Unmanned aerial vehicle 5g communication resource intelligent scheduling method based on base station large model

Abstract

The application belongs to the technical field of low-altitude economy, and particularly relates to a method for intelligently scheduling 5G communication resources of a UAV based on a base station large model, aiming at the problem of lag of existing 5G resource scheduling, the following scheme is proposed, which comprises the following steps: S1: a base station large model training step: constructing a multi-dimensional training data set, the multi-dimensional training data set comprising historical network operation data, full-amount data of a UAV flight trajectory, UAV business type annotation data, completely changing the hysteresis of traditional responsive scheduling, avoiding network congestion problems in high-density UAV scenarios from the root; accurately guaranteeing low-latency and high-reliability communication of high-priority businesses such as emergency rescue, simultaneously realizing efficient allocation of network resources, greatly improving resource utilization and overall network capacity; continuously improving scheduling accuracy and system risk resistance, providing stable communication support for large-scale UAV cooperative operation, and effectively enabling low-altitude economic development.

Description

Technical Field

[0001] This invention relates to the field of low-altitude economic technology, and in particular to a method for intelligent scheduling of 5G communication resources for unmanned aerial vehicles (UAVs) based on a large-scale base station model. Background Technology

[0002] Chinese patent application number 202311343165.4 discloses a cloud-edge collaborative resource scheduling method, system, and storage medium for unmanned aerial vehicles (UAVs). The method includes: an edge data center acquiring data information collected by the UAV; the edge data center preprocessing the acquired data; splitting the services corresponding to the preprocessed data; and using a resource optimization algorithm to perform cloud-edge collaborative resource scheduling on the split services to obtain an optimal scheduling scheme. The edge data center and the cloud center then complete the corresponding data processing according to the optimal scheduling scheme. This invention also discloses a system for implementing the cloud-edge collaborative resource scheduling method for UAV patrols, and a storage medium including the cloud-edge collaborative resource scheduling method for UAVs. This invention not only achieves cloud-edge collaborative resource scheduling for UAVs but also exhibits high reliability, practicality, and efficiency.

[0003] However, the cloud-edge collaborative resource scheduling method, system, and storage medium for drones also have some problems. For example, when a large number of drones fly in the same area at the same time, they will generate sudden and highly mobile traffic loads on the 5G network, which may lead to network congestion, increased latency, and connection interruption. The traditional 5G resource scheduling mechanism is "responsive", that is, it adjusts after network congestion occurs. This lag cannot meet the stringent requirements of drones for low latency and high reliability communication. Summary of the Invention

[0004] In view of the problem of lag in 5G resource scheduling in the background technology, this invention proposes an intelligent scheduling method for 5G communication resources of UAVs based on a large base station model.

