feature-failure mechanism mapping library

[0065] In the specific implementation, time-series multi-source data corresponding to the attenuation anomaly interval (such as image data, spectral data, and environmental stress parameter time-series data within the anomaly interval) are extracted from multi-source data. The extracted multi-source data of the anomaly interval is input into a trained multimodal deep neural network model (including a feature fusion layer) to obtain the attention weight matrix output by the feature fusion layer. Through attention weight visualization (generating an attention heatmap), image feature regions and spectral feature bands with weight values ​​higher than a preset threshold are selected as target image defect regions and target sensitive spectral bands, respectively. Simultaneously, based on the extracted time-series data of environmental stress parameters in the anomaly interval, a controlled variable experiment is designed: fix any two environmental stress parameters (such as temperature and light intensity), and only change a third parameter (such as relative humidity), comparing the changes in the device PCE attenuation rate under different parameter values; repeat this process to traverse all environmental stress parameters, and determine the environmental parameter with the strongest correlation to the attenuation anomaly interval as the target environmental influencing factor. Finally, the pre-defined multimodal feature-failure mechanism mapping library is retrieved, and the target image defect area, target sensitive spectral band, and target environmental influencing factors are input. The mapping library matches the failure phenomena and underlying failure mechanisms corresponding to each target feature through multi-dimensional indexing. The influence weights of each target factor are extracted, sorted from largest to smallest weight, and the priority ranking results of key factors are output. A complete attribution report of "feature-phenomenon-mechanism-weight" is also output, providing a basis for generating optimization suggestions in the future.

[0066] As can be seen, in this embodiment, the end-to-end design, which includes attenuation anomaly interval anchoring, attention weight visualization positioning, control variable attribution, and multimodal feature-failure mechanism mapping library fusion and sorting, not only achieves accurate positioning and priority ranking of key factors affecting the stability of photovoltaic devices, but also relies on underlying failure mechanisms rather than simple data statistical correlation in the attribution process, effectively avoiding the one-sidedness and misjudgment risk of traditional attribution methods. At the same time, this solution does not require device disassembly, significantly shortening the analysis cycle of key factors and significantly improving the efficiency and pertinence of stability optimization, providing reliable technical support for the R&D improvement and mass production quality control of various photovoltaic devices.

[0067] Step S107: Generate device optimization and improvement suggestions based on the key factors.

[0068] In some embodiments, generating device optimization and improvement suggestions based on the key factors includes: obtaining a preset optimization measure rule library, the optimization measure rule library including a mapping relationship between the key factors and the device optimization and improvement suggestions; obtaining corresponding device optimization and improvement suggestions based on the key factors and the optimization measure rule library; and determining the priority order of each device optimization and improvement suggestion based on the priority order of the key factors.

[0069] The optimization measure rule base is a structured knowledge base built on technical common sense, experimental optimization data and industrial application experience in the photovoltaic device field. It stores the one-to-one mapping relationship between "key factors and optimization improvement suggestions" and can also include auxiliary information such as the implementation difficulty, cost and expected effect of stability improvement of the suggestions, so as to select implementation according to actual needs. For example, for perovskite photovoltaic devices, the rule base can include mapping relationships such as "high humidity environment - optimize packaging process to improve moisture resistance" and "pinhole defects - adjust annealing process to reduce defects". For crystalline silicon photovoltaic devices, it can include mapping relationships such as "light-induced degradation - optimize doping process" and "surface oxidation - add anti-reflection coating".

[0070] In the specific implementation, the first step is to obtain a pre-defined rule library for optimization measures.

[0071] Next, based on the identified key factors of the photovoltaic device under test, corresponding optimization and improvement suggestions are retrieved from the optimization measure rule base. That is, through the multi-dimensional index of the rule base (such as device type, key factor type), specific optimization schemes that are suitable for each key factor are quickly matched. For example, if the key factor is "high humidity environment", the optimization suggestion of "using composite encapsulation layer + edge sealing enhancement" is matched; if the key factor is "edge pinhole defect", the optimization suggestion of "optimizing preparation annealing parameters" is matched.

[0072] Finally, based on the priority of key factors, the priority of each optimization and improvement suggestion is determined. The optimization suggestions corresponding to the primary key factors have the highest priority and should be implemented first to quickly solve the core attenuation problem. The optimization suggestions corresponding to the secondary key factors have the second priority, and the optimization suggestions corresponding to the auxiliary key factors have the lowest priority. For example, if the priority of key factors is "high humidity environment > edge pinhole defects > sensitive spectral bands", then the priority of optimization suggestions is "packaging process optimization > annealing process adjustment > material composition optimization".

