Satellite solar cell array data-driven modeling method based on working condition recognition

Abstract

The application provides a satellite solar cell array data-driven modeling method based on working condition recognition, belongs to the technical field of aerospace, and solves the difficulties in management and maintenance of a solar cell array and the deficiencies in a mechanism modeling method and a traditional data-driven modeling method. Modeling steps include: collecting telemetry data of an on-orbit satellite and performing preprocessing; adding labels to the preprocessed telemetry data, and dividing the labeled telemetry data into a training set, a validation set and a test set; establishing a data-driven model, training, validating and testing a CNN network in the data-driven model, and outputting working condition recognition results; re-dividing the telemetry data after feature selection and data cleaning into data sets under five working conditions according to the labels, and respectively training, validating and testing LSTM networks under different working conditions in the data-driven model; inputting the working condition recognition results output by the CNN network into corresponding LSTM networks, and acquiring predicted output parameters of the solar cell.

Description

TECHNICAL FIELD

[0001] The application relates to a satellite solar cell array data-driven modeling method based on working condition recognition and belongs to the technical field of aerospace. BACKGROUND

[0002] With the rapid development of aerospace technology, satellites play an important role in the fields of communication, weather forecasting, geological monitoring and deep space exploration. Most satellites use a solar cell array-battery combined power supply to power the load. The solar cell array is the only energy source for the satellite in orbit, so it has become one of the most important sub-modules of the satellite. Due to the long-term influence of a large amount of space radiation and temperature changes during the on-orbit operation of the solar cell array sailboard, the stability of the solar cell array is reduced. In order to ensure the stable operation of the satellite in orbit and avoid the failure of the entire satellite mission due to the failure of the solar cell array, a model of the solar cell array needs to be established to identify its operating conditions and analyze and predict its output, so as to realize predictive management and maintenance.

[0003] At present, the modeling method of the solar cell array includes mechanism modeling and data-driven modeling. Among them, the mechanism modeling method mainly includes photovoltaic equation and volt-ampere characteristic curve model. The photovoltaic equation involves many parameters and complex calculation. The ultraviolet loss factor, combination loss factor, irradiation loss factor and temperature coefficient need to be determined by referring to relevant literature. The accuracy depends on the acquisition of parameters, and there is a large error. The volt-ampere characteristic curve model is an implicit function that cannot be visualized. The model is difficult to analyze, and usually needs to use a computer for iterative derivation, which requires high computing power. The traditional data-driven modeling method mainly includes decision tree, vector machine or BP neural network method. Although it does not need to consider the complex model mechanism, for dynamic and variable solar cell array working conditions, a single model will lead to a decrease in model accuracy. Especially for the shadowed state and the unshadowed state of the solar cell array, the time experienced by the above states is short, and the data volume is much smaller than that of other states, which leads to a serious data class imbalance problem, making it difficult to realize effective management and maintenance of the solar cell array. SUMMARY

[0004] In order to solve the problems of difficult management and maintenance of the solar cell array, many parameters of the solar cell array mechanism modeling method are difficult to accurately acquire, the volt-ampere characteristic equation involves implicit function analysis difficulty, and the traditional data-driven modeling method has a single model with poor precision and data class imbalance, a satellite solar cell array data-driven modeling method based on working condition recognition is proposed. The steps include:

[0005] Step 1: collecting telemetry data of a satellite in orbit;

[0006] Step 2: performing feature selection, data cleaning and data sampling on the collected telemetry data to complete data preprocessing;

[0007] Step 3: Adding labels to the preprocessed telemetry data, and dividing the labeled telemetry data into a training set, a validation set, and a test set;

[0008] Step 4: Establishing a data-driven model, training, validating, and testing the CNN network in the data-driven model using the training set, the validation set, and the test set, respectively, and outputting the operating condition recognition result;

[0009] Step 5: Re-dividing the feature-selected and data-cleaned telemetry data into data sets under five operating conditions according to the labels, dividing each operating condition data set into a training set, a validation set, and a test set in proportion, and training, validating, and testing the LSTM network in the data-driven model under different operating conditions;

[0010] Step 6: Inputting the operating condition recognition result output by the CNN network into the corresponding LSTM network to obtain the predicted output parameter of the solar cell array.