[0005] This invention proposes an intelligent scheduling method for 5G communication resources of unmanned aerial vehicles (UAVs) based on a large-scale base station model, comprising the following steps: S1: Base Station Large Model Training Steps: Construct a multi-dimensional training dataset, which includes historical network operation data, full data of UAV flight trajectories, UAV service type labeled data, airspace environment data, and network congestion case data. Clean, normalize, and extract features from the multi-dimensional training dataset to obtain standardized training samples. Construct a base station large model based on the Transformer architecture. The base station large model includes a feature fusion layer, a prediction and inference layer, and a scheduling decision layer. The feature fusion layer is used to fuse UAV flight features, service features, and network operation features. The prediction and inference layer is used to complete traffic load prediction. The scheduling decision layer is used to output resource scheduling instructions. Iteratively train the base station large model using standardized training samples, and introduce L2 regularization mechanism to suppress model overfitting until the model prediction error is lower than a preset threshold, thus obtaining the trained base station large model. S2: Multi-dimensional traffic prediction steps: Build a drone airspace information collection terminal to collect real-time location, flight speed, flight heading, remaining battery power, mission type, and flight plan of all drones in the current airspace. Simultaneously, collect real-time network operation parameters of each 5G base station cell, including current bandwidth utilization, latency data, packet loss rate, and number of connections. Input the collected drone information and network operation parameters into the trained base station large model. The feature fusion layer completes the association and fusion of multi-dimensional features. The prediction inference layer outputs the communication traffic load prediction results for each 5G base station cell within the next 5-15 minutes. The traffic load prediction results include peak load, load duration, and load distribution range. S3: Predictive Resource Scheduling Steps: Based on traffic load prediction results, a preset intelligent scheduling strategy is initiated to complete the advance allocation and optimization of 5G network resources, specifically including: Dynamic network slicing configuration: Pre-defined drone service priority division rules divide drone services into four priority categories: emergency rescue, critical facility inspection, logistics transportation, and general entertainment. Emergency rescue has the highest priority, and general entertainment has the lowest priority. Based on the traffic load prediction results of each priority service, independent network slices are dynamically configured for different priority services, and corresponding transmission rates, latency thresholds, and reliability indicators are allocated to ensure the exclusive use of network slice resources for high-priority services. Dynamic bandwidth reservation: Based on the predicted peak traffic load of each cell, the bandwidth reservation amount of each cell is calculated. Bandwidth reservation amount = predicted peak traffic load × redundancy coefficient, and the redundancy coefficient is 1.1 to 1.3. Bandwidth is allocated in a coordinated manner for high traffic load cells and adjacent cells. The idle bandwidth of low traffic load cells is temporarily allocated to high traffic load cells to complete the bandwidth reservation configuration and generate bandwidth scheduling instructions. Handover strategy optimization: Based on UAV flight trajectory data and traffic load prediction results, the future flight paths and handover requirements of each UAV are predicted through a large base station model. The target cell for handover is selected, and the signal strength difference and transmission delay difference between the UAV and the target cell and the current serving cell are calculated. A handover trigger threshold is set. When the signal strength difference and transmission delay difference meet the handover trigger threshold, the resource reservation and handover parameter configuration of the target cell are completed in advance, and the optimized handover command is generated. S4: Real-time Closed-Loop Optimization Steps: A network performance monitoring module is built to collect real-time actual network performance data from each cell. This data includes actual traffic load, transmission latency, packet loss rate, connection stability, and resource utilization. The actual network performance data is compared with traffic load prediction results and resource scheduling target parameters to calculate the deviation value. When the deviation value exceeds a preset deviation threshold, the deviation data and actual network performance data are fed back to the base station's large model, initiating a model fine-tuning process to update the model's prediction parameters and scheduling strategies. This achieves continuous optimization of resource scheduling, enabling predictive intelligent scheduling of 5G communication resources based on the base station's large model. This completely changes the lag of traditional responsive scheduling, fundamentally avoiding network congestion problems in high-density drone scenarios. Through differentiated network slicing, dynamic bandwidth reservation, and switching strategy optimization, low-latency, high-reliability communication for high-priority services such as emergency rescue is accurately guaranteed, while simultaneously achieving efficient allocation of network resources, significantly improving resource utilization and overall network capacity. The model's real-time closed-loop optimization and abnormal emergency scheduling mechanism continuously improve scheduling accuracy and system resilience, providing stable communication support for large-scale drone collaborative operations and effectively empowering the development of the low-altitude economy.

[0006] Preferably, in S1, the specific composition of the multi-dimensional training dataset is as follows: historical network operation data includes bandwidth usage data, latency data, packet loss rate data, connection number change data, and network congestion occurrence time and duration data for each cell over the past 6 to 12 months; full data of UAV flight trajectories includes flight path data, position coordinate data, heading change data, and dwell time data for UAVs of different types, flight altitudes, and flight speeds in various airspace environments; UAV service type labeling data includes data packet size, transmission frequency, latency requirements, reliability requirements, and data transmission direction data for various services; airspace environment data includes wind speed, wind direction, precipitation, visibility, and airspace obstacle distribution data; network congestion case data includes the number of UAVs, service type distribution, network resource allocation, congestion mitigation measures, and effectiveness data when congestion occurs. This clarifies the specific composition of the multi-dimensional training dataset, covering multiple types of data such as network, UAVs, and airspace, providing comprehensive and accurate training samples for the base station large model, solidifying the foundation for model prediction and scheduling, and improving the model's ability to understand UAV communication behavior.

[0007] Preferably, in S1, the processing of the multi-dimensional training dataset specifically includes: using an outlier detection algorithm to remove outlier data from the dataset, including bandwidth usage values, latency values, and invalid flight trajectory data that exceed reasonable ranges; using a min-max normalization method to map various types of data to the 0-1 interval to eliminate dimensional differences; extracting key features through a feature extraction algorithm, where UAV flight features include flight speed variance, heading change frequency, and flight altitude gradient; service features include data packet transmission cycle and data volume fluctuation coefficient; network operation features include mean bandwidth usage and latency fluctuation; using a principal component analysis algorithm to perform dimensionality reduction on the extracted key features, retaining principal components with a cumulative contribution rate of not less than 95% as model input features; and optimizing the quality of training samples, eliminating data interference and dimensional differences, simplifying input features, improving the training efficiency and prediction accuracy of the base station large model, and reducing redundant calculations through outlier removal, normalization, and feature dimensionality reduction.

