Satellite spectrum resource dynamic allocation method and device based on time-space diagram Transform
By constructing a dynamic allocation model for satellite spectrum resources based on the spatiotemporal graph Transformer method, the problems of large prediction errors and poor interpretability in existing technologies are solved, and accurate prediction of future spectrum demand and executable dynamic scheduling are achieved.
Patent Information
- Application Number
- CN202511749525.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-26
- Publication Date
- 2026-03-06
AI Technical Summary
Existing technologies for predicting satellite spectrum resources suffer from limitations in long-term acquisition capabilities, difficulty in integrating multi-source heterogeneous features, and a lack of interpretability and executability, resulting in prediction results that cannot be directly used for resource scheduling.
The method based on the spatiotemporal graph Transformer is adopted. By constructing a spatiotemporal graph of satellite resources and combining the node relationships of satellites, beams and ground stations, orbital position coding, rain attenuation, ionospheric disturbances and surveillance masks are introduced to establish a multi-head attention mechanism, perform multi-task multi-quantile prediction, and output the uncertainty range of spectrum demand.
It achieves accurate prediction of future spectrum demand, reduces prediction errors, takes into account orbit periodicity and regulatory constraints, and improves the robustness and security of dynamic scheduling.
Smart Images
Figure CN121619006A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of satellite communication and wireless spectrum management technology, and in particular to a method and apparatus for dynamic allocation of satellite spectrum resources based on a spatiotemporal graph Transformer. Background Technology
[0002] With the development of low-Earth orbit satellite constellations and broadband satellite communication systems, satellite spectrum resources are characterized by high dynamism and strong uncertainty. The scarcity of spectrum resources is becoming increasingly prominent. Spectrum demand fluctuates dramatically with orbital transits, beam switching, weather conditions, and ground service load. Traditional prediction methods based on statistical modeling (such as ARIMA) and recurrent neural networks (such as LSTM and GRU) have the following shortcomings:
[0003] 1. Limited long-term capture capability: When predicting spectrum occupancy for the next 30 minutes to 1 hour, the recursive prediction error accumulates significantly.
[0004] 2. Difficulty in integrating multi-source heterogeneous features: Satellite ephemeris, rain attenuation, ionospheric disturbances, interference observations, etc., are difficult to model effectively.
[0005] 3. Lack of interpretability and feasibility: The prediction results do not take into account regulatory constraints and physical feasibility, which means that the predicted values cannot be directly used for resource scheduling.
[0006] Therefore, there is an urgent need to research a new method that can model spatiotemporal correlations, integrate physical priors, and output uncertainty intervals to support the dynamic allocation and efficient utilization of satellite spectrum. The Transformer model, due to its excellent performance in natural language processing and time series forecasting due to its self-attention mechanism, provides a new solution for the dynamic prediction and scheduling of satellite spectrum. Summary of the Invention
[0007] The purpose of this invention is to provide a method and apparatus for dynamic allocation of satellite spectrum resources based on spatiotemporal graph Transformer, which has low prediction error, high accuracy, high executability, and high robustness and security.
[0008] The technical solution to achieve the purpose of this invention is: a method for dynamic allocation of satellite spectrum resources based on spatiotemporal graph Transformer, comprising the following steps:
[0009] Step 1: Collect satellite resource data;
[0010] Step 2: Construct a spatiotemporal map of satellite resources;
[0011] Step 3: Establish a spatiotemporal graph Transformer prediction model;
[0012] Step 4: Use the spatiotemporal graph Transformer prediction model to perform multi-task multiquantile prediction;
[0013] Step 5: Use the prediction results to drive the dynamic scheduling of satellite resources.