[0073] After prioritizing the optimization suggestions, the photovoltaic devices can be optimized according to the core logic of priority-based implementation, effect verification iteration, and multi-factor collaborative optimization. That is, R&D resources are prioritized to implement the highest priority suggestions to solve the dominant degradation problem, avoiding resource dispersion. After completion, the aforementioned multi-modal data acquisition, model prediction, and stability rating methods are used for retesting. If the stability meets the standard, no further optimization is needed. If it does not meet the standard, subsequent suggestions are implemented in order of priority and retested one by one. For multi-factor coupled degradation scenarios, multi-measure collaborative adjustments can be carried out on the basis of single-factor optimization. Finally, the verified effective correspondence between "key factors - optimization measures - effect data" is updated to the optimization measure rule base to provide a reference for the optimization of similar devices.

[0074] Furthermore, in addition to rule bases, optimization recommendations can be generated automatically using machine learning methods. For example, a generative model can be trained based on historical optimization experimental data to output recommended improvement parameters when a specific degradation trajectory is observed. Different implementation methods all fall under the category of design changes aimed at the same goal.

[0075] As can be seen, in this embodiment, by accurately matching the rule base of optimization measures and correspondingly transmitting the priority of key factors, the targeted generation and orderly implementation of photovoltaic device stability optimization suggestions are realized. This not only avoids the problems of blind trial and error and resource dispersion in traditional optimization work, but also can quickly improve device stability by prioritizing the solution of dominant degradation factors, thereby significantly shortening the optimization cycle and reducing R&D costs.

[0076] This technical solution has the following technical advantages and effects: (1) Comprehensiveness: It integrates information from multiple sources such as microstructure, spectrum and electricity, encompassing the microscopic mechanism of material degradation and macroscopic performance changes, and realizes a comprehensive description of the behavior of the device throughout its entire life cycle. Compared with models that rely on only a single electrical parameter, it is more accurate and reliable.

[0077] (2) Foresight: It can predict the long-term stability and lifespan performance of devices in the early stages or even the manufacturing stage, and screen out potentially unstable devices or material formulations in advance, thereby reducing the cost of R&D trial and error.

[0078] (3) Guidance: The feedback module provides optimization suggestions, combining "prediction" and "optimization," which can directly guide researchers to improve the predicted weaknesses and increase the actual lifespan of the device. Traditional methods often can only analyze the causes after the fact, while this invention can provide early warnings and solutions in advance.

[0079] (4) Adaptability: The system has model scalability and can continuously optimize itself with the addition of new data; at the same time, it has a certain degree of universality and can be applied to the stability prediction of other new photovoltaic devices. Only the data acquisition and feature extraction modules of the corresponding modes need to be adjusted.

[0080] It should be noted that, in implementing the technical solution of this invention in detail, the model structure and process sequence can be adjusted according to actual needs. For example, the feature fusion module can be implemented using a Transformer encoder to achieve sequence feature alignment and fusion, or it can be implemented using a simpler method of concatenating features one by one and then fusing them through a fully connected network. In addition to LSTM, the time series model can also use the Transformer Time-Series model, sequence convolutional networks, etc. Environmental stress data can be used as model input features, or prediction curves under different stress conditions can be simulated for comparison. These modifications and variations are within the scope of protection of this application, provided that the overall concept remains unchanged.

[0081] The following specific embodiments are provided for illustration: Comparative Example: Electrical Single-Mode Data + Traditional Regression Model (XGBoost) This comparative example uses single-modal electrical performance data as input and employs the traditional machine learning method XGBoost to construct a performance prediction model. For detailed processing procedures, please refer to [link / reference needed]. Figure 2 , Figure 2 This is a schematic diagram of the comparative residual efficiency prediction process provided in the embodiments of this application, as shown below. Figure 2 As shown, the prediction process includes: Step S201, Data Input: Input electrical data such as Voc, Jsc, and FF; Step S202, Data preprocessing: Perform feature standardization on the electrical data to obtain a standardized dataset, and then divide it into training and validation sets; Step S203: Establish the XGBoost regression model; Step S204, Output: A predicted residual efficiency after 1000 hours of continuous operation under standard solar conditions; Step S205, Evaluation: Result feedback and optimization suggestions are generated.