[0008] Preferably, in S1, the training process of the base station large model specifically includes: dividing standardized training samples into training set, validation set, and test set in a 7:2:1 ratio; initializing model parameters, setting the learning rate to 0.001–0.005, the number of iterations to 100–200 rounds, and the batch size to 32–64; calculating the loss between the model's predicted value and the true value using the cross-entropy loss function, updating the model parameters using the Adam optimizer, verifying the model performance using the validation set every 10 rounds, and stopping the iteration when the validation set loss value does not decrease for 5 consecutive rounds; testing the model's performance using the test set, and determining that the model training is complete when the model's traffic load prediction error is below 5% and the scheduling decision accuracy is above 95%; if the preset performance indicators are not met, adjusting the model architecture and hyperparameters, restarting the training process, standardizing the model training process and parameter settings, and ensuring the model's generalization ability through hierarchical validation, overfitting suppression, and performance compliance detection, ensuring that the model's prediction error and scheduling accuracy meet the requirements, and improving the reliability of the scheduling method.

[0009] Preferably, in step S2, the airspace information acquisition terminal acquires data at a frequency of 1–5 Hz, covering the entire service airspace of the corresponding base station. The acquired UAV information and network operation parameters are transmitted to the base station's large model using an encrypted transmission method employing the AES-256 encryption algorithm. The traffic load prediction results also include the traffic proportion of each service type in each cell. The base station's large model enhances the correlation between UAV flight trajectory features and service features through an attention mechanism, improving the accuracy of traffic load prediction. The traffic prediction error for high-priority services is separately managed to ensure that its prediction error is less than 3%. Through high-frequency acquisition, encrypted transmission, and attention mechanism optimization, the transmission security and prediction correlation of UAV information and network data are improved, the prediction error of high-priority services is accurately managed, and the communication stability of high-priority services is guaranteed.

[0010] Preferably, in S3, the dynamic network slice configuration also includes a slice resource dynamic adjustment mechanism, which monitors the resource utilization of each network slice in real time. When the actual traffic of a certain priority service is lower than 70% of the predicted traffic and the duration exceeds 5 minutes, some idle resources of that slice are released and allocated to other high-demand slices. When the actual traffic of a certain priority service exceeds 120% of the predicted traffic, an emergency resource allocation process is initiated, prioritizing the allocation of idle resources from low-priority slices. If there are no idle resources in low-priority slices, the network bandwidth is temporarily expanded to ensure normal service transmission. The newly added slice resource dynamic adjustment mechanism allocates and reclaims slice resources as needed, avoids resource idleness and waste, quickly responds to service traffic fluctuations, takes into account the needs of various priority services, and improves the utilization rate of network slice resources.

[0011] Preferably, in S3, the bandwidth reservation calculation process also includes airspace priority weight adjustment, with preset priority weights for urban core areas, emergency rescue areas, and ordinary airspace at 1.3, 1.2, and 1.0 respectively. The bandwidth reservation amount = traffic load prediction peak × redundancy coefficient × airspace priority weight. For bandwidth allocation between adjacent cells, a greedy algorithm is used to select the optimal allocation scheme to ensure that the load of each cell is balanced after bandwidth allocation, and the transmission delay during the allocation process does not exceed 10ms, avoiding the impact of bandwidth allocation on network stability. The introduction of airspace priority weight, combined with the greedy algorithm to optimize bandwidth allocation, takes into account the needs of key airspaces, achieves cell load balance, controls bandwidth allocation delay, and ensures network transmission stability while improving bandwidth utilization.

[0012] Preferably, in S3, the handover strategy optimization also includes dynamic adjustment of handover timing and a multi-target cell alternative mechanism: based on the flight speed and signal strength change trend of the UAV, the handover trigger threshold is dynamically adjusted. The faster the flight speed, the more lenient the handover trigger threshold setting, and the handover process is started in advance; 2-3 alternative handover target cells are selected for each UAV. The alternative target cells are dynamically sorted according to real-time signal strength, resource occupancy rate and transmission latency. When the primary target cell has insufficient resources or unstable signal, the system immediately switches to the optimal alternative target cell to avoid connection interruption caused by handover failure. The handover timing and alternative cell mechanism are optimized to adapt to the mobile characteristics of UAVs, avoid handover risks in advance, reduce the probability of connection interruption, and ensure the continuity and reliability of communication during high-speed movement of UAVs.

[0013] Preferably, in S4, the real-time closed-loop optimization specifically includes short-term optimization and long-term optimization: Short-term optimization targets single scheduling deviations. When the deviation value exceeds a preset deviation threshold, the current resource scheduling parameters, including bandwidth allocation, slice resource configuration, and switching strategies, are immediately adjusted to quickly compensate for the deviation and ensure network performance meets standards. Long-term optimization, based on historical deviation data and network operation data, performs a comprehensive fine-tuning of the base station's large model weekly, updating the model's feature weights and prediction model to continuously improve the model's prediction accuracy and scheduling adaptability. Simultaneously, the parameter adjustments and network performance changes for each optimization are recorded to form an optimization log, providing data support for subsequent scheduling strategy optimization. Through closed-loop optimization of short-term deviation correction and long-term model fine-tuning, the model's prediction accuracy and scheduling adaptability are continuously improved, optimization experience is accumulated, and dynamic iteration of resource scheduling strategies is achieved to adapt to complex airspace communication scenarios.

