A satellite communication-oriented traffic prediction method, device, equipment and medium

Abstract

The application relates to the technical field of low-orbit satellite communication, and discloses a satellite communication-oriented traffic prediction method, device, equipment and medium, which are applied to a satellite-borne system, wherein the method comprises the following steps: receiving a self-description model package sent by a ground terminal in a standby area, and keeping an old model currently running in an active area to execute an inference task; obtaining original traffic data in a preset sliding window, determining a data correction factor of an adaptive ground model based on the difference between the original traffic data and statistical information; loading the data correction factor to the ground model to obtain a corrected new model; and in response to a model switching instruction, updating a global active model pointer from a memory address pointing to the old model to a memory address pointing to the new model by atomic operation. The technical scheme provided by the application can realize seamless hot updating of a model from the ground to the satellite without interrupting real-time inference services of the satellite-borne system.

Description

Technical Field

[0001] This application relates to the field of low-Earth orbit satellite communication technology, and in particular to a traffic prediction method, apparatus, device and medium for satellite communication. Background Technology

[0002] As low-Earth orbit satellite constellations become critical communication infrastructure, the highly dynamic nature of their satellite-to-ground links makes traffic prediction a core prerequisite for resource scheduling. However, existing technologies face multiple bottlenecks: isolated modeling of multi-scale features leads to fragmented and distorted prediction results, failing to reconstruct the complex physical nature of traffic; rigid model deployment and updates require service interruptions, easily causing inference jumps due to differences in data distribution, resulting in lost communication opportunities; lightweight design adopts a passive slimming strategy, sacrificing feature extraction capabilities in highly dynamic scenarios, making it difficult to guarantee prediction accuracy; and the lack of decision-oriented uncertainty quantification prevents scheduling strategies from being flexibly adjusted based on prediction confidence, failing to drive intelligent resource decision-making and severely restricting the improvement of intelligent operation and maintenance levels of satellite networks.

[0003] Therefore, how to achieve seamless hot updates of models from the ground to the satellite without interrupting the real-time inference service of the satellite system, and how to solve the model prediction failure caused by the inconsistency between the distribution of ground training data and satellite real-time data are the technical problems that urgently need to be solved. Summary of the Invention

[0004] This application provides a traffic prediction method, apparatus, device, and medium for satellite communications, which solves the technical problem of achieving seamless hot updates of models from the ground to the satellite without interrupting the real-time inference service of the satellite system, and solves the technical problem of model prediction failure caused by the inconsistency between the distribution of ground training data and satellite real-time data.

[0005] To achieve the above objectives, the main technical solutions adopted in this application include: In a first aspect, embodiments of this application provide a traffic prediction method for satellite communications, applied to a spaceborne system, the method comprising: In the standby area, a self-describing model package sent from the ground terminal is received, while in the active area, the currently running old model is maintained to perform inference tasks; wherein, the self-describing model package includes a trained ground model and statistical information from the training dataset; Obtain raw flow data within a preset sliding window, and determine a data correction factor to adapt to the ground model based on the difference between the raw flow data and the statistical information; The data correction factor is loaded into the ground model to obtain a new corrected model, which is then used to perform correction and inference tasks on the acquired real-time traffic data. In response to the model switching command, the global active model pointer is updated from pointing to the memory address of the old model to pointing to the memory address of the new model through an atomic operation, so that the subsequently received real-time traffic data can be input into the new model to perform inference tasks and obtain predicted traffic.

[0006] In one implementation, the step of acquiring raw traffic data within a preset sliding window and determining a data correction factor adapted to the ground model based on the difference between the raw traffic data and the statistical information includes: Obtain the raw traffic data within a preset sliding window, and determine the current traffic mean and current traffic standard deviation based on the raw traffic data; Extract the minimum value, maximum value and mean value of the training set from the statistical information, and determine the first correction factor corresponding to the current traffic mean. Extract the training set standard deviation from the statistical information and determine the second correction factor corresponding to the current traffic standard deviation; The first correction factor and the second correction factor are determined as data correction factors to adapt to the ground model.

[0007] In one implementation, the step of correcting the acquired real-time traffic data using the new model includes: Extract the minimum and maximum values ​​of the training set from the statistical information; The acquired real-time traffic data is normalized using the minimum and maximum values ​​of the training set to obtain normalized data. The normalized data is linearly transformed and corrected according to the data correction factor to obtain the corrected data.

[0008] In one implementation, after performing correction and inference tasks on the acquired real-time traffic data using the new model, the method further includes: Obtain historical real traffic data within a preset time period, and correct the historical real traffic data based on the data correction factor to obtain corrected historical data; The new model is used to infer the corrected historical data to obtain the average predicted flow rate. Obtain the mean and standard deviation of the predicted traffic output by the old model for the historical real traffic data; The absolute value of the difference between the mean flow predicted by the new model and the mean flow predicted by the old model is determined, and if the absolute value of the difference is less than the product of a preset threshold and the standard deviation of the flow predicted by the old model, the new model is determined to have been successfully corrected.

[0009] In one implementation, after inputting subsequently received real-time traffic data into the new model to perform the inference task, the method further includes: The subsequently received real-time traffic data is input into the new model to perform inference tasks, so as to obtain the predicted traffic including the mean and standard deviation of the predicted traffic of the new model. The ratio between the standard deviation of the predicted flow rate of the new model and the mean of the predicted flow rate of the new model is determined as the coefficient of variation; The current confidence level is determined based on the mapping relationship between the coefficient of variation and the preset confidence level table; The satellite channel bandwidth is determined based on the bandwidth scheduling strategy corresponding to the confidence level.

[0010] In one implementation, determining the satellite channel bandwidth according to the bandwidth scheduling strategy corresponding to the confidence level includes: If the confidence level is high confidence, an aggressive bandwidth allocation strategy is executed, and the satellite channel bandwidth is determined by the positive superposition of the mean traffic predicted by the new model and the standard deviation of the traffic predicted by the new model by a specified multiple. If the confidence level is low, a conservative bandwidth allocation strategy is implemented, and the satellite channel bandwidth is determined based on the difference between the mean traffic predicted by the new model and the standard deviation of the traffic predicted by the new model.

[0011] Secondly, embodiments of this application provide a traffic prediction method for satellite communications, applied to a ground station, the method comprising: A training dataset containing long-term historical traffic data and short-term historical traffic data is obtained. The long-term historical traffic data is encoded and predicted using a first prediction network to obtain a trend prediction sequence. The trend prediction value corresponding to the short-term historical traffic data is extracted from the trend prediction sequence, and the short-term historical traffic data, the trend prediction value and the time feature are concatenated to construct a concatenated vector. The concatenated vector is input into the second prediction network for feature extraction, and the prediction mean and prediction variance are output respectively through the dual fully connected layer split at the end of the second prediction network. Based on the actual observed flow rate, the predicted mean, and the predicted variance, a loss function is constructed, and the parameters of the first prediction network and the second prediction network are jointly optimized using the loss function to obtain a trained ground model. The trained ground model and the statistical information in the training dataset are bound and packaged to form a self-describing model package, which is then sent to the spaceborne system.

[0012] Thirdly, embodiments of this application provide a traffic prediction device for satellite communications, applied to a spaceborne system, the device comprising: The model receiving unit is used to receive the self-describing model package sent by the ground terminal in the standby area, and to maintain the currently running old model to perform inference tasks in the active area; wherein, the self-describing model package includes the trained ground model and statistical information in the training dataset; The factor determination unit is used to acquire the raw flow data within a preset sliding window, and determine the data correction factor adapted to the ground model based on the difference between the raw flow data and the statistical information. The model correction unit is used to load the data correction factor into the ground model to obtain a new corrected model, so as to perform correction and inference tasks on the acquired real-time traffic data through the new model; The model switching unit is used to respond to the model switching command by updating the global active model pointer from the memory address pointing to the old model to the memory address pointing to the new model through atomic operations, so as to input the subsequently received real-time traffic data into the new model to perform inference tasks and obtain predicted traffic.

