Satellite energy prediction method and apparatus, computer device, and storage medium

Abstract

The application relates to a satellite energy prediction method and device, computer equipment and a storage medium. State data and task plan data of a target satellite are acquired, a pre-trained hybrid prediction model is called based on the state data, a prediction SOC sequence of the target satellite in a prediction period is obtained, the remaining battery capacity in the prediction period is determined according to the prediction SOC sequence, and an energy budget value for load task planning is calculated according to the remaining battery capacity and the task plan data. Since the hybrid prediction model integrates ARIMA, CNN and LSTM, the advantages of different algorithms are fully utilized through structured division of labor and cooperation, so that the energy state and budget of the satellite in a complex and dynamic environment can be accurately predicted, reliable and forward-looking decision basis is provided for intelligent management of satellite energy, and long-term reliable operation and maximum service efficiency of the satellite can be ensured.

Description

Technical Field

[0001] This application relates to the field of aerospace telemetry and control technology, and in particular to a satellite energy prediction method, device, computer equipment, computer-readable storage medium, and computer program product. Background Technology

[0002] With the increasing complexity of satellite missions and the continuous extension of on-orbit operation time, higher demands are placed on the reliability, stability, and intelligence of satellite energy systems. Accurately predicting the satellite's energy state (such as battery state of charge, SOC) over a future period is a core prerequisite for ensuring safe satellite operation, optimizing mission scheduling, and improving energy utilization efficiency. Traditional technologies for predicting satellite energy state mainly fall into three categories:

[0003] The first category is physical modeling methods based on digital twin technology. This method constructs a high-fidelity digital twin, including a satellite orbital dynamics model, a solar azimuth calculation model, a solar panel power generation model, and energy consumption models of various platform subsystems, to simulate and extrapolate the satellite's future energy budget. Although this type of method has strong interpretability under ideal conditions, it heavily relies on accurate prior physical parameters and environmental models, resulting in high modeling complexity, high development and maintenance costs, and difficulty in effectively characterizing nonlinear and non-steady-state factors such as space environment disturbances (e.g., geomagnetic storms, high-energy particle events) and equipment aging and degradation. This leads to a significant decrease in prediction accuracy and robustness under real and complex operating conditions.

[0004] The second category involves methods that use a single statistical model, machine learning model, or deep learning model for prediction. Examples include traditional statistical / machine learning methods such as ARIMA and Support Vector Regression (SVR), or recurrent neural network structures such as LSTM and GRU. However, these methods generally have limitations in modeling capabilities: traditional statistical models struggle to capture the complex nonlinear dynamics in energy data; while single deep learning models, although possessing some nonlinear fitting ability, often fail to balance long-term trend modeling and local detail responses when processing multi-scale spatiotemporal signals that simultaneously contain long-term periodicity (such as orbital periods, seasonal variations, and solar activity cycles) and short-term transient characteristics (such as power spikes caused by load switching), leading to overfitting or underfitting problems and limited predictive performance.

[0005] The third type is the simple weighted average combination model, which involves weighting or averaging the outputs of multiple independent prediction models with fixed weights. While this method alleviates the bias of a single model to some extent, it lacks a dynamic adjustment mechanism for weight settings, fails to fully explore the complementary advantages of each sub-model under different time scales and operating modes, and cannot adaptively optimize the combination strategy based on real-time observation data. Therefore, the overall performance improvement is limited, making it difficult to meet the requirements of high-precision and high-robustness on-orbit energy prediction.

[0006] It is evident that traditional satellite energy forecasting technologies generally suffer from the following technical bottlenecks: insufficient forecast accuracy and robustness, making it difficult to cope with nonlinear noise caused by complex space environment disturbances and equipment degradation; insufficient feature extraction, lacking the ability to collaboratively model multi-scale spatiotemporal features (including linear trends, short-term abrupt changes, long-term periodic dependencies, etc.) in energy time series; and poor model adaptability, failing to achieve online learning and self-evolution when satellite operation modes switch or encounter unknown disturbances, highly dependent on human intervention, and with a low level of intelligence.

[0007] Therefore, there is an urgent need for a satellite energy prediction mechanism with strong generalization capabilities and online adaptive mechanisms to overcome the limitations of traditional technologies and improve the autonomy and reliability of satellite energy management. Summary of the Invention

[0008] Therefore, it is necessary to provide a satellite energy prediction method, device, computer equipment, computer-readable storage medium, and computer program product that can accurately predict the energy status and budget of satellites in complex and dynamic environments, in order to address the above-mentioned technical problems.

[0009] In a first aspect, this application provides a satellite energy prediction method, the method comprising:

[0010] Acquire the target satellite's status data and mission plan data, wherein the status data includes the target satellite's current energy sequence, orbital attitude sequence, and space environment data;

[0011] Based on the state data, a pre-trained hybrid prediction model is invoked to obtain the predicted SOC sequence of the target satellite in the prediction period. The hybrid prediction model is a pre-built and trained ARIMA-CNN-LSTM hybrid deep learning model.

[0012] The remaining battery charge for the prediction period is determined based on the predicted SOC sequence.

[0013] Based on the remaining battery power and the mission plan data, calculate the energy budget value for payload mission planning.

[0014] In one embodiment, the hybrid prediction model is trained as follows: Sample data is acquired and preprocessed to obtain multiple sample sequences; the sample data includes historical satellite state data; for each sample sequence, an initial ARIMA module is used to process it, obtaining a prediction sequence for the corresponding sample sequence, the prediction sequence including a linear prediction subsequence and a white noise residual subsequence; based on the sample sequence and the white noise residual subsequence, an initial CNN module is used to process it, obtaining a local feature subsequence; based on the local feature subsequence, an initial LSTM module is used for nonlinear prediction, obtaining a nonlinear prediction subsequence; an initial feature fusion module is used to fuse the white noise residual subsequence, the linear prediction subsequence, and the nonlinear prediction subsequence, obtaining a feature fusion sequence; based on the feature fusion sequence, a predicted SOC sample sequence corresponding to the sample sequence is determined; based on the predicted SOC sample sequence and pre-configured training parameters, the initial ARIMA module, the initial CNN module, the initial LSTM module, and the initial feature fusion module are jointly fine-tuned to obtain the trained hybrid prediction model.

[0015] In one embodiment, the sample sequence includes historical SOC samples; the step of processing each sample sequence using an initial ARIMA module to obtain a predicted sequence for the corresponding sample sequence includes: determining the model parameters of the initial ARIMA module based on the historical SOC samples in the sample sequence, and obtaining a historical SOC stationary sequence; separating a predictable linear trend portion and an unpredictable residual portion from the historical SOC stationary sequence based on the initial ARIMA module and the corresponding model parameters; if the residual portion is detected to be white noise, determining the linear trend portion as a linear prediction subsequence and the residual portion as a white noise residual subsequence to obtain a predicted sequence for the corresponding sample sequence.