[0014] Preferably, the scheduling method further includes an emergency scheduling step: when a sudden drone malfunction is detected, extreme weather occurs in the airspace, or the network suffers a malicious attack that causes a sharp decline in network performance, an emergency scheduling mechanism is immediately triggered to suspend the allocation of resources for ordinary entertainment services, and allocate all idle resources and some resources of low-priority services to emergency rescue and critical facility inspection services; at the same time, the emergency prediction mode of the base station's large model is activated, the prediction cycle is shortened to 1-3 minutes, the prediction frequency is increased, and the resource scheduling strategy is adjusted in real time until the network performance returns to normal or the abnormal situation is resolved; after the abnormal situation is resolved, the resource allocation of various services is gradually restored, transitioning to the regular scheduling mode to avoid large fluctuations in network performance. The addition of emergency scheduling allows for rapid response to sudden malfunctions and extreme situations, prioritizing high-priority services, shortening the prediction cycle, quickly restoring network performance, and improving the robustness and emergency response capability of the scheduling method.

[0015] The beneficial effects of this invention are: 1. From "Responsive" to "Predictive": This has enabled a paradigm shift in communication resource scheduling, moving from passive response to proactive management and fundamentally preventing network congestion. 2. Ensure communication for critical tasks: Through differentiated network slicing and resource reservation, the communication quality of high-priority tasks such as emergency rescue and critical facility inspection is ensured; 3. Improve overall network utilization: Intelligent scheduling enables network resources to be matched more accurately with business needs, avoiding resource waste and improving the capacity and efficiency of the entire 5G network. 4. Enables large-scale drone applications: Provides reliable communication support for the safe and efficient simultaneous flight of tens of thousands of drones in future cities; This invention relies on a large-scale base station model to achieve predictive intelligent scheduling of 5G communication resources, completely changing the lag of traditional reactive scheduling and fundamentally avoiding network congestion problems in high-density drone scenarios. Through differentiated network slicing, dynamic bandwidth reservation, and switching strategy optimization, it accurately ensures low-latency and high-reliability communication for high-priority services such as emergency rescue, while achieving efficient allocation of network resources, significantly improving resource utilization and overall network capacity. The model's real-time closed-loop optimization and abnormal emergency scheduling mechanism continuously improve scheduling accuracy and system resilience, providing stable communication support for large-scale drone collaborative operation and effectively empowering the development of the low-altitude economy. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the workflow proposed in this invention. Figure 2 This is a system block diagram proposed in this invention. Detailed Implementation

[0017] The present invention will be further explained below with reference to specific embodiments.

