A method for predicting the electrical performance of a solar cell

Abstract

This invention discloses a method for predicting the electrical performance of solar cells, comprising the following steps: acquiring electroluminescence image data of solar cells and their corresponding current and voltage characteristic data; constructing a cross-spatial axis hierarchical convolutional neural network to extract features from the electroluminescence image data, and introducing a cross-spatial axis attention module during the high-level semantic feature extraction process of the network to obtain enhanced feature representations; globally aggregating the enhanced feature representations and outputting the classification results of solar cells regarding current and voltage data through a classification head layer; employing a two-stage training strategy to extract high-dimensional feature vectors from the images, reducing their dimensionality through principal component analysis, and inputting them into an ensemble tree model for training; using the trained ensemble tree model to learn the nonlinear relationship between image features and actual current and voltage characteristic values, and predicting the continuous characteristics of photovoltaic cell current and voltage data; the electrical performance prediction method of this scheme has high computational efficiency and high prediction accuracy.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic cell electrical performance prediction technology, and specifically to a method for predicting the electrical performance of solar cells. Background Technology

[0002] Currently, achieving carbon peaking and carbon neutrality goals is of great significance for promoting the global energy structure transition from fossil fuels to clean energy. Among various renewable energy technologies, solar photovoltaic systems have become a key support for energy transition due to their mature technology and abundant resources. At present, crystalline silicon-based photovoltaic technology occupies a dominant position in the market, among which TOPCon solar cells have attracted much attention due to their high conversion efficiency and good process compatibility. The efficiency and lifespan of photovoltaic modules are directly related to the economic benefits of photovoltaic systems. Therefore, quality inspection and grading of solar cells during the production process are of great importance to ensuring module performance and reducing failure rates.

[0003] In existing photovoltaic cell quality testing methods, current-voltage characteristic testing is the most critical step, which can obtain core performance indicators such as open-circuit voltage, short-circuit current density, fill factor, and conversion efficiency. However, this method has inherent limitations: the contact measurement process can generate mechanical stress on the silicon wafer, increasing the risk of breakage of thin silicon wafers; at the same time, the insertion and removal of probes and the stabilization time limit the detection throughput, which is not conducive to the promotion and application of multi-wire or wireless cell technology.

[0004] Electroluminescence imaging technology has become an important means of quality inspection for photovoltaic cells due to its non-destructive, non-contact, and high spatial resolution characteristics. In recent years, researchers have attempted to combine machine learning methods to predict current and voltage parameters from electroluminescence images. Traditional machine learning methods rely on manually designed features and have limited generalization ability. Although convolutional neural network-based methods perform better, they have problems such as high computational complexity and limited sensing range, making it difficult to effectively capture long-distance dependencies and global features in electroluminescence images, resulting in low prediction accuracy.

[0005] Therefore, how to construct a model architecture that can effectively integrate local feature extraction and global context modeling capabilities to achieve high-precision non-contact prediction of current and voltage parameters from electroluminescent images has become a technical challenge that urgently needs to be solved in the field of photovoltaic cell quality inspection. Summary of the Invention

[0006] To address the aforementioned shortcomings of existing technologies, this invention provides a method for predicting the electrical performance of solar cells, which solves the problems of low computational efficiency and poor prediction accuracy in existing methods for predicting the electrical performance of photovoltaic cells based on electroluminescence images.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for predicting the electrical performance of a solar cell is provided, comprising the following steps: S1: Obtain electroluminescence image data of solar cells and their corresponding current and voltage characteristic data; and after preprocessing them, divide the dataset for the grading task and the regression prediction task. S2: Construct a cross-spatial axis hierarchical convolutional neural network to extract features from electroluminescent image data, and introduce a cross-spatial axis attention module in the high-level semantic feature extraction process of the network to obtain enhanced feature representation; S3: Perform global aggregation on the enhanced feature representation, and output the classification results of solar cell current and voltage data through the classification head layer; S4: A two-stage training strategy is adopted. After the first stage of classification training converges, the network parameters are frozen as a feature extractor to extract the high-dimensional feature vector of the image. After dimensionality reduction by principal component analysis, the vector is input into the ensemble tree model for training. S5: Utilize the trained ensemble tree model to learn the nonlinear relationship between image features and actual current and voltage characteristics, and perform continuous characteristic prediction on photovoltaic cell current and voltage data.

