Wireless network dynamic planning and optimizing method based on vehicle positioning information

Abstract

The invention relates to the cross technical field of mobile communication and intelligent traffic, in particular to a wireless network dynamic planning and optimizing method based on vehicle positioning information, which comprises the following steps: collecting full-dimensional data through multiple interfaces, and carrying out standardization processing to form a combined spatio-temporal data set of vehicle, network and traffic combination; based on the joint spatio-temporal data set, constructing a double-layer prediction model of the individual trajectory and the macroscopic flow, and outputting a network load thermodynamic diagram in a future preset duration range; constructing a network digital twin system covering the whole target area, verifying an execution effect of the optimization strategy in a future preset duration range, and outputting an optimal optimization strategy; based on the optimal optimization strategy, a standardized optimization instruction is issued, and multi-dimensional resource dynamic scheduling is carried out; and calculating a deviation ratio between the optimized network performance and the predicted performance through comparison of previous and later data, judging whether the deviation ratio reaches an iteration triggering condition or not, and if so, carrying out iteration operation. The method can adapt to long-term traffic scene change and business upgrade.

Description

Technical Field

[0001] This invention relates to the field of mobile communication and intelligent transportation technologies, specifically to a method for dynamic planning and optimization of wireless networks based on vehicle location information. Background Technology

[0002] With the rapid development of the intelligent connected vehicle industry, vehicles have been upgraded from traditional means of transportation to "mobile intelligent terminals" that require continuous and highly reliable network connections. Core functions of high-level autonomous driving, such as real-time environmental perception, in-vehicle high-definition video streaming, and remote vehicle control, place stringent demands on network performance.

[0003] However, existing wireless network planning methods have significant limitations and cannot adapt to the dynamic characteristics of vehicle-to-everything (V2X) networks. Specific shortcomings are as follows: The data dimension is too limited and the prediction accuracy is insufficient: the existing technology only relies on network traffic statistics for load prediction, without considering V2X high-precision positioning (error <1m) and vehicle motion status data, resulting in network traffic prediction errors generally >25%, and failing to identify "micro mobile hotspots formed by the temporary gathering of 10-20 vehicles", which can easily cause local communication congestion.

[0004] The optimization strategy is static and has poor dynamic adaptability: the existing method adopts the "fixed time period parameter adjustment" mode, which cannot cope with sudden traffic events. The vehicle switching failure rate is as high as 8% in high-speed scenarios, which seriously affects the continuity of autonomous driving business.

[0005] Lack of simulation verification and high trial-and-error costs: Existing methods directly apply the prediction results to the existing network adjustment without verifying the effectiveness of the strategy through virtual simulation. More than 30% of parameter adjustment operations cause short-term network outages due to "incompatibility with the existing network environment" (such as excessive adjustment of cell individual offset (CIO) leading to a large number of vehicle handover failures), significantly increasing the operator's operating costs.

[0006] In summary, traditional static or semi-static wireless network planning methods cannot solve the problems of "spatial mobile hotspots" and "temporal tidal loads" in vehicle-to-everything (V2X) networks. Summary of the Invention

[0007] The present invention aims to provide a method for dynamic planning and optimization of wireless networks based on vehicle positioning information, in order to solve the problems of existing network data having a single dimension, insufficient prediction accuracy, static optimization strategies, poor dynamic adaptability, lack of simulation verification, and high trial and error costs.

[0008] The wireless network dynamic planning and optimization method based on vehicle positioning information in this solution includes the following steps: Step 100: Collect full-dimensional data through multiple interfaces and standardize the data to form a joint spatiotemporal dataset that combines vehicles, networks, and traffic. Step 200: Construct a two-layer prediction model of individual trajectories and macro traffic based on the joint spatiotemporal dataset, and output a network load heat map within a preset time range in the future from the two-layer prediction model. Step 300: Construct a network digital twin system covering the entire target area, verify the execution effect of the optimization strategy within a preset time range through virtual simulation, and output the optimization strategy that meets the set conditions as the optimal optimization strategy. Step 400: Based on the optimal optimization strategy, issue standardized optimization instructions to perform multi-dimensional dynamic resource scheduling, including wireless parameter self-optimization, dynamic allocation of physical resource blocks, network slice reconfiguration, and enhanced mobility management. Step 500: Monitor the optimized network performance and vehicle positioning data in real time. Through the simulation verification in step 300 and the data comparison of the actual optimized configuration in step 400, calculate the deviation rate between the optimized network performance and the predicted simulation performance, and determine whether the deviation rate reaches the iteration trigger condition. If yes, perform the iteration operation according to the preset steps; otherwise, continue monitoring.

[0009] The beneficial effects of this plan are: By integrating multi-source data and employing a two-layer prediction model, network traffic prediction errors are reduced, enabling early identification of micro-mobile hotspots and replacing passive response with proactive optimization. Dynamic resource allocation optimizes load spatiotemporal balance, improves base station PRB utilization, and significantly reduces operator expenses. Predictive handover and coverage optimization enhance the success rate of high-speed scenario handover, reduce end-to-end latency for autonomous driving services, meet the critical business needs of vehicle-to-everything (V2X) communication, and ensure the continuity of critical services. A self-evolving closed-loop optimization system is constructed to adapt to long-term changes in traffic scenarios and business upgrades.