[0014] Furthermore, the collection of satellite resource data in step 1 is specifically as follows:
[0015] Step 1.1: Obtain historical statistics on spectrum occupancy, traffic throughput, and number of connections from the satellite network management system;
[0016] Step 1.2: Generate satellite position and elevation sequence from orbital dynamics TLE or Kepler element data;
[0017] Step 1.3: Extract rain attenuation, TEC index, and solar activity parameters from meteorological and ionospheric data sources;
[0018] Step 1.4: Obtain spectrum licensing and coexistence rules from the regulatory database.
[0019] Furthermore, the construction of the spatiotemporal map of satellite resources described in step 2 is as follows:
[0020] Step 2.1: Model the satellite, beam, and ground station as graph nodes;
[0021] Step 2.2: Model the inter-satellite links, neighboring beams, and visibility relationships as graph edges and add time-varying weights.
[0022] Furthermore, the establishment of the spatiotemporal graph Transformer prediction model described in step 3 is as follows:
[0023] Step 3.1, Temporal coding: A Transformer encoder with causal masking is used to model the historical sequence;
[0024] Step 3.2, Orbital Position Encoding: The spatial position obtained from satellite ephemeris parameters or TLE conversion is encoded as the temporal input of the Transformer to capture transit periodicity, as detailed below:
[0025] Suppose that the position of the satellite at time t is expressed in Cartesian coordinates calculated from TLE data as follows:
[0026]
[0027] The position encoding function is then defined as:
[0028]
[0029] in, For the satellite velocity vector, The elevation angle of the satellite relative to the ground station. It is a multilayer perceptron. and For learnable parameters, the encoding is the same as the standard Transformer position encoding. Merge into:
[0030]
[0031] in For adaptive fusion coefficients, These are the original input features;
[0032] Step 3.3, Gated Attention: Rain attenuation and ionospheric perturbations are embedded as gating signals into the attention weights to enhance the model's ability to perceive propagation fading, as detailed below:
[0033] In the spatiotemporal graph Transformer model, the core computational unit is the multi-head attention mechanism, whose basic form is as follows:
[0034]
[0035] in, , , These represent the query, key, and value vectors, respectively; the softmax function is used to ensure weight normalization. This is a scaling factor to prevent the vector dot product from becoming too large;
[0036] In multi-head attention mechanisms, the input is first projected onto multiple subspaces through different linear transformations, attention calculations are performed on each subspace, and then the results are concatenated and mapped uniformly.
[0037]
[0038] Where 𝐻 represents the number of attention heads. To output the mapping matrix, each attention head is calculated as follows:
[0039]
[0040] , , Let be the linear projection matrix of the i-th head;
[0041] Incorporating environmental factors as gating signals into attention calculations:
[0042]
[0043] in, It is the environment-gated attention function, calculated from environment parameters:
[0044]
[0045] This is a characteristic of rain attenuation. This is a characteristic of ionospheric disturbance. The flicker index is a characteristic feature. For the sigmoid function, and These are learnable parameters;
[0046] Step 3.4, Regulatory Mask: A spectrum usage rule mask is introduced into the attention calculation to ensure that the prediction results comply with ITU and regional spectrum allocation specifications, as detailed below:
[0047] Regulatory mask Defined as:
[0048]
[0049] Based on the standard MHA, environmental gating and regulatory constraints are introduced to form an improved attention mechanism:
[0050]
[0051] in, It is a propagation environment-gated attention function; It is a mask for regulation and coexistence. It is element-wise multiplication.
[0052] Furthermore, the multi-task, multi-quantile prediction using the spatiotemporal graph Transformer prediction model described in step 4 is as follows:
[0053] Step 4.1: Use the spatiotemporal graph Transformer prediction model to predict the occupancy rate, traffic volume, and interference probability of each beam and subband in the next K time slices;
[0054] Step 4.2: Simultaneously predict the three quantiles P10, P50, and P90 using quantile regression. The loss function for a single task is:
[0055]
[0056] in Quantile loss function:
[0057]
[0058] in For the true value, For the model pair quantiles The predicted value, For the target quantile, These correspond to P10, P50, and P90, respectively.