[0016] In one embodiment, the method further includes: if the residual portion is detected to be non-white noise, returning to the step of determining the model parameters of the initial ARIMA module based on historical SOC samples in the sample sequence and obtaining the historical SOC stationary sequence.

[0017] In one embodiment, the feature fusion sequence is represented by the following formula: ,in, Represents a linear prediction subsequence. Represents a white noise residual subsequence. Represents a nonlinear prediction subsequence. Represents the feature fusion sequence, and Indicates the fusion weight, and .

[0018] In one embodiment, after jointly fine-tuning the initial ARIMA module, initial CNN module, initial LSTM module, and initial feature fusion module, the method further includes: determining a validation sequence from the plurality of sample sequences, obtaining the prediction loss of the jointly fine-tuned model based on the validation sequence; performing multiple forward propagations on the validation sequence based on Monte Carlo Dropout to obtain an uncertainty quantization estimate; the uncertainty quantization estimate includes the mean, standard deviation, and confidence interval determined based on the propagation results of multiple forward propagations; if the jointly fine-tuned model meets the expected performance based on the prediction loss and the uncertainty quantization estimate, determining the jointly fine-tuned model as a trained hybrid prediction model; if the jointly fine-tuned model does not meet the expected performance based on the prediction loss and the uncertainty quantization estimate, returning to the step of processing each sample sequence using the initial ARIMA module to obtain the prediction sequence of the corresponding sample sequence.

[0019] In one embodiment, the mission planning data includes the basic energy consumption and orbital maneuvering energy consumption required for the target satellite to perform payload mission planning; the step of calculating the energy budget value for payload mission planning based on the remaining battery power and the mission planning data includes: determining a reserved safety margin for the target satellite to perform payload mission planning, the reserved safety margin representing the energy consumption required for unpredictable events; obtaining the sum of the reserved safety margin, the basic energy consumption, and the orbital maneuvering energy consumption; and calculating the difference between the remaining battery power and the sum, the difference representing the energy budget value for payload mission planning.

[0020] In one embodiment, determining the reserved safety margin for the target satellite's payload mission planning includes: obtaining the product of the basic energy consumption and a set weighting coefficient, and determining the product as the reserved safety margin for the target satellite's payload mission planning.

[0021] Secondly, this application also provides a satellite energy prediction device, the device comprising:

[0022] The data acquisition module is used to acquire the status data and mission plan data of the target satellite. The status data includes the current energy sequence, orbital attitude sequence and space environment data of the target satellite.

[0023] The SOC prediction module is used to call a pre-trained hybrid prediction model based on the state data to obtain the predicted SOC sequence of the target satellite in the prediction period. The hybrid prediction model is a pre-built and trained ARIMA-CNN-LSTM hybrid deep learning model.

[0024] The determination module is used to determine the remaining battery power for the prediction period based on the predicted SOC sequence.

[0025] The energy budget module is used to calculate the energy budget value for payload mission planning based on the remaining battery power and the mission plan data.

[0026] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the method described in the first aspect.

[0027] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described in the first aspect.

[0028] Fifthly, this application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the method described in the first aspect above.

[0029] The aforementioned satellite energy prediction method, apparatus, computer equipment, computer-readable storage medium, and computer program product acquire the target satellite's status data and mission plan data. Based on the status data, they invoke a pre-trained hybrid prediction model to obtain the predicted State of Charge (SOC) sequence of the target satellite during the prediction period. Based on the predicted SOC sequence, they determine the remaining battery power for the prediction period. Finally, based on the remaining battery power and mission plan data, they calculate the energy budget value for payload mission planning. Because this hybrid prediction model integrates ARIMA, CNN, and LSTM, it fully leverages the advantages of different algorithms through structured division of labor and collaboration. This allows for accurate prediction of the satellite's energy status and budget in complex and dynamic environments, providing a reliable and forward-looking decision-making basis for intelligent satellite energy management, ensuring long-term reliable satellite operation and maximizing service efficiency. Attached Figure Description

[0030] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0031] Figure 1 This is a flowchart illustrating a satellite energy prediction method in one embodiment;

[0032] Figure 2This is a flowchart illustrating the model training steps in one embodiment;

[0033] Figure 3 This is a schematic diagram illustrating the principle of a satellite energy prediction method in one embodiment;

[0034] Figure 4 This is a structural block diagram of a satellite energy prediction device in one embodiment;

[0035] Figure 5 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0036] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0037] It should be noted that the terms "first," "second," etc., used in this application can be used to describe various elements, but these elements are not limited by these terms. These terms are only used to distinguish the first element from the second element. The terms "comprising" and "having," and any variations thereof, used in this application, are intended to cover non-exclusive inclusion. The term "multiple" used in this application refers to two or more. The term "and / or" used in this application refers to one of the embodiments, or any combination of multiple embodiments.

[0038] In one embodiment, such as Figure 1 As shown, a satellite energy prediction method is provided. This embodiment illustrates the method by applying it to a computer device, and it may specifically include the following steps:

[0039] Step 102: Obtain the target satellite's status data and mission plan data.

[0040] The target satellite can be any satellite for which energy forecasting is currently required. Status data includes the target satellite's current energy sequence, orbital attitude sequence, and space environment data. For example, the energy sequence can include time series data such as the target satellite's battery SOC, charge-discharge ratio, current-discharge ratio, bus voltage, main error voltage, bus current, discharge capacity, charging capacity, discharge current, charging current, load current, array current, battery temperature, battery pressure, solar panel current, solar panel voltage, solar panel output power, and solar panel temperature.

[0041] The orbital attitude sequence can include the target satellite's position X component, position Y component, position Z component, velocity X component, velocity Y component, velocity Z component, yaw angle Y, pitch angle P, roll angle R, gyro X-axis angular velocity, gyro Y-axis angular velocity, gyro Z-axis angular velocity, gyro X-axis angular acceleration, gyro Y-axis angular acceleration, gyro Z-axis angular acceleration, solar visibility (state quantity), and solar incidence angle, etc.

[0042] Space environment data may include time series data such as geomagnetic AP, atmospheric damping coefficient CD, solar radiation pressure coefficient CR, solar activity index F10.7, and mean solar radiation flux. Mission planning data may include the basic energy consumption required for the target satellite to perform payload mission planning, orbital maneuver energy consumption, and mission planning energy consumption.

[0043] Specifically, when it is necessary to predict the energy of a target satellite, the first step is to obtain the satellite's status data and mission plan data, and then perform subsequent steps to make the prediction.

[0044] Step 104: Based on the state data, call the pre-trained hybrid prediction model to obtain the predicted SOC sequence of the target satellite in the prediction period.

[0045] The hybrid prediction model is a pre-built and trained deep learning model combining ARIMA (Autoregressive Integrated Moving Average), CNN (Convolutional Neural Network), and LSTM (Long Short-Term Memory Network). This model is trained and built based on ARIMA, CNN, and LSTM. The predicted SOC sequence is the SOC sequence predicted by this model based on the target satellite's state data.