[0008] Furthermore, step S1 specifically includes: S11: Normalize the electroluminescence image data and construct a dataset by adopting differentiated data augmentation strategies for images with different defect types; S12: The current and voltage characteristic data are processed using the median and mean absolute deviation normalization methods; S13: Divide the dataset into two independent subsets for training and validation of the grading task and the regression prediction task, respectively. S14: For the dataset of the regression prediction task, samples are filtered by setting parameter thresholds, and low-quality samples that do not meet the threshold conditions are moved to the dataset of the tiered task.

[0009] Furthermore, the cross-spatial-axis hierarchical convolutional neural network includes a block embedding layer for extracting initial feature representations of the input image and a hierarchical convolutional backbone network for further extracting multi-scale semantic representations; The hierarchical convolutional backbone network consists of four layers, which construct multi-scale feature representations through layer-by-layer downsampling and feature transformation. The first two layers are used to extract local texture features, while the last two layers are used to extract high-level semantic features and adaptively enhance spatial structure information.

[0010] Furthermore, step S2 specifically includes: S21: Input the preprocessed electroluminescence image data into the block embedding layer to extract features and obtain the initial feature representation FC2; S22: The initial feature representation FC2 enters the first layer of the hierarchical convolutional backbone network, and features are extracted through a 1-layer ConvNeXt convolutional layer to obtain the feature representation FC3; S23: Feature representation FC3 is downsampled to reduce spatial resolution and increase the number of channels to obtain feature representation FC4, which then enters the second stage of two ConvNeXt convolutional layers to obtain feature representation FC5; S24: Feature representation FC5 is then downsampled to reduce spatial resolution and increase the number of channels to obtain feature representation FC6. Subsequently, a cross-spatial axis attention layer is introduced in the middle of the 8 ConvNeXt convolutional layers in the third stage. After the last 4 ConvNeXt convolutional layers in this stage, feature representation FC9 is obtained. S25: Feature representation FC9 is finally passed through a downsampling layer to obtain feature representation FC10, and then enters the fourth level stage with two ConvNeXt convolutional layers to obtain feature representation FC11. Following this, a cross-spatial axis attention layer is introduced to obtain feature representation FC12; and overall, a pyramid-shaped hierarchical feature extraction structure is formed.

[0011] Furthermore, the cross-spatial axis attention module is embedded in the last two layers of the hierarchical convolutional backbone network. The cross-spatial axis attention module captures local features through basic depth convolution, and then performs one-dimensional convolution independently along the height axis and width axis. Through the concatenation decomposition of horizontal and vertical convolutions, the long-range dependencies in the two spatial dimensions are explicitly modeled to obtain adaptively enhanced spatial structure information.

[0012] Furthermore, the specific construction method of the cross-spatial axis attention module is as follows: First, local features are extracted through depthwise convolution; then, for each scale branch, a 1×k horizontal convolution is performed first, followed by a k×1 vertical convolution; the outputs of each scale branch are accumulated and fused, and then channel fusion is performed through a 1×1 convolution; the fused attention map is multiplied element-wise with the input features to obtain the attention-weighted output.

[0013] Furthermore, step S3 specifically includes: S31: Input the enhanced feature representation into the attention pooling module to perform weighted convergence of spatial dimensions and obtain the global semantic feature vector; S32: After the global semantic feature vector is mapped to a fixed-dimensional feature space by a convolutional layer, the fully connected classification layer transforms the global semantic feature vector and outputs the grading range of each solar cell.

[0014] Furthermore, the size and boundaries of the grading intervals are dynamically divided based on the current and voltage data distribution of all solar cells.