[0013] Fourthly, embodiments of this application provide a traffic prediction device for satellite communications, applied to a ground station, the device comprising: The first prediction unit is used to acquire a training dataset containing long-term historical traffic data and short-term historical traffic data, and to encode and predict the long-term historical traffic data using the first prediction network to obtain a trend prediction sequence. The data splicing unit is used to extract the trend prediction value corresponding to the short-term historical traffic data from the trend prediction sequence, and splice the short-term historical traffic data, the trend prediction value and the time feature to construct a splicing vector. The second prediction unit is used to input the spliced ​​vector into the second prediction network for feature extraction, and output the prediction mean and prediction variance respectively through the dual fully connected layer split at the end of the second prediction network. The model training unit is used to construct a loss function based on the actual observed flow rate, the predicted mean, and the predicted variance, and to jointly optimize the parameters of the first prediction network and the second prediction network using the loss function to obtain a trained ground model. The model sending unit is used to bind and package the trained ground model and the statistical information in the training dataset to form a self-describing model package and send it to the spaceborne system.

[0014] Fifthly, embodiments of this application provide a computer device, including: The system includes a memory and a processor, which are interconnected. The memory stores computer instructions, and the processor executes these computer instructions to perform the aforementioned traffic prediction method for satellite communications.

[0015] Sixthly, embodiments of this application provide a computer-readable storage medium storing computer instructions for causing a computer to execute the aforementioned traffic prediction method for satellite communication.

[0016] The technical solution provided by one or more embodiments of this application achieves a dual-zone isolation architecture by logically dividing the onboard system memory into an active zone and a backup zone. In the active zone, the old model continues to execute inference tasks to ensure zero interruption of satellite communication services. At the same time, in the backup zone, a self-describing model package containing the training model and training statistics sent by the ground end is received. Based on the difference between the original traffic data and statistics within a preset sliding window, a data correction factor is dynamically calculated. This correction factor is injected into the ground model to form a corrected new model. The new model can achieve seamless adaptation to the differences between the satellite and ground environments without modifying the weights through data distribution alignment. Finally, in response to the model switching command, the global active model pointer is instantly updated from the old model address to the new model address through atomic operations, so that subsequent real-time traffic data can be directly input into the new model to execute inference tasks. Thus, this embodiment achieves zero service interruption during the model update process, adaptive compensation for data distribution drift, and atomicity guarantee of the switching operation. It effectively solves the multiple constraints faced by model upgrades in the satellite environment, such as limited computing power preventing retraining, limited satellite-to-ground link bandwidth making frequent transmission difficult, and real-time requirements not allowing service jitter. It significantly improves the reliability, security, and autonomous operation and maintenance capabilities of on-orbit model iteration. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the specific embodiments of this application or the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0018] Figure 1 A flowchart illustrating a traffic prediction method for satellite communications applied to a spaceborne system, provided in an embodiment of this application; Figure 2 A flowchart of step S3 provided in the embodiments of this application; Figure 3 A flowchart of step S5 provided in an embodiment of this application; Figure 4A flowchart following step S5 is provided for an embodiment of this application; Figure 5 A flowchart following step S7 is provided for an embodiment of this application; Figure 6 A flowchart illustrating a traffic prediction method for satellite communications applied to ground stations, provided as an embodiment of this application; Figure 7 A block diagram of a traffic prediction device for satellite communications provided in this application embodiment; Figure 8 A block diagram of a traffic prediction device for satellite communications provided in this application embodiment; Figure 9 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0020] With the large-scale deployment of Low Earth Orbit (LEO) satellite constellations, satellite networks have transcended their traditional supplementary role in communications, becoming a critical infrastructure supporting global broadband access, real-time remote sensing data transmission, and emergency response. Compared to geostationary satellites, LEO satellites operate at altitudes of 500km to 1500km, orbiting the Earth at high speed with a period of approximately 90 minutes. This unique motion pattern results in short-period, highly dynamic, and drastic changes in the coverage relationship, Doppler shift, and propagation delay of satellite-to-ground links. Under these harsh conditions, network traffic is not a static process but rather a composite signal superimposed with orbital periodicity, user mobility, and sudden service interruptions. Therefore, achieving accurate perception and prediction of satellite network traffic is a core prerequisite for forward resource scheduling, avoiding congestion and packet loss, and maximizing link utilization. However, existing traffic prediction technologies, when faced with the unique characteristics of LEO satellite networks, exhibit multiple bottlenecks, including fragmented multi-scale features, rigid model iteration, insufficient lightweight capabilities, and weak decision support, severely restricting the improvement of intelligent operation and maintenance levels.

[0021] First, existing prediction methods generally suffer from isolated modeling of multi-scale features, leading to fragmented and distorted prediction results. Even current mainstream technologies employing temporal neural networks such as LSTM often train independently only for a single time granularity (e.g., simple hourly fine-grained predictions or daily trend predictions). If the focus is on short-term predictions, the model struggles to capture long-term macro-trends determined by constellation configurations (e.g., periodic remote sensing data backhaul missions), causing significant drift in the prediction baseline over time. If the focus is on long-term predictions, the model is slow to respond to short-term sudden fluctuations (e.g., instantaneous traffic spikes caused by onboard payload mode switching or emergency command injections), failing to support millisecond-level congestion control. More critically, existing architectures lack information interaction and coordination mechanisms between long-term and short-term features, failing to establish a closed loop where trends guide fluctuations and fluctuations feed back to trends. This results in predictions that cannot accurately reflect the complex physical nature of traffic flow, where long-term trends superimposed with short-term mutations.

[0022] Secondly, the rigid model deployment architecture makes it difficult to adapt to the dynamic evolution of the on-board environment, and the update process is accompanied by high-risk service interruptions. In actual operation, the service models, user distribution, and channel environment of satellite networks often undergo "conceptual drift," causing the performance of statically deployed models to degrade over time. However, existing technologies typically burn models as static files or deploy them on onboard equipment, and model updates must be completed through telemetry, tracking, and command (TT&C) and restarting the inference service. In resource-constrained onboard embedded systems, this process often introduces service vacuum periods of seconds or even minutes. For low-Earth orbit satellite communications with overshoot windows of only a few minutes, this means the complete loss of valuable communication opportunities. Furthermore, at the moment of switching between old and new models, if the statistical distribution differences between training data and real-time data are not smoothly adapted, mismatches in normalization parameters (such as maximum / minimum values) can easily cause numerical gaps or drastic jumps in the inference output, thereby misleading upper-layer scheduling strategies and causing network instability.

[0023] Furthermore, existing lightweight designs often employ a "passive slimming" strategy, severely sacrificing the model's representational capabilities in highly dynamic scenarios. Limited by the computing power and power consumption constraints of embedded processors such as onboard ARM processors, existing solutions typically achieve lightweighting by simply reducing the depth or width of the neural network. While this crude compression method reduces computational overhead, it directly weakens the model's ability to extract complex nonlinear features from highly dynamic traffic. Especially at critical junctures such as satellites entering or exiting shadow zones, switching beams, or encountering ionospheric scintillation, traffic data often exhibits drastic non-steady-state changes. In these situations, the "passively slimmed-down" model suffers from insufficient expressive power, resulting in a sharp drop in prediction accuracy. This makes it impossible to provide a reliable basis for link adaptive adjustments and fails to meet the stringent accuracy requirements of high-reliability communication.

[0024] Finally, existing methods generally lack a decision-oriented uncertainty quantification mechanism, leading to a disconnect between prediction results and resource scheduling. Traditional prediction models typically output only a deterministic point estimate, neglecting to assess the confidence level of the prediction. In automated operation and maintenance of satellite networks, the upper-level resource scheduling module faces an isolated prediction value and cannot determine whether the current prediction falls within a high-confidence range (such as a stable flight phase) or a low-confidence range (such as a phase of drastic environmental change). This puts scheduling strategies in a dilemma: they dare not aggressively allocate bandwidth to improve utilization when confidence is high, nor dare they conservatively reserve resources to ensure critical data transmission when confidence is low. The lack of quantified output of prediction uncertainty means that traffic prediction can only serve as a reference curve and cannot truly drive closed-loop intelligent resource decision-making.

[0025] In summary, how to achieve seamless hot updates of models from the ground to the satellite without interrupting the real-time inference service of the satellite system, and how to solve the model prediction failure caused by the inconsistency between the distribution of ground training data and satellite real-time data are the technical problems that urgently need to be solved.

