CIGS photovoltaic panel-based life prediction method for decline of each layer of material

Abstract

The invention belongs to the technical field of material life prediction, and particularly discloses a life prediction method based on decline of each layer of material of a CIGS photovoltaic panel, and the method comprises the following steps: S1, building baseline data and a CIGS degradation mode library; s2, obtaining a time sequence data set; s3, training a random forest model; s4, encoding degradation type labels into numerical features, splicing the numerical features with physicochemical data features, calculating a minimum cumulative distance, and determining a confidence level and a degradation level to obtain an input feature set of the LSTM model; s5, constructing an LSTM (Long Short Term Memory) model; and S6, life prediction. According to the service life prediction method based on decline of each layer of material of the CIGS photovoltaic panel, each layer of material is analyzed through multi-source data fusion and a machine learning model, parameters which have the largest influence on the service life can be rapidly positioned, prediction errors are reduced, along with accumulation of outdoor actual measurement data, the model can be continuously iteratively optimized, and the service life of the CIGS photovoltaic panel is greatly prolonged. And the prediction precision is further improved.

Description

Technical Field

[0001] This invention relates to the field of material lifetime prediction technology, and in particular to a method for predicting lifetime based on the degradation of materials in each layer of a CIGS photovoltaic panel. Background Technology

[0002] Copper indium gallium selenide (CIGS) photovoltaic panels have entered commercial production and are being applied on a large scale in multiple application fields. CIGS technology, with its excellent low-light performance, lightweight, and flexible properties, has irreplaceable advantages in building-integrated photovoltaics (BIPV) and mobile energy, and its future market prospects are broad. However, its short commercialization time (less than 10 years of large-scale mass production), lack of long-term outdoor reliability data, and lack of clear lifespan predictions, coupled with the inability of traditional lifespan assessments to simulate the coupling effects of complex outdoor environments, result in low prediction accuracy and long testing cycles, directly hindering its market penetration.

[0003] In existing technologies, the lifespan of CIGS photovoltaic panels (typically 10-25 years) is generally shorter than that of crystalline silicon photovoltaic panels (approximately 25-30 years). Due to its flexible, non-silicon structural characteristics, its aging mechanism differs significantly from that of crystalline silicon photovoltaic panels. Furthermore, influenced by factors such as material properties, manufacturing processes, usage environment, and encapsulation technology, its lifespan and annual degradation rate, among other core performance indicators, differ from those of crystalline silicon photovoltaic panels. Moreover, the core internal cause of the shortened lifespan of CIGS photovoltaic panels is the degradation of the physicochemical properties of its multilayer thin films (CIGS absorber layer, CdS buffer layer, ZnO window layer) and encapsulation structure. Therefore, there is an urgent need for a lifespan prediction method based on the degradation of each layer of materials in CIGS photovoltaic panels. Summary of the Invention

[0004] The purpose of this invention is to provide a lifespan prediction method based on the degradation of materials in each layer of CIGS photovoltaic panels. By using multi-source data fusion and machine learning models to analyze each layer of materials, the method can quickly locate the parameters that have the greatest impact on lifespan, thereby reducing prediction errors. Furthermore, with the accumulation of outdoor measured data, the model can be continuously iterated and optimized to further improve prediction accuracy.

[0005] To achieve the above objectives, this invention provides a lifetime prediction method based on the degradation of materials in each layer of a CIGS photovoltaic panel, comprising the following steps: S1. Establish baseline data and CIGS degradation mode library, including physicochemical data such as CIGS crystallinity, ZnO transmittance, interfacial adhesion, temperature coefficient, CIGS element ratio, CdS layer S content, encapsulation layer water vapor permeability, and ambient humidity. S2. Data acquisition, outlier handling, missing value imputation, normalization, time series alignment, to obtain a time series dataset; S3. Train the random forest model, output the importance scores of the physicochemical data, and filter the physicochemical data features as input to the model; S4. Encode the degradation type label into numerical features, and then concatenate it with the physical and chemical data features as static features. The length is uniform and standardized. Calculate the minimum cumulative distance through the dynamic time warping algorithm to determine the confidence level and degradation level, and obtain the input feature set of the LSTM model. S5. Construct an LSTM model, including an input layer, three hidden layers, a fully connected layer, and an output layer; S6. Lifetime Prediction: Set a failure threshold and predict the failure time point as the time point when the lifetime first falls below the failure threshold to obtain the remaining lifetime.