[0015] Furthermore, step S4 specifically includes: Phase 1: Train a cross-spatial axis hierarchical convolutional neural network with a classification head layer, and use the cross-entropy loss function to learn the mapping relationship from images to efficiency tiers; Second stage: After the training in the first stage converges, the classification head layer is replaced with the feature projection head layer, and the parameters of the entire backbone network are frozen and used as a feature extractor. The feature extractor extracts the high-dimensional feature vectors of the image, and then reduces them to a low-dimensional space through principal component analysis. The reduced features are then input into the ensemble tree model for training.

[0016] Furthermore, the prediction method is based on non-contact measurement using electroluminescence images, and is used for quality inspection and performance classification in the photovoltaic cell production process.

[0017] The beneficial effects of this invention are as follows: 1. This scheme introduces a cross-spatial axis attention module into the high-level semantic stage of the hierarchical convolutional backbone network and adopts a method of horizontal and vertical strip convolutional decomposition to explicitly model the long-range dependencies in two spatial dimensions. This enables the model to effectively extract global spatial features of electroluminescent images without significantly increasing the computational burden, thereby obtaining more discriminative deep feature representations and laying a solid foundation for subsequent accurate prediction.

[0018] 2. This scheme employs a two-stage hybrid training strategy of "classification first, feature regression later". In the first stage, the model learns the common characteristics of solar cells of different performance levels through a classification task, giving the backbone network strong generalization ability and robustness. In the second stage, the backbone network is frozen as a feature extractor, and the extracted high-dimensional deep features are reduced in dimensionality by principal component analysis and then input into the ensemble tree model for regression prediction. This approach combines the powerful automatic feature extraction capability of deep learning with the advantages of ensemble tree models in processing tabular data and fitting nonlinear relationships, effectively improving the prediction accuracy of continuous current and voltage characteristics.

[0019] 3. This scheme dynamically divides the grading intervals based on the true statistical distribution of current and voltage data of all solar cell samples; thus more realistically reflecting the performance distribution characteristics of batch products, making the grading results more meaningful for practical production guidance; at the same time, by setting parameter thresholds for the dataset of the regression prediction task to filter samples, low-quality or marginal samples are assigned to the grading task to optimize the training quality of the regression model.

[0020] 4. This solution uses non-contact prediction based on electroluminescence images, which fundamentally avoids the damage to solar cells caused by contact measurements and enables rapid online evaluation of the electrical performance of solar cells. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly described below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. The above and other objects, features, and advantages of the present invention will become clearer through the accompanying drawings. The same reference numerals indicate the same parts in all the drawings. The drawings are not intentionally drawn to scale to actual dimensions; the focus is on illustrating the main points of the invention.

[0022] Figure 1 This is a flowchart illustrating the solar cell electrical performance prediction method of this scheme.

[0023] Figure 2 This is a diagram of the network model structure for this scheme.

[0024] Figure 3 The graph shows the accuracy and coefficient of determination results obtained from the experiment on the conversion efficiency of this scheme.

[0025] Figure 4 This is a graph showing the results of the specific grading and regression prediction information for the conversion efficiency of this scheme. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0027] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0028] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0029] Furthermore, the terms "first," "second," etc., are used only to distinguish descriptions and should not be interpreted as indicating or implying relative importance.

[0030] like Figure 1 and Figure 2As shown, this scheme is a method for predicting the electrical performance of solar cells based on cross-spatial axis hierarchical convolutional networks and ensemble tree models, including the following steps: S1: Preprocess the raw electroluminescence image data and current-voltage characteristic data, process the label data using the MMAD normalization method, and divide the dataset for the classification and regression prediction tasks; specifically including: Pixel values ​​of electroluminescence image data were normalized to a mean of 0.5 and a standard deviation of 0.5. Differentiated data augmentation strategies were employed for images with different defect types to construct training datasets. For tiered tasks, a custom interval binning strategy was designed based on the statistical distribution of key indicators such as efficiency, discretizing continuous physical performance indicators into multiple intervals. For regression prediction tasks targeting real current and voltage performance, the MMAD normalization method was used to process the prediction target based on its distribution characteristics. The MMAD normalization formula is as follows:

[0031] in, Indicates standardization efficiency. Indicates the original efficiency. The median represents the original efficiency. This represents the mean absolute deviation of the original efficiency. This represents the median of the target variable. It represents the mean absolute deviation of the target variable.