[0059] Multi-task joint loss function:
[0060]
[0061] The loss functions for each subtask are defined as follows: This represents the loss in the spectrum occupancy prediction task; This indicates the loss from the business volume forecasting task; This represents the loss of the interference probability prediction task; , , , which is a weighting coefficient used to balance the relative importance of the losses of each task.
[0062] Furthermore, the dynamic scheduling of satellite resources driven by the prediction results described in step 5 is as follows:
[0063] Step 5.1: Input the prediction results into the scheduling module, and use the chance-constrained programming method to generate frequency allocation, bandwidth adjustment, and power control schemes; the chance-constrained scheduling problem based on the prediction interval can be expressed as:
[0064]
[0065] Constraints:
[0066]
[0067] in, For the benefit of beamforming in frequency band 5, For the capacity of frequency band 𝑓, Interference between beams 𝑏 and 𝑏′ The maximum allowable interference threshold, and The threshold for the allowed probability of violation; This is a binary decision variable, indicating whether beam 𝑏 is allocated frequency band 𝑓;
[0068] Step 5.2: Return to step 1 to perform rolling updates, thereby achieving closed-loop linkage between prediction and scheduling.
[0069] A satellite spectrum resource dynamic allocation device based on spatiotemporal graph Transformer is provided. The device is used to implement the aforementioned satellite spectrum resource dynamic allocation method. The device includes a data acquisition module, a feature processing module, a prediction module, and a scheduling module.
[0070] The data acquisition module is used to acquire spectrum, orbit, environmental and regulatory data;
[0071] The feature processing module is used to construct the spatiotemporal graph input;
[0072] The prediction module outputs future spectrum demand predictions based on the spatiotemporal graph Transformer.
[0073] The scheduling module generates a dynamic spectrum allocation scheme based on prediction results and chance constraints.
[0074] A mobile terminal includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the aforementioned method for dynamic allocation of satellite spectrum resources.
[0075] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the aforementioned method for dynamic allocation of satellite spectrum resources.
[0076] Compared with the prior art, the significant advantages of this invention are: (1) it can accurately predict future spectrum demand and reduce prediction error; (2) it takes into account orbit periodicity, propagation environment and regulatory constraints to ensure that the results are executable; (3) it outputs prediction intervals to improve the robustness and security of dynamic scheduling. Attached Figure Description
[0077] Figure 1 This is a schematic diagram of the satellite spectrum resource dynamic allocation device based on the spatiotemporal diagram Transformer of the present invention.
[0078] Figure 2 This is a flowchart illustrating the dynamic allocation method for satellite spectrum resources based on the spatiotemporal graph Transformer of this invention. Detailed Implementation
[0079] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.
[0080] like Figure 1 As shown, the present invention provides a satellite spectrum resource dynamic allocation device based on spatiotemporal graph Transformer, comprising a data acquisition module, a feature processing module, a prediction module, and a scheduling module;
[0081] The data acquisition module is used to acquire spectrum, orbit, environmental and regulatory data;
[0082] The feature processing module is used to construct the spatiotemporal graph input;
[0083] The prediction module outputs future spectrum demand predictions based on the spatiotemporal graph Transformer.
[0084] The scheduling module generates a dynamic spectrum allocation scheme based on prediction results and chance constraints.
[0085] like Figure 2 As shown, the present invention provides a method for dynamic allocation of satellite spectrum resources based on a spatiotemporal graph Transformer, comprising the following steps:
[0086] Step 1: Collect satellite resource data;
[0087] Step 2: Construct a spatiotemporal map of satellite resources;
[0088] Step 3: Establish a spatiotemporal graph Transformer prediction model;
[0089] Step 4: Use the spatiotemporal graph Transformer prediction model to perform multi-task multiquantile prediction;
[0090] Step 5: Use the prediction results to drive the dynamic scheduling of satellite resources.