[0082] This comparative example uses a single mode containing only electrical performance parameters (Voc, Jsc, FF, etc., a total of 12 features) as input. The goal is to predict the residual efficiency of a perovskite module after 1000 hours of continuous operation under a standard solar load (≈1000W / m²). XGBoost regression modeling and validation were used under a unified data partitioning and evaluation protocol, yielding validation set results of R² = 0.76, MAE = 2.2%, and RMSE = 2.9%. The RMSE is significantly higher than the MAE, indicating the presence of a small number of long-tailed error samples.

[0083] This comparative example utilizes electrical testing (IV parameter / efficiency time series) to drive traditional machine learning, neglecting the coupling between multiple parameters / modalities. It struggles to capture lifetime-related micro-defects and interface states (requiring image information), material phase transitions / defect state evolution (requiring spectral information), and the nonlinear degradation of environmental stress-time (requiring time series information). Furthermore, it lacks interpretable observation and feedback control capabilities regarding internal mechanisms, thus limiting overall prediction accuracy and robustness. While this comparative example can serve as a baseline, it falls short of meeting the requirements for high-precision, interpretable, and optimizable engineering evaluation of perovskite device stability / lifetime.

[0084] Example 1: Image + Electrical Data → CNN Feature Extraction + XGBoost Ensemble Prediction This embodiment provides a method for predicting the stability of perovskite devices based on the fusion of electroluminescence (EL) images and electrical test data. The system includes an EL imaging camera, an electrical parameter acquisition unit (IV test module), a synchronous acquisition and timing alignment module, and a data processing and model calculation unit (processor, memory, and model library). The method flowchart is shown below. Figure 3 As shown, during the acquisition phase, EL images and electrical parameters (e.g., Voc, Jsc, FF, etc., totaling 12 dimensions) of the device in its operating state are acquired simultaneously. In the preprocessing phase, grayscale normalization and denoising are performed on the EL images, with optional size unification and lightweight enhancement. Missing values ​​are filled and standardized for the electrical parameters to form standardized samples. Subsequently, a convolutional neural network (CNN) is used to extract deep features from the EL images, outputting 128-dimensional image features. This 128-dimensional vector is then fused with the aforementioned 12-dimensional electrical features to obtain a comprehensive feature vector (140 dimensions) for each sample. Using this comprehensive feature as input, a gradient boosting decision tree regression model (XGBoost) is constructed and trained through K-fold cross-validation and hyperparameter optimization (number of trees, maximum depth, learning rate, etc.), aiming to achieve the residual efficiency after 1000 hours of continuous operation under 1 Sun (1000W / m²) conditions. After training, the model is applied to the test samples, outputting the corresponding stability prediction results.

[0085] Data and Results (Validation Set): Under the same data partitioning and evaluation protocol as the control, this embodiment obtained validation set metrics of R² = 0.87, MAE = 1.3%, and RMSE = 1.6%; training time was approximately 30 minutes. This method, while maintaining a simple workflow, utilizes information such as EL image defect / brightness distribution extracted by CNN and electrical macroscopic parameters for fusion modeling, which can effectively reduce the average error and suppress fluctuations caused by long-tailed samples. The XGBoost-based tree model also provides a moderate level of interpretability, such as feature importance ranking, facilitating R&D and process parameter tuning references.

[0086] Example 2: Image + Spectrum (PL) + Electrical Data → Transformer Cross-Modal Modeling This embodiment provides a three-modal fusion prediction method for perovskite devices. The system may include: an EL imaging camera, a PL spectrometer, an electrical parameter acquisition unit (IV testing / online monitoring), a synchronous acquisition and registration module, and a data processing and model calculation unit. By synchronously acquiring EL images, their spatially corresponding PL spectra, and electrical parameters (e.g., Voc, Jsc, FF, etc.), integrated acquisition and alignment of multi-source data is achieved (e.g., ...). Figure 4 ).

[0087] Data processing and feature construction: Gray-level normalization, denoising, and size unification are performed on the EL image; baseline correction, normalization, and (optionally) peak / integral intensity and other derived feature extraction are performed on the PL spectrum; missing values ​​are imputed and normalized for electrical parameters. In some embodiments, a CNN or lightweight preprocessing network is used to extract 128-dimensional image features from the EL image; a one-dimensional encoding network / autoencoder is used to extract 100-dimensional spectral features from the PL spectrum; and the electrical features are 12-dimensional. The three parts of features are constructed into tokens according to modality, and concatenated with modality embedding / position encoding to form a feature sequence, which is used as the input of the Transformer.