[0032] The image data is divided into two independent training-validation subsets: train1 / val1 is used for training and validating the tiered model, while train2 / val2 is used for training and evaluating the traditional machine learning regression model. A parameter threshold is set for the regression task dataset to filter samples and remove low-quality samples that do not meet the threshold conditions to the tiered task training set.

[0033] S2: Construct a cross-spatial axis hierarchical convolutional neural network. Input the preprocessed electroluminescence image data into the block embedding layer of the cross-spatial axis hierarchical convolutional neural network to extract initial feature representations. Further extract multi-scale semantic representations through the hierarchical convolutional backbone network. S3: In the high-level semantic feature extraction process of the last two stages of the hierarchical convolutional backbone network, a cross-spatial axis attention module is introduced to achieve adaptive enhancement of spatial structure information, thereby obtaining enhanced representations of key features. Steps S2 and S3 specifically include: First, the preprocessed graph data is input into the block embedding layer to extract features, resulting in the initial feature representation FC2. The hierarchical convolutional backbone network consists of four levels. The initial feature representation FC2 enters the first stage, and features are extracted through a single ConvNeXt convolutional layer to obtain the feature representation FC3. Feature representation FC3 is downsampled to reduce spatial resolution and increase the number of channels to obtain feature representation FC4, which then enters the second stage with two ConvNeXt convolutional layers to obtain feature representation FC5; Feature representation FC5 is then downsampled to reduce spatial resolution and increase the number of channels to obtain feature representation FC6. Subsequently, a cross-spatial axis attention layer is introduced in the middle of the 8 ConvNeXt convolutional layers in the third stage. After the last 4 ConvNeXt convolutional layers in this stage, feature representation FC9 is obtained. Feature representation FC9 is finally passed through a downsampling layer to obtain feature representation FC10, which then enters the fourth stage with two ConvNeXt convolutional layers to obtain feature representation FC11. Following this, a cross-spatial axis attention layer is introduced to obtain feature representation FC12. Overall, a pyramid-shaped hierarchical feature extraction structure is formed, in which the first two stages mainly realize the extraction of local texture features, while the latter two stages realize the extraction of high-level semantic features and adaptive enhancement of spatial structure information.

[0034] The cross-spatial axis attention modules are embedded in the last two layers of the hierarchical convolutional backbone network. The cross-spatial axis attention modules capture local features through basic depth convolution, and then perform one-dimensional convolutions independently along the height and width axes. Through the concatenation decomposition of horizontal and vertical convolutions, the long-range dependencies in the two spatial dimensions are explicitly modeled to obtain adaptively enhanced spatial structure information.

[0035] The specific construction method of the cross-spatial axis attention module is as follows: First, local features are captured through basic depthwise convolution; for each scale branch, one-dimensional convolution is performed independently along the height axis and width axis, first performing 1*k horizontal convolution and then k*1 vertical convolution; the outputs of each scale branch are accumulated and fused, and then channel fusion is performed through 1*1 convolution; the fused attention map is multiplied element-wise with the input features to obtain the attention-weighted output; the cross-spatial axis attention module is embedded in the middle position of the third stage of the network to capture the spatial dependencies of the mid-level features; after being embedded in the fourth stage of the network, global context information is established.

[0036] S4: Input the enhanced feature representation into the attention pooling module for global aggregation, establish the mapping relationship between the global semantic feature vector of the graph data and the current and voltage data, and output the classification results of the current and voltage data corresponding to the solar cell samples through the classification head layer; S5: The size and boundaries of the grading intervals are determined by the current and voltage data distribution of all solar cell samples, thus achieving the task of grading solar cell samples. Steps S4 and S5 specifically include: An attention pooling module with learnable query vectors is used to weight and converge spatial dimensions to obtain a global semantic feature vector; The number of channels is mapped to a fixed-dimensional feature space through 1*1 convolution. The feature vector is then transformed by a fully connected classification layer to output the efficiency grading range for each battery sample.