[0026] According to an embodiment of this application, a traffic prediction method for satellite communication is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0027] This embodiment provides a traffic prediction method for satellite communications. Figure 1 A flowchart illustrating a traffic prediction method for satellite communications applied to a spaceborne system, as provided in this application embodiment, is shown below. Figure 1 As shown, the process includes the following steps: Step S1: Receive the self-describing model package sent by the ground terminal in the standby area, and maintain the currently running old model to perform inference tasks in the active area; wherein, the self-describing model package includes the trained ground model and statistical information from the training dataset.

[0028] Specifically, to achieve seamless upgrades to onboard models, the system logically divides the memory space into two independent regions: the active region and the standby region. The active region loads the current version of the old model. It independently occupies a set of memory resources and computing threads, continuously receiving real-time traffic data, executing inference tasks, and outputting bandwidth scheduling instructions. Regardless of what data writing or model loading operations are taking place in the standby region, the operation of the active region remains completely undisturbed. This ensures that satellite communication services will not experience millisecond-level interruptions or jitter during updates.

[0029] The spare area acts as a pre-loading area. The ground station no longer sends a single model weight file via the satellite-to-ground link, but a self-describing model package.

[0030] Step S3: Obtain the raw flow data within the preset sliding window, and determine the data correction factor for the adapted ground model based on the difference between the raw flow data and statistical information.

[0031] Specifically, the onboard system first maintains a preset sliding window in memory to capture and cache the current raw traffic data in real time, thereby calculating real-time statistics reflecting the current channel characteristics (such as the current mean and variance). Then, these real-time statistics are compared item by item with the training statistics carried in the self-describing model package, which represent the historical benchmark, to accurately quantify the drift magnitude of the two in the data distribution. Based on this difference, a set of exclusive data correction factors is calculated through specific mathematical mapping (such as affine transformation parameter calculation). This set of factors does not modify the model's own weights, but rather adjusts the scale and bias of the input data to forcibly pull the distribution characteristics of the real-time traffic data back to the training distribution range that the model is most familiar with. This ensures that the ground model can still maintain high-precision inference capabilities when facing the complex and ever-changing traffic environment on the satellite, achieving seamless migration where the model remains stationary and the data adapts.

[0032] Step S5: Load the data correction factor into the ground model to obtain a new corrected model, and use the new model to perform correction and inference tasks on the acquired real-time traffic data.

[0033] Specifically, the onboard system does not physically modify the model files. Instead, it dynamically injects the data correction factors calculated in the previous step into the inference of the ground model, upgrading it from a simple static predictor to a dynamic adaptor. When real-time traffic data is input, this corrected new model first uses the correction factors to align the data distribution—that is, it eliminates the distribution drift caused by differences between the satellite and ground environments through linear transformation, disguising the real-time data as a distribution pattern familiar to the model during training—before feeding it into the neural network core for inference. This mechanism cleverly bypasses the high cost of onboard retraining, ensuring that the satellite can continuously cope with the ever-changing onboard communication environment using a high-precision model trained on the ground without interrupting service.

[0034] In step S7, in response to the model switching instruction, the global active model pointer is updated from pointing to the memory address of the old model to pointing to the memory address of the new model through an atomic operation, so as to input the subsequently received real-time traffic data into the new model to perform inference tasks and obtain predicted traffic.

[0035] Specifically, the new model completes shadow inference in the spare area, and the output difference between it and the old model is within the allowable range (i.e., it passes the consistency check). The onboard system maintains a global active model pointer, and all real-time inference requests determine which model instance to use by reading this pointer. An atomic write instruction is executed, instantly changing the value of the global active model pointer from the old model to the new model. That is, within the switching time window, the pointer value is either the old address or the new address; there will never be an intermediate state (such as a null pointer or a partially written address), and the inference thread will not read an inconsistent state. After the pointer switch is complete, the processing flow changes immediately, directly calling the new model to execute the inference task. For example, the onboard system executes an atomic pointer swap instruction, changing the global model handle g_ActiveModel from pointing to the old model address 0x1000_0000 to pointing to the new model address 0x2000_0000.

[0036] This embodiment provides a traffic prediction method for satellite communication. It achieves a dual-zone isolation architecture by logically dividing the onboard system memory into an active zone and a backup zone. In the active zone, the old model continues to execute inference tasks to ensure zero interruption of satellite communication services. Simultaneously, in the backup zone, a self-describing model package containing the training model and training statistics sent from the ground is received. A data correction factor is dynamically calculated based on the difference between the original traffic data and statistics within a preset sliding window. This correction factor is injected into the ground model to form a corrected new model. This new model can seamlessly adapt to differences between the satellite and ground environments without modifying weights through data distribution alignment. Finally, in response to a model switching command, the global active model pointer is instantly updated from the old model address to the new model address via an atomic operation, allowing subsequent real-time traffic data to be directly input into the new model for inference tasks. Therefore, this embodiment achieves zero service interruption during model updates, adaptive compensation for data distribution drift, and atomicity guarantees for switching operations. It effectively solves multiple constraints faced by model upgrades in a satellite environment, such as limited computing power preventing retraining, limited satellite-to-ground link bandwidth hindering frequent transmission, and real-time requirements prohibiting service jitter. This significantly improves the reliability, security, and autonomous operation and maintenance capabilities of on-orbit model iteration.

[0037] Figure 2 The flowchart for step S3 provided in the embodiments of this application may include the following steps: Step S31: Obtain the raw traffic data within the preset sliding window, and determine the current traffic mean and current traffic standard deviation based on the raw traffic data.

[0038] Satellite communication environments are dynamic. Ground models are trained based on historical data (training datasets), and their internal normalization parameters (such as maximum and minimum values) are fixed. However, when satellites are in orbit, the raw traffic data collected in real time may experience overall mean drift (e.g., increased overall traffic) or fluctuations (increased variance) due to seasonal changes, sudden traffic surges, etc. If the fixed parameters of the old model are directly used to process new data, the numerical distribution of the input model will be inconsistent with that during training, leading to prediction failure. To address this, this embodiment utilizes a sliding window to monitor data distribution in real time, calculates compensation bias and scaling factors, and performs secondary correction on the normalized data, achieving data adaptation while the model remains unchanged.

[0039] Specifically, the onboard system maintains a preset sliding window (e.g., containing traffic data from the most recent 24 hours) to capture current real-time traffic characteristics. The average current traffic (μ) of the data within this window is calculated. live ) and current flow standard deviation ( ).

[0040] Step S33: Extract the minimum value, maximum value and mean value of the training set from the statistical information, and determine the first correction factor corresponding to the current traffic mean value.

[0041] Specifically, statistical information from ground training is read from the self-describing model package: the minimum value of the training set x min Maximum value of training set x max and training set mean μ train .

[0042] Calculate the first correction factor (i.e., the compensation bias β): The first correction factor is used to measure the difference between the current real-time mean and the training set benchmark in the normalized space.

[0043] Step S35: Extract the training set standard deviation from the statistical information and determine the second correction factor corresponding to the current flow standard deviation.

[0044] Specifically, the standard deviation of the training set is read from statistical information. Calculate the second correction factor (i.e., the scaling factor γ): The second correction factor is used to measure the ratio between the current real-time fluctuation and the training set fluctuation.

[0045] Step S37: Determine the first correction factor and the second correction factor as data correction factors to adapt to the ground model.

[0046] Specifically, the two factors (β and γ) calculated above are determined as data correction factors for adapting the ground model.

[0047] This embodiment acquires raw traffic data in real time by setting a preset sliding window in the satellite system, dynamically calculates the current traffic mean and standard deviation to capture the real-time data distribution characteristics, and determines a first correction factor and a second correction factor based on the training set statistics (minimum, maximum, mean, and standard deviation) in the self-describing model package. This allows the satellite system to perform secondary adaptive correction on the normalized real-time data without retraining or updating the ground model. The first correction factor compensates for the overall traffic level shift caused by seasonal changes, service growth, etc., by quantifying the degree of drift of the current traffic mean relative to the training set mean in the normalized space. The second correction factor corrects the variance scaling difference caused by sudden service or environmental dynamic changes by measuring the ratio of the current fluctuation amplitude to the training set standard deviation, thereby mapping the real-time data distribution back to the input space consistent with the model training. Thus, this method effectively overcomes the mismatch between the dynamic evolution of data distribution in the satellite communication environment and the fixed normalized parameters of the ground model, realizing a lightweight on-orbit deployment strategy of model immobility and data self-adaptation, avoiding the bandwidth pressure and latency overhead caused by frequent satellite-to-ground transmission of model update packages.