[0006] Preferably, in S1, the CIGS degradation mode library includes degradation type labels, feature thresholds, and temporal variation patterns, wherein the degradation type labels include CIGS elemental segregation, ZnO window layer oxidation, encapsulation layer aging, and interface peeling.

[0007] Preferably, in S1, non-core power generation areas at the edges and corners of the component are sampled, and the cuts are sealed with sealant. The glass substrate is peeled off, the surface encapsulation layer is removed, the crystallinity of CIGS is calculated by X-ray diffraction, and the CIGS element ratio is determined by inductively coupled plasma mass spectrometry.

[0008] Preferably, S2 is as follows: Lossless indicators are collected once per quarter. Lossy indicators are collected once every 6 months in the first year after the new component is introduced, and once a year from the second year onwards. Outliers are replaced with the moving average of the previous 3 data for the indicator. If data for a certain quarter is missing, it is supplemented by linear interpolation. Indicators of different dimensions are mapped to the [0,1] interval. All indicators are integrated by timestamp to form a time series dataset of components, time, and multi-dimensional features.

[0009] Preferably, in S3, at least the top 5 physicochemical data features with the highest scores are selected as model inputs.

[0010] Preferably, in S4, the standardized calculation formula is as follows: ; in, x For input data, μ Let σ be the mean and σ be the standard deviation.

[0011] Preferably, in S4, the confidence level and degradation level are determined as follows: Similarity = 1 - (minimum cumulative distance / sequence length). The closer the similarity is to 1, the higher the matching degree. Confidence level ,in sim For the current similarity, con f∈[0,1]; con f ≥0.8: High confidence match, determine the degradation type label, and classify it as moderate degradation; 0.6≤con f <0.8: Medium confidence match, judged as mild degradation, marked as suspected degradation type; con f <0.6: Low confidence match, judged as no significant degradation.

[0012] Preferably, in S5, the number of neurons in the fully connected layer is 16, and the activation function is ReLU.

[0013] Preferably, in S5, the number of neurons in the three hidden layers are 128, 64, and 32, respectively, the activation function is tanh, and the dropout is 0.2.

[0014] Preferably, in S6, the failure threshold is set to 0.7, the predicted failure time is the time when it first falls below the failure threshold, and the remaining lifespan is equal to the current years of operation plus (predicted failure time / 4) - (number of quarters of operation / 4).

[0015] Therefore, the present invention employs the above-mentioned lifetime prediction method based on the degradation of materials in each layer of a CIGS photovoltaic panel, and the beneficial effects are as follows: This invention uses sample data to represent the characteristics of the entire batch, avoiding large-scale damage. Sampling and destructive testing avoid the core power generation area of ​​the module. It utilizes the physicochemical changes of each layer of CIGS photovoltaic panels, and analyzes each layer of materials through multi-source data fusion and machine learning models. LSTM networks are good at capturing the long-term time-series trends of physicochemical indicators, accurately capturing the temporal and pattern-based nature of CIGS degradation, and can quickly locate the parameters that have the greatest impact on lifespan, avoiding misjudgments. At the same time, with the accumulation of outdoor test data, the model can be continuously iterated and optimized to further improve prediction accuracy.

[0016] The data acquisition in this invention uses conventional instruments with a quarterly acquisition frequency, eliminating the need for high-frequency real-time monitoring, thus reducing operation and maintenance costs. It is suitable for maintenance and long-term life monitoring of CIGS photovoltaic panels during use and is highly practical.