[0037] S6: A two-stage training strategy is adopted. After the first stage of classification training converges, the entire backbone network is frozen as a feature extractor, and the classification head layer is removed and replaced with a feature projection head. High-dimensional feature vectors of the images are extracted, reduced in dimensionality by principal component analysis, and then input into the ensemble tree model for training. Specifically, this includes: Phase 1: Training a cross-spatial axis hierarchical convolutional neural network with a classification head layer, using the cross-entropy loss function to learn the mapping relationship from images to efficiency tiers; the optimizer is AdamW, the learning rate scheduling strategy is ReduceLROnPlateau, and the learning rate is multiplied by 0.5 when the validation loss does not decrease for 3 consecutive rounds; multiple regularization methods are used to prevent overfitting, including random depth, classification head dropout, and differential data augmentation. Second stage: After the training in the first stage converges, the classification head layer is replaced with the feature projection head layer, and the parameters of the entire backbone network are frozen and used as a feature extractor. The feature extractor extracts the high-dimensional feature vector of the image, maps the feature dimension from D to 0.75D, and then inputs the dimensionality-reduced features into the ensemble tree model for training.

[0038] S7: Utilize the trained ensemble tree model to learn the nonlinear relationship between the dimensionality reduction features of the graph data and the actual current and voltage characteristics, thereby enabling continuous characteristic prediction of photovoltaic cell current and voltage data.

[0039] The following is a specific example to illustrate the solar cell electrical performance prediction method of this scheme: The dataset used in this study comes from a photovoltaic module manufacturer, from which 70,529 electroluminescence images of TOPCon solar cells were collected. These images include defect-free (OK) samples and defective (NG) samples, along with their corresponding current and voltage parameters. Defect types are labeled based on visual characteristics or underlying causes, including grid line breakage, black core, scratches, dirt, star-shaped cracks, microcracks, and black edges. These seven defect types account for approximately 8% of the total. Two primary evaluation metrics are used: accuracy and coefficient of determination. To measure the effectiveness of the method of the present invention.

[0040] Step 1: Regarding the conversion efficiency Eff, one of the characteristics of current and voltage, the median Eff was found to be 26.1% and the mean absolute deviation was 0.23% by statistically analyzing the current and voltage parameters of all samples. The Eff parameter of each image sample was then normalized using MMAD normalization. The MMAD normalization formula is as follows:

[0041] The dataset was divided into seven bins based on the median and mean absolute deviation of the efficiency parameter Eff: one “reject” bin (<25.4%) and six bins with a width of 0.2% between 25.4% and 26.6%; the “reject” bins accounted for 1.8% of the total electroluminescence-current-voltage data pairs.

[0042] Step 2: Perform preprocessing on the collected raw electroluminescence image samples, and divide the training-validation set according to the grading task and regression prediction task.

[0043] The specific steps for preprocessing and partitioning the raw data are as follows: 1. Differentiated data augmentation strategies are adopted for images with different defect types to construct a training dataset; OK sample transformation includes image scaling to 256*256 pixels, grayscale conversion to three channels, random horizontal flip probability of 0.5, random vertical flip probability of 0.5, and pixel value normalization to the range of -1 to 1; NG sample transformation adds random rotation of ±20 degrees and random affine transformation including translation range of ±10% and scaling range of 0.9 to 1.1 on the basis of OK sample transformation; while test set transformation only includes image scaling, grayscale conversion and pixel value normalization without random augmentation.

[0044] 2. A mutually exclusive partitioning strategy is adopted to divide the dataset into two mutually exclusive subsets: "ranking" and "regression". The dataset for the ranking task accounts for 80% and the dataset for the regression prediction task accounts for 20%. The random seed is 42. The threshold filtering mechanism sets the threshold for the efficiency parameter Eff to 25.4% and adopts the greater than or equal to mode. Samples that do not meet the condition are moved from the regression prediction task to the ranking task training set.

[0045] Step 3: For the preprocessed and partitioned dataset, perform block embedding and hierarchical convolutional backbone network to extract multi-scale semantic representations, and introduce cross-spatial axis attention modules at specific locations in its high-level semantic stage.

