Ultra-low orbit satellite in-orbit ignition storage battery discharge depth prediction method

Abstract

The invention discloses a method for predicting the discharge depth of an in-orbit ignition storage battery of an ultra-low orbit satellite, and the method comprises the following steps: firstly, extracting n pieces of characteristic data related to the work of a satellite and a storage battery in a preset time period from the telemetry data of the satellite, and constructing a data set; setting a machine learning regression algorithm; feature data in the training set are used as input, a prediction target is used as output, the mean square error is made to be minimum through iteration, and an optimization model is obtained; then taking the feature data in the test set as model input, predicting the output of the storage battery, and comparing and evaluating a prediction result with a prediction target in the test set to obtain a final regression model; and performing prediction by using the final regression model to obtain a storage battery discharge voltage sequence, and further predicting the discharge depth of the satellite storage battery. According to the method, a reliable basis can be quickly provided for planning, formulating and evaluating the ignition strategy, and latest telemetry data can be added, so that the subsequent prediction accuracy is improved.

Description

Technical Field

[0001] The present invention relates to the field of aerospace engineering technology, mainly to the field of satellite on-orbit autonomous control technology, and specifically to a method for predicting the discharge depth of an on-orbit ignition battery for an ultra-low-orbit satellite. Background Art

[0002] Ultra-low Earth Orbit (ULEO) satellites orbit at altitudes below 300 kilometers. As orbital altitude decreases, atmospheric density increases, and the impact of aerodynamic drag and interference torque on the satellite becomes increasingly significant. To maintain the satellite's orbital altitude and effectively conduct operations, the electric propulsion system requires increasingly frequent ignitions. While each ignition of the electric propulsion system is short during orbit maintenance, it rapidly and significantly consumes onboard energy. When igniting in illuminated areas, the solar array provides sufficient energy for ignition and other operations, and also replenishes the onboard batteries in a timely manner. In shadowed areas, the solar array lacks energy, forcing the satellite to rely solely on onboard batteries for power. Improper ignition timing, duration, and location in shadowed areas can lead to excessive battery discharge, shortening battery life and, in turn, the satellite's operational life. If the battery's final discharge voltage falls below the safety design threshold, the satellite may trigger an on-orbit safety strategy, automatically shutting off power to non-critical equipment to ensure satellite safety, resulting in unnecessary losses.

[0003] Furthermore, due to the low orbital depth of ultra-low-orbit satellites and the short visibility time of ground-based measurement and control, manual command injection cannot meet the requirements of high-frequency, high-precision satellite control. This requires ground-based planning and development of a future ignition strategy. Therefore, while maintaining the same operating mode, it is particularly important to reasonably and accurately predict the changes in battery depth of discharge after ignition over a period of time, thereby assessing the safety of the ignition strategy.

[0004] Satellites transmit massive amounts of telemetry data daily, providing information on the supply voltage, operating current, and temperature of various onboard devices, batteries, solar arrays, and more. In-depth analysis of this data can also reveal changes in the satellite's operating environment, such as whether the ignition location is in a shadowed or illuminated area, or near or far from the pole. These data are crucial for developing and evaluating ignition strategies. However, the sheer volume of telemetry data, its numerous features, and the rich variety of data types make manual analysis and judgment alone prohibitively labor-intensive and difficult to provide timely responses to ignition strategy development. Furthermore, human error can lead to the overlooking or omission of relevant feature data, resulting in inaccurate evaluation and inference results. Summary of the Invention

[0005] The purpose of the present invention is to overcome the shortcomings of the existing technology and propose a method for predicting the discharge depth of the on-orbit ignition battery of an ultra-low-orbit satellite. The method uses satellite telemetry data to build a model to predict the discharge depth of the on-orbit battery, which can quickly provide a reliable basis for the planning, formulation and evaluation of the ignition strategy. The model can also be corrected and updated according to the latest telemetry data to improve the prediction accuracy.

[0006] In order to achieve the above object, the technical solutions specifically adopted by the present invention are as follows:

[0007] A method for predicting the discharge depth of an on-orbit ignition battery of an ultra-low-orbit satellite comprises the following steps:

[0008] Step 1: Extract n feature data related to the operation of the satellite and the battery within a predetermined time period from the satellite telemetry data, and use the actual output voltage value of the battery discharge process in the telemetry data as the prediction target to construct a data set;

[0009] Step 2: Preprocess the constructed dataset and divide it into training set and test set in proportion. The test set is only used for model evaluation and does not participate in training.