[0018] Reference Figure 1-2 , Example This embodiment proposes an intelligent scheduling method for 5G communication resources of unmanned aerial vehicles (UAVs) based on a large-scale base station model, which includes the following steps: S1: Base Station Large Model Training Steps: Construct a multi-dimensional training dataset, including historical network operation data, full UAV flight trajectory data, UAV service type labeled data, airspace environment data, and network congestion case data. Clean, normalize, and extract features from the multi-dimensional training dataset to obtain standardized training samples. Build a base station large model based on the Transformer architecture. The base station large model includes a feature fusion layer, a prediction and inference layer, and a scheduling decision layer. The feature fusion layer fuses UAV flight features, service features, and network operation features. The prediction and inference layer performs traffic load prediction, and the scheduling decision layer outputs resource scheduling. Instructions: Iteratively train the large-scale base station model using standardized training samples, introducing L2 regularization to suppress overfitting until the model prediction error falls below a preset threshold, resulting in a fully trained large-scale base station model. The specific composition of the multi-dimensional training dataset is as follows: Historical network operation data includes bandwidth usage, latency, packet loss rate, connection number changes, and network congestion occurrence and duration data for each cell over the past 6-12 months; Full-scale UAV flight trajectory data includes flight path data, position coordinate data, heading change data, and dwell time data for different types of UAVs at different altitudes and speeds in various airspace environments; (The last sentence appears to be incomplete and possibly refers to a separate dataset.) The human-machine service type labeling data includes data packet size, transmission frequency, latency requirements, reliability requirements, and data transmission direction for various services; airspace environment data includes wind speed, wind direction, precipitation, visibility, and airspace obstacle distribution data; network congestion case data includes the number of drones, service type distribution, network resource allocation, congestion mitigation measures, and their effects during congestion. The processing of the multi-dimensional training dataset specifically includes: using outlier detection algorithms to remove outlier data from the dataset, including bandwidth usage values, latency values, and invalid flight trajectory data exceeding reasonable ranges; and using the min-max normalization method to map various data types to... The range of 0 to 1 is used to eliminate dimensional differences. Key features are extracted using a feature extraction algorithm. UAV flight features include flight speed variance, heading change frequency, and flight altitude gradient. Service features include data packet transmission cycle and data volume fluctuation coefficient. Network operation features include mean bandwidth utilization and latency fluctuation. Principal component analysis is used to reduce the dimensionality of the extracted key features, retaining principal components with a cumulative contribution rate of no less than 95% as model input features. The training process of the base station large model specifically includes: dividing standardized training samples into training, validation, and test sets in a 7:2:1 ratio; initializing model parameters and setting the learning rate to 0.001–0.005, the number of iterations is 100-200 rounds, and the batch size is 32-64; the cross-entropy loss function is used to calculate the loss between the model's predicted values ​​and the actual values, and the model parameters are updated through the Adam optimizer. Every 10 rounds, the model performance is validated using a validation set. If the validation set loss value does not decrease for 5 consecutive rounds, the iteration stops. The model performance is tested using a test set. When the model's traffic load prediction error is below 5% and the scheduling decision accuracy is above 95%, the model training is considered complete. If the preset performance indicators are not met, the model architecture and hyperparameters are adjusted, and the training process is restarted. S2: Multi-dimensional Traffic Prediction Steps: A drone airspace information collection terminal is set up to collect real-time data on the location, flight speed, flight heading, remaining battery power, mission type, and flight plan of all drones in the current airspace. Simultaneously, real-time network operation parameters of each 5G base station cell are collected, including current bandwidth utilization, latency, packet loss rate, and connection count. The collected drone information and network operation parameters are input into the trained base station large model. Multi-dimensional feature fusion is achieved through a feature fusion layer. The prediction and inference layer outputs the communication traffic load prediction results for each 5G base station cell within the next 5-15 minutes. The test results include peak load, load duration, and load distribution range. The airspace information acquisition terminal collects data at a frequency of 1–5 Hz, covering the entire service airspace of the corresponding base station. The collected UAV information and network operation parameters are transmitted to the base station's large model using an encrypted transmission method employing the AES-256 encryption algorithm. The traffic load prediction results also include the traffic share of each service type in each cell. The base station's large model enhances the correlation between UAV flight trajectory characteristics and service characteristics through an attention mechanism, improving the accuracy of traffic load prediction. The traffic prediction error for high-priority services is separately managed to ensure that its prediction error is below 3%. S3: Predictive Resource Scheduling Steps: Based on traffic load prediction results, a preset intelligent scheduling strategy is initiated to complete the advance allocation and optimization of 5G network resources, specifically including: Dynamic network slicing configuration: Pre-defined drone service priority division rules divide drone services into four priority categories: emergency rescue, critical facility inspection, logistics transportation, and general entertainment. Emergency rescue has the highest priority, and general entertainment has the lowest priority. Based on the traffic load prediction results of each priority service, independent network slices are dynamically configured for different priority services, and corresponding transmission rates, latency thresholds, and reliability indicators are allocated to ensure the exclusive use of network slice resources for high-priority services. Dynamic bandwidth reservation: Based on the predicted peak traffic load of each cell, the bandwidth reservation amount of each cell is calculated. Bandwidth reservation amount = predicted peak traffic load × redundancy coefficient, and the redundancy coefficient is 1.1 to 1.3. Bandwidth is allocated in a coordinated manner for high traffic load cells and adjacent cells. The idle bandwidth of low traffic load cells is temporarily allocated to high traffic load cells to complete the bandwidth reservation configuration and generate bandwidth scheduling instructions. Handover strategy optimization: Based on UAV flight trajectory data and traffic load prediction results, a large-scale base station model is used to predict the future flight paths and handover needs of each UAV, select target cells for handover, calculate the signal strength difference and transmission delay difference between the UAV and the target cell and the current serving cell, and set a handover trigger threshold. When the signal strength difference and transmission delay difference meet the handover trigger threshold, resource reservation and handover parameter configuration for the target cell are completed in advance, and optimized handover instructions are generated. Dynamic network slice configuration also includes a dynamic adjustment mechanism for slice resources, which monitors the resource utilization of each network slice in real time. When the actual traffic of a certain priority service is lower than 70% of the predicted traffic and the duration exceeds 5 minutes, some idle resources of that slice are released and allocated to other high-demand slices. When the actual traffic of a certain priority service exceeds 120% of the predicted traffic, an emergency resource allocation process is initiated, prioritizing the allocation of idle resources from low-priority slices. If there are no idle resources in low-priority slices, network bandwidth is temporarily expanded to ensure normal service transmission. The bandwidth reservation calculation process also includes airspace priority weight adjustment, with preset priority weights of 1.3, 1.2, and 1.0 for urban core areas, emergency rescue areas, and ordinary airspace, respectively. The bandwidth reservation amount = peak traffic load prediction × redundancy coefficient × airspace priority weight. For bandwidth allocation between adjacent cells, a greedy algorithm is used to select the optimal allocation scheme to ensure that the load of each cell is balanced after bandwidth allocation, and the transmission latency during the allocation process does not exceed 10ms, avoiding the impact of bandwidth allocation on network stability. The handover strategy optimization also includes dynamic adjustment of handover timing and a multi-target cell alternative mechanism: based on the flight speed and signal strength change trend of the UAV, the handover trigger threshold is dynamically adjusted. The faster the flight speed, the more lenient the handover trigger threshold setting, and the handover process is started in advance. Two to three alternative handover target cells are selected for each UAV. The alternative target cells are dynamically sorted according to real-time signal strength, resource occupancy rate, and transmission latency. When the primary target cell has insufficient resources or unstable signal, the UAV immediately switches to the optimal alternative target cell to avoid connection interruption caused by handover failure. S4: Real-time Closed-Loop Optimization Steps: A network performance monitoring module is built to collect real-time actual network performance data from each cell. This data includes actual traffic load, transmission latency, packet loss rate, connection stability, and resource utilization. The actual network performance data is compared with traffic load prediction results and resource scheduling target parameters to calculate the deviation. When the deviation exceeds a preset deviation threshold, the deviation data and actual network performance data are fed back to the base station's large model, initiating a model fine-tuning process. This updates the model's prediction parameters and scheduling strategies, achieving continuous optimization of resource scheduling. Real-time closed-loop optimization specifically includes short-term and long-term optimization: Short-term optimization targets single scheduling deviations. When the deviation exceeds a preset deviation threshold, the current resource scheduling parameters are immediately adjusted, including bandwidth allocation, slice resource configuration, and handover strategies, to quickly compensate for the deviation and ensure network performance meets standards. Long-term optimization is based on historical deviation data and network operation data. A comprehensive fine-tuning of the base station's large model is performed weekly, updating the model's feature weights and prediction model to continuously improve the model's prediction accuracy and scheduling adaptability. Simultaneously, the parameter adjustments and network performance changes for each optimization are recorded, forming an optimization log to provide data support for subsequent scheduling strategy optimization. The scheduling method also includes an emergency scheduling step: when a sudden drone malfunction is detected, extreme weather occurs in the airspace, or the network suffers a malicious attack that causes a sharp decline in network performance, the emergency scheduling mechanism is immediately triggered. This suspends the allocation of resources for ordinary entertainment services and allocates all idle resources and some resources of low-priority services to emergency rescue and critical facility inspection services. At the same time, the emergency prediction mode of the base station's large model is activated, shortening the prediction cycle to 1-3 minutes, increasing the prediction frequency, and adjusting the resource scheduling strategy in real time until network performance returns to normal or the abnormal situation is resolved. After the abnormal situation is resolved, the resource allocation of various services is gradually restored, transitioning to the normal scheduling mode to avoid significant fluctuations in network performance.