[0046] The specific steps for constructing a hierarchical convolutional neural network that integrates cross-spatial axis attention are as follows: 1. Set the network configuration for the preprocessed Eff image data: 3 input channels, 96 initial embedding dimensions, 4 MLP expansion ratio, and 4 image patch size and stride; extract features from the input patch embedding layer to obtain the initial feature representation FC2 with batch size * 96 * 64 * 64.

[0047] 2. The hierarchical convolutional backbone network consists of four layers, with the number of ConvNeXt convolutional layers in each layer being 1, 2, 8, and 2 respectively. The initial feature representation FC2 enters the first-stage convolutional block to extract features, resulting in feature representation FC3. Feature representation FC3 is downsampled to reduce spatial resolution and increase the number of channels, resulting in batch *192*32*32 feature representation FC4. It then enters the second stage convolutional block to obtain feature representation FC5. FC5 is then downsampled to reduce spatial resolution and increase the number of channels, resulting in batch *384*16*16 feature representation FC6.

[0048] 3. After the fourth layer of the third-stage convolutional block, a cross-spatial axis attention layer is inserted. After processing, a downsampling and merging layer is used to obtain the batch *768*8*8 feature representation FC10. Then, the fourth-stage convolutional block is entered, and a cross-spatial axis attention layer is inserted at the end to obtain the feature representation FC12. Overall, a pyramid-shaped hierarchical feature extraction structure is formed. The first two stages mainly realize the extraction of local texture features, while the latter two stages realize the extraction of high-level semantic features and adaptive enhancement of spatial structure information.

[0049] 4. The construction of the cross-spatial axis attention layer first captures local features through basic depthwise convolution; for each scale branch, one-dimensional convolution is performed independently along the height and width axes, first performing 1*k horizontal convolution and then k*1 vertical convolution; the outputs of each scale branch are accumulated and fused, and then channel fusion is performed through 1*1 convolution; the fused attention map is multiplied element-wise with the input features to obtain the attention-weighted output; the cross-spatial axis attention module is embedded in the middle position of the third stage of the network to capture the spatial dependencies of the mid-layer features, and after the fourth stage of the network, global context information is established.

[0050] Step 4: Globally aggregate features through attention pooling layer, followed by the classification head fully connected layer to output the photovoltaic cell efficiency Eff classification results.

[0051] The specific steps for the final training of the model are as follows: 1. An attention pooling module with learnable query vectors is adopted. The learnable query parameters are 1*1*channel number in shape. Both key projection and value projection use 1*1 convolution. The key and value are reconstructed as batch*HW*channel number, and the query is expanded as batch*1*channel number. The spatial dimensions are weighted and converged to obtain the global semantic feature vector. The number of channels is mapped to a fixed-dimensional feature space through 1*1 convolution. The feature vector is then transformed by a fully connected classification layer to output the efficiency grading range for each battery sample.

[0052] 2. Train a hierarchical convolutional network classifier that integrates cross-spatial axis attention: The trainer is configured with a batch size of 32, a validation set ratio of 20%, and a learning rate of 4e-4; the classification network is trained using the cross-entropy loss function, the optimizer is AdamW, the learning rate scheduler strategy is ReduceLROnPlateau, and the learning rate is multiplied by 0.5 when the validation loss does not decrease for 3 consecutive rounds; the model is tiered on the test features, the accuracy evaluation index is calculated, and the model is saved.

[0053] 3. For example Figure 3 As shown in Figure a, the comparison results of the proposed method with four classic convolutional neural networks on three accuracy metrics in the conversion efficiency ranking task are presented. The comparison results show that the proposed method performs better in all accuracy metrics and can meet the needs of practical applications. The specific ranking results and the sample size distribution corresponding to each rank are shown in Figure a. Figure 4 As shown in Figure a, it can be seen that the vast majority of samples were correctly classified.

[0054] Step 5: Freeze the backbone network as a feature extractor to extract graph data features and realize regression prediction of the current-voltage efficiency parameter Eff.

[0055] The specific steps for feature extraction and ensemble tree regression model training are as follows: 1. After classification training, all backbone network parameters are frozen and used as feature extractors. The original fully connected layer of the classification head is replaced with the feature projection head layer, and the feature dimension is mapped from D to 0.75D. For batch feature extraction, the feature extractor is set to evaluation mode, the gradient traversal data loader is not calculated, and all batch features are vertically stacked to obtain an N*960 feature matrix.