[0010] Step 3: Construct a battery discharge depth model, wherein the battery discharge depth model adopts a machine learning regression algorithm;

[0011] Step 4: Take the feature data in the training set as input and the predicted target as output, and obtain the preliminary predicted battery voltage output through model operation processing. Calculate the mean square error between the predicted result and the true value of the battery voltage output, and adjust the model parameters through algorithms (such as stochastic gradient descent) to gradually reduce the mean square error value so that the model prediction gradually approaches the true value. After multiple iterations, a model f is obtained. t ; Then use the feature data in the test set as the model f t Input, get the new model f t The predicted output is compared with the predicted target in the test set, and the model is evaluated by the mean square error. If the model is underfitting (poor performance in both the training set and the test set) or overfitting (good performance in the training set but poor performance in the test set), the model f needs to be adjusted. t The complexity of the model can be adjusted, key parameters such as the learning rate and regularization coefficient can be adjusted, or the model can be replaced. If the effect is good, it is considered the final regression model F';

[0012] Step 5: Apply the final regression model F' to make predictions, obtain the battery discharge voltage series, and then predict the discharge depth of the satellite battery.

[0013] Preferably, the data set is preprocessed by removing incomplete telemetry data and omitting telemetry data during non-ignition periods.

[0014] In the above technical solution, due to the relatively short satellite tracking and control time and the need for other remote control operations, there is a possibility that the time allocated for telemetry download is insufficient, resulting in incomplete ignition process data included at the end of the downloaded telemetry data. There is also a risk that the tracking and control signal will be interrupted near the end of the download due to factors such as viewing angle and weather, resulting in the loss of the tail telemetry data. Therefore, the raw data needs to be preprocessed to first remove incomplete telemetry data and secondly omit telemetry data from non-ignition periods.

[0015] Preferably, the data set is divided into a training set and a test set in a ratio of 8:2 according to time sequence.

[0016] In the above technical solution, there are two ways to segment the dataset: one is to randomly select data to form the training set and the other is to maintain the data order and segment it according to time. Although from a training perspective, randomly selecting data results in a relatively smaller training error, ignition is a continuous process, and the change in battery discharge voltage is a time-dependent sequence. Randomly selecting data will lose this characteristic, resulting in reduced prediction accuracy. Therefore, the dataset segmentation method adopts the latter method, which maintains the order of the data in the original dataset.

[0017] Preferably, the machine learning regression algorithm is any one of a support vector machine regression algorithm, a decision tree regression algorithm, a random forest regression algorithm, and an xgboost regression algorithm.

[0018] Preferably, in step 5, in the prediction stage, the method for constructing the prediction input data set is: based on the feature data of the last m ignition processes in the historical telemetry data, the prediction input data set is constructed in a weighted manner, and the expression is as follows:

[0019]

[0020] Among them, X′ is the generated prediction input feature dataset; w i is the weight, w i ∈(0,1],(w1+w2+...+w m ) = 1; m is the number of the last few ignitions. These m ignitions have the working state closest to the battery output of the ignition process required to be predicted. Considering the amount of calculation, m = 3 to 5; X i is the i-th historical telemetry data during the last m ignitions; σ is a random number between [-0.2, +0.2], which makes the generated feature data different from the known data in the training set and also represents the random fluctuation during operation.

[0021] Preferably, in step 5, the discharge depth of the satellite battery is predicted using the following expression:

[0022] D′=G(V′)

[0023] Where V′ refers to the battery discharge voltage series, D′ refers to the predicted change in battery capacity under battery voltage changes, and G(.) refers to the mapping function, which is fitted based on the battery factory data.

[0024] In the above technical solution, based on the predicted capacity change, it can be determined whether the battery capacity after ignition discharge meets the safety requirements when the subsequent ignition mode remains unchanged, thereby determining whether the expected ignition strategy needs to be adjusted.

[0025] As a preference, obtain the latest satellite telemetry data and extract the battery voltage change data V after ignition. new , and compare it with the previous prediction result V' to evaluate the prediction accuracy. At the same time, before making the next prediction, the new telemetry data is added to the original dataset data_set to form a new dataset, so that the model can be corrected and updated in the subsequent training process.