[0011] Preferably, the telemetry data of the on-orbit satellite collected in step 1 specifically includes illumination state data, earth shadow state data, entry shadow state data, exit shadow state data, and position keeping state data.

[0012] Preferably, the steps of feature selection, data cleaning, and data sampling on the collected telemetry data in step 2 include:

[0013] Step 2.1: Extracting features related to the operating state of the solar cell array from the collected telemetry data of the on-orbit satellite as training features of the model, wherein the features related to the operating state of the solar cell array include but are not limited to solar cell array temperature, analog solar sensor current, solar cell subarray voltage, and illumination / earth shadow state;

[0014] Step 2.2: Removing missing values from the feature-selected telemetry data, introducing special labels to the earth shadow state data, and separately using the Z-score method to process, using the Z-score method to process abnormal values of non-special label data, replacing abnormal values with upper and lower threshold values of 3 times the standard deviation, completing data cleaning, wherein the non-special label data includes illumination state data, entry shadow state data, exit shadow state data, and position keeping state data;

[0015] Step 2.3: Taking the entry shadow state data as a reference, the amount of entry shadow state data is n, the amounts of earth shadow state and exit shadow state data are 2n, the amount of illumination state data before sampling entry shadow state is n, the amount of illumination state data after sampling exit shadow state is n, and the position keeping state data is sampled one by one, completing data preprocessing, and saving and exporting the preprocessed telemetry data as time series data sets according to the time dimension.

[0016] Preferably, the step 3 of adding labels to the pretreated telemetry data specifically includes: obtaining labels according to the "light / eclipse state" feature, when the light / eclipse state is 0, the label is a light state; when the light / eclipse state is 1, the label is an eclipse state; the label of the data before the eclipse state and after the light state is an eclipse-in state; the label of the data after the eclipse state and before the light state is an eclipse-out state; and obtaining a label of a position keeping mode according to a feature of an attitude orbit control mode.

[0017] Preferably, the CNN network in step 4 is composed of 4-6 convolutional layer structures and 1-2 fully connected layer structures, each convolutional layer is composed of a one-dimensional convolutional layer, a batch normalization layer, a nonlinear activation function layer and a maximum pooling layer, and each convolutional layer structure is sequentially connected, and the fully connected layer is connected after the last convolutional layer structure and is used to convert data into a final classification category number.

[0018] Preferably, the step 4 of training, verifying and testing the CNN network in the data-driven model using the training set, the verification set and the test set specifically includes:

[0019] The training set is used to train the CNN network, the verification set is used to verify the structure parameters, weights and biases of the CNN network, the network parameters are optimized and saved, the test set is used to test the accuracy of the solar cell array working condition classification and recognition, if the accuracy is lower than a preset value, the model is retrained until the accuracy is greater than the preset value;

[0020] The expression of the accuracy of the solar cell array working condition classification and recognition is:

[0021]

[0022] In formula (1), TP i is the number of samples correctly predicted as category i, C is the number of working conditions, and N is the total number of samples.

[0023] Preferably, the step 5 of redividing the telemetry data after feature selection and data cleaning into data sets of five working conditions according to the labels includes:

[0024] Step 5.1: dividing the time series data set into data sets in five working conditions according to the labels of the telemetry data "light state, eclipse state, eclipse-in state, eclipse-out state and position keeping state", wherein the data sets in the five working conditions include a light state data set, an eclipse state data set, an eclipse-in state data set, an eclipse-out state data set and a position keeping state data set;

[0025] Step 5.2: Determine the time step tw of each working condition corresponding data set, and generate the input sequence and corresponding target value using the sliding window method, where the input sequence is the tw previous time step data, and the target value is the tw+1 data. Starting from the beginning of the time series, slide forward one time step each time until the end of the sequence, complete the creation of fixed-length input and output sequences.