[0091] As a specific example, the collection of satellite resource data in step 1 is as follows:
[0092] Step 1.1: Obtain historical statistics on spectrum occupancy, traffic throughput, and number of connections from the satellite network management system;
[0093] Step 1.2: Generate satellite position and elevation sequence from orbital dynamics TLE or Kepler element data;
[0094] Step 1.3: Extract rain attenuation, TEC index, and solar activity parameters from meteorological and ionospheric data sources;
[0095] Step 1.4: Obtain spectrum licensing and coexistence rules from the regulatory database.
[0096] As a specific example, step 2, which involves constructing a spatiotemporal map of satellite resources, is as follows:
[0097] Step 2.1: Model the satellite, beam, and ground station as graph nodes;
[0098] Step 2.2: Model the inter-satellite links, neighboring beams, and visibility relationships as graph edges and add time-varying weights.
[0099] As a specific example, step 3, which involves establishing a spatiotemporal graph Transformer prediction model, is as follows:
[0100] Step 3.1, Temporal coding: A Transformer encoder with causal masking is used to model the historical sequence;
[0101] Step 3.2, Orbital Position Encoding: The spatial position obtained from satellite ephemeris parameters or TLE conversion is encoded as the temporal input of the Transformer to capture transit periodicity, as detailed below:
[0102] Suppose that the position of the satellite at time t is expressed in Cartesian coordinates calculated from TLE data as follows:
[0103]
[0104] The position encoding function is then defined as:
[0105]
[0106] in, For the satellite velocity vector, The elevation angle of the satellite relative to the ground station. It is a multilayer perceptron. and As learnable parameters, this encoding is similar to the standard Transformer position encoding. Merge into:
[0107]
[0108] Where 𝛼∈[0,1] are the adaptive fusion coefficients, These are the original input features;
[0109] Step 3.3, Gated Attention: Rain attenuation and ionospheric perturbations are embedded as gating signals into the attention weights to enhance the model's ability to perceive propagation fading, as detailed below:
[0110] In the spatiotemporal graph Transformer model, the core computational unit is the multi-head attention mechanism, whose basic form is as follows:
[0111]
[0112] in, , , These represent the query, key, and value vectors, respectively; the softmax function is used to ensure weight normalization. This is a scaling factor to prevent the vector dot product from becoming too large;
[0113] In multi-head attention mechanisms, the input is first projected onto multiple subspaces through different linear transformations, attention calculations are performed on each subspace, and then the results are concatenated and mapped uniformly.
[0114]
[0115] Where 𝐻 represents the number of attention heads. To output the mapping matrix, each attention head is calculated as follows:
[0116]
[0117] , , Let be the linear projection matrix of the i-th head;
[0118] Incorporating environmental factors as gating signals into attention calculations:
[0119]
[0120] in, It is the environment-gated attention function, calculated from environment parameters:
[0121]
[0122] This is a characteristic of rain attenuation. This is a characteristic of ionospheric disturbance. The flicker index is a characteristic feature. For the sigmoid function, and These are learnable parameters;
[0123] Step 3.4, Regulatory Mask: A spectrum usage rule mask is introduced into the attention calculation to ensure that the prediction results comply with ITU and regional spectrum allocation specifications, as detailed below:
[0124] Regulatory mask Defined as:
[0125]
[0126] Based on the standard MHA, environmental gating and regulatory constraints are introduced to form an improved attention mechanism:
[0127]
[0128] in, It is a propagation environment-gated attention function; It is a mask for regulation and coexistence. It is element-wise multiplication.
[0129] As a specific example, step 4 describes using a spatiotemporal graph Transformer prediction model for multi-task multi-quantile prediction, as follows:
[0130] Step 4.1: Use the spatiotemporal graph Transformer prediction model to predict the occupancy rate, traffic volume, and interference probability of each beam and subband in the next K time slices;
[0131] Step 4.2: Simultaneously predict the three quantiles P10, P50, and P90 using quantile regression. The loss function for a single task is:
[0132]
[0133] in Quantile loss function:
[0134]
[0135] in For the true value, For the model pair quantiles The predicted value, For the target quantile, These correspond to P10, P50, and P90, respectively.