[0048] Figure 3 The flowchart for step S5 provided in the embodiments of this application may include the following steps: Step S51: Extract the minimum and maximum values ​​of the training set from the statistical information.

[0049] Specifically, after the spaceborne system loads the self-describing model package, it first reads static statistical information from the package, including the minimum value of the training set (x). min ) and the maximum value of the training set (x) max Regardless of changes in the onboard environment, these two values ​​are read-only and fixed within the model package, representing the data distribution range during model training.

[0050] Step S53: Normalize the acquired real-time traffic data using the minimum and maximum values ​​of the training set to obtain normalized data.

[0051] Specifically, collect current real-time traffic data (denoted as x). raw ), directly apply the fixed scale extracted in step S51 for Min-Max normalization: This step converts the physical unit of traffic (e.g., Mbps) into a dimensionless value acceptable to the model (typically between 0 and 1). However, due to changes in the real-time onboard environment (e.g., overall traffic drift), x at this point... norm_old Although the values ​​are within the legal range, their distribution center may have deviated from the comfort zone during model training. Therefore, it is only intermediate data and cannot be directly fed into the model.

[0052] Step S55: Perform linear transformation correction on the normalized data according to the data correction factor to obtain the corrected data.

[0053] Specifically, using the data correction factors (i.e., compensation bias β and scaling factor γ) calculated in the previous round, the normalized data x is... norm_old Perform an affine transformation (linear transformation) to obtain the corrected data (denoted as x) for the final fitted model. norm_new ): After processing, x norm_new The statistical distribution (mean and variance) was forcibly disguised as the data distribution during model training. The corrected data x... norm_new Input into the new model for inference.

[0054] Through steps S51 to S55, the spaceborne system reshapes the real-time data through mathematical transformations without altering the model weights. This allows a fixed model trained on the ground to adapt to the constantly shifting data environment on the satellite, ensuring the accuracy of the inference output.

[0055] This embodiment extracts the minimum and maximum values ​​of the training set fixed within the self-describing model package in the spaceborne system as normalization benchmarks to normalize real-time traffic data, completing the initial conversion from physical units to dimensionless values. Furthermore, it utilizes data correction factors to perform affine transformation correction on the normalized data, forcing the statistical distribution of the corrected data back to the comfort zone of the model training. Thus, this method achieves dynamic adaptation of real-time data distribution to the fixed model input space through mathematical transformations without modifying model weights or retraining the network. This effectively solves the data distribution offset problem caused by overall traffic drift and fluctuation amplitude changes in the satellite communication environment, avoids bandwidth consumption and latency overhead caused by frequent model updates via the space-to-ground link, and significantly improves the long-term inference accuracy and on-orbit autonomous adaptability of the ground-trained model in the spaceborne environment.

[0056] Figure 4 The flowchart provided for the embodiments of this application after step S5 may include the following steps: Step S61: Obtain historical real traffic data within a preset time period, and correct the historical real traffic data based on the data correction factor to obtain corrected historical data.

[0057] Specifically, to verify the reliability of the new model, we cannot directly use real-time traffic data for testing. Instead, we select real traffic data that has occurred within a preset time period (e.g., the past hour) as the test set. We extract this historical real traffic data (denoted as x). history Using the currently calculated data correction factors (β and γ), and following the logic of steps S53 to S55, adjust x... history Normalization and linear transformation corrections are performed to obtain the corrected historical data.

[0058] Step S63: Infer the corrected historical data using the new model to obtain the average predicted flow rate of the new model.

[0059] Specifically, the corrected historical data is input into a new model loaded in the backup area. At this time, the new model is in shadow mode; its output is not sent to the business end but is only recorded in the background. The mean traffic predicted by the new model for this period of historical real traffic data is obtained, denoted as μ. new .

[0060] Step S65: Obtain the mean and standard deviation of the predicted traffic from the old model based on historical real traffic data.

[0061] Specifically, it reads the historical output records of the old model (the model that is currently working stably) running in the active area within the same time period.

[0062] Old model predicted mean (μ) old ): This is the predicted value currently used by the system, representing the current business baseline.

[0063] Old model predicts standard deviation (σ) old This is the old model's assessment of its own predictive uncertainty.

[0064] Step S67: Determine the absolute value of the difference between the mean flow predicted by the new model and the mean flow predicted by the old model. If the absolute value of the difference is less than the product of a preset threshold and the standard deviation of the flow predicted by the old model, the new model is determined to have been successfully corrected.

[0065] Specifically, the difference between the predictions from the old and new models is calculated and compared with a dynamic threshold: |μ new -μ old ∣<λ×σ old λ is a preset threshold (e.g., set to 3), representing how many standard deviations are allowed to deviate.

[0066] While the new model's predictions may differ from the old model, the differences are within the normal range of statistical fluctuations. This indicates that the data correction factor is effective, and the new model has successfully adapted to the current distribution drift. The new model is deemed successfully corrected, triggering subsequent atomic switching operations to migrate the new model from the standby area to the active area, immediately taking over the business.

[0067] Conversely, if the correction fails, it indicates a failure. This usually means a drastic shift in the concept (e.g., a sudden abnormal traffic surge), and simple linear correction is no longer sufficient to compensate for the distribution differences. The system should be blocked from switching, the old model should be kept running, and alarm logs should be generated to await manual intervention from the ground, thereby avoiding business disruptions or interruptions caused by model switching.

[0068] This embodiment selects historical real traffic data that has occurred within a preset time period as a security test set. After normalization and linear transformation correction using the current data correction factor, the data is input into a new model in the backup area for shadow pattern inference. Simultaneously, the predicted average traffic output by the new model is dynamically compared with the predicted average traffic output and standard deviation of the old model in the active area for the same historical data period. The uncertainty of the old model's own quantitative assessment is used as an adaptive threshold benchmark to determine whether the difference between the predictions of the new and old models is within the statistically normal fluctuation range. Thus, this method achieves security verification and risk isolation before the new model goes live without interfering with current business operations. When the predicted deviation of the corrected new model is within a preset threshold multiple of the standard deviation, the data correction factor is confirmed to be effective and the new model has successfully adapted to the current distribution drift, triggering an atomic switch operation for seamless takeover. When the difference exceeds the dynamic threshold, a severe conceptual drift is determined, the system actively blocks the switch and maintains the stable operation of the old model, while generating an alarm awaiting ground intervention. This effectively avoids business shocks or service interruptions caused by model switching, significantly improving the security, reliability, and autonomous decision-making capability of the on-orbit model update process of the spaceborne system.

[0069] Figure 5 The flowchart provided for the embodiments of this application after step S7 may include the following steps: Step S81: Input the subsequently received real-time traffic data into the new model to perform inference tasks, so as to obtain the predicted traffic including the mean of the predicted traffic and the standard deviation of the predicted traffic.

[0070] Specifically, after the model hot-swap is complete, the new model officially takes over the business. For real-time traffic data, the model outputs two key metrics: The new model predicts the mean flow rate μ t : Represents a point prediction of the flow rate at a future time.

[0071] The new model predicts the standard deviation of flow σ t (Right now : Represents the model's estimate of the uncertainty of the current prediction result (i.e., the range of fluctuation of the predicted value).

[0072] Step S83: The ratio between the standard deviation of the predicted flow rate of the new model and the mean of the predicted flow rate of the new model is determined as the coefficient of variation.

[0073] Specifically, because the magnitude of the mean flow predicted by the new model varies drastically over time (e.g., low flow at night and high flow during the day), directly using the standard deviation of the predicted flow as a metric is inaccurate. Therefore, this step calculates the coefficient of variation (U0). t The formula is as follows: U t =σ t / μ t U t It is a dimensionless statistic that reflects the degree of dispersion under a unit mean. t The larger the value, the more drastic the current fluctuation relative to the predicted flow, and the lower the reliability of the prediction.