[0088] Model and Training: A multi-head self-attention Transformer model is employed to perform cross-modal modeling and fusion of image, spectral, and electrical tokens, outputting performance prediction results. In some embodiments, a multi-task objective is adopted: 1000h residual efficiency is used as the regression label, and stability level is used as the classification label. The training process maintains a consistent partitioning and evaluation protocol with the control group, employing K-fold cross-validation and hyperparameter optimization.

[0089] Data and Results (Validation Set): Under the same conditions as the control, the validation set metrics for the three-modal + Transformer model in this embodiment are R² = 0.91, MAE = 1.0%, and RMSE = 1.2%; training time is approximately 1.5 hours. Compared to the single-modal baseline, the above results demonstrate the simultaneous convergence of the mean error and the long-tailed error.

[0090] The Transformer's self-attention mechanism can adaptively focus on the correlation between degradation-related local image defect patterns (such as bright spots / crack regions), PL spectrum peak / intensity changes (reflecting nonradiative recombination, defect states, etc.), and electrical indices. In some embodiments, attention weights / channel contributions can be visualized to locate image regions or spectral bands that contribute significantly to prediction, for defect diagnosis and process optimization decisions. This embodiment improves prediction accuracy, robustness, and interpretability, and features a unified model structure and a simple inference chain, facilitating engineering deployment and expansion.

[0091] Example 3: Image + Integrated Spectroscopy + Electrical + Time Series Data → Transformer + LSTM Hybrid Model and Feedback Proposal System This embodiment focuses on health monitoring and degradation prediction throughout the entire lifecycle of perovskite solar cells. Based on multimodal data, it further integrates time-series information and introduces a feedback suggestion module to construct a closed-loop intelligent prediction system (e.g., Figure 5The system may include: an EL imaging camera (optionally acquiring EL / IR images), a spectral module (integrating PL and Raman spectrometers), an electrical acquisition system (IV, efficiency, impedance, etc.), and a recording unit for environmental and historical operating conditions; in conjunction with synchronous acquisition, spatial registration, and temporal alignment modules, a unified data flow of image / spectral / electrical / environmental time series is formed, as shown in the flowchart. Figure 5 As shown.

[0092] Data processing and feature construction: Gray-level normalization, denoising, and size unification are performed on the EL image; baseline correction, normalization, and extraction of derived features such as peak position / peak intensity are performed on the PL spectrum; missing values ​​are imputed and standardized for electrical and environmental parameters; a sliding window and sampling alignment (window length > 24 points) are used for the historical sequence. In some embodiments, lightweight CNN / encoding networks are used to extract image features and spectral features respectively, which are then concatenated with electrical / environmental features to form a multimodal feature sequence (comprehensive dimension > 300) for the current time t.

[0093] Model Structure and Training: A Transformer sub-network is used to perform cross-modal modeling and fusion of the multimodal feature sequences at the current time t (multi-head self-attention); an LSTM sub-network is used to dynamically model the time series of historical windows, capturing nonlinearity and lag effects. The outputs of the two sub-networks converge in a fusion layer and are connected to: ① the efficiency prediction curve over time (regression head); ② the health status index (optional); ③ the remaining lifetime (RUL) (regression head). Training uses the same data partitioning and evaluation protocol as the control group, employing K-fold cross-validation and hyperparameter optimization (number of layers / heads / hidden dimensions, number of LSTM units, learning rate, and early stopping strategy, etc.). The training time in this embodiment is approximately 2 hours (including sequence iterations).

[0094] Validation set results: Under the same evaluation conditions, this embodiment achieved R² = 0.94, MAE = 0.7%, and RMSE = 0.9%. Compared with the single-modal baseline, both the mean error and the long-tailed error converged further, and the time-related mismatch was effectively suppressed.

[0095] Feedback and suggestion closed loop: Based on the prediction results and attention / time series contribution analysis, the feedback module automatically generates operating condition or maintenance suggestions and returns them to the R&D / control end through the interface, realizing a closed loop of "prediction → suggestion → adjustment → data update". For example, when Raman / PL indicates that the active layer shows peak position / intensity changes related to non-radiative recombination, the system can suggest optimizing the encapsulation or reducing temperature and humidity; when image features show that specific defect morphology dominates degradation, it can suggest adjusting the preparation parameters to reduce the defect incidence. The new data after feedback is incorporated into subsequent training or incremental updates to improve the model's adaptability to complex operating conditions.