[0019] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for intelligent scheduling of 5G communication resources for unmanned aerial vehicles (UAVs) based on a large-scale base station model, characterized in that, Includes the following steps: S1: Base station large model training steps: Construct a multi-dimensional training dataset, which includes historical network operation data, full data of UAV flight trajectories, UAV service type labeling data, airspace environment data and network congestion case data. Clean, normalize and extract features from the multi-dimensional training dataset to obtain standardized training samples. A large base station model is constructed based on the Transformer architecture. The large base station model includes a feature fusion layer, a prediction and inference layer, and a scheduling decision layer. The feature fusion layer is used to fuse UAV flight features, service features, and network operation features. The prediction and inference layer is used to complete traffic load prediction. The scheduling decision layer is used to output resource scheduling instructions. The large base station model is iteratively trained using standardized training samples. An L2 regularization mechanism is introduced to suppress model overfitting until the model prediction error is lower than a preset threshold, thus obtaining the trained large base station model. S2: Multi-dimensional traffic prediction steps: Build a drone airspace information collection terminal to collect real-time location, flight speed, flight heading, remaining battery power, mission type, and flight plan of all drones in the current airspace. Simultaneously, collect real-time network operation parameters of each 5G base station cell, including current bandwidth utilization, latency data, packet loss rate, and number of connections. Input the collected drone information and network operation parameters into the trained base station large model. The feature fusion layer completes the association and fusion of multi-dimensional features. The prediction inference layer outputs the communication traffic load prediction results for each 5G base station cell within the next 5-15 minutes. The traffic load prediction results include peak load, load duration, and load distribution range. S3: Predictive Resource Scheduling Steps: Based on traffic load prediction results, a preset intelligent scheduling strategy is initiated to complete the advance allocation and optimization of 5G network resources, specifically including: Dynamic network slicing configuration: Pre-defined drone service priority division rules divide drone services into four priority categories: emergency rescue, critical facility inspection, logistics transportation, and general entertainment. Emergency rescue has the highest priority, and general entertainment has the lowest priority. Based on the traffic load prediction results of each priority service, independent network slices are dynamically configured for different priority services, and corresponding transmission rates, latency thresholds, and reliability indicators are allocated to ensure the exclusive use of network slice resources for high-priority services. Dynamic bandwidth reservation: Based on the predicted peak traffic load of each cell, the bandwidth reservation amount of each cell is calculated. Bandwidth reservation amount = predicted peak traffic load × redundancy coefficient, and the redundancy coefficient is 1.1 to 1.