[0074] Step S85: Determine the current confidence level based on the mapping relationship between the coefficient of variation and the preset confidence level table.

[0075] Specifically, the calculated U t The continuous values ​​are compared with a pre-set confidence level table to convert them into discrete confidence levels. The pre-set confidence level table is shown below: Step S87: Determine the satellite channel bandwidth according to the bandwidth scheduling strategy corresponding to the confidence level.

[0076] In a preferred embodiment, if the confidence level is high confidence, an aggressive bandwidth allocation strategy is executed, and the satellite channel bandwidth is determined by the positive superposition of the mean of the new model predicted traffic and the standard deviation of the new model predicted traffic by a specified multiple. If the confidence level is low, a conservative bandwidth allocation strategy is implemented, and the satellite channel bandwidth is determined based on the difference between the mean traffic predicted by the new model and the standard deviation of the traffic predicted by the new model.

[0077] Specifically, based on the confidence level determined in step S85, the corresponding bandwidth calculation formula is selected to determine the final bandwidth B allocated to the satellite channel. alloc .

[0078] Scenario 1: High confidence (aggressive strategy), when U t When the value is ≤0.1, the model considers the current environment stable and the predicted value highly reliable. To handle potential minor fluctuations and maximize user experience, the system adopts a strategy of mean averaging plus positive redundancy. Calculation formula: B alloc =μt +2σ t .

[0079] Scenario 2: Low confidence (conservative strategy), when U t When the value is greater than 0.3, the model considers the current region to be highly dynamic (such as the edge of satellite switching), and the prediction error may be large. To prevent excessive bandwidth allocation leading to resource waste (because the predicted value may be artificially high) or insufficient allocation due to artificially low predicted values, a baseline strategy is adopted. Calculation formula: B alloc =μ t -σ t .

[0080] This embodiment inputs real-time traffic data into a new model to perform inference tasks, simultaneously obtaining the predicted traffic mean and predicted traffic standard deviation, and determining the ratio of the two as a dimensionless coefficient of variation to eliminate measurement bias caused by fluctuations in the absolute value of traffic. Then, based on the mapping relationship between the coefficient of variation and a pre-set confidence level table, the continuous uncertainty quantification value is transformed into discrete confidence levels. Finally, differentiated bandwidth scheduling strategies are triggered according to different confidence levels. This achieves closed-loop control of prediction uncertainty from quantitative assessment to hierarchical decision-making and then to differentiated resource scheduling, effectively solving the problem of rigid bandwidth allocation caused by drastic dynamic changes in traffic and complex satellite handover edge scenarios in satellite communication environments. It improves channel resource utilization efficiency in stable scenarios and significantly enhances the adaptive bandwidth scheduling capability and on-orbit autonomous decision-making level of the onboard system.

[0081] This embodiment also provides another traffic prediction method for satellite communications. Figure 6 A flowchart illustrating a traffic prediction method for satellite communications applied to ground stations, as provided in this application embodiment, is shown below. Figure 6 As shown, the process includes the following steps: Step S101: Obtain a training dataset containing long-term historical traffic data and short-term historical traffic data, and use the first prediction network to encode and predict the long-term historical traffic data to obtain a trend prediction sequence.

[0082] Specifically, we acquire training data, which is divided into long-term and short-term historical data.

[0083] Long-term historical traffic data refers to time series with a long span (such as the past 30 days or longer) used to capture periodicity (such as weekly cycles, monthly cycles) and long-term growth / decline trends.

[0084] Short-term historical flow data refers to recent data (such as the past hour) that is close to the forecast time, and is used to capture sudden fluctuations and real-time conditions.

[0085] The first prediction network uses a first LSTM network as a long-term trend encoder for the input sequence T.long Temporal encoding is performed, and the implicit trend components are extracted through state propagation across multiple layers of neurons. The goal of the first prediction network is not to predict the traffic in the next minute, but rather to predict the underlying trend value over a relatively long period. A network of length P is defined. long The prediction time window is set, for example, 15 days. The first prediction network outputs a trend prediction sequence T. pred , indicating the future P long The theoretical flow trend within a day, assuming no sudden short-term disruptions (such as sudden congestion or transient failures): Step S103: Extract the trend prediction value corresponding to the short-term historical traffic data from the trend prediction sequence, and concatenate the short-term historical traffic data, trend prediction value and time features to construct a concatenation vector.

[0086] Specifically, since the first prediction network outputs a long sequence, it's necessary to extract the trend values ​​corresponding to the timestamps of short-term historical traffic data. This involves constructing an input vector containing rich information and then concatenating the vectors. It can be represented as: in, This is short-term historical traffic data; This is a trend forecast value; Embedded vectors for time features (such as hours, days of the week, and whether it is a holiday) are used to help the model understand the time context.

[0087] By splicing the data together, the second prediction network can simultaneously see the current actual traffic, the theoretical trend traffic, and the time attribute, thus more accurately determining whether the current situation is a normal fluctuation or an abnormal change.

[0088] In step S105, the concatenated vector is input into the second prediction network for feature extraction, and the predicted mean and predicted variance are output through the dual fully connected layers split at the end of the second prediction network.

[0089] Specifically, the second prediction network uses a second LSTM network as a short-term fluctuation predictor to predict the concatenated vector. Encoding is performed. Unlike traditional deterministic regression networks, this embodiment improves the output layer structure of the second prediction network, enabling it to perform probabilistic prediction and uncertainty quantification.

[0090] At the end of the second prediction network, the traditional single-path fully connected layer is replaced with a split dual-path fully connected layer. This structure outputs two key statistics in parallel, representing the "numerical" and "confidence" of the traffic prediction, respectively: First output (predicted mean μ)t : Corresponds to the expected value of the traffic forecast. This branch typically uses a linear activation function to directly output the mean of the traffic forecast for future times.

[0091] Second output (prediction variance) or logarithmic variance This corresponds to the uncertainty (fluctuation range) in flow forecasting. To ensure the stability of numerical calculations and prevent negative variance, this embodiment preferably outputs the logarithmic variance. The variance is then restored through exponential operations (Exp).

[0092] Step S107: Based on the actual observed flow rate, predicted mean, and predicted variance, a loss function is constructed, and the parameters of the first prediction network and the second prediction network are jointly optimized using the loss function to obtain the trained ground model.

[0093] Specifically, during the training phase, in order to enable the network to automatically learn when to be confident and when to be conservative, this embodiment abandons the traditional mean squared error (MSE) loss function and instead adopts the negative log-likelihood (NLL) loss function.

[0094] Assume the actual observed flow rate is y t The model output predicts a mean of μ. t The prediction variance is Then the loss function L NLL Defined as: First item (Variance Penalty Term): This term penalizes excessively large variance. It prevents the model from lazily outputting infinitely large variance in order to reduce prediction error, and forces the model to converge the variance as much as possible.

[0095] Second item (Weighted Error Term): This term is the prediction error weighted by the inverse of the variance.

[0096] When the prediction error (y t -μ t ) 2 When the variance is large (e.g., due to sudden traffic surges), the model is forced to increase its variance in order to minimize the total loss. This is to reduce the contribution of this item to the total loss.

[0097] When the prediction error is small (in a stable environment), the model will be forced to reduce its variance. This is to improve the confidence level of the prediction.

[0098] Through this unsupervised adaptive calibration mechanism, the variance of the model output is reduced. It can accurately reflect the reliability of current prediction results.

[0099] Step S109: The trained ground model and the statistical information in the training dataset are bound and packaged to form a self-describing model package and sent to the spaceborne system.

[0100] Specifically, using massive amounts of historical traffic data, the prediction model (i.e., a complete model including the first and second prediction networks and a dual-path output layer) is trained, validated, and fine-tuned using quantization. The ground station will then provide the trained ground model with all training parameters (weights W and biases b) of the first prediction network, the second prediction network, and the dual-path fully connected layers, as well as statistical information from the training dataset (minimum value of the training set). Maximum value of training set training set mean Training set variance The data is bound and packaged to form a self-describing model package, which is then sent to the spaceborne system.

[0101] It should be noted that, in order to address the hardware bottleneck of limited computing power of onboard ARM processors or FPGAs and the lack of dedicated AI acceleration chips, this embodiment adopts a quantization-based perceptual training strategy to adaptively fine-tune and compress the converged model.