[0046] In one scenario, before processing the pre-trained hybrid prediction model based on the state data, the time series data in the aforementioned state data can be deduplicated, have missing values ​​and outliers processed, and then normalized to unify all data to the same timestamp. For example, missing values ​​and outliers can be handled using K-Nearest Neighbor (KNN) multivariate chain equation interpolation, and normalization can be performed using Min-Max Scaling (a linear transformation method).

[0047] Then, based on the time sequence, the various time series of the state data are sampled to obtain the input data for the hybrid prediction model. For example, sampling can be performed every minute. For each sampling time point, the parameter values ​​corresponding to the time series of the aforementioned state data can be extracted to obtain a multivariate time series data, which serves as the input data for the model. This multivariate time series data can include N parameter values ​​at T time points (i.e., each time point corresponds to N parameter values, such as SOC, voltage, current, temperature, pressure, etc., for each time point). This standardizes the model input, and after processing the input data, the model outputs the predicted SOC sequence of the target satellite within the prediction period. This sequence includes the SOC value of the target satellite at each time point within the prediction period (e.g., minutes, hours, days, weeks, or months). The SOC value is used to characterize the proportion of the battery's remaining charge to its nominal capacity at the corresponding time point.

[0048] Step 106: Determine the remaining battery capacity for the predicted cycle based on the predicted SOC sequence.

[0049] Since the State of Charge (SOC) value represents the proportion of the battery's remaining capacity to its nominal capacity at a given time point, the remaining battery capacity at that time point can be calculated based on the SOC value at each time point in the predicted SOC sequence and the battery's nominal capacity. For example, if the SOC value at time t is... The battery's nominal capacity is Then the remaining battery charge at time t .

[0050] Step 108: Calculate the energy budget value for payload mission planning based on the remaining battery power and mission plan data.

[0051] The energy budget value represents the energy constraints that can be used for payload mission planning. Mission plan data can include the basic energy consumption required for the target satellite to perform payload mission planning (i.e., the energy consumption required to maintain the basic operation of the satellite within the mission plan) and orbital maneuver energy consumption (representing the energy consumption required for orbit maintenance and collision avoidance maneuvers within the mission plan). Specifically, after completing the prediction of the SOC sequence for each prediction period, the energy budget value used for payload mission planning can be calculated based on the corresponding remaining battery power and mission plan data.

[0052] For example, a safety margin can first be determined for the target satellite's payload mission planning. This safety margin characterizes the energy consumption required for unpredictable events within the mission plan. Unpredictable events such as sudden collision avoidance and abnormal component power consumption will generate additional power consumption. Sudden collision avoidance maneuvers may additionally increase attitude control power consumption. Common component anomalies include heater failure, reaction wheel stall or saturation (magnetic torque compensation), safe mode entry, and communication link anomaly retry. This additional power consumption is difficult to calculate precisely and can be considered as adjustable power to cope with anomalies, but it does not represent daily consumption. Therefore, it is usually necessary to cover it by reserving a safety margin. The safety margin is usually a certain percentage of the basic energy consumption. For example, the safety margin can be the product of the basic energy consumption and a set weighting coefficient. .in, This represents the basic energy consumption, with 0.2 being a set weighting coefficient. It is understood that the weighting coefficient can be configured according to the actual scenario, and this embodiment does not impose any limitations on it.

[0053] Then, obtain the sum of the reserved safety margin, basic energy consumption, and orbital maneuver energy consumption, and calculate the difference between the remaining battery power and this sum. The difference represents the energy budget value used for payload mission planning.

[0054] For example, energy budget value ).in, This indicates the remaining battery capacity, determined based on the predicted SOC value at the corresponding time point. Indicates basic energy consumption. Indicates energy consumption for orbital maneuvers. This indicates a safety margin reserved. Through dynamic energy budgeting, it can support intelligent energy management and energy-constrained payload mission planning for satellites, thereby ensuring long-term reliable operation and maximizing service efficiency.

[0055] The aforementioned satellite energy prediction method acquires the target satellite's state of energy and mission plan data. Based on the state data, a pre-trained hybrid prediction model is invoked to obtain the predicted State of Charge (SOC) sequence for the target satellite during the prediction period. The remaining battery power for the prediction period is determined based on the predicted SOC sequence. Finally, based on the remaining battery power and mission plan data, an energy budget value for payload mission planning is calculated. Because this hybrid prediction model integrates ARIMA, CNN, and LSTM, it leverages the advantages of different algorithms through structured division of labor and collaboration. This allows for accurate prediction of the satellite's energy state and budget in complex and dynamic environments, providing a reliable and forward-looking decision-making basis for intelligent satellite energy management. This ensures long-term reliable satellite operation and maximizes service efficiency.

[0056] In one exemplary embodiment, such as Figure 2As shown, the hybrid prediction model in step 104 can be trained using the following method, the training process of which may specifically include:

[0057] Step 202: Obtain sample data, preprocess the sample data to obtain multiple sample sequences.

[0058] The sample data includes historical data on satellites, such as energy sequences, orbital and attitude sequences, and space environment data from different historical periods. For example, the historical state data of a satellite may include time series of data such as battery SOC, charge-discharge ratio, current-discharge ratio, bus voltage, main error voltage, bus current, discharge capacity, charge capacity, discharge current, charging current, load current, array current, battery temperature, battery pressure, solar panel current, solar panel voltage, solar panel output power, solar panel temperature, position X component, position Y component, position Z component, velocity X component, velocity Y component, velocity Z component, yaw angle Y, pitch angle P, roll angle R, gyro X-axis angular velocity, gyro Y-axis angular velocity, gyro Z-axis angular velocity, gyro X-axis angular acceleration, gyro Y-axis angular acceleration, gyro Z-axis angular acceleration, solar visibility (state quantity), solar incidence angle, geomagnetic AP, atmospheric damping coefficient CD, solar radiation pressure coefficient CR, solar activity index F10.7, mean solar radiation flux, payload mission planned energy consumption, orbital maneuver planned energy consumption, and basic energy consumption of the satellite platform.

[0059] Specifically, by defining a multivariate time series data dictionary structure, where one timestamp corresponds to multiple parameter values, and based on the aforementioned data preprocessing, multiple sample sequences are obtained. For example, data preprocessing may include:

[0060] Step 1, Time Alignment. Input the multivariate raw time series and read it into memory according to the dictionary structure.

[0061] Step 2, handle outliers. For example, you can use... Statistical methods are used to detect and process outliers.

[0062] Step 3: Fill in the missing values. For example, use K-nearest neighbor (KNN) multivariate chain equation interpolation.

[0063] Step 4: Create derived features. These include, but are not limited to, time-domain features, power-related features, orbit-related features, Fourier transform features (to extract periodic information), and hysteresis features (for ARIMA components).

[0064] Step 5: Data Standardization and Normalization. Different standardization strategies are used for different statistical characteristics of the original sequences. If the data distribution is unknown or uniform, Min-Max is used; if the data is approximately normal, Z-Score is used; if there are many outliers, Robust is used.

