A high-low orbit inter-satellite spectrum access method based on spectrum prediction
By optimizing satellite handover strategies through spectrum prediction and deep reinforcement learning, the problem of real-time adjustment of access conditions in high and low orbit satellite communication was solved, achieving dynamic beam switching and stability of continuous communication.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHINA INST OF RADIO PROPAGATION
- Filing Date
- 2025-02-17
- Publication Date
- 2026-07-07
AI Technical Summary
Existing methods for accessing high and low orbit satellite spectrum cannot adjust access conditions in real time, resulting in interference with high orbit satellite channels and failing to guarantee continuous communication stability. In particular, they cannot minimize the number of beam switching when communication demands change dynamically.
A spectrum prediction-based approach is adopted, which uses a CNN-LSTM model to predict future spectrum utilization and combines deep reinforcement learning to optimize satellite handover strategies. The DQN algorithm is used to minimize the number of satellite handovers to ensure communication stability.
It enables real-time beam switching when communication needs change dynamically, avoids interference with high-orbit satellite channels, and ensures the stability of continuous communication and efficient channel utilization.
Smart Images
Figure CN120049943B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of inter-satellite spectrum access, and specifically relates to a high- and low-orbit inter-satellite spectrum access method based on spectrum prediction in this field. Background Technology
[0002] The connection time between high-speed low-Earth orbit (LEO) satellites and ground equipment such as ground terminals and gateways is short, making it impossible for a single satellite to effectively transmit complete, continuous data. Conversely, because geostationary orbit (GSO) satellites are stationary relative to the ground, they can serve as relay satellites for LEO satellites, used for data forwarding. LEO satellites can utilize the idle frequency bands of GSO satellites for access, thereby achieving efficient data transmission.
[0003] Currently, research on spectrum access methods for high and low orbit satellites is relatively limited, and most methods are static, requiring pre-planning of beam switching strategies. For example, existing patent document CN117856876B discloses a distributed cooperative communication system and method between high and low orbit satellites. This system includes: a high-orbit satellite; a low-orbit communication unit installed on the high-orbit satellite, used to send synchronization signals, cooperative relay timing tables, and phase synchronization completion identifiers to the high-orbit communication unit, and to receive data transmission requests, test signals, and data to be transmitted from the high-orbit communication unit; a cooperative relay calculation unit installed on the high-orbit satellite, used to calculate the cooperative relay timing table; multiple low-orbit satellites; a high-orbit communication unit installed on the low-orbit satellite, used to send data transmission requests, perform signal synchronization tests, send test signals, and send data to be transmitted; and an inter-satellite interconnection unit installed on the low-orbit satellite, used for data transmission between low-orbit satellites. This existing technology can solve the problems of weak transmission capacity of a single low-orbit satellite and transmission discontinuity caused by the relative motion between high and low orbit satellites, improving the capacity and efficiency of satellite communication.
[0004] However, with the dynamic changes in communication demands, access conditions need to be adjusted in real time for beam switching. Existing methods mainly suffer from the following problems:
[0005] (1) Existing access methods usually use the channel utilization of satellite beams as known data, which cannot cope with the frequency band interference or sudden situations that low-orbit satellites may encounter during actual access. The channel frequency utilization may change in the future, causing access to the channels that high-orbit satellites are using, which will interfere with GSO satellites.
[0006] (2) The current main optimization goal of access is the longest overall communication time, without considering the fewest number of handovers, thus failing to guarantee the stability of continuous communication. Summary of the Invention
[0007] The technical problem to be solved by this invention is:
[0008] The purpose of this invention is to provide a high-low orbit inter-satellite spectrum access method based on spectrum prediction, in order to solve the technical problems of existing access methods, such as the inability to adjust access conditions in real time for beam switching, resulting in access to channels used by high-orbit satellites and causing interference to GSO satellites, and the failure to consider the minimum number of switching times, thus failing to guarantee the stability of continuous communication.
[0009] The technical solution adopted by the present invention to solve the above-mentioned technical problems is as follows:
[0010] An improved method for inter-satellite spectrum access in high and low orbits based on spectrum prediction includes the following steps:
[0011] Step 1: Normalize the historical spectrum data of high-orbit GSO satellites, dividing the spectrum data into training and testing sets:
[0012] Historical spectrum data of GSO satellites for the past L channels within the historical time period t0∈[1,T0]. After normalization, the data range is scaled to the interval [0, 1]. The normalization formula is shown below:
[0013]
[0014] In the above formula, x l ' represents the normalized spectrum result of the l-th channel, l∈[1,L], max(x l ) and min(x l () represents the maximum and minimum values of the l-th channel;
[0015] The normalized spectral data was divided into training and testing sets in an 8:2 ratio.