[0017] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0018] Figure 1 This is a flowchart of an embodiment of the present invention, which is a method for predicting the lifetime of CIGS photovoltaic panels based on the degradation of materials in each layer. Detailed Implementation

[0019] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0020] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0021] Working Principle: CIGS photovoltaic panels consist of multiple thin films (CIGS absorber layer, CdS buffer layer, ZnO window layer) and encapsulation structure. The CIGS absorber layer, as the core power generation layer, is prone to elemental segregation (Cu / In / Ga / Se) due to environmental factors, leading to decreased crystallinity and reduced carrier mobility. Long-term irradiation or high temperatures can generate deep-level defects, accelerating photogenerated carrier recombination and causing a sharp drop in efficiency. The CdS buffer layer is prone to hydrolysis (especially in high-humidity environments), resulting in sulfur loss and an increase in PN junction interface states. The ZnO window layer's transmittance drops from 90% to below 75% due to oxidation or impurity deposition, reducing incident light absorption. Insufficient adhesion between the Mo back electrode and the CIGS absorber layer, and between CIGS and the CdS buffer layer, makes them prone to peeling during temperature and humidity cycling, directly causing module failure. Long-term exposure to ultraviolet radiation and high temperatures causes the encapsulation material to yellow, become embrittled, and experience increased moisture permeability, leading to internal moisture corrosion within the module.

[0022] Example 1: like Figure 1 As shown, a lifetime prediction method based on the degradation of materials in each layer of a CIGS photovoltaic panel includes the following steps: S1. Establish baseline data and a CIGS degradation pattern library: Measure the physicochemical data of the same batch of new CIGS modules (at least 3 modules, with edge and corner areas selected for each module) and add them to the database. Physicochemical data includes CIGS crystallinity, ZnO transmittance, interfacial adhesion, temperature coefficient, CIGS elemental ratio (Ga / (Ga+In)), CdS layer S content, encapsulation layer water vapor permeability, and ambient humidity. The CIGS degradation pattern library includes degradation type labels, characteristic thresholds, and temporal variation patterns, as shown in Table 1.

[0023] Table 1 CIGS Degradation Mode Library

[0024] The testing methods for CIGS crystallinity, ZnO transmittance, interfacial adhesion, temperature coefficient, CIGS elemental ratio (Ga / (Ga+In)), S content in the CdS layer, water vapor permeability of the encapsulation layer, and ambient humidity are as follows: CIGS crystallinity and (Ga / (Ga+In): This refers to the non-core power generation areas at the edges and corners of the module, with an area of ​​only 1 cm². 2After sampling, the cut was sealed with sealant, the glass substrate was peeled off (to avoid damaging the CIGS layer), and the surface encapsulation layer (EVA / POE) was removed. X-ray diffraction (XRD) was used to determine and record the diffraction patterns (focusing on the characteristic peaks of the chalcopyrite phase: 2θ = 28.5°, 47.4°, and 56.2°), and the crystallinity was calculated. Elemental content was determined by inductively coupled plasma mass spectrometry (ICP-MS).

[0025] ZnO transmittance, interfacial adhesion, and S content in the CdS layer: Samples were taken from the edge of the component, and the ZnO transmittance was measured using a UV-Vis spectrophotometer. Interfacial adhesion was measured using a scratch tester. The S content was determined using salt spray testing and X-ray photoelectron spectroscopy.

[0026] Temperature coefficient: Select complete components and measure the temperature coefficient using a high and low temperature environmental chamber and an IV characteristic tester.

[0027] Encapsulation layer water vapor permeability: A sample of the encapsulation layer was taken from the edge of the component, and the permeability was measured using a water vapor transmission rate tester.

[0028] Ambient humidity: Measured within 1 meter of the component installation point using temperature and humidity sensors.

[0029] S2. Data acquisition, outlier handling, missing value imputation, normalization, and time series alignment are performed to obtain a time series dataset with quarterly time units.

[0030] Specifically, non-destructive indicators are collected once per quarter, and destructive indicators (including CIGS crystallinity, CIGS element ratio, interface adhesion, and CdS layer S content) are collected once every 6 months in the first year after the new component is installed, and once a year from the second year onwards. Outliers are replaced with the moving average of the previous 3 data for the indicator. If data for a certain quarter is missing, it is supplemented by linear interpolation. Indicators of different dimensions are mapped to the [0,1] interval, and all indicators are integrated by timestamp to form a time-series dataset of components, time, and multi-dimensional features.