[0102] The process is clearly divided into the following two steps: Phase 1: Standard high-precision training (until convergence). The model is trained using FP32 (32-bit floating-point) full-precision data until the loss function converges and the model achieves optimal performance on the validation set. At this point, the model is in an "ideal floating-point environment".

[0103] The second stage: Quantization-aware fine-tuning (the last few epochs). After model convergence, the final few training epochs are performed. During these epochs, simulated quantization nodes (pseudo-quantization operators) are inserted into the computation graph. During forward propagation, the process of quantizing weights and activations from FP32 to INT8 and then dequantizing them back to FP32 is simulated, artificially introducing uniform quantization noise. This allows the network weights to automatically adjust to this noise through backpropagation, thus achieving robustness of the INT8 model while maintaining FP32 inference accuracy.

[0104] After the above fine-tuning, the model file was converted into an INT8 quantized model. To retain critical computational accuracy while compressing file size, this embodiment employs a heterogeneous hybrid precision scheme: Core layer quantization (INT8): The weight tensors and activation tensors of computationally intensive layers such as LSTM and fully connected layers are compressed from 32-bit floating-point numbers to 8-bit integers. With almost no loss in prediction accuracy, the model's inference speed on ARM Cortex-A series processors is increased by more than 3 times, and the storage volume is reduced by approximately 75% (from 4 bytes / parameter to 1 byte / parameter), significantly reducing the bandwidth pressure on satellite-to-ground link transmission of model packets and onboard storage requirements.

[0105] Key layer retention (FP32): Only the normalized computation layers (scaling layers) at the input / output ends are retained at FP32 precision.

[0106] This embodiment provides a traffic prediction method for satellite communication. Through hierarchical decoupling of long-term and short-term historical traffic data, a first prediction network extracts cross-period trend baseline components. The predicted trend value is then concatenated with short-term actual traffic and time characteristics and input into a second prediction network. This allows the second prediction network to simultaneously perceive the theoretical trend trajectory, the current actual state, and the time context, effectively distinguishing between normal fluctuations and abrupt changes. Furthermore, a dual-path fully connected layer outputs the predicted mean and variance in parallel, and the two networks are jointly optimized based on a negative log-likelihood loss function. This enables the model to automatically converge the variance in stable scenarios to improve prediction confidence, and adaptively increase the variance during sudden traffic surges to quantify uncertainty, achieving synergistic optimization of prediction accuracy and reliability assessment. Finally, by binding and packaging the model with training statistics into a self-describing model package and sending it to the satellite system, the problem of network traffic probability prediction and lightweight deployment under limited satellite computing power is solved.

[0107] The present invention will be further described in detail below with reference to specific embodiments. This embodiment takes the power supply link traffic management of a low-Earth orbit broadband communication satellite constellation as the application scenario. The onboard computer adopts an ARM Cortex-A72 processor and runs a lightweight embedded Linux operating system.

[0108] Step 1: Ground Training and Collaborative Model Building 1. Data Acquisition and Preprocessing: Historical traffic data from the past year or more was collected at ground stations. Long-term historical traffic data (for training long-term models) and short-term historical traffic data (for training short-term models) were constructed separately. Min-Max normalization was performed on both datasets, and their respective minimum values ​​(x) were calculated. min and maximum value x max .

[0109] 2. Training the trend component extraction model: A first LSTM network (long-term trend encoder) with an input time step of 45 days and a prediction time step of 15 days is trained. This network consists of two LSTM layers with a hidden layer dimension of 32. Early stopping is used during training. The model focuses on learning the macroscopic envelope of the flow, and its output is denoted as the trend prediction sequence T. pred .

[0110] 3. Couple short-term prediction model training: Train a second LSTM network (short-term fluctuation predictor) with an input time step of 48 hours, a prediction step of 12 hours, and a hidden layer dimension of 24.

[0111] At each step of training the second LSTM network, the trend prediction value of the first LSTM network at the corresponding time point is... As an additional feature dimension, it is used in conjunction with short-term historical traffic data. Time Feature t Concatenate them into a 3D input vector: It should be noted that the time feature is... t Including but not limited to: hour of day, whether it is a weekend / holiday, sine / cosine values ​​of satellite orbit phase angle, etc., to help the model capture periodic patterns.

[0112] The top fully connected layer of the second LSTM network is mapped to two neurons, corresponding to the predicted mean μt and the log-variance, respectively. Backpropagation is performed using the negative log-likelihood loss function: By minimizing this loss, the network automatically learns to operate in environments with drastic traffic fluctuations. Deviation The variance of the output is greater in the far region.

[0113] 4. Quantitative perception fine-tuning: After the model converges, a final number of training epochs are performed. During these epochs, simulated quantization nodes (pseudo-quantization operators) are inserted into the computation graph. During forward propagation, the process of quantizing weights and activations from FP32 to INT8 and then dequantizing them back to FP32 is simulated, artificially introducing uniform quantization noise. This allows the network weights to automatically adjust to this noise through backpropagation, thereby maintaining the robustness of the INT8 model while maintaining the inference accuracy of FP32, thus adapting to the integer arithmetic units of onboard ARM processors.

[0114] Step 2: Model Packaging and On-board Hot Migration 1. Self-describing model package generation: The ground station generates the quantized model structure file, weight file, and statistical information recorded during training (x... min ,x max ,μ train , The process involves packaging and serializing the data to generate a binary model file, model_package_v2.bin.

[0115] 2. Monitoring and control betting and background silent loading: When the satellite passes overhead, the ground station uploads model_package_v2.bin to the onboard storage via a high-speed power supply link. Upon detecting the new file, the onboard model management daemon starts a low-priority background thread.

[0116] When the system is idle, this thread loads the new model weights into a pre-allocated spare memory buffer (address assumed to be 0x2000_0000).

[0117] After loading is complete, the daemon reads the statistical information of the new model package and compares it with the preset sliding window statistics (the current average traffic value μ of the past 24 hours) maintained by the current onboard inference module. live and current flow standard deviation Compare them.

[0118] 3. Normalized mapping adaptation and atom switching: Suppose that a shift in real-time traffic distribution is detected due to recent business growth. To prevent normalization bias, the daemon automatically calculates data correction factors for the normalization parameters, including a scaling factor γ and a compensation bias β.

[0119] Specifically, during inference preprocessing, the original Min-Max normalized value x is... norm_old Corrected to the expected input x of the new model norm_new : Shadow reasoning and consistency verification: Before the official switchover, the system preprocesses recent historical real data (e.g., the past hour) using calculated γ and β, and inputs this data into the new model for silent inference (shadow mode). The system compares the difference between the mean traffic predicted by the new model and the mean predicted by the old model. If the consistency condition is met: |μ new -μ old ∣<λ×σ old If the value of λ is a preset threshold coefficient (e.g., 3), the verification is considered successful. If the difference is huge, it indicates that the data distribution has undergone a drastic conceptual shift, the correction fails, the system reports an error, and the old model continues to run.

[0120] Atomic Switchover: After successful verification, the onboard system executes an atomic pointer swap instruction, changing the global model handle g_ActiveModel from pointing to the old model address 0x1000_0000 to pointing to the new model address 0x2000_0000. All subsequent traffic prediction requests are immediately responded to by the new model. Throughout the switchover process, the onboard master scheduling task is unaware of any service interruptions or abnormal jumps in traffic prediction values.

[0121] Step 3: Online Reasoning and Confidence Decision Making 1. Real-time Feature Extraction: The onboard system triggers predictive inference every 15 minutes. It extracts long-term historical traffic data X from the past 48 hours and reads the corresponding trend prediction values ​​for that time period. .

[0122] 2. Model Output Analysis: The concatenated vector is input into the second prediction network of the current activity, and the output is the predicted mean sequence {μ} for the next 12 hours. t+1 ,…,μ t+12} and predicted variance sequence .

[0123] 3. Scheduling Decision Linkage: The system transmits the prediction results to the onboard queue scheduling module. The scheduling module then uses the coefficient of variation U... t The confidence level is calculated based on the magnitude of the value, and the following logic is executed: Scenario A (High Confidence): If the coefficient of variation U at the current time... t If the value is below 0.1 (indicating the satellite is in a stable flight phase, channel conditions are good, and prediction reliability is high), an aggressive allocation strategy is adopted. Allocate bandwidth B. alloc Set to cover upper side fluctuations: B alloc =μ t +2σ t This strategy has a high probability of covering sudden traffic spikes and maximizing the bandwidth utilization of the forward link.