[0016] Step 2: Construct a multi-channel spectrum prediction model based on CNN-LSTM, set the initial parameter values of the model, optimize the hyperparameters of the model using the training set to obtain the optimal model; apply the test set to the trained optimal model, perform inverse normalization on the predicted values to make them the same order of magnitude as the true values, and output the prediction results, prediction error curve, and fitting curve of the true values and predicted values to evaluate the prediction quality of the optimal model.
[0017] Step 3: Obtain other variables during the access process:
[0018] This includes the remaining visible time and coverage information of LEO satellites within the GSO satellite beam;
[0019] Step 4: Establish an optimization model:
[0020] The total number of LEO satellites is defined as K = {1, 2, ..., K}, and the set of available GSO satellite beams in the system is N = {1, 2, ..., N}. This indicates the coverage of GSO satellite beam n for LEO satellite k at time t:
[0021]
[0022] At time [t] u ,t u+1 The set of GSO satellite beams covering the area is as follows:
[0023]
[0024] LEO satellite k being served in beam n:
[0025]
[0026] The total bandwidth of each beam is divided into L channels of equal bandwidth. The channel frequency usage is obtained from step 2. During the transmission of each LEO satellite, each user can only use one channel, and the idle channel usage of the entire beam cannot exceed the number of channels L. The overall channel budget constraint is:
[0027]
[0028] Viewing the LEO satellite handover issue as a process within each coverage area from The optimization problem of selecting a GSO satellite beam to serve LEO satellite k is shown below, with the average number of LEO satellite handovers as follows:
[0029]
[0030] HO k This indicates the number of satellite beam switching times for LEO satellite k. If the service relationship between GSO satellite beam n and LEO satellite k changes with the coverage area from [t...] u ,t u+1 ) changes to [t u+1 ,t u+2 ),So HO k Increase by 1;
[0031] Establish the following optimization model:
[0032]
[0033] In the above formula, This represents two states of user association between LEO satellite k and GSO satellite n;
[0034] Step 5: Design and optimize the access algorithm. Use the remaining LEO visibility time obtained in Step 1 and the spectrum prediction results in Step 3 as input data for the DQN model to perform empirical learning and optimize and update the network parameters.
[0035] Step 6: Use the future spectrum usage results predicted in Step 2 and other factor variables in Step 3 as inputs to the DQN model. The DQN model outputs the satellite frequency bands to be accessed.
[0036] Furthermore, step 2 specifically includes the following steps:
[0037] Step 21, initialize the CNN-LSTM multi-channel spectrum prediction model:
[0038] First, the historical spectrum data is organized into a format that the model can use and imported into the model;
[0039] Secondly, in the CNN-LSTM model, the CNN analyzes the frequency band utilization and the correlation features with other frequency bands, then the LSTM analyzes the temporal features of the spectrum, and finally the FC layer maps the high-dimensional feature vector output by the LSTM layer to a lower-dimensional representation space to retain effective information. At the same time, a Dropout layer is added after the CNN network layer.
[0040] Finally, output the model's prediction results;
[0041] Step 22, train the model using the training set:
[0042] The training set from step 1 is input into the CNN-LSTM network to predict the spectrum utilization of the beam in the future time period, and the hyperparameters of the network are continuously updated through backpropagation.
[0043] Step 23: Validate the model using the test set and perform data inverse normalization for verification and comparison.
[0044] Using the test set from step 1 as input to the model, the model's output is as follows:
[0045]
[0046] T1 represents the predicted total future time, and X1 represents the historical spectrum data of L channels in the next T1 time slots. This represents the spectrum prediction data of the l-th channel of the GSO beam channel received by the spectrum sensing satellite at a future time t1.