[0026] Preferably, the feature dimension of the LSTM network in step 5 includes but is not limited to solar array temperature, analog solar sensor current, solar cell subarray voltage, illumination / eclipse state, and the target feature of the LSTM network is solar array current. Each LSTM network consists of 2-6 layers of LSTM subnetwork layer structure and one fully connected layer, wherein the LSTM subnetwork layer structure contains a certain number of hidden layer units and a nonlinear activation function, the LSTM subnetwork layer structure is connected in sequence, the fully connected layer structure contains a forward propagation and a backward propagation process, the fully connected layer is connected after the last LSTM subnetwork layer structure, and the output size of the fully connected layer is 1, which is used to fit the output parameter of the solar array.

[0027] Preferably, the LSTM networks under five working conditions in step 5 include: an LSTM network under illumination state, an LSTM network under eclipse state, an LSTM network under entering eclipse state, an LSTM network under exiting eclipse state, and an LSTM network under position keeping state. Training, validating and testing the LSTM networks under five working conditions in the data-driven model specifically include:

[0028] The LSTM network under illumination state is trained using the illumination state training set, the structure parameters, weights and biases of the LSTM network under illumination state are verified using the illumination state validation set, the network parameters are optimized and saved, and the fitting effect of the selected solar array target feature under illumination state is verified using the illumination state test set.

[0029] The LSTM network under eclipse state is trained using the eclipse state training set, the structure parameters, weights and biases of the LSTM network under eclipse state are verified using the eclipse state validation set, the network parameters are optimized and saved, and the fitting effect of the selected solar array target feature under eclipse state is verified using the eclipse state test set.

[0030] The LSTM network under entering eclipse state is trained using the entering eclipse state training set, the structure parameters, weights and biases of the LSTM network under entering eclipse state are verified using the entering eclipse state validation set, the network parameters are optimized and saved, and the fitting effect of the selected solar array target feature under entering eclipse state is verified using the entering eclipse state test set.

[0031] The LSTM network under the shadowing state condition is trained using the shadowing state training set, the structure parameters, weights and biases of the LSTM network under the shadowing state condition are verified using the shadowing state verification set, the network parameters are optimized and saved, and the fitting effect of the shadowing state selected solar cell array target feature is tested using the shadowing state test set.

[0032] The LSTM network under the position keeping state condition is trained using the position keeping state training set, the structure parameters, weights and biases of the LSTM network under the position keeping state condition are verified using the position keeping state verification set, the network parameters are optimized and saved, and the fitting effect of the position keeping state selected solar cell array target feature is tested using the position keeping state test set.

[0033] Preferably, the predicted output parameter of the solar cell array in step 6 specifically comprises:

[0034] According to the working condition recognition result of the CNN network, the corresponding LSTM network is selected for prediction, the future solar cell array output is predicted for 5-20 time steps, and the error index of the data-driven model is calculated, if the error index is greater than a preset threshold, the CNN network and the LSTM network are retrained until the error index is ≤ the preset threshold, and the predicted output of the solar cell array is obtained.

[0035] The expression of the error index is:

[0036]

[0037] In formulas (2) and (3), N is the total number of samples, y i is the true value of the i th sample, is the model output value of the i th sample, MAE is the mean absolute error, and MSE is the mean square error.

[0038] The beneficial effects of the present application are:

[0039] (1) The present application proposes a working condition recognition method for a satellite solar cell array, which can realize accurate working condition recognition through a simple network structure, and can be used for state monitoring and early diagnosis of faults of the satellite solar cell array, and has an important role in adjusting data-driven decision and optimizing performance.

[0040] (2) The present application predicts the output parameter on the basis of solar cell array working condition recognition, and uses multiple models to fit the data characteristics under different working conditions, which effectively reduces the complexity of the model and increases the accuracy of the model compared with directly performing regression and fitting of the output parameter. Although the output parameters under different working conditions are obtained by different models, the output results of the models can be recombined into a time sequence according to the original time sequence of the data, so as to realize the on-orbit synchronization and perception of the satellite solar cell array.