[0065] Step 6: Construct supervised learning samples, i.e., construct sample sequences.

[0066] If the original data contains a total of time points L s There are N features (1 point per minute) at each time point (e.g., N parameter values ​​such as SOC, voltage, current, temperature, and pressure). The sliding window length is L. win At each time point, the window size is N×L win (Number of window features), prediction step size is L pre Sliding step size L sli For each time point t, take [tL] win The features of the interval [t-1] are taken as input, and the features of the interval [t,t+L] are taken as input. pre The SOC value in the interval [tL] is used as the output label. Each sample covers the interval [tL]. win ,t+L pre -1], generating a total of time-overlapping samples floor((L s -L win -L pre ) / L sli )+1. When L sli When =1, the maximum number of samples (L) can be obtained. s -L win -L pre +1) items.

[0067] During training, the input to the hybrid model is a batch of sample sequences of batch_size, and the predicted value (fit value) and label value of each sample sequence are used as the input to the loss function.

[0068] For example, suppose the historical SOC sequence is 1000 in length (1000 points from 0 to 999 on the timeline). If the sliding window size is 100, the prediction step size is 10, and the sliding step size is 1 (a sliding step size of 1 allows for the acquisition of the most overlapping samples), then starting from the beginning of the time series (point 0), the SOCs of the first 100 time points (0-99) are taken as the first input sample sequence, and the predicted values ​​for time points 100-109 are used as the labels (i.e., outputs) for the corresponding sample sequences. Then, the window slides by 1 step (i.e., 1 time point), taking the SOCs of time points 1-100 as the second input sample sequence, and the predicted values ​​for time points 101-110 as the labels for the corresponding sample sequences. This process continues until the end of the window reaches the end of the time series, resulting in the following sample sequence:

[0069] The first sample input is [0:99], and the output is [100:109].

[0070] The second sample input is [1:100], and the output is [101:110].

[0071] The third sample input is [2:101], and the output is [102:111].

[0072] ...

[0073] The 891st sample input is [890:989], and the output is [990:999].

[0074] Since we need to predict 10 future time points, the starting position of the last window should be the (1000-100-10+1)=891st time point, i.e., the 891st time point. In this way, a total of 891 sample sequences can be obtained.

[0075] In one scenario, when there are a sufficient number of sample sequences, to achieve better model training results, the sample sequences can be divided into training, validation, and test sets according to time sequence. When the number of samples reaches millions, the ratio of training, validation, and test sets can be 98:1:1.

[0076] Step 204: For each sample sequence, process it using the initial ARIMA module to obtain the predicted sequence of the corresponding sample sequence.

[0077] The sample sequence includes historical SOC samples. The prediction sequence includes a linear prediction subsequence and a white noise residual subsequence. In this embodiment, the model parameters of the initial ARIMA module can be determined based on the historical SOC samples in the sample sequence, and a historical SOC stationary sequence can be obtained. Based on the initial ARIMA module and the corresponding model parameters, the predictable linear trend part and the unpredictable residual part are separated from the historical SOC stationary sequence. If the residual part is detected as white noise, the linear trend part is determined as the linear prediction subsequence, and the residual part is determined as the white noise residual subsequence, thus obtaining the prediction sequence for the corresponding sample sequence. If the residual part is detected as not being white noise, the steps of determining the model parameters of the initial ARIMA module based on the historical SOC samples in the sample sequence and obtaining the historical SOC stationary sequence are returned. This achieves the training of the ARIMA module.

[0078] For example, the training process for ARIMA may specifically include:

[0079] Step 1: Model Initialization and Parameter Selection. Input historical SOC samples from the training set. Use the `auto_arima` toolkit to automatically determine the model parameters (p, d, q) and output a stationary SOC sequence. For example, the unit root test (ADF) can be used to check if the historical SOC sequence is stationary. If no unit root exists, the sequence is stationary and can be directly used for modeling. For instance, the `auto_arima` toolkit can be used to automatically determine the model parameters (3, 1, 2) and output a stationary historical SOC sequence. If a unit root exists, the sequence is non-stationary. In this case, d-order differencing is used to transform the non-stationary sequence into a stationary one, allowing ARIMA to better fit the data and output a stationary historical SOC sequence.

[0080] Step 2: Perform linear prediction using ARIMA. Obtain the ARIMA model from Python's statsmodels library and set the model parameters (p, d, q). Input the historical SOC stationary sequence obtained above, and output the predicted sequence for the next h steps. . For linear prediction subsequences, The white noise residual subsequence is a pure random process with zero mean, constant variance, and no autocorrelation, representing unpredictable random perturbations. Whether it is white noise can be tested using the Ljung-Box test. The presence of white noise indicates that all predictable information has been extracted; therefore, the predicted sequence obtained above is determined as the predicted sequence for the corresponding sample sequence. If... If it is not white noise, it means that not all predictable information has been fully extracted, and we need to go back to step 1.

[0081] Step 3: Process the prediction results. Iterate through all sample sequences and generate ARIMA predictions (predicted mean, predicted standard deviation, first-step predicted value, and last-step predicted value) as features for each sequence. This completes the pre-training of the ARIMA module.

[0082] Step 206: Based on the sample sequence and the white noise residual subsequence, the initial CNN module is used for processing to obtain the local feature subsequence.

[0083] The input to the initial CNN module includes the aforementioned white noise residual subsequence. The input sequence consists of sample sequences from the training set (excluding SOC). The initial CNN module converts the input sequence into a fixed-length vector sequence, which is then extracted through multiple convolutional layers. Specific local temporal patterns can be captured, and local patterns at different scales can be obtained by adjusting the size and number of convolution kernels to obtain local feature subsequences.

[0084] For example, the initial CNN module may include three one-dimensional convolutional layers at different scales, operating in parallel. For instance, the scales of the three convolutional layers could be (10, 16), (30, 32), and (60, 48), with prediction strides of 10, 30, and 60, and the number of channels being 16, 32, and 48, respectively. The initial CNN module treats the multivariate input sequence as multiple one-dimensional "channels" and performs convolution operations through the one-dimensional convolutional layers. Each convolutional kernel is responsible for extracting a specific local temporal pattern. Subsequently, pooling layers compress the features, preserving salient information. The three convolutional layers at different scales operate in parallel to capture patterns at different temporal scales. The feature map sliding stride is set to strides=1, and padding is set to 'same' to maintain the feature map size. The pooling layer, MaxPooling1D (pool_size=2), is used to reduce dimensionality and enhance the translation invariance of the features. The optimizer Adam has a short-term prediction gradient decay of beta1=0.9, a long-term prediction gradient decay of beta2=0.999, and a weight_decay=1e-4. It works in conjunction with cosine annealing or the ReduceLROnPlateau scheduler to accelerate the convergence process and output local feature subsequences.