[0124] Scenario B (low confidence): If the coefficient of variation U at the current time... t For values ​​higher than 0.3 (e.g., when the satellite has just entered a high-latitude polar region, ionospheric scintillation is present, or a feed link switch is imminent, resulting in high prediction uncertainty), a conservative allocation strategy is adopted. The allocated bandwidth Balloc is set to ensure downstream fluctuations: B alloc =μ t -σ t This strategy prioritizes ensuring that critical telemetry and signaling data are not lost even under the worst-case channel conditions, thus guaranteeing basic network connectivity.

[0125] As can be seen from the above specific implementation methods, the present invention not only achieves accurate prediction of low-Earth orbit satellite network traffic, but also constructs a smart sensing closed loop with self-evolution capability for highly dynamic low-Earth orbit satellite networks through a trend-guided collaborative coupling mechanism, a zero-interruption hot migration architecture, and uncertainty quantification for decision-making.

[0126] Accordingly, please refer to Figure 7 A block diagram of a traffic prediction device for satellite communication provided in this application embodiment, applied to a spaceborne system, the device includes: The model receiving unit 101 is used to receive the self-describing model package sent by the ground terminal in the standby area and maintain the currently running old model to perform inference tasks in the active area; wherein, the self-describing model package includes the trained ground model and statistical information in the training dataset; The factor determination unit 103 is used to acquire the raw flow data within a preset sliding window and determine the data correction factor for adapting the ground model based on the difference between the raw flow data and statistical information. The model correction unit 105 is used to load the data correction factor into the ground model to obtain a new corrected model, so as to perform correction and inference tasks on the acquired real-time traffic data through the new model; The model switching unit 107 is used to respond to the model switching instruction by updating the global active model pointer from the memory address pointing to the old model to the memory address pointing to the new model through atomic operations, so as to input the subsequently received real-time traffic data into the new model to perform inference tasks and obtain predicted traffic.

[0127] In some optional implementations, the factor determination unit 103 includes: Obtain the raw traffic data within the preset sliding window, and determine the current traffic mean and current traffic standard deviation based on the raw traffic data; Extract the minimum, maximum, and mean values ​​of the training set from the statistical information, and determine the first correction factor corresponding to the current average traffic. Extract the training set standard deviation from the statistical information and determine the second correction factor corresponding to the current flow standard deviation; The first and second correction factors were determined as data correction factors for adapting the ground model.

[0128] In some alternative implementations, the model correction unit 105 includes: Extract the minimum and maximum values ​​of the training set from the statistical information; The acquired real-time traffic data is normalized using the minimum and maximum values ​​of the training set to obtain normalized data. The normalized data is corrected by performing a linear transformation based on the data correction factor to obtain the corrected data.

[0129] In some alternative implementations, after the model correction unit 105, the apparatus is further configured to: Obtain historical real traffic data within a preset time period, and correct the historical real traffic data based on the data correction factor to obtain corrected historical data; By reasoning from the corrected historical data using the new model, the average predicted flow rate of the new model is obtained; Obtain the mean and standard deviation of the predicted traffic from the old model based on historical real traffic data; Determine the absolute value of the difference between the mean flow rate predicted by the new model and the mean flow rate predicted by the old model. If the absolute value of the difference is less than the product of a preset threshold and the standard deviation of the flow rate predicted by the old model, the new model is considered to have been successfully corrected.

[0130] In some alternative implementations, after the model switching unit 107, the apparatus is further configured to: The subsequently received real-time traffic data is input into the new model to perform inference tasks, so as to obtain the predicted traffic including the mean and standard deviation of the predicted traffic of the new model. The ratio between the standard deviation of the predicted flow rate of the new model and the mean of the predicted flow rate of the new model is determined as the coefficient of variation; The current confidence level is determined based on the mapping relationship between the coefficient of variation and the pre-set confidence level table; The satellite channel bandwidth is determined based on the bandwidth scheduling strategy corresponding to the confidence level.

[0131] In some alternative embodiments, the apparatus is also used for: If the confidence level is high, an aggressive bandwidth allocation strategy is implemented, and the satellite channel bandwidth is determined by the positive superposition of the mean of the new model's predicted traffic and the standard deviation of the new model's predicted traffic by a specified multiple. If the confidence level is low, a conservative bandwidth allocation strategy is implemented, and the satellite channel bandwidth is determined based on the difference between the mean traffic predicted by the new model and the standard deviation of the traffic predicted by the new model.

[0132] Accordingly, please refer to Figure 8 A block diagram of a traffic prediction device for satellite communication provided in this application embodiment, applied to a ground station, the device comprising: The first prediction unit 201 is used to acquire a training dataset containing long-term historical traffic data and short-term historical traffic data, and to use the first prediction network to encode and predict the long-term historical traffic data to obtain a trend prediction sequence. The data splicing unit 203 is used to extract the trend prediction value corresponding to the short-term historical traffic data from the trend prediction sequence, and splice the short-term historical traffic data, trend prediction value and time features to construct a splicing vector. The second prediction unit 205 is used to input the concatenated vector into the second prediction network for feature extraction, and output the prediction mean and prediction variance respectively through the dual fully connected layer split at the end of the second prediction network. The model training unit 207 is used to construct a loss function based on the actual observed flow rate, the predicted mean, and the predicted variance, and to jointly optimize the parameters of the first prediction network and the second prediction network using the loss function to obtain a trained ground model. The model sending unit 209 is used to bind and package the trained ground model and statistical information in the training dataset to form a self-describing model package and send it to the spaceborne system.

[0133] Further functional descriptions of the above modules and units are the same as those in the corresponding embodiments described above, and will not be repeated here.

[0134] In this embodiment, a traffic prediction device for satellite communication is presented in the form of a functional unit. Here, a unit refers to an ASIC (Application Specific Integrated Circuit) circuit, a processor and memory that execute one or more software or fixed programs, and / or other devices that can provide the above functions.

[0135] Please see Figure 9 , Figure 9 This application provides a schematic diagram of the structure of a computer device, as shown in the embodiment of the present application. Figure 9 As shown, the computer device includes one or more processors 10, memory 20, and interfaces for connecting the components, including high-speed interfaces and low-speed interfaces. The components communicate with each other via different buses and can be mounted on a common motherboard or otherwise installed as needed. The processors can process instructions executed within the computer device, including instructions stored in or on memory to display graphical information of a GUI on external input / output devices (such as display devices coupled to the interfaces). In some alternative implementations, multiple processors and / or multiple buses can be used with multiple memories and multiple memory modules, if desired. Similarly, multiple computer devices can be connected, each providing some of the necessary operations (e.g., as a server array, a group of blade servers, or a multiprocessor system). Figure 9 Take a processor 10 as an example.

[0136] Processor 10 may be a central processing unit, a network processor, or a combination thereof. Processor 10 may further include a hardware chip. The hardware chip may be an application-specific integrated circuit (ASIC), a programmable logic device (PLD), or a combination thereof. The programmable logic device may be a complex programmable logic device (CAMP), a field-programmable gate array (FPGA), a general-purpose array logic (GPA), or any combination thereof.

[0137] The memory 20 stores instructions executable by at least one processor 10 to cause the at least one processor 10 to perform the method shown in the above embodiments.

[0138] The memory 20 may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created based on the use of the computer device. Furthermore, the memory 20 may include high-speed random access memory and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some alternative embodiments, the memory 20 may optionally include memory remotely located relative to the processor 10, and these remote memories may be connected to the computer device via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0139] The memory 20 may include volatile memory, such as random access memory; the memory may also include non-volatile memory, such as flash memory, hard disk or solid-state drive; the memory 20 may also include a combination of the above types of memory.

[0140] The computer device also includes a communication interface 30 for communicating with other devices or communication networks.