[0136] Multi-task joint loss function:
[0137]
[0138] The loss functions for each subtask are defined as follows: This represents the loss in the spectrum occupancy prediction task; This indicates the loss from the business volume forecasting task; This represents the loss of the interference probability prediction task; , , , which is a weighting coefficient used to balance the relative importance of the losses of each task.
[0139] As a specific example, step 5, which uses the prediction results to drive the dynamic scheduling of satellite resources, is as follows:
[0140] Step 5.1: Input the prediction results into the scheduling module, and use the chance-constrained programming method to generate frequency allocation, bandwidth adjustment, and power control schemes; the chance-constrained scheduling problem based on the prediction interval can be expressed as:
[0141]
[0142] Constraints:
[0143]
[0144] in, For the benefit of beamforming in frequency band 5, For the capacity of frequency band 𝑓, Interference between beams 𝑏 and 𝑏′ The maximum allowable interference threshold, and The threshold for the allowed probability of violation; This is a binary decision variable, indicating whether beam 𝑏 is allocated frequency band 𝑓;
[0145] Step 5.2: Return to step 1 to perform rolling updates, thereby achieving closed-loop linkage between prediction and scheduling.
[0146] The present invention also provides a mobile terminal, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the aforementioned dynamic allocation method for satellite spectrum resources.
[0147] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the program, when executed by a processor, implements the steps in the aforementioned method for dynamic allocation of satellite spectrum resources.
[0148] The above are merely preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.
Claims
1. A satellite spectrum resource dynamic allocation method based on a space-time graph Transformer, characterized in that, The method comprises the following steps: Step 1, collecting satellite resource data; Step 2, constructing a space-time graph of satellite resources; Step 3, establishing a space-time graph Transformer prediction model; Step 4, using the space-time graph Transformer prediction model for multi-task multi-quantile prediction; Step 5, using the prediction results to drive the dynamic scheduling of satellite resources.
2. The method of claim 1, wherein the method is based on a spatiotemporal graph Transformer. The satellite resource data collected in step 1 is as follows: Step 1.1, obtaining historical statistical data of spectrum occupancy, traffic throughput and connection number from a satellite network management system; Step 1.2, generating satellite position and elevation sequences from TLE or Kepler element data of orbital mechanics; Step 1.3, extracting rain attenuation, TEC index and solar activity parameters from meteorological and ionospheric data sources; Step 1.4, obtaining spectrum licenses and coexistence rules from a regulatory database.
3. The method of claim 2, wherein the method is based on a spatiotemporal graph Transformer. The space-time graph of satellite resources constructed in step 2 is as follows: Step 2.1, modeling satellites, beams and ground stations as graph nodes; Step 2.2, modeling inter-satellite links, adjacent beams and visibility relationships as graph edges, and adding time-varying weights.