[0096] Transformer captures intermodal correlations (local defects-spectral peaks-electrical parameters), while LSTM characterizes temporal evolution (short-term fluctuations and long-term decay coexist). The two complement each other to achieve a more comprehensive characterization of degradation paths and lifetimes. Through attention hotspots and time-series attribution maps, image regions, key spectral segments, and time segments that contribute significantly to predictions are visualized, providing a basis for locating performance bottlenecks and formulating process / operation strategies.

[0097] Table 2 compares the differences among the four modeling schemes in terms of data sources, model structure, output objectives, and validation set metrics. Table 2 is as follows: Table 2 Performance Comparison of Multimodal and Temporal Modeling Schemes with Comparative Examples

[0098] From left to right, the data modalities gradually expand from single electrical to image + PL / Raman spectroscopy + environmental time series; the model transitions from XGBoost to CNN+XGBoost and Transformer, and in "Example 3" a Transformer+LSTM and feedback module are introduced.

[0099] With the introduction of modal and temporal information and the improvement of model capabilities, R² continuously increased from 0.76 to 0.94, while MAE gradually decreased from 2.2% to 0.7% and RMSE gradually decreased from 2.9% to 0.9%, demonstrating the synchronous convergence of average error and long-tail error. Simultaneously, interpretability improved from "weak" to "strong," and a prediction-suggestion-control closed loop was formed in "Example 3." These results indicate that multimodal and time-series fusion can more comprehensively characterize device states and improve engineering usability.

[0100] The above primarily describes the solutions of the embodiments of this application from the perspective of the method execution process. It is understood that, in order to achieve the above functions, the server includes the corresponding hardware structure and / or software modules for executing each function. Those skilled in the art should readily recognize that, based on the units and algorithm steps of the examples described in the embodiments provided herein, this application can be implemented in hardware or a combination of hardware and computer software. Whether a function is executed in hardware or by computer software driving hardware depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0101] This application embodiment can divide the server into functional units according to the above method example. For example, each function can be divided into different functional units, or two or more functions can be integrated into one processing module. The integrated unit can be implemented in hardware or as a software program module. It should be noted that the unit division in this application embodiment is illustrative and only represents a logical functional division, while other division methods may be used in actual implementation.

[0102] In the case of using integrated units, please refer to Figure 6 , Figure 6 A functional unit structure block diagram of a photovoltaic device stability prediction device based on a multimodal model provided in this application embodiment is shown below. Figure 6 As shown, the photovoltaic device stability prediction device 6 includes: The acquisition unit 601 is used to acquire multi-source data of the photovoltaic device under test, wherein the multi-source data includes at least two of the following: image data, spectral data, initial photoelectric data, and environmental stress parameter sequence. Processing unit 602 is configured to perform modal adaptation preprocessing on the multi-source data to obtain a multi-modal sample set; input the training sample set of the multi-modal sample set into a trained multi-modal deep neural network model to output a high-dimensional feature vector representing the overall device state; the multi-modal deep neural network model includes parallel modal sub-networks and a feature fusion layer, each modal sub-network is used to extract features from the training sample set, and the feature fusion layer is used to fuse the features extracted by each modal sub-network to obtain the high-dimensional feature vector; input the high-dimensional feature vector into a time series prediction sub-model to obtain the predicted photoelectric conversion efficiency (PCE) values ​​at each time point in the entire life cycle of the photovoltaic device under test and plot the performance degradation curve; generate a stability rating for the photovoltaic device under test based on the performance degradation curve, the stability rating including stable and unstable states; detect that the stability rating is unstable, determine the key factors affecting the stability of the device based on the performance degradation curve and the multi-source data, the key factors including at least one of defect regions, sensitive spectral bands, and environmental influencing factors; and generate device optimization and improvement suggestions based on the key factors.