[0141] This application also provides a computer-readable storage medium. The methods described in this application can be implemented in hardware or firmware, or implemented as recordable on a storage medium, or implemented as computer code downloaded over a network and originally stored on a remote storage medium or a non-transitory machine-readable storage medium and subsequently stored on a local storage medium. Thus, the methods described herein can be processed by software stored on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. The storage medium can be a magnetic disk, optical disk, read-only memory, random access memory, flash memory, hard disk, or solid-state drive, etc.; further, the storage medium can also include combinations of the above types of memory. It is understood that computers, processors, microprocessor controllers, or programmable hardware include storage components capable of storing or receiving software or computer code. When the software or computer code is accessed and executed by the computer, processor, or hardware, the methods shown in the above embodiments are implemented.

[0142] The systems, devices, modules, or units described in the above embodiments can be implemented by computer chips or entities, or by products with certain functions. A typical implementation device is a computer. Specifically, a computer can be, for example, a personal computer, laptop computer, cellular phone, camera phone, smartphone, personal digital assistant, media player, navigation device, email device, game console, tablet computer, wearable device, or any combination of these devices.

[0143] For ease of description, the above devices are described separately by function as various units. Of course, in implementing this application, the functions of each unit can be implemented in one or more software and / or hardware.

[0144] Those skilled in the art will understand that the embodiments of this application can be provided as methods or apparatus. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0145] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatuses, and devices according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0146] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0147] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0148] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0149] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the apparatus embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0150] The above description is merely an embodiment of this application and is not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

[0151] Although embodiments of this application have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of this application, and such modifications and variations all fall within the scope defined by the appended claims.

Claims

1. A traffic prediction method for satellite communications, characterized in that, Applied to spaceborne systems, the method includes: In the standby area, a self-describing model package sent from the ground terminal is received, while in the active area, the currently running old model is maintained to perform inference tasks; wherein, the self-describing model package includes a trained ground model and statistical information from the training dataset; Obtain raw flow data within a preset sliding window, and determine a data correction factor to adapt to the ground model based on the difference between the raw flow data and the statistical information; The data correction factor is loaded into the ground model to obtain a new corrected model, which is then used to perform correction and inference tasks on the acquired real-time traffic data. In response to the model switching command, the global active model pointer is updated from pointing to the memory address of the old model to pointing to the memory address of the new model through an atomic operation, so that the subsequently received real-time traffic data can be input into the new model to perform inference tasks and obtain predicted traffic.

2. The method according to claim 1, characterized in that, The step of acquiring raw traffic data within a preset sliding window and determining a data correction factor adapted to the ground model based on the difference between the raw traffic data and the statistical information includes: Obtain the raw traffic data within a preset sliding window, and determine the current traffic mean and current traffic standard deviation based on the raw traffic data; Extract the minimum value, maximum value and mean value of the training set from the statistical information, and determine the first correction factor corresponding to the current traffic mean. Extract the training set standard deviation from the statistical information and determine the second correction factor corresponding to the current traffic standard deviation; The first correction factor and the second correction factor are determined as data correction factors to adapt to the ground model.

3. The method according to claim 1, characterized in that, The step of correcting the acquired real-time traffic data using the new model includes: Extract the minimum and maximum values ​​of the training set from the statistical information; The acquired real-time traffic data is normalized using the minimum and maximum values ​​of the training set to obtain normalized data. The normalized data is linearly transformed and corrected according to the data correction factor to obtain the corrected data.

4. The method according to claim 1, characterized in that, After performing correction and inference tasks on the acquired real-time traffic data using the new model, the method further includes: Obtain historical real traffic data within a preset time period, and correct the historical real traffic data based on the data correction factor to obtain corrected historical data; The new model is used to infer the corrected historical data to obtain the average predicted flow rate. Obtain the mean and standard deviation of the predicted traffic output by the old model for the historical real traffic data; The absolute value of the difference between the mean flow predicted by the new model and the mean flow predicted by the old model is determined, and if the absolute value of the difference is less than the product of a preset threshold and the standard deviation of the flow predicted by the old model, the new model is determined to have been successfully corrected.

5. The method according to claim 1, characterized in that, After inputting subsequently received real-time traffic data into the new model to perform the inference task, the method further includes: The subsequently received real-time traffic data is input into the new model to perform inference tasks, so as to obtain the predicted traffic including the mean of the new model's predicted traffic and the standard deviation of the new model's predicted traffic. The ratio between the standard deviation of the predicted flow rate of the new model and the mean of the predicted flow rate of the new model is determined as the coefficient of variation; The current confidence level is determined based on the mapping relationship between the coefficient of variation and the preset confidence level table; The satellite channel bandwidth is determined based on the bandwidth scheduling strategy corresponding to the confidence level.

6. The method according to claim 5, characterized in that, The step of determining the satellite channel bandwidth according to the bandwidth scheduling strategy corresponding to the confidence level includes: If the confidence level is high confidence, an aggressive bandwidth allocation strategy is executed, and the satellite channel bandwidth is determined by the positive superposition of the mean traffic predicted by the new model and the standard deviation of the traffic predicted by the new model by a specified multiple. If the confidence level is low, a conservative bandwidth allocation strategy is implemented, and the satellite channel bandwidth is determined based on the difference between the mean traffic predicted by the new model and the standard deviation of the traffic predicted by the new model.

7. A traffic prediction method for satellite communications, characterized in that, Applied to ground stations, the method includes: A training dataset containing long-term historical traffic data and short-term historical traffic data is obtained. The long-term historical traffic data is encoded and predicted using a first prediction network to obtain a trend prediction sequence. The trend prediction value corresponding to the short-term historical traffic data is extracted from the trend prediction sequence, and the short-term historical traffic data, the trend prediction value and the time feature are concatenated to construct a concatenated vector. The concatenated vector is input into the second prediction network for feature extraction, and the prediction mean and prediction variance are output respectively through the dual fully connected layer split at the end of the second prediction network. Based on the actual observed flow rate, the predicted mean, and the predicted variance, a loss function is constructed, and the parameters of the first prediction network and the second prediction network are jointly optimized using the loss function to obtain a trained ground model. The trained ground model and the statistical information in the training dataset are bound and packaged to form a self-describing model package, which is then sent to the spaceborne system.

8. A traffic prediction device for satellite communication, characterized in that, The device, used in spaceborne systems, includes: The model receiving unit is used to receive the self-describing model package sent by the ground terminal in the standby area, and to maintain the currently running old model to perform inference tasks in the active area; wherein, the self-describing model package includes the trained ground model and statistical information in the training dataset; The factor determination unit is used to acquire the raw flow data within a preset sliding window, and determine the data correction factor adapted to the ground model based on the difference between the raw flow data and the statistical information. The model correction unit is used to load the data correction factor into the ground model to obtain a new corrected model, so as to perform correction and inference tasks on the acquired real-time traffic data through the new model; The model switching unit is used to respond to the model switching command by updating the global active model pointer from the memory address pointing to the old model to the memory address pointing to the new model through atomic operations, so as to input the subsequently received real-time traffic data into the new model to perform inference tasks and obtain predicted traffic.

9. A traffic prediction device for satellite communication, characterized in that, Applied to ground stations, the device includes: The first prediction unit is used to acquire a training dataset containing long-term historical traffic data and short-term historical traffic data, and to encode and predict the long-term historical traffic data using the first prediction network to obtain a trend prediction sequence. The data splicing unit is used to extract the trend prediction value corresponding to the short-term historical traffic data from the trend prediction sequence, and splice the short-term historical traffic data, the trend prediction value and the time feature to construct a splicing vector. The second prediction unit is used to input the spliced ​​vector into the second prediction network for feature extraction, and output the prediction mean and prediction variance respectively through the dual fully connected layer split at the end of the second prediction network. The model training unit is used to construct a loss function based on the actual observed flow rate, the predicted mean, and the predicted variance, and to jointly optimize the parameters of the first prediction network and the second prediction network using the loss function to obtain a trained ground model. The model sending unit is used to bind and package the trained ground model and the statistical information in the training dataset to form a self-describing model package and send it to the spaceborne system.

10. A computer device, characterized in that, include: A memory and a processor are communicatively connected, the memory storing computer instructions, and the processor executing the computer instructions to perform the traffic prediction method for satellite communication as described in any one of claims 1 to 7.

11. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing a computer to execute the traffic prediction method for satellite communications as described in any one of claims 1 to 7.

Images