[0031] S3. Train the random forest model, output the importance scores of the physicochemical data, and filter the physicochemical data features as model inputs. In this embodiment, the top 5 physicochemical data features with the highest scores are selected as model inputs, namely, Ga / (Ga+In) ratio > ZnO transmittance > water vapor permeability of the encapsulation layer > ambient humidity > CIGS crystallinity.

[0032] S4. The degradation type labels are encoded as numerical features (no degradation (0), elemental segregation (1), window layer oxidation (2), encapsulation aging (3), interface peeling (4)), and then used as static features to be concatenated with 5 types of dynamic physicochemical data features. The length of the input feature set is unified, that is, the short real-time sequence is padded to 12 time points using linear interpolation. Then, the standardization is performed to eliminate the difference in units, and the minimum cumulative distance is calculated by the dynamic time warping (DTW) algorithm to determine the confidence level and degradation level, thus obtaining the input feature set of the LSTM model.

[0033] Specifically, the standardized calculation formula is as follows: ; in, x For input data, μ Let σ be the mean and σ be the standard deviation.

[0034] The confidence level and degradation level are determined as follows: Similarity = 1 - (minimum cumulative distance / sequence length). The closer the similarity is to 1, the higher the matching degree. Confidence level ,in sim For the current similarity, con f ∈[0,1]; con f ≥0.8: High confidence match, identifies degradation type label (such as element segregation), and judges it as moderate degradation; 0.6≤con f <0.8: Medium confidence match, judged as mild degradation, marked as suspected degradation type; con f <0.6: Low confidence match, judged as no significant degradation.

[0035] The S5 and LSTM models are constructed, and the model structure parameters are designed as shown in Table 2.

[0036] Table 2 Structural Parameters

[0037] The hyperparameter settings are as follows: Optimizer: Adam optimizer, initial learning rate = 0.001; Loss function: Mean Squared Error (MSE), adapted for continuous value prediction; Number of iterations: 300 rounds, with an early stopping mechanism (training stops if the validation set loss does not decrease for 20 consecutive rounds). Batch size: 32.

[0038] S6. Lifetime prediction: Set failure threshold = 0.7 (corresponding to a degradation level of 30%, at which point the component power decay is ≥25%, and it loses its economic value). The predicted failure time point is the time point when it first falls below the failure threshold. Remaining lifetime = (current years of operation) + (predicted failure time point / 4) - (number of quarters of operation / 4).

[0039] Example 2: Based on the lifespan prediction method for the degradation of each layer of CIGS photovoltaic panels in Example 1, a CIGS photovoltaic panel in service at a certain location was used as the baseline data: CIGS crystallinity 96.2%, ZnO transmittance 91.5%, Ga / (Ga+In) = 0.32, and encapsulation layer water vapor permeability 0.08 g / (m³). 2 • 24h), average ambient humidity baseline 60%, designed life 25 years, operating time 3 years, time series collected quarterly.

[0040] The input feature set of the LSTM model is shown in Table 3 below.

[0041] Table 3 Input Feature Set

[0042] With a confidence level of 0.86, the model is moderately degraded, matching the oxidation of the ZnO window layer. Key degradation indicators are: ZnO transmittance currently 78.9% (a decrease of 13.8% from the baseline), annual degradation rate 1.15%; Ga / (Ga+In) ratio currently 0.55 (a deviation of 71.9% from the baseline), predicted remaining lifetime: 18.2 years. The model accuracy was evaluated using mean absolute error (MAE) and root mean square error (RMSE). Test results showed MAE ≤ 0.8 years, RMSE ≤ 1.0 years, and the average MAE of the 5-fold cross-validation was ≤ 0.8 years.

[0043] Therefore, this invention adopts the above-mentioned lifetime prediction method based on the degradation of each layer of CIGS photovoltaic panels. By analyzing each layer of materials through multi-source data fusion and machine learning models, the parameters with the greatest impact on lifetime can be quickly located, the prediction error is reduced, and the model can be continuously iterated and optimized with the accumulation of outdoor measured data, further improving the prediction accuracy.