[0085] Step 208: Based on the local feature subsequence, use the initial LSTM module to perform nonlinear prediction to obtain the nonlinear prediction subsequence.

[0086] Specifically, the local feature subsequences extracted by the CNN are input into the LSTM layer. The output of the last time step is passed through a fully connected layer and mapped to a non-linear prediction subsequence for the next h steps. .

[0087] In this embodiment, training the LSTM module may specifically include:

[0088] Step 1, Model Architecture. For example, the LSTM model has an input sequence batch size of 64, a prediction step size of 60 (default), a hidden layer size of 256, 2 LSTM layers (num_layers), batch_first=True, bidirectional=True, and a dropout layer probability of 0.3. The optimizer Adam has short-term prediction gradient decay beta1=0.9, long-term prediction gradient decay beta2=0.999, and a gradient clipping threshold of 1.03. The LSTM model sets return_sequences=True for the first layer and return_sequences=False for the last layer, thus outputting a non-linear prediction subsequence for the next h steps. .

[0089] Step 2, Forward Prediction. From the CNN path to the LSTM path, an attention mechanism can be optionally applied, and the output of the last time step is taken as the LSTM feature.

[0090] Step 3: Calculate the hybrid loss. Set the main loss weights (MSE+MAE), Huber loss weights, and physical constraint loss weights. Physical constraint losses include SOC boundary [0,1] constraints, smoothness constraints (first-order difference), and charging rate constraints (the maximum charging rate per minute should conform to battery characteristics). Calculate the hybrid loss based on the weight allocation.

[0091] Step 4, iterative training. This stage involves 100 training epochs (stage1_epochs=100), training the ARIMA and CNN-LSTM models independently. The data is then retrieved iteratively. The training set contains sample sequences (excluding SOC). In each training round, forward propagation, loss calculation, backpropagation (updating model parameters), gradient pruning (to prevent gradient explosion), learning rate scheduling (adjusted every 100 batches), cosine annealing learning rate scheduling, and calculation of various evaluation metrics (such as Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Mean Absolute Percentage Error (MAPE) are used. The results of the evaluation metric calculations and their applications are then analyzed. Monitor the situation and adjust and optimize model parameters as needed. Record and save historical training loss values, MSE values, learning rates, etc., to complete the training of the LSTM module.

[0092] Step 210: The initial feature fusion module is used to fuse the white noise residual subsequence, the linear prediction subsequence, and the nonlinear prediction subsequence to obtain the feature fusion sequence.

[0093] The feature fusion sequence can be represented by the following formula:

[0094]

[0095] in, Represents a linear prediction subsequence. Represents a white noise residual subsequence. Represents a nonlinear prediction subsequence. Represents the feature fusion sequence, and This represents the fusion weights, which can be dynamically adjusted based on the data fitting results to ensure that the fitted value best approximates the label value. .

[0096] Step 212: Based on the feature fusion sequence, determine the predicted SOC sample sequence of the corresponding sample sequence.

[0097] Specifically, based on the above feature fusion sequence The weight matrix is ​​fine-tuned through a fully connected layer to better fit the actual data distribution and optimize the model's predictive performance. Finally, the predicted SOC values ​​for the next h steps are output, yielding the predicted SOC sample sequence for the corresponding sample sequence.

[0098] Step 214: Perform joint fine-tuning on each module to obtain the trained hybrid prediction model.

[0099] Specifically, the initial ARIMA module, initial CNN module, initial LSTM module, and initial feature fusion module can be jointly fine-tuned based on the predicted SOC sample sequence and pre-configured training parameters to obtain the trained hybrid prediction model.

[0100] Understandably, based on the above adjustments, a complete ARIMA-CNN-LSTM hybrid model can be created. Then, all layers can be unfrozen, training parameters configured, and the model can be trained iteratively using a smaller learning rate to fine-tune and optimize the parameters. For example, the main parameter configuration for this training stage might include: data parameters {'lookback':120,'forecast_horizon':10,'batch_size':64,'train_val_split':0.98}, model parameters {'input_dim':48,'hidden_dim':256,'num_lstm_layers':2,'dropout_rate':0.3,'time_steps':60}, training parameters {'stage1_epochs':100, 'stage2_epochs':50, 'learning_rate':1e-3,'weight_decay':1e-4,'h':1}, and loss function weights {'mse_weight':0.7, The training parameters are set as follows: `mae_weight`: 0.2, `physical_weight`: 0.1; `early_stopping_patience`: 20, `min_delta`: 1e-3; `lr_scheduler`: cosine, `warmup_epochs`: 5; and `online_learning`: True, `adaptation_frequency`: 100, `online_learning_rate`: 1e-5. The training parameters are then adjusted based on the model evaluation results. Finally, multi-timescale joint prediction is performed, sequentially predicting short time ranges (1 hour, h=1), medium time ranges (3 hours, h=3), and long time ranges (24 hours, h=24), outputting the final SOC prediction value.

[0101] In one scenario, a joint fine-tuning strategy using phased training can be employed. For example, ARIMA can be trained separately to obtain residuals, and then the number of input channels for the CNN-LSTM can be fixed. Model optimization can use a two-stage learning rate: a smaller learning rate (1e-4) for the ARIMA stage and a larger learning rate (1e-3) for the CNN-LSTM stage. Rolling prediction can be used, first allowing the model to learn to predict the short-term future, and then gradually increasing the prediction step size by h steps to achieve joint fine-tuning of the model.

[0102] In one scenario, a joint fine-tuning strategy combining training and optimization can be employed. For example, the loss between the predicted and labeled values ​​can be minimized using the backpropagation algorithm, with Huber Loss as the loss function. Since Huber Loss is less sensitive to outliers than the mean squared error (MSE), it is more robust to satellite telemetry data, thus ensuring differentiability while reducing the impact of outliers.

[0103] In one scenario, a joint fine-tuning strategy using hyperparameter search can also be employed. For example, Bayesian optimization or grid search can be used to focus on optimizing the combination of LSTM unit number and Dropout probability.

[0104] In one scenario, a two-level update mechanism can be established. For example, short-term online fine-tuning requires deploying an online update module. Upon receiving new satellite telemetry data, the weights of the CNN-LSTM part in the hybrid model are updated online using gradient descent, while the ARIMA part is periodically refitted, and the fusion weights are dynamically adjusted to respond to short-term changes. When the error of the hybrid model exceeds a threshold for multiple consecutive prediction periods, retraining the structure and optimizing parameters is triggered to address long-term drift. When selecting training data, the weights for data from the last three months are set to 0.7, while the weights for data from three months ago are set to 0.3. Training is then restarted from the old model parameters to accelerate convergence. This two-level update mechanism balances computational efficiency and model performance, enabling the prediction model to continuously adapt to changes in satellite performance and achieve "self-evolution."