[0047] Furthermore, step 5 specifically includes:
[0048] The optimization access problem is transformed into a multi-agent reinforcement learning optimization problem based on stochastic game theory. The key definition of multi-agent reinforcement learning is as follows:
[0049] Agent: LEO satellite k∈K, takes action at each step and causes a transition in the covered state;
[0050] State: This represents the state of the k-th LEO satellite agent at time t, determined by the covering satellite. Available satellite channels and remaining visible time of the satellite composition;
[0051] Action: This represents the action of the k-th LEO satellite agent. Indicates whether LEO satellite k is served by GSO satellite beam n at time t;
[0052] Reward: This represents the reward for the k-th LEO satellite agent, used to describe the state. implement The immediate reward after an action, assuming the agent does not know the reward functions of other agents, but can obtain the actions of other agents, is defined as follows:
[0053]
[0054] When LEO satellite agent k selects a GSO satellite beam to cover the satellite, but that beam does not serve LEO satellites, an instantaneous switch occurs.
[0055] When LEO satellite agent k selects a GSO satellite beam to cover the satellite, but the beam channel of that GSO satellite is overloaded,
[0056]
[0057] When LEO satellite agent k selects a GSO satellite beam that covers and serves LEO satellites, and the satellite channel is sufficient for the LEO satellites it accesses,
[0058] Furthermore, a deep neural network is constructed using convolutional layers and fully connected layers. This deep neural network consists of 6 2D convolutional layers, one unfolded layer, and one fully connected layer. The 6 convolutional layers are used to extract features from the state information. The kernel size of each convolutional layer is set to 3x3, and the activation function is the LeakyReLU function. The filter output values of the 6 convolutional layers are set to 8, 16, 32, 64, 16, and 8, respectively. The unfolded layer unfolds the feature matrix learned by the convolutional layers into a vector, which is then input into the fully connected layer to map the value Q of each behavior. The number of neurons in the fully connected layer is consistent with the number of behaviors in the behavior space.
[0059] The beneficial effects of this invention are:
[0060] This invention employs a spectrum prediction-based inter-satellite spectrum access method for both high and low Earth orbits. First, it uses predicted channel occupancy as the data input for access. Simultaneously, it utilizes deep reinforcement learning to dynamically adjust the available channels for LEO constellations to access GSO satellite beams. Addressing the issue of complete data transmission from a single satellite, the optimization objective is set to minimize the number of satellite handovers. The DQN algorithm is used to achieve optimal long-term access performance. This invention is suitable for dynamically changing communication needs, enabling real-time adjustment of access conditions for beam switching. It addresses potential frequency band interference or sudden situations encountered by low-Earth orbit satellites during actual access, preventing interference from high-Earth orbit satellites using channels that could otherwise be used by LEO satellites when channel usage changes. Furthermore, this invention ensures the stability of continuous communication.
[0061] The method disclosed in this invention obtains abnormal frequency usage results by predicting future times of known spectrum and comparing the results with known information. The spectrum prediction used can forecast future channel utilization of satellite beams to address potential frequency band interference or sudden situations that low-Earth orbit satellites may encounter during access, avoiding access to channels already in use by high-Earth orbit satellites and preventing interference with communication quality.
[0062] The method of this invention uses the results of abnormal frequency usage and the remaining visible time as constraints. For the problem of complete transmission of single satellite data in a scenario, the optimization objective is set to minimize the number of satellite handovers, so as to ensure the stability of continuous communication quality and achieve the overall long-term optimal dynamic access effect. Attached Figure Description
[0063] Figure 1 This is a flowchart of the method of the present invention;
[0064] Figure 2 This is a diagram of a multi-channel spectrum prediction model based on CNN-LSTM;
[0065] Figure 3This is a diagram illustrating the optimization of access using deep reinforcement learning.
[0066] Figure 4 This is a DQN network diagram. Detailed Implementation
[0067] To make the objectives, technical solutions, and advantages of this invention clearer, the following description is provided in conjunction with the appendix. Figure 1-4 The present invention will be further described in detail with reference to the embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.
[0068] Example 1 discloses a high-low orbit inter-satellite spectrum access method based on spectrum prediction. First, the predicted channel occupancy is used as input data. Simultaneously, a deep reinforcement learning method is employed to dynamically adjust the available channels for LEO constellation access to GSO satellite beams. Addressing the issue of complete data transmission for individual satellites, the optimization objective is set to minimize the number of satellite handovers. The DQN algorithm is used to achieve optimal long-term access performance.