[0041] (3) This invention establishes a model based on a data-driven method, which does not involve the working mechanism of the solar cell array and is different from traditional data methods. Only the collected raw telemetry data needs to be input into the algorithm, and the functions of data preprocessing, calculation, output, comparison, etc. can all be automatically realized by the algorithm, without requiring relevant professional knowledge from ground operators. Attached Figure Description

[0042] Figure 1 A flowchart illustrating a satellite solar array data-driven modeling method based on operational condition identification provided by this invention;

[0043] Figure 2 A schematic diagram illustrating the principle of a satellite solar array data-driven modeling method based on operational condition identification provided by this invention.

[0044] Figure 3 This is a schematic diagram of the CNN network structure provided by the present invention. Detailed Implementation

[0045] Combination Figures 1-3 This implementation method is described as follows: Figure 1 and Figure 2 As shown, the steps of the satellite solar cell array data-driven modeling method based on operating condition identification described in this embodiment include:

[0046] S1: Collect telemetry data and perform data preprocessing;

[0047] This implementation method uses telemetry data from the solar array of a certain type of high-orbit satellite collected from March 19, 2023 to December 20, 2023 as the acquired data. The acquired telemetry data includes the spring equinox, summer solstice, autumn equinox, winter solstice, and other dates or time periods. The data preprocessing specifically includes:

[0048] S101: Feature selection is performed on telemetry data. Training features include the temperature of the solar array at different sampling points, the state of the illuminated / shaded areas, the voltage of all solar array subarrays, and the current of the analog solar sensor at different sampling points. The target features are manually labeled tags.

[0049] S102: Remove missing values in telemetry data. Due to the long distance of the transmission link and the existence of certain electromagnetic interference, a small number of training features of telemetry data have abnormal values. The abnormal values are detected using the Z-score method and replaced with the upper and lower threshold of 3 standard deviations. The Z-score method is a data standardization method mainly used to measure the distance of a data point from the mean standard deviation. By converting the data to a standard normal distribution (mean 0, standard deviation 1), the dimensional differences between different data sets are eliminated, facilitating comparison or subsequent analysis.

[0050] S103: Independently sample data for each day of data. For example, the telemetry data on the autumn equinox contains 61 time steps in the earth shadow state. Then, sample 122 time steps at equal intervals for the earth shadow state and the out-of-shadow state. Sample 61 time steps at equal intervals before and after the illumination state in the in-shadow state. After sampling the telemetry data in all time periods, save and export the new time series data set in the time dimension;

[0051] S2: Add labels and divide the training set, validation set, and test set.

[0052] The illumination state in the preprocessed data set is labeled 0, the earth shadow state is labeled 1, the in-shadow state is labeled 2, the out-of-shadow state is labeled 3, and the position maintaining state is labeled 4. The labeled telemetry data set is divided into a training set, a validation set, and a test set in a ratio of 14:3:3.

[0053] S3: Build a data-driven model, train the CNN network, and verify the structure parameters, weights, and biases of the network.

[0054] The CNN network in the embodiment is a convolutional neural network composed of 5 convolutional layer structures and 2 fully connected layer structures. The input channel number of the network is 1, and the output channel number is 5. The first to third convolutional layer structures are composed of one-dimensional convolutional layers, batch normalization layers, nonlinear activation function ReLU layers, and maximum pooling layers. The fourth and fifth convolutional layer structures are composed of one-dimensional convolutional layers, batch normalization layers, and nonlinear activation function ReLU layers. The output channel number of the mth convolutional layer is consistent with the input channel number of the m+1 convolutional layer. The batch normalization channel number of the mth layer is consistent with the output channel number of the mth convolution. The convolution kernel size, step, and padding size can be designed according to the data dimension. The network structure of the CNN is shown in Figure 3 The input dimension of the model is the batch size, the input channel number, and the sequence length. After processing by the convolutional layer structure and the fully connected layer structure, the output of the model is the batch size and the number of classifications.

[0055] The training of the CNN network and the verification of the structure parameters, weights and biases of the network specifically include: training the CNN network using the training set, verifying the structure parameters, weights and biases of the CNN network using the verification set, optimizing the network parameters and saving.

[0056] S4: verifying and outputting the identification result of the solar cell array working condition;

[0057] The embodiment uses the test set to verify the accuracy of the solar cell array working condition classification identification, and if the accuracy is lower than a preset value, the model is retrained until the accuracy is greater than the preset value.