[0056] 2. For each sample in the regression task training set, extract high-dimensional feature vectors through the reconstructed network, reduce the dimensionality to a low-dimensional space through the principal component analysis module, fit and transform on the training features, and only transform on the test features, and save the PCA model; traverse each model configuration, train the ensemble tree model on the dimensionality-reduced training features, predict on the test features, calculate the R² evaluation index, and save the model.

[0057] 3. For example Figure 3 As shown in Figure b, the results of comparing the proposed method with four classic convolutional neural networks combined with ensemble tree models in the conversion efficiency regression prediction task are presented; the results show that the proposed method performs best overall in terms of the coefficient of determination; furthermore, Figure 4b shows the specific prediction results and the distribution of the conversion efficiency of each sample prediction, where only 0.1% of the samples have a prediction deviation exceeding the set threshold.

[0058] In summary, this scheme introduces a cross-spatial-axis attention module into the high-level semantic stage of the hierarchical convolutional backbone network and employs a cascaded decomposition of horizontal and vertical strip convolutions to explicitly model the long-range dependencies in two spatial dimensions. This enables the model to effectively extract global spatial features from electroluminescent images without significantly increasing computational burden, thereby obtaining more discriminative deep feature representations and laying a solid foundation for subsequent accurate predictions. Simultaneously, a two-stage hybrid training strategy is employed. In the first stage, a classification task allows the model to learn the common features of solar cells of different performance levels, giving the backbone network strong generalization ability and robustness. In the second stage, the backbone network is frozen as a feature extractor, and the extracted high-dimensional deep features are dimensionality-reduced using principal component analysis and then input into the ensemble tree model for regression prediction. This approach combines the powerful automatic feature extraction capabilities of deep learning with the advantages of ensemble tree models in handling tabular data and fitting nonlinear relationships, effectively improving the prediction accuracy of continuous current and voltage characteristics.

[0059] Although the specific embodiments of the invention have been described in detail with reference to the accompanying drawings, this should not be construed as limiting the scope of protection of this patent; various modifications and variations that can be made by a person skilled in the art without inventive effort within the scope described in the claims are still within the scope of protection of this patent.

Claims

1. A method for predicting the electrical performance of a solar cell, characterized in that, Includes the following steps: S1: Obtain electroluminescence image data of solar cells and their corresponding current and voltage characteristic data; and after preprocessing them, divide the dataset for the grading task and the regression prediction task. S2: Construct a cross-spatial axis hierarchical convolutional neural network to extract features from electroluminescent image data, and introduce a cross-spatial axis attention module in the high-level semantic feature extraction process of the network to obtain enhanced feature representation; S3: Perform global aggregation on the enhanced feature representation, and output the classification results of solar cell current and voltage data through the classification head layer; S4: A two-stage training strategy is adopted. After the first stage of classification training converges, the network parameters are frozen as a feature extractor to extract the high-dimensional feature vector of the image. After dimensionality reduction by principal component analysis, the vector is input into the ensemble tree model for training. S5: Utilize the trained ensemble tree model to learn the nonlinear relationship between image features and actual current and voltage characteristics, and perform continuous characteristic prediction on photovoltaic cell current and voltage data.

2. The method for predicting the electrical performance of solar cells according to claim 1, characterized in that, Step S1 specifically includes: S11: Normalize the electroluminescence image data and construct a dataset by adopting differentiated data augmentation strategies for images with different defect types; S12: The current and voltage characteristic data are processed using the median and mean absolute deviation normalization methods; S13: Divide the dataset into two independent subsets for training and validation of the grading task and the regression prediction task, respectively. S14: For the dataset of the regression prediction task, samples are filtered by setting parameter thresholds, and low-quality samples that do not meet the threshold conditions are moved to the dataset of the tiered task.