[0026] The present invention has the following characteristics and beneficial effects:

[0027] A dataset constructed using real satellite telemetry data offers flexible feature combinations and broad coverage. The resulting model accurately reflects changes in battery discharge depth during ignition while maintaining a satellite's orbital altitude. This provides a reliable basis for developing orbital maintenance strategies for ultra-low-orbit satellites. By continuously incorporating the latest telemetry data, the model can be revised and updated during the next training session to improve prediction accuracy.

[0028] In addition, by accurately predicting the battery's output voltage and calculating the battery's discharge depth, it is possible to accurately determine whether ignition poses a safety hazard to the satellite, thereby reducing safety hazards in satellite operation and increasing the satellite's lifespan in ultra-low orbit. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 Schematic diagram of a flow chart of an embodiment of the present invention.

[0030] Figure 2 Schematic diagram of data segmentation in an embodiment of the present invention.

[0031] Figure 3 This is a comparison diagram of the voltage change and remaining capacity predictions obtained using historical telemetry data as input in an embodiment of the present invention.

[0032] Figure 4 This is a diagram showing the battery voltage output and remaining capacity prediction effect in an embodiment of the present invention.

[0033] Figure 5 This is a comparison and prediction effect diagram of the battery voltage and actual telemetry value in an embodiment of the present invention. DETAILED DESCRIPTION

[0034] The present invention is described in detail below in conjunction with specific embodiments. The following examples will help those skilled in the art to further understand the present invention, but are not intended to limit the present invention in any form. It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments can be combined with each other.

[0035] A method for predicting the depth of discharge of batteries during in-orbit ignition of ultra-low-orbit satellites is described. Actual telemetry data from an ultra-low-orbit satellite is used to construct a dataset, and a model is trained using the XGBoost regression algorithm. The implementation process of the invention is described below. The constraints are: ignition cannot cause the battery depth of discharge to exceed 20%. That is, after ignition, the battery's remaining capacity must be at least 80%. Based on battery product characteristics, this translates to an output voltage of no less than 27.2V. When the battery output voltage exceeds 28.4V, the remaining capacity is 100%.

[0036] Specific methods, such as Figure 1 As shown, the following steps are included:

[0037] Step 1: Extract n characteristic data related to satellite and battery operation within a predetermined time period from satellite telemetry data. These include: the voltage, current, and temperature of each solar array related to battery charging; the voltage, current, and temperature of the propulsion system and payload equipment related to battery discharge during ignition; and the bus voltage and current. Bus voltage and current not only reflect the power consumption of the propulsion system and payload equipment, but also the energy consumption of other electronic systems and components (such as integrated electronics, flywheel, star sensor, and sun sensor) that maintain normal satellite operation. The operating status and energy consumption of the payload equipment and other electronic devices before and during propulsion ignition directly affect the final output voltage and remaining capacity of the battery after ignition. The mission schedule of the platform and payload varies across different orbits, resulting in variations in these operating status and energy consumption. Adding these relevant features provides more detail to the algorithm, improving model prediction accuracy. A dataset is constructed using these n characteristic data as input and the actual output voltage of the battery during discharge as the prediction target.

[0038] Specifically, according to the working relationship between the onboard equipment, components and batteries, such as the battery power supply relationship, battery charging relationship, etc., in the time period T{t0,t1,t2,...,t m} select n features X{X0,X1,X2,...,X n-1}, the voltage V{v0,v1,v2,...,v m} is the prediction target, and the following data set data_set is constructed.

[0039] Step 2: Preprocess the constructed dataset and divide it into training set and test set in proportion.

[0040] Understandably, due to the relatively short satellite tracking and control timeframes and the need for other remote control operations, there's a chance that insufficient time is allocated for telemetry downloads, resulting in incomplete data on the ignition process at the end of the downloaded telemetry data. There's also the possibility that near the end of the download, due to viewing angles, weather conditions, or other factors, the tracking and control signal is interrupted, leading to loss of the tail telemetry data. Therefore, preprocessing of the raw data is necessary to first remove incomplete telemetry data and secondly, omit telemetry data from non-ignition periods.

[0041] The preprocessed data are shown in Table 1.

[0042] Table 1 Dataset

[0043]

[0044]

[0045] The above dataset is constructed from real telemetry data. Each feature in the table is correlated with battery voltage, current, temperature, and other parameters, though the strength of the correlation varies. The dataset data_set contains the actual mapping from feature combinations X to predicted targets V: V = F(X).