[0105] In an exemplary embodiment, after jointly fine-tuning the initial ARIMA module, initial CNN module, initial LSTM module, and initial feature fusion module, the method may further include: determining a validation sequence from multiple sample sequences; obtaining the prediction loss of the jointly fine-tuned model based on the validation sequence; performing multiple forward propagations on the validation sequence using Monte Carlo Dropout to obtain an uncertainty quantification estimate; wherein the uncertainty quantification estimate includes the mean, standard deviation, and confidence interval determined based on the propagation results of multiple forward propagations; if the jointly fine-tuned model meets the expected performance based on the prediction loss and uncertainty quantification estimate, determining the jointly fine-tuned model as the trained hybrid prediction model; if the jointly fine-tuned model does not meet the expected performance based on the prediction loss and uncertainty quantification estimate, returning to the step of processing each sample sequence using the initial ARIMA module to obtain the prediction sequence for the corresponding sample sequence to train the model until the model meets the expected performance, thereby obtaining the final trained model. This achieves model validation and evaluation, improving the model's accuracy and robustness.

[0106] For example, the above model validation and evaluation may specifically include: First, forward chain identical distribution validation can be used based on sample sequences in the training set, test set, and validation set. That is, the model is run on each of the three datasets, and the prediction loss of key indicators is calculated, which may specifically include the root mean square loss (RMSE) and the mean absolute error loss (MAE), thereby validating the model's performance under different data distributions.

[0107] Then, all Dropout layers and Monte Carlo Dropout predictions are enabled, and multiple forward propagations are performed to obtain an uncertainty quantification estimate. Monte Carlo Dropout applies Dropout to the inference stage (i.e., Dropout is not disabled), thereby achieving Bayesian approximate inference, ensuring that each forward propagation produces slightly different prediction results due to the randomness of Dropout. After multiple repetitions (e.g., N=100 times), a set of predicted values ​​is obtained. By calculating the mean, standard deviation, and confidence interval of the predicted values, an uncertainty quantification estimate is obtained.

[0108] Then, by using the aforementioned quantitative estimation of prediction loss and uncertainty, it is determined whether the current model meets the expected performance. If it does, the current model after joint fine-tuning is determined to be the trained hybrid prediction model. If it does not meet the expectations, the model learning rate, Dropout rate, and fusion weights are adjusted. and Then, return to the training phase to retrain and adjust until the expected performance is met, and obtain the final model.

[0109] In one scenario, all historical models from the training process can be saved for subsequent data analysis or learning purposes.

[0110] In one scenario, an online learning buffer (maximum 2 hours of data) can be established to receive real-time data. After data preprocessing, the data is input into the hybrid prediction model for prediction, outputting the predicted SOC value for the next hour. The model is then trained online based on the actual SOC value for the next hour to improve its robustness. Alternatively, real-world data can be collected for online adaptive learning. For example, the hybrid model can be updated every 10 days of data collected. Mini-batch training data can be created, and gradient updates and iterations can be performed using a small learning rate. Based on the loss evaluation results, the model parameters are updated online.

[0111] In one exemplary embodiment, and in combination with Figure 3 To further explain the above-mentioned satellite energy prediction method, such as... Figure 3 As shown, it can specifically include:

[0112] 1) Data Acquisition and Preprocessing Stage

[0113] Specifically, target satellite status data and mission plan data can be collected. Status data may include the target satellite's current energy sequence, orbital attitude sequence, and space environment data. Preprocessing may include deduplication, handling of missing and outlier values, and normalization to obtain the corresponding input dataset (such as a multivariate sequence). For details, please refer to the detailed description in the above embodiments; this will not be repeated in this embodiment.

[0114] 2) ARIMA linear prediction stage

[0115] Then, based on the SOC sequences in the input dataset, the ARIMA model is used to capture the linear trend and short-term dependence in the data to obtain the linear prediction sequence. .in, For linear prediction subsequences, It is a white noise residual subsequence.

[0116] 3) CNN-LSTM nonlinear prediction stage

[0117] Based on white noise residual subsequence For multivariate sequences, CNNs are used to extract local features and patterns from time series data. LSTM time series modeling is employed to learn... The long-term dependencies in the sequence are identified, and nonlinear predictions are performed to obtain the nonlinear prediction subsequence. .

[0118] Furthermore, the nonlinear prediction subsequence is then processed through a feature fusion layer. , white noise residual subsequence and linear prediction subsequence Perform fusion processing to obtain the feature fusion sequence. .

[0119] 4) Energy Budgeting Phase

[0120] Specifically, the feature fusion sequence can be processed through a fully connected layer. Fine-tuning the weight matrix is ​​performed to better fit the actual data distribution and optimize the model's predictive performance, ultimately outputting the predicted SOC sequence for the next h steps.

[0121] After the fully connected layer outputs the final predicted sequence It can perform dynamic energy budgeting, providing support for intelligent energy management of satellites and energy-constrained payload mission planning, ensuring long-term reliable operation of satellites and maximizing service efficiency.

[0122] For example, several variable sequences can be selected based on the application scenario and computing device performance to perform short-term (minutes / hours) and long-term (days / weeks) predictions, and the output prediction results can be... It is the core data for satellite energy intelligent decision-making.

[0123] For example, short-term forecasts can be based on the current SOC and an accurate orbital model to predict the next orbital period. The input channels of a CNN include The weighted fusion includes time series such as bus voltage, main error voltage, bus current, discharge capacity, charging capacity, discharge current, charging current, load current, array current, battery temperature, battery pressure, solar panel current, solar panel voltage, solar panel output power, solar panel temperature, position X component, position Y component, position Z component, velocity X component, velocity Y component, velocity Z component, yaw angle Y, pitch angle P, roll angle R, gyro X-axis angular velocity, gyro Y-axis angular velocity, gyro Z-axis angular velocity, gyro X-axis angular acceleration, gyro Y-axis angular acceleration, gyro Z-axis angular acceleration, solar visibility, solar incidence angle, payload mission planned energy consumption, orbital maneuver planned energy consumption, and satellite basic energy consumption. The default settings are as follows during weighted fusion. and .

[0124] Long-term forecasts focus on orbital changes over the next few days or weeks and the energy balance trend under planned tasks (charge-discharge ratio curve). The input channels of the CNN include... The time series includes the following parameters from a multivariate sequence: charge / discharge ratio, current charge / discharge ratio, main error voltage, solar panel output power, position X component, position Y component, position Z component, velocity X component, velocity Y component, velocity Z component, yaw angle Y, pitch angle P, roll angle R, solar visibility, solar incidence angle, geomagnetic AP, atmospheric damping coefficient CD, solar radiation pressure coefficient CR, solar activity index F10.7, mean solar radiation flux, payload mission planned energy consumption, orbital maneuver planned energy consumption, and satellite basic energy consumption. The default settings are used during weighted fusion. and .

[0125] After each forecast cycle is completed, a dynamic energy budget value can be generated for each satellite. Energy constraints used for payload mission planning (less than) For example, energy budget values It can be calculated using the following formula:

[0126] );

[0127] in, This indicates the remaining battery capacity, determined based on the predicted SOC value at the corresponding time point. Indicates basic energy consumption. Indicates energy consumption for orbital maneuvers. This indicates that a safety margin has been reserved.