[0058] The expression of the accuracy of the solar cell array working condition classification identification is:

[0059]

[0060] In formula (1), TP is the number of samples correctly predicted as class i, C is the number of working conditions, and N is the total number of samples. i

[0061] S5: creating fixed-length input and output sequences from the time series data according to the five different working conditions, and dividing them into five groups of training set, verification set and test set respectively;

[0062] According to the labels of the telemetry data, i.e., illumination state, earth shadow state, entry shadow state, exit shadow state and position keeping state, the embodiment redivides the data after feature selection and data cleaning into five new data sets. Input and output sequences are created for each LSTM network to facilitate the training of time series data, wherein the LSTM network is a recurrent neural network. The method of creating input and output sequences specifically includes:

[0063] S501: extracting consecutive time steps from the input data to form a sequence of a fixed time window tw as input, i.e., the required historical data of the model, and the time window of the embodiment is 100.

[0064] S502: determining the number of future steps that the model needs to predict, and the prediction length of the embodiment is 10 time steps.

[0065] S503: traversing the input data segmentation sequence, starting from the beginning of the time series, and sliding the index i by one time step each time until L-tw, L being the sequence length of the data set.

[0066] S504: in each iteration, extracting an input feature sequence of length tw from the current index i.

[0067] S505: extracting the part from the end of the current window to the prediction length in the target feature as an output feature sequence. ​

[0068] S506: The input sequence and output sequence under each working condition are divided into training set, validation set and test set according to the ratio of 14:3:3, and minimum-maximum normalization is performed. Among them, the training features are consistent with the training set of the CNN network, and the target features are the solar cell array current.

[0069] S6: Five LSTM networks are trained respectively, and the structure parameters, weights and biases of the network are verified;

[0070] In this embodiment, the LSTM network is composed of a 3-layer LSTM subnetwork layer structure and a full connection layer structure, and adopts a multi-element feature prediction unit feature. The LSTM subnetwork layer structure contains 50 hidden layer units. The training process of the LSTM subnetwork layer structure follows the following formula:

[0071]

[0072] h (t) =o (t) ·tanh(c (t) ) (4);

[0073] In formula (2)-(4), i (t) is the input gate, f (t) is the forget gate, o (t) is the output gate, is the input new information, σ is a nonlinear activation function sigmoid, W is a matrix of weights, h (t-1) is the output information at t-1, c (t-1) is the state of information at t-1, and c (t) is the information state at t. In the iterative update process of the model, the input is the input information x (t) at t, the output is the output information h (t) at t, and c (t-1) and c (t) participate in the iteration of the model as historical state information, and determine the allocation of the model to the data at t-1 and t;

[0074] The LSTM networks under five working conditions are trained using the corresponding training set, the structure parameters, weights and biases of the network are verified using the validation set data, the network parameters are optimized and saved, and finally the fitting effect of the target features of the solar cell array is verified using the test set.

[0075] S7: Verify and predict the output parameters of the solar cell array;

[0076] For different working conditions, the LSTM network corresponding to the characteristics of the working conditions is used for prediction, the trained LSTM model is loaded, and the test set sequence corresponding to the working condition is input, the output is generated and inverse normalized, and the mean absolute error and mean square error are used to evaluate the performance of the model;

[0077] The expression of the error index is:

[0078]

[0079] In formulas (5) and (6), N is the total number of samples, y i is the true value of the ith sample, is the model output value of the ith sample, MAE is the mean absolute error, and MSE is the mean square error.

[0080] The above is only a preferred embodiment of the present application, and is not intended to limit the present application in any form. Although the present application has been disclosed as above with a preferred embodiment, it is not intended to limit the present application. Any person skilled in the art can make some changes or modifications to the above disclosed technical content without departing from the scope of the technical solution of the present application, and any simple modification, equivalent replacement and improvement of the above embodiments within the scope of the technical solution of the present application, the technical essence of the present application, and the spirit and principles of the present application, all of which are within the protection scope of the present application.