[0069] This method proposes using a CNN-LSTM network to predict anomalies in planned spectrum, preventing abnormal spectrum access issues in future access processes and improving anti-interference capabilities. The designed CNN-LSTM spectrum prediction network can achieve high-accuracy prediction of spectrum data and detect spectrum occupancy anomalies. The spectrum prediction results are used as a condition for deep reinforcement learning access, and the optimization objective of deep reinforcement learning is set to minimize the number of satellite handovers. The DQN algorithm is used for deep reinforcement learning to ensure the continuity of communication services.
[0070] This method employs deep learning to address scenarios where low-Earth orbit (LEO) satellite constellations and high-Earth orbit (HEO) satellites coexist, with LEO satellites using HEO satellites as relays for dynamic communication with ground gateway stations. LEO satellites calculate future channel occupancy of GSO satellite beams through spectrum prediction and obtain their remaining visibility time under the GSO beam by calculating their own ephemeris. Using channel occupancy and remaining visibility time as constraints, deep reinforcement learning is used for iterative optimization to provide the optimal access strategy for LEO satellites to access GSO satellite beams. This aims to minimize the number of inter-satellite handovers during the entire access process, ensuring the continuity of communication services.
[0071] like Figure 1 As shown, the specific steps include the following:
[0072] Step 1: Normalize the historical spectrum data of high-orbit GSO satellites, dividing the spectrum data into training and testing sets:
[0073] Normalized data can accelerate model convergence and reduce training time. Specifically, within the historical time period t0∈[1,T0], the historical spectrum data of GSO satellites from the past L channels are normalized. After normalization, the data range is scaled to the interval [0, 1]. The normalization formula is shown below:
[0074]
[0075] In the above formula, x l ' represents the normalized spectrum result of the l-th channel, l∈[1,L], max(x l ) and min(x l () represents the maximum and minimum values of the l-th channel;
[0076] The normalized spectral data was divided into training and testing sets in an 8:2 ratio.
[0077] Step 2: Construct a multi-channel spectrum prediction model based on CNN-LSTM, set the initial parameter values of the model, optimize the hyperparameters of the model using the training set to obtain the optimal model; apply the test set to the trained optimal model, perform inverse normalization on the predicted values to make them the same order of magnitude as the true values, and output the prediction results, prediction error curve, and fitting curve of the true values and predicted values to evaluate the prediction quality of the optimal model.
[0078] Step 21, initialize the CNN-LSTM multi-channel spectrum prediction model:
[0079] like Figure 2 As shown, the basic structure of the CNN-LSTM-based multi-channel spectrum prediction model includes input, CNN-LSTM model, and output;
[0080] First, the historical spectrum data is organized into a format that the model can use and imported into the model;
[0081] Secondly, the CNN in the CNN-LSTM model analyzes the frequency band utilization and the correlation features with other frequency bands, then the LSTM analyzes the temporal features of the spectrum, and finally the FC layer maps the high-dimensional feature vector output by the LSTM layer to a lower-dimensional representation space to retain effective information. At the same time, a Dropout layer is added after the CNN network layer to enhance robustness and prevent the model from overfitting.
[0082] Finally, output the model's prediction results;
[0083] Step 22, train the model using the training set:
[0084] The training set from step 1 is input into the CNN-LSTM network to predict the spectrum utilization of the beam in the future time period, and the hyperparameters of the network are continuously updated through backpropagation.
[0085] Step 23: Validate the model using the test set and perform data inverse normalization for verification and comparison.
[0086] Using the test set from step 1 as input to the model, the model's output is as follows:
[0087]
[0088] T1 represents the predicted total future time, and X1 represents the historical spectrum data of L channels in the next T1 time slots. This represents the spectrum prediction data of the l-th channel of the GSO beam channel received by the spectrum sensing satellite at a future time t1.
[0089] Step 3: Obtain other variables during the access process:
[0090] This includes the remaining visible time and coverage information of LEO satellites within the GSO satellite beam;
[0091] Using the LEO onboard computing module, based on its own TLE data and the GSO satellite beam position, the remaining visibility time of LEO relative to the GSO satellite beam is calculated. Since the GSO satellite is stationary relative to the Earth, its beam position is also fixed. Therefore, the positional relationship between LEO and GSO can be used to calculate the visibility time range and coverage of LEO for each GSO beam.