[0128] The above embodiments are based on signal decomposition and ensemble learning theory, which transforms complex satellite energy multivariate sequences. Decomposed into .in, It is a linear component, fitted by the ARIMA model; The nonlinear residual components are learned by the CNN-LSTM network and output as predictions. The two modules were optimized separately, and the results were then linearly weighted and fused for output. This allows the model to approximate real-world energy dynamics with greater efficiency and learning capabilities. Through the complementary advantages of hybrid models, compared to single or physical models, its RMSE and MAPE are significantly reduced, resulting in a substantial improvement in prediction accuracy. It exhibits stronger adaptability to scenarios such as satellite performance degradation and abrupt changes in the space environment, demonstrating high prediction stability. Furthermore, it can provide high-precision predictions for the next few hours to days, saving valuable decision-making time for satellite mission planning and energy security management, demonstrating high foresight and practicality.

[0129] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages in other steps. It is understood that the steps in different embodiments can be freely combined as needed, and all non-contradictory solutions formed by such combinations are within the scope of protection of this application.

[0130] Based on the same inventive concept, this application also provides a satellite energy prediction device for implementing the satellite energy prediction method described above. The solution provided by this device is similar to the solution described in the above method; therefore, the specific limitations in one or more satellite energy prediction device embodiments provided below can be found in the limitations of the satellite energy prediction method described above, and will not be repeated here.

[0131] In one exemplary embodiment, such as Figure 4 As shown, a satellite energy prediction device is provided, including: a data acquisition module 402, a SOC prediction module 404, a determination module 406, and an energy budget module 408, wherein:

[0132] Data acquisition module 402 is used to acquire the status data and mission plan data of the target satellite. The status data includes the current energy sequence, orbital attitude sequence and space environment data of the target satellite.

[0133] SOC prediction module 404 is used to call a pre-trained hybrid prediction model based on the state data to obtain the predicted SOC sequence of the target satellite in the prediction period. The hybrid prediction model is a pre-built and trained ARIMA-CNN-LSTM hybrid deep learning model.

[0134] The determination module 406 is used to determine the remaining battery power for the prediction period based on the predicted SOC sequence.

[0135] The energy budget module 408 is used to calculate the energy budget value for payload task planning based on the remaining battery power and the task plan data.

[0136] In an exemplary embodiment, the apparatus further includes a model training module, configured to: acquire sample data; preprocess the sample data to obtain multiple sample sequences; the sample data includes historical satellite state data; for each sample sequence, process it using an initial ARIMA module to obtain a prediction sequence for the corresponding sample sequence, the prediction sequence including a linear prediction subsequence and a white noise residual subsequence; based on the sample sequence and the white noise residual subsequence, process it using an initial CNN module to obtain a local feature subsequence; based on the local feature subsequence, perform nonlinear prediction using an initial LSTM module to obtain a nonlinear prediction subsequence; fuse the white noise residual subsequence, the linear prediction subsequence, and the nonlinear prediction subsequence using an initial feature fusion module to obtain a feature fusion sequence; based on the feature fusion sequence, determine a predicted SOC sample sequence corresponding to the sample sequence; and jointly fine-tune the initial ARIMA module, the initial CNN module, the initial LSTM module, and the initial feature fusion module according to the predicted SOC sample sequence and pre-configured training parameters to obtain a trained hybrid prediction model.

[0137] In an exemplary embodiment, the sample sequence includes historical SOC samples; the model training module is further configured to: determine the model parameters of the initial ARIMA module based on the historical SOC samples in the sample sequence, and obtain a historical SOC stationary sequence; based on the initial ARIMA module and the corresponding model parameters, separate a predictable linear trend portion and an unpredictable residual portion from the historical SOC stationary sequence; if the residual portion is detected to be white noise, determine the linear trend portion as a linear prediction subsequence, determine the residual portion as a white noise residual subsequence, and obtain a prediction sequence for the corresponding sample sequence.

[0138] In an exemplary embodiment, the model training module is further configured to: if the residual part is detected to be non-white noise, return to the step of determining the model parameters of the initial ARIMA module based on historical SOC samples in the sample sequence and obtaining the historical SOC stationary sequence.

[0139] In an exemplary embodiment, the feature fusion sequence is represented by the following formula: ,in, Represents a linear prediction subsequence. Represents a white noise residual subsequence. Represents a nonlinear prediction subsequence. Represents the feature fusion sequence, and Indicates the fusion weight, and .

[0140] In an exemplary embodiment, the model training module is further configured to: determine a validation sequence from the plurality of sample sequences; obtain the prediction loss of the jointly fine-tuned model based on the validation sequence; perform multiple forward propagations on the validation sequence based on Monte Carlo Dropout to obtain an uncertainty quantification estimate; the uncertainty quantification estimate includes the mean, standard deviation, and confidence interval determined based on the propagation results of multiple forward propagations; if the jointly fine-tuned model meets the expected performance based on the prediction loss and the uncertainty quantification estimate, determine the jointly fine-tuned model as the trained hybrid prediction model; if the jointly fine-tuned model does not meet the expected performance based on the prediction loss and the uncertainty quantification estimate, return to the step of processing each sample sequence using the initial ARIMA module to obtain the prediction sequence of the corresponding sample sequence.

[0141] In an exemplary embodiment, the mission planning data includes the basic energy consumption and orbital maneuvering energy consumption required for the target satellite to perform payload mission planning; the energy budget module is further configured to: determine a reserved safety margin for the target satellite to perform payload mission planning, the reserved safety margin representing the energy consumption required for unpredictable events; obtain the sum of the reserved safety margin, the basic energy consumption, and the orbital maneuvering energy consumption; and calculate the difference between the remaining battery power and the sum, the difference representing the energy budget value used for payload mission planning.

[0142] In an exemplary embodiment, the energy budget module is further configured to: obtain the product of the basic energy consumption and the set weight coefficient, and determine the product as the reserved safety margin for the target satellite to perform payload mission planning.

[0143] Each module in the aforementioned satellite energy prediction device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0144] In one exemplary embodiment, a computer device is provided, the internal structure of which can be as shown in the figure. Figure 5As shown, the computer device includes a processor, memory, input / output interfaces, a communication interface, a display unit, and an input device. The processor, memory, and input / output interfaces are connected via a system bus, and the communication interface, display unit, and input device are also connected to the system bus via the input / output interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The input / output interfaces are used for exchanging information between the processor and external devices. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, Near Field Communication (NFC), or other technologies. When the computer program is executed by the processor, it implements a satellite energy prediction method. The display unit is used to form a visually visible image and can be a display screen, a projection device, or a virtual reality imaging device. The display screen can be an LCD screen or an e-ink screen. The input device of the computer device can be a touch layer covering the display screen, or buttons, trackballs, or touchpads set on the casing of the computer device, or external keyboards, touchpads, or mice, etc.