Claims

1. A data-driven modeling method for satellite solar cell arrays based on operating condition identification, characterized in that, The steps of the satellite solar array data-driven modeling method based on operating condition identification include: Step 1: Collect telemetry data from the satellites in orbit; Step 2: Perform feature selection, data cleaning, and data sampling on the collected telemetry data to complete data preprocessing; Step 3: Add labels to the preprocessed telemetry data and divide the labeled telemetry data into training set, validation set and test set; Step 4: Establish a data-driven model, and use the training set, validation set and test set to train, validate and test the CNN network in the data-driven model, and output the working condition recognition results. Step 5: Based on the labels, the telemetry data after feature selection and data cleaning are re-divided into datasets for five different operating conditions. The datasets for each operating condition are divided into training, validation, and test sets according to the proportions, and the LSTM networks in the data-driven model under different operating conditions are trained, validated, and tested respectively. The datasets for the five operating conditions include illumination status dataset, ground shadow status dataset, shadow entry status dataset, shadow exit status dataset, and position holding status dataset. Step 6: Input the working condition recognition results output by the CNN network into the corresponding LSTM network to obtain the predicted output parameters of the solar cell array.

2. The satellite solar cell array data-driven modeling method based on operating condition identification according to claim 1, characterized in that, The telemetry data collected in step 1 from the on-orbit satellite specifically includes illumination status data, ground shadow status data, shadow entry status data, shadow exit status data, and position holding status data.

3. The satellite solar array data-driven modeling method based on operating condition identification according to claim 1, characterized in that, Step 2, which involves feature selection, data cleaning, and data sampling of the collected telemetry data, includes: Step 2.1: Extract features related to the operating status of the solar array from the telemetry data collected from the on-orbit satellite as training features for the model. The features related to the operating status of the solar array include, but are not limited to, solar array temperature, analog solar sensor current, solar subarray voltage, and illumination / shadow status. Step 2.2: Remove missing values ​​from the telemetry data after feature selection, introduce special labels for the shadow status data and process them separately using the Z-score method, and process outliers for the data without special labels using the Z-score method, replacing outliers with upper and lower limits of 3 times the standard deviation to complete data cleaning. The data without special labels include illumination status data, shadow entry status data, shadow exit status data and position holding status data. Step 2.3: Using the shadow entry state data as a baseline, the amount of shadow entry state data is n, the amount of shadow entry state and shadow exit state data sampled at equal intervals is 2n, the amount of illumination state data before the shadow entry state is sampled is n, and the amount of illumination state data after the shadow exit state is n. The position-maintaining state data is sampled one by one to complete the data preprocessing. The preprocessed telemetry data is saved according to the time dimension and exported as a time series dataset.

4. The satellite solar array data-driven modeling method based on operating condition identification according to claim 1, characterized in that, Step 3, which involves adding labels to the preprocessed telemetry data, specifically includes: obtaining labels based on the "illuminance / shadow state" feature; when the illumination / shadow state is 0, the label is illumination state; when the illumination / shadow state is 1, the label is shadow state; the data before the shadow state and after the illumination state is labeled as entering shadow state; the data after the shadow state and before the illumination state is labeled as exiting shadow state; and obtaining the label for the position holding mode based on the attitude and orbit control working mode features.

5. The satellite solar array data-driven modeling method based on operating condition identification according to claim 1, characterized in that, In step 4, the CNN network consists of 4-6 convolutional layers and 1-2 fully connected layers. Each convolutional layer consists of a one-dimensional convolutional layer, a batch normalization layer, a non-linear activation function layer, and a max pooling layer. Each convolutional layer is connected in sequence, and the fully connected layer is connected after the last convolutional layer to transform the data into the final number of classification categories.

6. The satellite solar array data-driven modeling method based on operating condition identification according to claim 1, characterized in that, Step 4, which involves training, validating, and testing the CNN network in the data-driven model using the training, validation, and test sets, specifically includes: The CNN network is trained using the training set, and the structural parameters, weights, and biases of the CNN network are verified using the validation set. The network parameters are optimized and saved. The accuracy of the solar cell array condition classification and recognition is tested using the test set. If the accuracy is lower than the preset value, the model is retrained until the accuracy is greater than the preset value. The expression for the accuracy of solar cell array condition classification and identification is: （1）； In formula (1), To correctly predict the category i The number of samples, C For the number of working conditions, N The total number of samples.