[0092] Step 4: Establish an optimization model to achieve the optimal result of minimizing the number of long-term outages, aiming to minimize the average number of handovers while improving the channel utilization efficiency of the LEO satellite system.
[0093] The total number of LEO satellites is defined as K = {1, 2, ..., K}, and the set of available GSO satellite beams in the system is N = {1, 2, ..., N}. This indicates the coverage of GSO satellite beam n for LEO satellite k at time t:
[0094]
[0095] At time [t] u ,t u+1 The set of GSO satellite beams covering the area is as follows:
[0096]
[0097] LEO satellite k being served in beam n:
[0098]
[0099] The total bandwidth of each beam is divided into L channels of equal bandwidth. The channel frequency usage is obtained from step 2. During the transmission of each LEO satellite, each user can only use one channel, and the idle channel usage of the entire beam cannot exceed the number of channels L. The overall channel budget constraint is:
[0100]
[0101] Viewing the LEO satellite handover issue as a process within each coverage area from The optimization problem of selecting a GSO satellite beam to serve LEO satellite k is shown below, with the average number of LEO satellite handovers as follows:
[0102]
[0103] HO k This indicates the number of satellite beam switching times for LEO satellite k. If the service relationship between GSO satellite beam n and LEO satellite k changes with the coverage area from [t...] u ,t u+1 ) changes to [t u+1 ,t u+2 ),So HO k Increase by 1;
[0104] The following optimization model is established to achieve the optimal result for the number of long-term outages, aiming to optimize service-related metrics. The goal is to minimize the average number of handovers within a time period T, while simultaneously improving the channel utilization efficiency of the LEO satellite system. The entire optimization problem is formulated as follows:
[0105]
[0106] In the above formula, This represents two states of user association between LEO satellite k and GSO satellite n; L is the constraint of the total satellite beam channel.
[0107] Step 5: Design and optimize the access algorithm. Use the remaining LEO visibility time obtained in Step 1 and the spectrum prediction results in Step 3 as input data for the DQN model to perform empirical learning and optimize and update the network parameters.
[0108] The optimization access problem is transformed into a multi-agent reinforcement learning optimization problem based on stochastic game theory. The LEO satellite handover optimization problem is essentially a generalized K-agent game problem because the agents in the system have both cooperative and competitive relationships. Figure 3As shown, the key definitions of Multi-Agent Reinforcement Learning (MARL) are as follows:
[0109] Agent: LEO satellite k∈K, takes action at each step and causes a transition in the covered state;
[0110] State: This represents the state of the k-th LEO satellite agent at time t, determined by the covering satellite. Available satellite channels and remaining visible time of the satellite composition;
[0111] Action: This represents the action of the k-th LEO satellite agent. Indicates whether LEO satellite k is served by GSO satellite beam n at time t;
[0112] Reward: This represents the reward for the k-th LEO satellite agent, used to describe the state. implement Instantaneous reward after the action This represents the joint action of all agents at time t. Assume that the agents do not know the reward functions of the other agents, but they can obtain the actions of the other agents. The reward function is defined as follows:
[0113]
[0114] When LEO satellite agent k selects a GSO satellite beam to cover the satellite, but that beam does not serve LEO satellites, an instantaneous switch occurs.
[0115] When LEO satellite agent k selects a GSO satellite beam to cover the satellite, but the beam channel of that GSO satellite is overloaded,
[0116] When LEO satellite agent k selects a GSO satellite beam that covers and serves LEO satellites, and the satellite channel is sufficient for the LEO satellites it accesses,
[0117] Representing the visible time, the reward value is a positive integer only if there is no immediate handover and the satellite's channel budget is sufficient. Otherwise, it is a negative integer. The positive integer reward is equal to the remaining visible time, because the longer the remaining visible time, the lower the probability of a future handover.
[0118] To better learn the latent relationship function between state information and each behavior and reduce computational complexity, deep neural networks are constructed using convolutional layers and fully connected layers, such as... Figure 4 As shown, this deep neural network consists of 6 2D convolutional layers, one unfolded layer, and one fully connected layer. The 6 convolutional layers are used to extract features from the state information. The kernel size of each convolutional layer is set to 3x3, and the activation function is the LeakyReLU function. The filter output values of the 6 convolutional layers are set to 8, 16, 32, 64, 16, and 8 respectively. The unfolded layer unfolds the feature matrix learned by the convolutional layers into a vector, which is then input into the fully connected layer to map the value Q of each behavior. The number of neurons in the fully connected layer is consistent with the number of behaviors in the behavior space.