[0046] Based on the above technical solution, in this embodiment, the historical telemetry data of a certain ultra-low orbit satellite between February 28 and March 14, 2025 is sorted out, the timestamps therein are converted into date and time format, and incomplete data is checked and eliminated.

[0047] Based on the degree of association with the battery, 32 types of data are extracted from the telemetry data as associated features, with each feature being a column. The time series related to orbital control ignition during this period are extracted, totaling 3647 rows, forming a two-dimensional data table containing 32 × 3647 = 116704 data points, as shown in Table 2, which is the prepared data set data_set.

[0048] Table 2. Dataset preparation using telemetry data

[0049]

[0050]

[0051] It's important to note that there are two ways to split a dataset: one is to randomly select data to form the training set and the other is to maintain the data order and split it by time. Although from a training perspective, randomly selecting data results in a relatively lower training error, ignition is a continuous process, and the change in battery discharge voltage is a time-dependent sequence. Randomly selecting data loses this characteristic, resulting in reduced prediction accuracy. Therefore, the latter method is used for dataset splitting, which maintains the order of the data in the original dataset.

[0052] Therefore, in this embodiment, the time sequence of the data in the telemetry dataset is kept unchanged, and the data is divided into a training set accounting for 80% and a test set accounting for 20%.

[0053] Then we get the training set tra i n_set:

[0054] {(x 0,t0 ,x 1,t0 ,...,x n-1,t0 ,v t0 ),(x 0,t1 ,x 1,t1 ,...,x n-1,t1 ,v t1 ),...,(x 0,tj ,

[0055] x 1,tj ,...,x n-1,tj ,v tj )}

[0056] Test set test_set:

[0057] {(x 0,tj+1 ,x 1,tj+1 ,...,x n-1,tj+1 ,v t0 ),(x 0,tj+2 ,x 1,tj+2 ,...,x n-1,tj+2 ,v tj+2 ),...,(x 0,tm ,x 1,tm ,...,x n-1,tm ,v tm )}

[0058] Step 3: Construct a battery discharge depth model, wherein the battery discharge depth model adopts a machine learning regression algorithm.

[0059] In this embodiment, the machine learning regression algorithm used is the xgboost (eXtreme Gradient Boosting) regression algorithm. Xgboost is an integrated machine learning algorithm based on a decision tree. Xgboost introduces a regularization term based on the gradient boosting decision tree (GBDT) to prevent model overfitting; uses Taylor quadratic expansion terms to optimize the loss function to provide calculation accuracy; and improves the storage structure to support parallel computing.

[0060] In this embodiment, after adjusting the parameters of the xgboost regression algorithm, the parameters of the xgboost regression algorithm are selected as follows:

[0061] Number of decision trees: 180;

[0062] The maximum depth that a decision tree can reach during training is 3.

[0063] Sampling rate of each tree: 0.8;

[0064] The learning rate is 0.1.

[0065] Step 4: Take the feature data in the training set as input and the predicted target as output, and obtain the preliminary predicted battery voltage output through model operation processing. Calculate the mean square error between the predicted result and the true value of the battery voltage output, and adjust the model parameters through algorithms (such as stochastic gradient descent) to gradually reduce the mean square error value so that the model prediction gradually approaches the true value. After multiple iterations, a model f is obtained. t ; Then use the feature data in the test set as the model f t Input, get the new model f t The predicted output is compared with the predicted target in the test set, and the model is evaluated by the mean square error. If the model is underfitting (poor performance in both the training set and the test set) or overfitting (good performance in the training set but poor performance in the test set), the model f needs to be adjusted. t The complexity of the model can be adjusted, key parameters such as the learning rate and regularization coefficient can be adjusted, or the model can be replaced. If the effect is good, it is considered the final regression model F'.

[0066] Specifically, the feature values in the training set train_set are used as input, and the targets in the training set train_set are used as output. A model f is obtained by training. t ,

[0067] {v t0 ,v t1 ,...,v tj}=f t ((x 0,t0 ,x1,t0 ,...,x n-1,t0 ),(x 0,t1 ,x 1,t1 ,...,

[0068] x n-1,t1 ),...,(x 0,tj ,x 1,tj ,...,x n-1,tj ))

[0069] Take the feature value and target pair f of the test set test_set t The test evaluation is carried out using the mean square error as the evaluation method, which is defined as:

[0070]

[0071] If there is no underfitting or overfitting problem, the obtained model is regarded as the final regression model F'.