[0044] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A lifetime prediction method based on the degradation of materials in each layer of a CIGS photovoltaic panel, characterized in that, Includes the following steps: S1. Establish baseline data and CIGS degradation mode library, including physicochemical data such as CIGS crystallinity, ZnO transmittance, interfacial adhesion, temperature coefficient, CIGS element ratio, CdS layer S content, encapsulation layer water vapor permeability, and ambient humidity. S2. Data acquisition, outlier handling, missing value imputation, normalization, time series alignment, to obtain a time series dataset; S3. Train the random forest model, output the importance scores of the physicochemical data, and filter the physicochemical data features as input to the model; S4. Encode the degradation type label into numerical features, and then concatenate it with the physical and chemical data features as static features. The length is uniform and standardized. Calculate the minimum cumulative distance through the dynamic time warping algorithm to determine the confidence level and degradation level, and obtain the input feature set of the LSTM model. S5. Construct an LSTM model, including an input layer, three hidden layers, a fully connected layer, and an output layer; S6. Lifetime Prediction: Set a failure threshold and predict the failure time point as the time point when the lifetime first falls below the failure threshold to obtain the remaining lifetime.

2. The lifetime prediction method based on the degradation of materials in each layer of a CIGS photovoltaic panel according to claim 1, characterized in that, In S1, the CIGS degradation mode library includes degradation type labels, feature thresholds, and temporal variation patterns. The degradation type labels include CIGS elemental segregation, ZnO window layer oxidation, encapsulation layer aging, and interface peeling.

3. The lifetime prediction method based on the degradation of materials in each layer of a CIGS photovoltaic panel according to claim 1, characterized in that, In S1, non-core power generation areas at the edges and corners of the component are sampled, and the cuts are sealed with sealant. The glass substrate is peeled off, the surface encapsulation layer is removed, the crystallinity of CIGS is calculated by X-ray diffraction, and the CIGS element ratio is determined by inductively coupled plasma mass spectrometry.

4. The lifetime prediction method based on the degradation of materials in each layer of a CIGS photovoltaic panel according to claim 1, characterized in that, S2 specifically refers to: Lossless indicators are collected once per quarter. Lossy indicators are collected once every 6 months in the first year after the new component is introduced, and once a year from the second year onwards. Outliers are replaced with the moving average of the previous 3 data for the indicator. If data for a certain quarter is missing, it is supplemented by linear interpolation. Indicators of different dimensions are mapped to the [0,1] interval. All indicators are integrated by timestamp to form a time series dataset of components, time, and multi-dimensional features.

5. The lifetime prediction method based on the degradation of materials in each layer of a CIGS photovoltaic panel according to claim 1, characterized in that, In S3, at least the top 5 physicochemical data features with the highest scores are selected as model inputs.

6. The lifetime prediction method based on the degradation of materials in each layer of a CIGS photovoltaic panel according to claim 1, characterized in that, In S4, the standardized calculation formula is as follows: ； in, x For input data, μ Let σ be the mean and σ be the standard deviation.

7. The lifetime prediction method based on the degradation of materials in each layer of a CIGS photovoltaic panel according to claim 1, characterized in that, In S4, the confidence level and degradation level are determined as follows: Similarity = 1 - (minimum cumulative distance / sequence length). The closer the similarity is to 1, the higher the matching degree. Confidence level ,in sim For the current similarity, con f ∈[0,1]; con f ≥0.8: High confidence match, determine the degradation type label, and classify it as moderate degradation; 0.6≤con f <0.8: Medium confidence match, judged as mild degradation, marked as suspected degradation type; con f <0.6: Low confidence match, judged as no significant degradation.

8. The lifetime prediction method based on the degradation of materials in each layer of a CIGS photovoltaic panel according to claim 1, characterized in that, In S5, the fully connected layer has 16 neurons and the activation function is ReLU.

9. The lifetime prediction method based on the degradation of materials in each layer of a CIGS photovoltaic panel according to claim 1, characterized in that, In S5, the number of neurons in the three hidden layers are 128, 64, and 32, respectively, the activation function is tanh, and the dropout is 0.

2.

10. The lifetime prediction method based on the degradation of materials in each layer of a CIGS photovoltaic panel according to claim 1, characterized in that, In S6, the failure threshold is set to 0.7, the predicted failure time is the time when it first falls below the failure threshold, and the remaining lifespan is equal to the current years of operation + (predicted failure time / 4) - (number of quarters of operation / 4).

Images