[0119] Step 6: Use the future spectrum usage results predicted in Step 2 and other factor variables in Step 3 as inputs to the DQN model. The DQN model outputs the satellite frequency bands to be accessed.
[0120] Verification has shown that the method proposed in this invention solves the technical problem raised in this invention, and practical application has verified the technical effects and practicality claimed in this invention.
[0121] The method described in this invention has been verified through simulation experiments and practical applications, demonstrating the technical effects claimed by this invention.
[0122] The algorithm (method) proposed in this invention is the underlying technical core of this invention. Based on the algorithm (method) proposed in this invention, a high- and low-orbit inter-satellite spectrum access system based on spectrum prediction is developed using a programming language. This system has program modules corresponding to the steps of the above technical solution, and executes the steps in the above-mentioned high- and low-orbit inter-satellite spectrum access method based on spectrum prediction when running.
[0123] The developed system (software) computer program is stored on a computer-readable storage medium. This computer program is configured to implement the steps of the aforementioned high- and low-orbit inter-satellite spectrum access method based on spectrum prediction when invoked by a processor. In other words, the invention is materialized on a carrier, becoming a computer program product.
[0124] A high- and low-Earth orbit inter-satellite spectrum access device based on spectrum prediction, the device comprising at least one processor and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to execute the aforementioned high- and low-Earth orbit inter-satellite spectrum access method based on spectrum prediction.
[0125] Various implementations of the systems and techniques described herein can be implemented in digital electronic circuit systems, integrated circuit systems, application-specific integrated circuits (ASICs), computer hardware, firmware, software, and / or combinations thereof. These various implementations may include: implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transferring data and instructions to the storage system, the at least one input device, and the at least one output device.
[0126] The computational programs (also referred to as programs, software, software applications, or code) of this invention include machine instructions of a programmable processor and can be implemented using high-level procedural and / or object-oriented programming languages, and / or assembly / machine languages. As used herein, the terms "machine-readable medium" and "computer-readable medium" refer to any computer program product, device, and / or apparatus (e.g., disk, optical disk, memory, programmable logic device PLD) for providing machine instructions and / or data to a programmable processor, including machine-readable media that receive machine instructions as machine-readable signals. The term "machine-readable signal" refers to any signal for providing machine instructions and / or data to a programmable processor.
[0127] It should be understood that the various processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this application can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this application can be achieved, they are all within the protection scope of this invention.
Claims
1. A method for inter-satellite spectrum access in high and low orbits based on spectrum prediction, characterized in that, The method includes the following steps: Step 1: Normalize the historical spectrum data of high-orbit GSO satellites, dividing the spectrum data into training and testing sets: exist Historical spectrum data of GSO satellites over the past L channels within a historical period. After normalization, the data range is scaled to the interval [0, 1]. The normalization formula is shown below: In the above formula, Indicates the first The normalized spectrum results for each channel , and Representing the The maximum and minimum values of the spectrum data for each channel; The normalized spectral data was divided into training and testing sets in an 8:2 ratio. Step 2: Construct a multi-channel spectrum prediction model based on CNN-LSTM, set the initial parameter values of the model, optimize the hyperparameters of the model using the training set to obtain the optimal model; apply the test set to the trained optimal model, perform inverse normalization on the predicted values to make them the same order of magnitude as the true values, and output the prediction results, prediction error curve, and fitting curve of the true values and predicted values to evaluate the prediction quality of the optimal model. Step 3: Obtain other variables during the access process: This includes the remaining visible time and coverage information of LEO satellites within the GSO satellite beam; Step 4: Establish an optimization model: The total number of LEO satellites is defined as The available GSO satellite beam set in the system is as follows: , This indicates the coverage of GSO satellite beam n for LEO satellite k at time t: The set of GSO satellite beams covering the area at time t is as follows: Whether LEO satellite k is served by satellite beam n is as follows: The total bandwidth of each beam is divided into L channels of equal bandwidth. The channel frequency usage is obtained from step 2. During the transmission of each LEO satellite, each user can only use one channel, and the number of idle channels in the entire beam cannot exceed the number of channels L. The overall channel budget constraint is: Viewing the LEO satellite handover issue as a process within each coverage area from The optimization problem of selecting a GSO satellite beam to serve LEO satellite k is shown below, with the average number of LEO satellite handovers as follows: This represents the number of satellite beam switching times for LEO satellite k. If the service relationship between GSO satellite beam n and LEO satellite k changes from time t to time t+1 as the coverage area changes, then... , Increase by 1; Establish the following optimization model: In the above formula, This represents two states of user association between LEO satellite k and GSO satellite beam n; Step 5: Design and optimize the access algorithm. Use the remaining LEO visibility time obtained in Step 3 and the spectrum prediction results in Step 2 as input data for the DQN model to perform empirical learning and optimize and update the network parameters. Step 6: Use the future spectrum usage results predicted in Step 2 and other factor variables in Step 3 as inputs to the DQN model. The DQN model outputs the satellite frequency bands to be accessed.