3. Bandwidth is allocated in a coordinated manner for high traffic load cells and adjacent cells. The idle bandwidth of low traffic load cells is temporarily allocated to high traffic load cells to complete the bandwidth reservation configuration and generate bandwidth scheduling instructions. Handover strategy optimization: Based on UAV flight trajectory data and traffic load prediction results, the future flight paths and handover requirements of each UAV are predicted through a large base station model. The target cell for handover is selected, and the signal strength difference and transmission delay difference between the UAV and the target cell and the current serving cell are calculated. A handover trigger threshold is set. When the signal strength difference and transmission delay difference meet the handover trigger threshold, the resource reservation and handover parameter configuration of the target cell are completed in advance, and the optimized handover command is generated. S4: Real-time closed-loop optimization steps: Build a network performance monitoring module to collect real-time actual network performance data of each cell. The actual network performance data includes actual traffic load, transmission latency, packet loss rate, connection stability, and resource utilization. Compare the actual network performance data with the traffic load prediction results and resource scheduling target parameters, calculate the deviation value, and when the deviation value exceeds the preset deviation threshold, feed the deviation data and actual network performance data back to the base station big model, start the model fine-tuning process, update the model's prediction parameters and scheduling strategy, and achieve continuous optimization of resource scheduling.

2. The intelligent scheduling method for UAV 5G communication resources based on a large base station model according to claim 1, characterized in that, In S1, the multi-dimensional training dataset is specifically composed as follows: Historical network operation data includes bandwidth usage data, latency data, packet loss rate data, connection number change data, and network congestion occurrence time and duration data for each cell over the past 6 to 12 months; Full data of UAV flight trajectories includes flight path data, position coordinate data, heading change data, and dwell time data for UAVs of different types, at different flight altitudes, and at different flight speeds in various airspace environments; UAV service type labeling data includes data packet size, transmission frequency, latency requirements, reliability requirements, and data transmission direction for various services; Airspace environment data includes wind speed, wind direction, precipitation, visibility, and airspace obstacle distribution data; Network congestion case data includes the number of UAVs, service type distribution, network resource allocation, congestion mitigation measures, and their effects when congestion occurs.

3. The intelligent scheduling method for UAV 5G communication resources based on a large base station model according to claim 1, characterized in that, In S1, the processing of the multi-dimensional training dataset specifically includes: using an outlier detection algorithm to remove outlier data from the dataset, including bandwidth usage values, latency values, and invalid flight trajectory data that exceed reasonable ranges; using a min-max normalization method to map various types of data to the 0-1 interval to eliminate dimensional differences; extracting key features through a feature extraction algorithm, where UAV flight features include flight speed variance, heading change frequency, and flight altitude gradient; service features include data packet transmission cycle and data volume fluctuation coefficient; network operation features include mean bandwidth usage and latency fluctuation value; and using a principal component analysis algorithm to perform dimensionality reduction on the extracted key features, retaining principal components with a cumulative contribution rate of not less than 95% as model input features.

4. The intelligent scheduling method for UAV 5G communication resources based on a large base station model according to claim 1, characterized in that, In S1, the training process of the large base station model specifically includes: dividing standardized training samples into training, validation, and test sets in a 7:2:1 ratio; initializing model parameters, setting the learning rate to 0.001–0.005, the number of iterations to 100–200 rounds, and the batch size to 32–64; calculating the loss between the model's predicted and actual values ​​using the cross-entropy loss function, updating model parameters using the Adam optimizer, validating the model performance using the validation set every 10 rounds, and stopping the iteration when the validation set loss value does not decrease for 5 consecutive rounds; testing the model's performance using the test set, and determining that the model training is complete when the model's traffic load prediction error is below 5% and the scheduling decision accuracy is above 95%; if the preset performance indicators are not met, adjusting the model architecture and hyperparameters, and restarting the training process.