[0103] As can be seen, in this embodiment, multi-source data of the photovoltaic device under test is collected. The multi-source data includes at least two of the following: image data, spectral data, initial photoelectric data, and environmental stress parameter sequences. Modal adaptation preprocessing is performed on each of the multi-source data to obtain a multi-modal sample set. The training sample set from the multi-modal sample set is input into a trained multi-modal deep neural network model, which outputs a high-dimensional feature vector representing the overall device state. The multi-modal deep neural network model includes parallelly configured modal subnetworks and a feature fusion layer. Each modal subnetwork is used to extract features from the training sample set, and the feature fusion layer is used to extract features from each modal subnetwork. The extracted features are fused to obtain a high-dimensional feature vector. This high-dimensional feature vector is then input into a time-series prediction sub-model to obtain the predicted photoelectric conversion efficiency (PCE) values ​​for each time point throughout the entire lifespan of the photovoltaic device under test, and to plot performance degradation curves. Based on the performance degradation curves, a stability rating for the photovoltaic device under test is generated, including stable and unstable states. If an unstable state is detected, key factors affecting device stability are identified based on the performance degradation curves and multi-source data. These key factors include at least one of defect regions, sensitive spectral bands, and environmental factors. Based on these key factors, device optimization and improvement suggestions are generated. Therefore, compared to existing technologies, multi-modal data fusion and time-series prediction technologies enable accurate prediction and stability rating of photovoltaic device performance throughout its entire lifespan. This allows for the identification of core influencing factors of unstable states and the generation of targeted optimization suggestions, significantly shortening the stability assessment cycle, reducing testing costs, and providing efficient technical support for device R&D and mass production optimization.

[0104] In some embodiments, the image data is used to reflect the grain morphology and defect distribution of the photovoltaic device under test, the spectral data is used to reflect the material structure and stress information of the device, the initial photoelectric data includes open-circuit voltage Voc, short-circuit current density Jsc, fill factor FF, and initial photoelectric conversion efficiency PCE, and the environmental stress parameter sequence includes temperature T, relative humidity RH, and light intensity during testing or use; wherein, the modal subnetwork includes an image subnetwork, a spectral subnetwork, and a numerical subnetwork, the image subnetwork uses a convolutional neural network (CNN) or a lightweight CNN to extract image features, the image features include texture and structural features, the spectral subnetwork uses a spectral analysis network or a coding network to extract spectral features, the spectral features include spectral peak positions and intensity features, and the numerical subnetwork uses a fully connected network or a convolutional network to extract numerical features of the initial photoelectric data and the environmental stress parameter sequence.

[0105] In some embodiments, the feature fusion layer employs a multi-head self-attention mechanism of a Transformer network to achieve cross-modal feature association modeling through self-attention weight calculation; or, a multimodal variational autoencoder (VAE) is employed to encode each modal feature into jointly distributed latent variables, and then generate a fused high-dimensional feature vector through a decoding process.

[0106] In some embodiments, the time series prediction sub-model employs a Long Short-Term Memory (LSTM) network or an improved structure thereof, the improved structure including a bidirectional LSTM and a gated recurrent unit (GRU). The specific process by which the time series prediction sub-model generates the predicted power conversion efficiency (PCE) values ​​for each time point throughout the entire lifespan of the photovoltaic device under test and plots the performance degradation curve includes: inputting the high-dimensional feature vector at time t and the corresponding environmental stress parameters into the LSTM or its improved structure; processing the input data at time t through multiple memory gate units to output the predicted device performance value at time t+1; and processing the data at time t+1... The environmental stress parameters at time t+1 are integrated with the performance prediction value at time t+1 to form the input data at time t+1. t is updated to t+1, and the steps "inputting the high-dimensional feature vector at time t and the corresponding environmental stress parameters into the LSTM or its improved structure" and subsequent steps are repeated iteratively to obtain the PCE prediction values ​​at times t+2, t+3...t+n, until the predicted PCE value decreases to a preset device failure threshold, covering the entire lifespan. The PCE prediction values ​​at each time point within the entire lifespan are correlated sequentially along the time axis to form a continuous performance degradation curve.

[0107] In some embodiments, before generating a stability rating for the photovoltaic device under test based on the performance degradation curve, the processing unit 602 is further configured to: calculate the coefficient of determination R², mean absolute error MAE, and root mean square error RMSE on the validation set of the multimodal sample set to verify the model prediction accuracy.

[0108] In some embodiments, in determining key factors affecting device stability based on the performance degradation curve and the multi-source dataset, abnormal degradation intervals in the performance degradation curve are identified, including time points of abrupt decay rate changes, accelerated degradation intervals, and final failure threshold time points; multi-source data corresponding to the abnormal intervals of the degradation curve are acquired, and the target image defect region and target sensitive spectral band that contribute most to the degradation are visualized and located through the attention weights output by the feature fusion layer; the correlation between the environmental stress parameter and the abnormal intervals of the degradation curve is analyzed using the controlled variable method to determine the target environmental influencing factors that dominate the degradation; and the target image defect region, the target sensitive spectral band, and the target environmental influencing factors are fused using a preset multimodal feature-failure mechanism mapping library to output the priority ranking of the key factors.