7. The satellite solar array data-driven modeling method based on operating condition identification according to claim 1, characterized in that, Step 5, which involves reclassifying the feature-selected and data-cleaned telemetry data into datasets for five different operating conditions based on labels, includes the following steps: Step 5.1: Divide the time series dataset into five types of datasets based on the labels of the telemetry data: "illuminance status, shadow status, shadow entry status, shadow exit status, and position holding status"; Step 5.2: Determine the time step of the dataset corresponding to each working condition. tw The sliding window method is used to generate the input sequence and the corresponding target value, where the input sequence is the previous one. tw The data is at time step number n, and the target value is the nth time step. tw With one data point added, starting from the beginning of the time series, slide forward one time step at a time until the end of the series, thus creating a fixed-length input and output sequence.

8. The satellite solar array data-driven modeling method based on operating condition identification according to claim 1, characterized in that, In step 5, the feature dimensions of the LSTM network include, but are not limited to, solar array temperature, analog solar sensor current, solar subarray voltage, and illumination / shadow status. The target feature of the LSTM network is the solar array current. Each LSTM network consists of 2-6 LSTM subnetwork layer structures and one fully connected layer. The LSTM subnetwork layer structure contains a number of hidden layer units and nonlinear activation functions. The LSTM subnetwork layer structures are connected sequentially. The fully connected layer structure includes forward propagation and backward propagation processes. The fully connected layer is connected after the last LSTM subnetwork layer structure. The output of the fully connected layer is 1, which is used to fit the output parameters of the solar array.

9. The satellite solar array data-driven modeling method based on operating condition identification according to claim 1, characterized in that, Step 5 involves five LSTM networks under different operating conditions: illumination, shadow, shadow entry, shadow exit, and position-preserving. Training, validating, and testing these LSTM networks in the data-driven model specifically include: The LSTM network under illumination conditions is trained using the illumination condition training set, and the structural parameters, weights, and biases of the LSTM network under illumination conditions are verified using the illumination condition validation set. The network parameters are optimized and saved, and the fitting effect of the target features of the solar cell array under illumination conditions is tested using the illumination condition test set. The LSTM network under the shadow state working condition is trained using the shadow state training set. The structural parameters, weights and biases of the LSTM network under the shadow state working condition are verified using the shadow state validation set. The network parameters are optimized and saved. The fitting effect of the shadow state target features of solar cell array is tested using the shadow state test set. The LSTM network under the shadow state working condition is trained using the shadow state training set. The structural parameters, weights and biases of the LSTM network under the shadow state working condition are verified using the shadow state validation set. The network parameters are optimized and saved. The fitting effect of the target features of the solar cell array under the shadow state is tested using the shadow state test set. The LSTM network under the shadow state working condition is trained using the shadow state training set, and the structural parameters, weights and biases of the LSTM network under the shadow state working condition are verified using the shadow state validation set. The network parameters are optimized and saved, and the fitting effect of the target features of the solar cell array under the shadow state is tested using the shadow state test set. The LSTM network under the position-holding state condition is trained using the position-holding state training set. The structural parameters, weights, and biases of the LSTM network under the position-holding state condition are verified using the position-holding state validation set. The network parameters are optimized and saved. The fitting effect of the position-holding state target features of the solar cell array is tested using the position-holding state test set.

10. A satellite solar array data-driven modeling method based on operating condition identification according to claim 1, characterized in that, Step 6, obtaining the predicted output parameters of the solar cell array, specifically includes: Based on the working condition recognition results of the CNN network, the corresponding LSTM network is selected for prediction. The predicted output of the future solar cell array is 5-20 time steps. The error index of the data-driven model is calculated. If the error index is greater than the preset threshold, the CNN network and LSTM network are retrained until the error index is less than or equal to the preset threshold, and the predicted output of the solar cell array is obtained. The expression for the error index is: （2）； （3）； In formulas (2) and (3), N The total number of samples, For the first i The true value of each sample For the first i The model output value for each sample. MAE The mean absolute error, MSE This represents the mean square error.

Images