2. The high- and low-orbit inter-satellite spectrum access method based on spectrum prediction according to claim 1, characterized in that: Step 2 specifically includes the following steps: Step 21, initialize the CNN-LSTM multi-channel spectrum prediction model: First, the historical spectrum data is organized into a format that the model can use and imported into the model; Secondly, in the CNN-LSTM model, the CNN analyzes the frequency band utilization and the correlation features with other frequency bands, then the LSTM analyzes the temporal features of the spectrum, and finally the FC layer maps the high-dimensional feature vector output by the LSTM layer to a lower-dimensional representation space to retain effective information. At the same time, a Dropout layer is added after the CNN network layer. Finally, output the model's prediction results; Step 22, train the model using the training set: The training set from step 1 is input into the CNN-LSTM network to predict the spectrum utilization of the beam in the future time period, and the hyperparameters of the network are continuously updated through backpropagation. Step 23: Validate the model using the test set and perform data inverse normalization for verification and comparison. Using the test set from step 1 as input to the model, the model's output is as follows: T1 represents the predicted total future time. This represents the historical spectrum data of L channels in the next T1 time slots. This represents the number of GSO beam channels received by the spectrum sensing satellite at a future time t1. Spectrum prediction data for each channel.
3. The high- and low-orbit inter-satellite spectrum access method based on spectrum prediction according to claim 1, characterized in that: Step 5 specifically includes: The optimization access problem is transformed into a multi-agent reinforcement learning optimization problem based on stochastic game theory. The key definition of multi-agent reinforcement learning is as follows: Agent: LEO satellite k∈K, takes action at each step and causes a transition in the covered state; State: , representing the state of the k-th LEO satellite agent at time t, determined by the covering satellite Available satellite channels and remaining visible time of the satellite composition; Action: This represents the action of the k-th LEO satellite agent. ; among them The meaning is whether LEO satellite k is served by GSO satellite beam n at time t; Reward: This represents the reward for the k-th LEO satellite agent, used to describe the state. implement The immediate reward after an action, assuming the agent does not know the reward functions of other agents, but can obtain the actions of other agents, is defined as follows: When LEO satellite agent k selects a GSO satellite beam to cover the satellite, but that beam does not serve LEO satellites, an instantaneous switch occurs. ; When LEO satellite agent k selects a GSO satellite beam to cover the satellite, but the beam channel of that GSO satellite is overloaded, When LEO satellite agent k selects a GSO satellite beam that covers and serves LEO satellites, and the satellite channel is sufficient for the LEO satellites it accesses, .
4. The high- and low-orbit inter-satellite spectrum access method based on spectrum prediction according to claim 3, characterized in that: A deep neural network is constructed using convolutional layers and fully connected layers. This deep neural network consists of 6 2D convolutional layers, one unfolded layer, and one fully connected layer. The 6 convolutional layers are used to extract features from the state information. The kernel size of each convolutional layer is set to 3x3, and the activation function is the LeakyReLU function. The output values of the filters in the 6 convolutional layers are set to 8, 16, 32, 64, 16, and 8, respectively. The unfolded layer unfolds the feature matrix learned by the convolutional layers into a vector, which is then input into the fully connected layer to map the value Q of each behavior. The number of neurons in the fully connected layer is consistent with the number of behaviors in the behavior space.
5. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program configured to, when invoked by a processor, implement the steps of the high- and low-orbit inter-satellite spectrum access method based on spectrum prediction as described in any one of claims 1-4.