[0145] Those skilled in the art will understand that Figure 5 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0146] In one exemplary embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0147] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0148] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0149] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0150] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0151] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0152] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A satellite energy prediction method, characterized in that, The method includes: Acquire the target satellite's status data and mission plan data, wherein the status data includes the target satellite's current energy sequence, orbital attitude sequence, and space environment data; Based on the state data, a pre-trained hybrid prediction model is invoked to obtain the predicted SOC sequence of the target satellite in the prediction period. The hybrid prediction model is a pre-built and trained ARIMA-CNN-LSTM hybrid deep learning model. The remaining battery charge for the prediction period is determined based on the predicted SOC sequence. Based on the remaining battery power and the mission plan data, calculate the energy budget value for payload mission planning; The hybrid prediction model was trained using the following method: Acquire sample data, preprocess the sample data to obtain multiple sample sequences; the sample data includes historical satellite status data; For each of the sample sequences, the initial ARIMA module is used for processing to obtain the prediction sequence of the corresponding sample sequence. The prediction sequence includes a linear prediction subsequence and a white noise residual subsequence. Based on the sample sequence and the white noise residual subsequence, the initial CNN module is used for processing to obtain the local feature subsequence; Based on the local feature subsequence, nonlinear prediction is performed using the initial LSTM module to obtain the nonlinear prediction subsequence; The white noise residual subsequence, the linear prediction subsequence, and the nonlinear prediction subsequence are fused using an initial feature fusion module to obtain a feature fusion sequence. Based on the feature fusion sequence, a predicted SOC sample sequence corresponding to the sample sequence is determined; Based on the predicted SOC sample sequence and the pre-configured training parameters, the initial ARIMA module, initial CNN module, initial LSTM module and initial feature fusion module are jointly fine-tuned to obtain the trained hybrid prediction model. The mission plan data includes the basic energy consumption and orbital maneuvering energy consumption required for the target satellite to perform payload mission planning; the step of calculating the energy budget value for payload mission planning based on the remaining battery power and the mission plan data includes: determining the reserved safety margin for the target satellite to perform payload mission planning, the reserved safety margin representing the energy consumption required for unpredictable events; obtaining the sum of the reserved safety margin, the basic energy consumption, and the orbital maneuvering energy consumption; and calculating the difference between the remaining battery power and the sum, the difference representing the energy budget value for payload mission planning.

2. The method according to claim 1, characterized in that, The sample sequence includes historical SOC samples; For each of the sample sequences, the initial ARIMA module is used for processing to obtain the predicted sequence for the corresponding sample sequence, including: Based on the historical SOC samples in the sample sequence, the model parameters of the initial ARIMA module are determined, and the historical SOC stationary sequence is obtained. Based on the initial ARIMA module and the corresponding model parameters, the predictable linear trend part and the unpredictable residual part are separated from the historical SOC stationary sequence. If the residual portion is detected to be white noise, the linear trend portion is determined as a linear prediction subsequence, and the residual portion is determined as a white noise residual subsequence, thus obtaining the prediction sequence of the corresponding sample sequence.

3. The method according to claim 2, characterized in that, The method further includes: If the residual is detected to be non-white noise, the process returns to the steps of determining the model parameters of the initial ARIMA module based on historical SOC samples in the sample sequence and obtaining historical SOC stationary sequences.

4. The method according to claim 1, characterized in that, The feature fusion sequence is represented by the following formula: in, Represents a linear prediction subsequence. Represents a white noise residual subsequence. Represents a nonlinear prediction subsequence. Represents the feature fusion sequence, and Indicates the fusion weight, and .

5. The method according to claim 1, characterized in that, After jointly fine-tuning the initial ARIMA module, initial CNN module, initial LSTM module, and initial feature fusion module, the method further includes: A validation sequence is determined from the plurality of sample sequences, and the prediction loss of the jointly fine-tuned model based on the validation sequence is obtained. The validation sequence is propagated multiple times using Monte Carlo Dropout to obtain an uncertainty quantification estimate; the uncertainty quantification estimate includes the mean, standard deviation, and confidence interval determined based on the propagation results of multiple forward propagations; If the model after joint fine-tuning meets the expected performance based on the predicted loss and the uncertainty quantification estimate, the model after joint fine-tuning is determined to be the trained hybrid prediction model. If the model after joint fine-tuning fails to meet the expected performance based on the predicted loss and the uncertainty quantification estimate, the process returns to the step of processing each sample sequence using the initial ARIMA module to obtain the predicted sequence for the corresponding sample sequence.

6. The method according to claim 1, characterized in that, The determination of the reserved safety margin for the target satellite's payload mission planning includes: Obtain the product of the basic energy consumption and the set weight coefficient, and determine the product as the reserved safety margin for the target satellite to perform payload mission planning.

7. A satellite energy prediction device, characterized in that, The device includes: The data acquisition module is used to acquire the status data and mission plan data of the target satellite. The status data includes the current energy sequence, orbital attitude sequence and space environment data of the target satellite. The SOC prediction module is used to call a pre-trained hybrid prediction model based on the state data to obtain the predicted SOC sequence of the target satellite in the prediction period. The hybrid prediction model is a pre-built and trained ARIMA-CNN-LSTM hybrid deep learning model. The determination module is used to determine the remaining battery power for the prediction period based on the predicted SOC sequence. The energy budget module is used to calculate the energy budget value for payload task planning based on the remaining battery power and the task plan data. The model training module is used to acquire sample data, preprocess the sample data to obtain multiple sample sequences, the sample data including historical satellite state data; for each sample sequence, the initial ARIMA module is used to process it to obtain a prediction sequence for the corresponding sample sequence, the prediction sequence including a linear prediction subsequence and a white noise residual subsequence; based on the sample sequence and the white noise residual subsequence, the initial CNN module is used to process it to obtain a local feature subsequence; based on the local feature subsequence, the initial LSTM module is used to perform nonlinear prediction to obtain a nonlinear prediction subsequence; the initial feature fusion module is used to fuse the white noise residual subsequence, the linear prediction subsequence, and the nonlinear prediction subsequence to obtain a feature fusion sequence; based on the feature fusion sequence, the predicted SOC sample sequence corresponding to the sample sequence is determined; based on the predicted SOC sample sequence and pre-configured training parameters, the initial ARIMA module, the initial CNN module, the initial LSTM module, and the initial feature fusion module are jointly fine-tuned to obtain a trained hybrid prediction model; The mission planning data includes the basic energy consumption and orbital maneuvering energy consumption required for the target satellite to perform payload mission planning; the energy budget module is further used to: determine the reserved safety margin for the target satellite to perform payload mission planning, the reserved safety margin representing the energy consumption required for unpredictable events; obtain the sum of the reserved safety margin, the basic energy consumption, and the orbital maneuvering energy consumption; calculate the difference between the remaining battery power and the sum, the difference representing the energy budget value used for payload mission planning.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images