[0109] In some embodiments, in generating device optimization and improvement suggestions based on the key factors, the processing unit 602 is further configured to: obtain a preset optimization measure rule library, the optimization measure rule library including a mapping relationship between the key factors and the device optimization and improvement suggestions; obtain corresponding device optimization and improvement suggestions based on the key factors and the optimization measure rule library; and determine the priority order of each device optimization and improvement suggestion based on the priority order of the key factors.

[0110] Please see Figure 7 , Figure 7 This application provides a schematic diagram of the structure of a server, as shown in the embodiment of the present application. Figure 7 As shown, the server 10 includes a processor 101, a memory 103, a communication interface 102, and a computer program 1031. The computer program 1031 is stored in the memory 103 and configured to be executed by the processor 101. The program includes a method for performing a photovoltaic device stability prediction method based on a multimodal model as described in the above embodiments.

[0111] This application provides a computer-readable storage medium storing a computer program / instructions thereon, which, when executed by a processor, implement the steps of any possible embodiment of the method.

[0112] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

[0113] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0114] In the several embodiments provided in this application, it should be understood that the disclosed apparatus can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of the units described above is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical or other forms.

[0115] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0116] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0117] If the aforementioned integrated units are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage device (CMD). Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a memory and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned memory includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0118] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage device, which may include: a flash drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk, etc.

[0119] The embodiments of this application have been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of this application. The description of the above embodiments is only for the purpose of helping to understand the method and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for predicting the stability of photovoltaic devices based on a multimodal model, characterized in that, The method includes: Collect multi-source data from the photovoltaic device under test, including at least two of the following: image data, spectral data, initial photoelectric data, and environmental stress parameter sequences. Modality adaptation preprocessing is performed on the multi-source data to obtain a multimodal sample set; The training sample set from the multimodal sample set is input into the trained multimodal deep neural network model, which outputs a high-dimensional feature vector representing the state of the integrated device. The multimodal deep neural network model includes parallel modal subnetworks and a feature fusion layer. Each modal subnetwork is used to extract features from the training sample set, and the feature fusion layer is used to fuse the features extracted by each modal subnetwork to obtain the high-dimensional feature vector. The high-dimensional feature vector is input into the time series prediction sub-model to obtain the predicted photoelectric conversion efficiency (PCE) values ​​at each time point in the entire life cycle of the photovoltaic device under test and to plot the performance degradation curve. A stability rating for the photovoltaic device under test is generated based on the performance degradation curve, and the stability rating includes a stable state and an unstable state. If the stability rating is detected as the unstable state, the key factors affecting the stability of the device are determined based on the performance degradation curve and the multi-source data. The key factors include at least one of the following: defect region, sensitive spectral band and environmental factors. Based on the aforementioned key factors, device optimization and improvement suggestions are generated.

2. The method according to claim 1, characterized in that, The image data is used to reflect the grain morphology and defect distribution of the photovoltaic device under test, the spectral data is used to reflect the material structure and stress information of the device, the initial photoelectric data includes open circuit voltage Voc, short circuit current density Jsc, fill factor FF and initial photoelectric conversion efficiency PCE, and the environmental stress parameter sequence includes temperature T, relative humidity RH and light intensity during the test or use. The modal subnetwork includes an image subnetwork, a spectral subnetwork, and a numerical subnetwork. The image subnetwork uses a convolutional neural network (CNN) or a lightweight CNN to extract image features, which include texture and structural features. The spectral subnetwork uses a spectral analysis network or a coding network to extract spectral features, which include peak positions and intensity features. The numerical subnetwork uses a fully connected network or a convolutional network to extract numerical features of the initial photoelectric data and the environmental stress parameter sequence.

3. The method according to claim 1, characterized in that, The feature fusion layer employs a multi-head self-attention mechanism of the Transformer network, achieving cross-modal feature correlation modeling through self-attention weight calculation; or, A multimodal variational autoencoder (VAE) is used to encode the features of each modality into jointly distributed latent variables, and then generate a fused high-dimensional feature vector through the decoding process.