5. The intelligent scheduling method for 5G communication resources of unmanned aerial vehicles based on a large base station model according to claim 1, characterized in that, In S2, the airspace information acquisition terminal collects data at a frequency of 1–5 Hz, covering the entire service airspace of the corresponding base station. The collected UAV information and network operation parameters are transmitted to the base station's large model using an encrypted transmission method employing the AES-256 encryption algorithm. The traffic load prediction results also include the traffic proportion of each service type in each cell. The base station's large model enhances the correlation between UAV flight trajectory features and service features through an attention mechanism, improving the accuracy of traffic load prediction. The traffic prediction error for high-priority services is separately managed to ensure that its prediction error is less than 3%.

6. The intelligent scheduling method for UAV 5G communication resources based on a large base station model according to claim 1, characterized in that, In S3, the dynamic network slice configuration also includes a slice resource dynamic adjustment mechanism, which monitors the resource utilization of each network slice in real time. When the actual traffic of a certain priority service is lower than 70% of the predicted traffic and the duration exceeds 5 minutes, some idle resources of that slice are released and allocated to other high-demand slices. When the actual traffic of a certain priority service exceeds 120% of the predicted traffic, an emergency resource allocation process is initiated, prioritizing the allocation of idle resources from low-priority slices. If there are no idle resources in low-priority slices, the network bandwidth is temporarily expanded to ensure normal service transmission.

7. The intelligent scheduling method for UAV 5G communication resources based on a large base station model according to claim 1, characterized in that, In S3, the bandwidth reservation calculation process also includes airspace priority weight adjustment. The priority weights of the urban core area, emergency rescue area, and ordinary airspace are preset to be 1.3, 1.2, and 1.0, respectively. The bandwidth reservation amount = traffic load prediction peak × redundancy coefficient × airspace priority weight. For bandwidth allocation between adjacent cells, a greedy algorithm is used to select the optimal allocation scheme to ensure that the load of each cell is balanced after bandwidth allocation, and the transmission delay during the allocation process does not exceed 10ms, so as to avoid affecting network stability due to bandwidth allocation.

8. The intelligent scheduling method for UAV 5G communication resources based on a large base station model according to claim 1, characterized in that, In S3, the handover strategy optimization also includes dynamic adjustment of handover timing and a multi-target cell alternative mechanism: based on the flight speed and signal strength change trend of the UAV, the handover trigger threshold is dynamically adjusted. The faster the flight speed, the more lenient the handover trigger threshold setting, and the handover process is started in advance. For each drone, select 2 to 3 alternative target cells for handover. Based on real-time signal strength, resource utilization, and transmission latency, dynamically sort the alternative target cells. When the primary target cell has insufficient resources or unstable signal, immediately switch to the optimal alternative target cell to avoid connection interruption caused by handover failure.

9. The intelligent scheduling method for UAV 5G communication resources based on a large base station model according to claim 1, characterized in that, In S4, the real-time closed-loop optimization specifically includes short-term optimization and long-term optimization: short-term optimization targets single scheduling deviations. When the deviation value exceeds the preset deviation threshold, the current resource scheduling parameters are immediately adjusted, including bandwidth allocation, slice resource configuration and switching strategy, to quickly make up for the deviation and ensure that network performance meets the standards. Long-term optimization is based on historical deviation data and network operation data. The base station big model is comprehensively fine-tuned every week, updating the model's feature weights and prediction model to continuously improve the model's prediction accuracy and scheduling adaptability. At the same time, the parameter adjustments and network performance changes of each optimization are recorded to form an optimization log, which provides data support for subsequent scheduling strategy optimization.

10. The intelligent scheduling method for UAV 5G communication resources based on a large base station model according to claim 1, characterized in that, The scheduling method also includes an emergency scheduling step: when a sudden drone malfunction is detected, extreme weather occurs in the airspace, or the network suffers a malicious attack that causes a sharp decline in network performance, the emergency scheduling mechanism is immediately triggered. This suspends the allocation of resources for ordinary entertainment services and allocates all idle resources and some resources of low-priority services to emergency rescue and critical facility inspection services. At the same time, the emergency prediction mode of the base station's large model is activated, shortening the prediction cycle to 1-3 minutes, increasing the prediction frequency, and adjusting the resource scheduling strategy in real time until network performance returns to normal or the abnormal situation is resolved. After the abnormal situation is resolved, the resource allocation of various services is gradually restored, transitioning to the normal scheduling mode to avoid significant fluctuations in network performance.

Images