[0072] In this embodiment, after the training is completed, the evaluation results for the test set and the entire historical data are shown in Table 3:

[0073] Table 3 Best model evaluation results

[0074] Serial number project Evaluation results (MSE) 1 The training set is used as input, and the test set battery output voltage is used as comparison 0.005294 2 The entire historical data is used as input and the battery output voltage is used as comparison 0.002184

[0075] Step 5: Apply the final regression model F' to perform prediction and obtain the battery discharge voltage series.

[0076] Specifically, such as Figure 2 As shown in the figure, according to the subsequent expected ignition duration, the prediction input data set X' is constructed:

[0077] X'={(x' 0,tj+1 ,x' 1,tj+1 ,...,x' n-1,tj+1 ),(x' 0,tj+2 ,x' 1,tj+2 ,...,

[0078] x' n-1,tj+2 ),...,(x' 0,tk ,x' 1,tk ,...,x' n-1,tk )}.

[0079] Based on the initially planned ignition strategy, such as ignition position, duration, and other parameters, and based on the data X of the last m ignition processes in the historical telemetry data, the prediction input X' is constructed as follows:

[0080]

[0081] Among them, X′ is the generated prediction input feature dataset; w i is the weight, w i ∈(0,1],(w1+w2+...+w m )=1; m is the number of the last few ignitions, and these m ignitions have the working state closest to the battery output of the desired predicted ignition process. Considering the amount of calculation, in this embodiment, m=3~5; X i is the i-th historical telemetry data during the last m ignitions; σ is a random number between [-0.2, +0.2], which makes the generated feature data different from the known data in the training set and also represents the random fluctuation during operation.

[0082] Use F' to make predictions and get the expected output, which is the battery discharge voltage sequence V':

[0083] V'={v'0,v'1,v'2,...,v' k}=F'(X').

[0084] Furthermore, based on the relationship between battery voltage and capacity, the discharge depth of the satellite battery can be predicted.

[0085] Specifically, the discharge depth of the satellite battery is predicted using the following expression:

[0086] D′=G(V′)

[0087] Where V' refers to the battery discharge voltage series, D' refers to the predicted change in battery capacity under battery voltage changes, and G(.) refers to the mapping function, which is fitted based on the battery factory data. The function is fitted based on the "capacity-voltage" data of the batch battery pack used in the satellite:

[0088]

[0089] Wherein, G(V) is the battery capacity (mAh), V is the battery output voltage (V); a, b, and k are all constants: a = -138389; b = 8318; k = -0.02927.

[0090] pass Figure 4 As can be seen, in this embodiment, based on the initially planned ignition strategy, such as ignition position and duration, prediction input data is constructed based on the data from the last m = 5 ignition processes in historical telemetry data. The predicted battery output voltage and remaining capacity for the next ignition are shown in the figure below. The predicted results indicate that the minimum remaining capacity of the battery after ignition exceeds 80%, meaning that the depth of discharge is less than 20%, satisfying the constraints and indicating a safe ignition strategy.

[0091] Finally, according to the technical solution provided by this embodiment, prediction and verification are carried out through real historical data, such as Figure 3 As shown, during several ignitions between March 7 and March 10, the battery depth of discharge exceeded the 20% constraint, reaching a minimum of 28%. This exceeded expectations and failed to meet safety constraints. After March 10, the ignition strategy was adjusted, and the battery depth of discharge subsequently met the constraint.

[0092] The above results show that the prediction accuracy of the best model obtained through training is relatively high. The best model is saved and deployed in the ground system for future use.

[0093] like Figure 5 As shown in Figure 2, three ignition telemetry data on March 16, 2025 are obtained and compared with the model prediction data.

[0094] The comparison between the model prediction results and the actual telemetry values shows that the battery voltage changes predicted by the model are consistent with the actual telemetry values, and can be used as a reference for the formulation and evaluation of ignition strategies within a limited time when ultra-low orbit satellites are performing orbit control on orbit. Figure 5 As can be seen, the mean square error (MSE) between the predicted and actual telemetry values is slightly larger than the mean square error of the training results. This also reflects the fact that satellite operating conditions vary with each ignition, leading to slight variations in battery output voltage. Newly acquired telemetry data should be added to historical data so that the model can be corrected and updated during the next training and prediction process, improving subsequent prediction accuracy.