3. The method for predicting the electrical performance of solar cells according to claim 1, characterized in that, The cross-spatial-axis hierarchical convolutional neural network includes a block embedding layer for extracting initial feature representations of the input image and a hierarchical convolutional backbone network for further extracting multi-scale semantic representations; The hierarchical convolutional backbone network comprises four levels, constructing multi-scale feature representations through layer-by-layer downsampling and feature transformation. The first two levels are used to extract local texture features, while the last two levels are used to extract high-level semantic features and adaptively enhance spatial structure information.

4. The method for predicting the electrical performance of solar cells according to claim 3, characterized in that, Step S2 specifically includes: S21: Input the preprocessed electroluminescence image data into the block embedding layer to extract features and obtain the initial feature representation FC2; S22: The initial feature representation FC2 enters the first layer of the hierarchical convolutional backbone network, and features are extracted through a 1-layer ConvNeXt convolutional layer to obtain the feature representation FC3; S23: Feature representation FC3 is downsampled to reduce spatial resolution and increase the number of channels to obtain feature representation FC4, which then enters the second stage of two ConvNeXt convolutional layers to obtain feature representation FC5; S24: Feature representation FC5 is then downsampled to reduce spatial resolution and increase the number of channels to obtain feature representation FC6. Subsequently, a cross-spatial axis attention layer is introduced in the middle of the 8 ConvNeXt convolutional layers in the third stage. After the last 4 ConvNeXt convolutional layers in this stage, feature representation FC9 is obtained. S25: Feature representation FC9 is finally passed through a downsampling layer to obtain feature representation FC10, and then enters the fourth level stage with two ConvNeXt convolutional layers to obtain feature representation FC11. Following this, a cross-spatial axis attention layer is introduced to obtain feature representation FC12; and overall, a pyramid-shaped hierarchical feature extraction structure is formed.

5. The method for predicting the electrical performance of solar cells according to claim 4, characterized in that, The cross-spatial axis attention modules are embedded in the last two layers of the hierarchical convolutional backbone network. The cross-spatial axis attention modules capture local features through basic depth convolution, and then perform one-dimensional convolutions independently along the height axis and width axis. Through the concatenation decomposition of horizontal and vertical convolutions, the long-range dependencies in the two spatial dimensions are explicitly modeled to obtain adaptively enhanced spatial structure information.

6. The method for predicting the electrical performance of solar cells according to claim 5, characterized in that, The specific construction method of the cross-spatial axis attention module is as follows: First, local features are extracted through depthwise convolution; then, for each scale branch, a 1×k horizontal convolution is performed first, followed by a k×1 vertical convolution; the outputs of each scale branch are accumulated and fused, and then channel fusion is performed through a 1×1 convolution; the fused attention map is multiplied element-wise with the input features to obtain the attention-weighted output.

7. The method for predicting the electrical performance of solar cells according to claim 1, characterized in that, Step S3 specifically includes: S31: Input the enhanced feature representation into the attention pooling module to perform weighted convergence of spatial dimensions and obtain the global semantic feature vector; S32: After the global semantic feature vector is mapped to a fixed-dimensional feature space by a convolutional layer, the fully connected classification layer transforms the global semantic feature vector and outputs the grading range of each solar cell.

8. The method for predicting the electrical performance of a solar cell according to claim 7, characterized in that, The size and boundaries of the grading intervals are dynamically divided based on the current and voltage data distribution of all solar cell samples.

9. The method for predicting the electrical performance of a solar cell according to claim 1, characterized in that, Step S4 specifically includes: Phase 1: Train a cross-spatial axis hierarchical convolutional neural network with a classification head layer, and use the cross-entropy loss function to learn the mapping relationship from images to efficiency tiers; Second stage: After the training in the first stage converges, the classification head layer is replaced with the feature projection head layer, and the parameters of the entire backbone network are frozen and used as a feature extractor. The feature extractor extracts the high-dimensional feature vectors of the image, and then reduces them to a low-dimensional space through principal component analysis. The reduced features are then input into the ensemble tree model for training.

10. The method for predicting the electrical performance of a solar cell according to claim 1, characterized in that, The prediction method is based on non-contact measurement using electroluminescence images and is used for quality inspection and performance grading in the photovoltaic cell production process.

Images