4. The method of claim 3, wherein the method is based on a spatiotemporal graph Transformer. The space-time graph Transformer prediction model established in step 3 is as follows: Step 3.1, time series encoding: using a causal mask Transformer encoder to model historical sequences; Step 3.2, orbital position encoding: encoding the spatial position obtained by converting satellite ephemeris parameters or TLE into the time series input of the Transformer, so as to capture the periodicity of the transit, which is as follows: Let the position of the satellite at time t be represented by the Cartesian coordinates calculated from the TLE data: ； Then the position encoding function is defined as: ； wherein, is the satellite velocity vector, is the satellite elevation angle with respect to the ground station, is a multi-layer perceptron, and are learnable parameters, encoding and standard Transformer positional encoding fused as: ； wherein is an adaptive fusion coefficient, is the original input feature; Step 3.3, gated attention: embedding rain attenuation and ionospheric disturbance as a gating signal into the attention weight to enhance the model's ability to perceive propagation fading, which is as follows: In the space-time graph Transformer model, the core calculation unit is the multi-head attention mechanism, whose basic form is as follows: ； where, , , denote query, key and value vectors respectively; softmax function is used to normalize the weights; is a scaling factor to prevent the inner product of vectors from being too large; In the multi-head attention mechanism, the input is first projected into multiple subspaces through different linear transformations, performs attention calculation respectively, and then maps uniformly after splicing: ； where H is the number of attention heads, is the output mapping matrix, and each attention head calculation method is: ； , , linear projection matrix for the i-th head; Embedding the propagation environment factors as a gating signal into the attention calculation: ； where, is the propagation environment-gated attention function, computed from the environment parameters: ； is a rain fade characteristic, is an ionospheric disturbance characteristic, is a scintillation index characteristic, is a sigmoid function, and are learnable parameters; Step 3.4, regulatory mask: introducing a mask of spectrum usage rules in attention calculation to ensure that the prediction results comply with the International Telecommunication Union and regional spectrum allocation specifications, which is as follows: Regulatory mask is defined as: ； On the basis of the standard MHA, environmental gating and regulatory constraints are introduced to form an improved attention mechanism: ； wherein, is a propagation environment gated attention function; is a regulatory coexistence mask, is an element-wise multiplication.
5. The method of claim 4, wherein, The multi-task multi-quantile prediction using the space-time graph Transformer prediction model in step 4 is as follows: Step 4.1, using the space-time graph Transformer prediction model to predict the occupancy rate, traffic volume and interference probability of each beam and each sub-band in the next K time slices; Step 4.2, using quantile regression to simultaneously predict P10, P50 and P90 three quantile points, and the loss function of a single task is as follows: ； wherein is a quantile loss function: ； wherein is a true value, is a predicted value of the model for the quantile , and is a target quantile, corresponding to P10, P50, P90, respectively. Multi-task joint loss function: ； wherein each sub-task loss function is defined as follows: denotes the loss of the spectrum occupancy prediction task; denotes the loss of the traffic volume prediction task; denotes the loss of the interference probability prediction task; , , are weight coefficients for balancing the relative importance of each task loss.
6. The method of claim 5, wherein the space-time graph Transformer is configured to: The dynamic scheduling of satellite resources using the prediction results in step 5 is as follows: Step 5.1, input the prediction result into the scheduling module, adopt the chance constraint programming method to generate the frequency allocation, bandwidth adjustment and power control scheme; the chance constraint scheduling problem based on the prediction interval can be expressed as: ； Constraint condition: ； wherein, is the yield of beam b at frequency band f, is the capacity of frequency band f, is the interference between beams b and b’, is the allowed maximum interference threshold, and is the allowed violation probability threshold; is a binary decision variable indicating whether beam b is assigned frequency band f; Step 5.2, return to step 1 for rolling update, realize the closed-loop linkage of prediction-scheduling.
7. A satellite spectrum resource dynamic allocation device based on a space-time graph Transformer, characterized in that, The device is used for realizing the satellite spectrum resource dynamic allocation method in any one of claims 1-6, and the device comprises a data acquisition module, a feature processing module, a prediction module and a scheduling module; The data acquisition module is used for acquiring spectrum, orbit, environment and supervision data; The feature processing module is used for constructing a space-time graph input; The prediction module outputs future spectrum demand prediction based on the space-time graph Transformer; The scheduling module generates a dynamic spectrum allocation scheme based on the prediction result and the chance constraint optimization.
8. A mobile terminal comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the program to realize the satellite spectrum resource dynamic allocation method in any one of claims 1-6.
9. A computer-readable storage medium having stored thereon a computer program, characterized in that, The program is executed by the processor to realize the steps in the satellite spectrum resource dynamic allocation method in any one of claims 1-6.