[0095] The above shows and describes the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions are merely preferred examples of the present invention and are not intended to limit the present invention. Various changes and improvements may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and improvements fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for predicting the discharge depth of an on-orbit ignition battery of an ultra-low-orbit satellite, characterized in that: The steps include: Step 1: Extract n feature data related to the satellite and battery operation within a predetermined time period from the satellite telemetry data, and use the battery discharge voltage in the telemetry data as the prediction target to construct a data set; Step 2: Preprocess the constructed dataset and divide it into training set and test set in proportion; Step 3: Construct a battery discharge depth model, wherein the battery discharge depth model adopts a machine learning regression algorithm; Step 4: Take the feature data in the training set as input and the predicted target as output, and obtain a model f through multiple iterations. t ; Then use the feature data in the test set as the model f t The input is the predicted target, and the output is iterated. The minimum mean square error is obtained by the mean square error evaluation method, and then the final regression model F' is obtained; Step 5: Apply the final regression model F' to make predictions, obtain the battery discharge voltage series, and then predict the discharge depth of the satellite battery.

2. The method for predicting the discharge depth of an on-orbit ignition battery of an ultra-low-orbit satellite according to claim 1, wherein: The associated data include: the voltage, current, and temperature of each solar array related to battery charging, the voltage, current, and temperature of the propulsion system and load equipment related to battery discharging during ignition, as well as the bus voltage and bus current.

3. The method for predicting the discharge depth of an on-orbit ignition battery of an ultra-low-orbit satellite according to claim 1, wherein: The data set is preprocessed by removing incomplete telemetry data and omitting telemetry data during non-ignition periods.

4. The method for predicting the discharge depth of an on-orbit ignition battery of an ultra-low-orbit satellite according to claim 1, wherein: The data set is divided into a training set and a test set in a ratio of 8:2 according to the time sequence.

5. The method for predicting the discharge depth of an on-orbit ignition battery of an ultra-low-orbit satellite according to claim 1, wherein: The machine learning regression algorithm is any one of a support vector machine regression algorithm, a decision tree regression algorithm, a random forest regression algorithm, and an xgboost regression algorithm.

6. The method for predicting the discharge depth of an on-orbit ignition battery of an ultra-low-orbit satellite according to claim 1, wherein: In step 5, during the prediction phase, the prediction input data set is constructed by obtaining the feature data of the last m ignition processes in the historical telemetry data as a basis and constructing the prediction input data set in a weighted manner. The expression is as follows: Among them, X′ is the generated prediction input feature dataset; w i is the weight, w i ∈(0,1],(w1+w2+...+w m )=1; m is the number of the last few ignitions; X i is the i-th historical telemetry data during the last m ignitions; σ is a random number between [-0.2, +0.2], which makes the generated feature data different from the known data in the training set and also represents the random fluctuation during operation.

7. The method for predicting the discharge depth of an on-orbit ignition battery of an ultra-low-orbit satellite according to claim 6, characterized in that: In step 5, the discharge depth of the satellite battery is predicted using the following expression: D′=G(V′) Where V′ refers to the battery discharge voltage series, D′ refers to the predicted change in battery capacity under battery voltage changes, and G(.) refers to the mapping function, which is fitted based on the battery factory data.

8. The method for predicting the discharge depth of an on-orbit ignition battery of an ultra-low-orbit satellite according to claim 1, wherein: In step 4, the model f is obtained through multiple iterations. t The method is: The feature data in the training set is used as input and the predicted target is used as output. The battery discharge depth model is used for calculation and processing to obtain a preliminary predicted battery voltage output. The mean square error between the predicted result and the true value of the battery voltage output is calculated. The model parameters are then adjusted using the stochastic gradient descent algorithm to gradually reduce the mean square error value, allowing the battery discharge depth model prediction to gradually approach the true value. Multiple iterations are performed until convergence occurs.

9. The method for predicting the discharge depth of an on-orbit ignition battery of an ultra-low-orbit satellite according to claim 1, wherein: In step 4, the final regression model F' is obtained by: Use the feature data in the test set as model f t Input, get the new model f t The prediction output of the new prediction output is compared with the prediction target in the test set, and the model is evaluated by the mean square error. According to the evaluation results, the model f is adjusted t The complexity of the training set is adjusted, the learning rate and the regularization coefficient are adjusted until the predicted output fits the predicted target in the test set, and the final trained model f is trained. t It is regarded as the final regression model F'.

Images