4. The method according to claim 1, characterized in that, The time series prediction sub-model uses a Long Short-Term Memory (LSTM) network or an improved structure thereof, the improved structure including a bidirectional LSTM and a gated recurrent unit (GRU). The specific process by which the time series prediction sub-model generates the predicted power conversion efficiency (PCE) values ​​for each time point throughout the entire lifespan of the photovoltaic device under test and plots the performance degradation curve includes: The high-dimensional feature vector at time t and the corresponding environmental stress parameters are input into the LSTM or its improved structure. The input data at time t is processed by multiple memory gating units, and the predicted device performance value at time t+1 is output. The environmental stress parameters at time t+1 are integrated with the performance prediction values ​​at time t+1 to form the input data at time t+1. Update t to t+1, and repeat the steps "input the high-dimensional feature vector at time t and the corresponding environmental stress parameters into the LSTM or its improved structure" and subsequent steps, iterating sequentially to obtain the PCE prediction values ​​at times t+2, t+3...t+n until the predicted PCE value is reduced to the preset device failure threshold, covering the entire life cycle; The predicted PCE values ​​at each time point throughout the entire life cycle are correlated sequentially along the time axis to form a continuous performance degradation curve.

5. The method according to claim 1, characterized in that, Before generating the stability rating of the photovoltaic device under test based on the performance degradation curve, the method further includes: The coefficient of determination R², mean absolute error (MAE), and root mean square error (RMSE) are calculated on the validation set of the multimodal sample set to verify the model's prediction accuracy.

6. The method according to claim 3, characterized in that, The determination of key factors affecting device stability based on the performance degradation curve and the multi-source dataset includes: Identify the abnormal degradation intervals in the performance degradation curve, including the time point of sudden change in degradation rate, accelerated degradation interval, and final failure threshold time point; Obtain multi-source data corresponding to the abnormal interval of the attenuation curve, and visualize and locate the target image defect region and target sensitive spectral band that contribute the most to the attenuation through the attention weight output by the feature fusion layer; The correlation between the environmental stress parameters and the abnormal range of the attenuation curve was analyzed using the controlled variable method to determine the target environmental influencing factors that dominate the attenuation. By combining a pre-defined multimodal feature-failure mechanism mapping library, the defect region of the target image, the sensitive spectral band of the target, and the environmental influencing factors of the target are fused, and the priority ranking of the key factors is output.

7. The method according to claim 6, characterized in that, The method of generating device optimization and improvement suggestions based on the key factors includes: Obtain a preset optimization rule base, which includes the mapping relationship between the key factors and the device optimization and improvement suggestions; Based on the key factors and the optimization measure rule base, obtain corresponding device optimization and improvement suggestions; The priority order of the device optimization and improvement suggestions is determined based on the priority order of the key factors.

8. A photovoltaic device stability prediction device based on a multimodal model, characterized in that, The device includes: The acquisition unit is used to acquire multi-source data of the photovoltaic device under test, wherein the multi-source data includes at least two of the following: image data, spectral data, initial photoelectric data, and environmental stress parameter sequence. The processing unit is configured to perform modal adaptation preprocessing on the multi-source data to obtain a multi-modal sample set; input the training sample set from the multi-modal sample set into a trained multi-modal deep neural network model, and output a high-dimensional feature vector representing the overall device state; the multi-modal deep neural network model includes parallel modal sub-networks and a feature fusion layer, each modal sub-network is used to extract features from the training sample set, and the feature fusion layer is used to fuse the features extracted by each modal sub-network to obtain the high-dimensional feature vector; input the high-dimensional feature vector into a time series prediction sub-model to obtain the predicted photoelectric conversion efficiency (PCE) values ​​at each time point throughout the entire life cycle of the photovoltaic device under test and plot the performance degradation curve; generate a stability rating for the photovoltaic device under test based on the performance degradation curve, the stability rating including stable and unstable states; if the stability rating is detected to be unstable, determine the key factors affecting device stability based on the performance degradation curve and the multi-source data, the key factors including at least one of defect regions, sensitive spectral bands, and environmental influencing factors; and generate device optimization and improvement suggestions based on the key factors.

9. A server, characterized in that, It includes a processor and a memory, wherein the memory stores a computer program, and the processor executes the step instructions of the method as described in any one of claims 1-7 when it invokes the computer program in the memory.

10. A computer-readable storage medium, characterized in that, It stores a computer program / instruction thereon, which, when executed by a processor, implements the steps of the method as described in any one of claims 1-7.

Images