[0010] Furthermore, in step 100, the standardization process is as follows: Data cleaning was performed to remove outliers with GNSS positioning errors greater than 1m and invalid network data with a BLER greater than 10%. Linear interpolation was used to fill in missing traffic flow data for ≤1 minute. Timestamp alignment uses network-side data timestamps as a benchmark and linear interpolation to unify all dimensions of data to the same time dimension. Data association involves linking vehicle motion data with user-level network data based on the vehicle's unique ID, and linking joint data with road grids based on latitude and longitude coordinates, forming a five-dimensional structured joint spatiotemporal dataset that includes vehicle ID, time, location, network status, and traffic scenario.

[0011] The beneficial effect is that data processing can unify the data format of different interfaces, providing a foundation for subsequent processing.

[0012] Furthermore, in step 200, the two-layer prediction model includes: Individual trajectory prediction: For key vehicles, a long short-term memory network model combined with the Adam optimizer is used to predict and output the latitude and longitude sequence within a preset time range in the future; Macro-traffic prediction: The coverage area is divided into grids of preset size. The uniformity of network coverage is maintained within each grid. Different models are configured according to different road scenarios to perform macro-traffic prediction, generating vehicle density and network service demand for each grid, and presenting them in the form of a color heatmap.

[0013] The beneficial effects are: starting from the prediction of individual vehicles and the overall macro traffic flow, it is possible to monitor target vehicles from details to the whole, thereby improving the accuracy of subsequent optimization.

[0014] Furthermore, in step 200, the road scenario includes highway scenarios and urban arterial road scenarios. When it is a highway scenario, a combination model of LSTM and graph neural network is used, and the road topology is used to capture the movement pattern of vehicle groups and predict the vehicle density within the grid. When it is an arterial road scenario, a spatiotemporal convolutional network combined with an attention mechanism is used to assign weights to real-time traffic events and predict macro traffic flow.

[0015] The beneficial effects are: different traffic scenarios can be predicted using different model combinations to accurately predict the flow of dynamically moving vehicles under different conditions.

[0016] Furthermore, in step 200, the graph neural network assigns influence weights to different road elements, including service areas / toll stations, interchanges, and the number of main lanes. The LSTM first outputs the trajectory of a single vehicle for a preset future time. The graph neural network binds the road topology weights to the grid, combines the individual trajectories of the LSTM, counts the number of vehicles entering each grid, and then corrects the data according to the influence weights to output the final grid vehicle density. The attention mechanism assigns influence weights to different events, satisfying an influence weight ≥ 0.6.

[0017] The beneficial effect is that by dynamically adjusting the weights for different scenarios, it can adapt to changes in traffic scenarios and continuously ensure prediction accuracy.

[0018] Furthermore, in step 300, the construction process of the network digital twin system is as follows: Import electronic maps, base station engineering parameters, and wireless propagation models, and import the output network load heat map and real-time traffic event data into the twin system to construct a virtual network environment that maps 1:1 to the physical world.

[0019] The beneficial effect is that twin systems built based on multiple data sources can facilitate simulation and optimization design.

[0020] Furthermore, in step 300, Monte Carlo simulation is used to simulate the execution effect of different optimization strategies within a future preset time range, and at least two optimization strategies are designed for potential congestion areas. The simulation results are then verified according to KPI evaluation indicators. The setting conditions are: simultaneously satisfying latency ≤ 20ms, disconnection rate ≤ 0.5%, and SINR ≥ 10dB.

[0021] The beneficial effect is that simulation design of optimization strategies within a preset time range can improve the accuracy of the optimization strategies.

[0022] Furthermore, in step 400, the multi-dimensional resource dynamic scheduling follows a priority order of first consolidating basic guarantees and then performing fine-grained optimization, performing coverage optimization adjustments; then performing load balancing adjustments; then performing network slice reconfiguration; and finally implementing mobility management enhancements.

[0023] The beneficial effects are: by setting the priority order of scheduling, it is possible to avoid network fluctuations caused by operational conflicts, ensure stable operation, and guarantee the stability of network operation.

[0024] Furthermore, in step 500, the iteration triggering condition is: a single deviation rate > a preset threshold, or the deviation rates of three consecutive time slices are all greater than the preset threshold.

[0025] The beneficial effect is that triggering iteration when the deviation is large can improve the adaptability of the model to changing scenarios.

[0026] Furthermore, in step 500, the preset step is as follows: The two-layer prediction model was retrained based on the latest full-dimensional data collected before the iteration. The two-layer prediction model was fully retrained, including adjusting the number of hidden layer nodes for individual trajectory prediction (256-512 nodes) and optimizing the road topology weights and traffic event weights for macro traffic prediction. Adjust the wireless parameter self-optimization parameters, and combine the deviation source location before iteration to fine-tune the core configuration of wireless parameter self-optimization in step 400 in a scenario-based manner. The deviation sources include unrelieved congestion and coverage blind spots. After iteration, the effect is verified by continuously monitoring for two time slices, simultaneously collecting actual network performance KPIs and vehicle positioning data, and comparing the simulated performance values ​​output by the retrained two-layer prediction model. If the deviation rate between the actual performance and the simulated value is less than or equal to the preset threshold, and the core KPIs meet the standards, the verification is successful, and the regular monitoring process is resumed. If the deviation rate is still greater than the preset threshold or the KPIs do not meet the standards, the process of retraining the two-layer prediction model, adjusting the wireless parameter self-optimization parameters, and verifying the effect after iteration is repeated until the effect meets the standards.

[0027] The beneficial effect is that by retraining the model and adjusting and optimizing the parameters, the system can be adapted to changes in the scenario over a long period of time. Attached Figure Description

[0028] Figure 1 This is a flowchart of an embodiment of the wireless network dynamic planning and optimization method based on vehicle positioning information of the present invention; Figure 2 This is a schematic diagram of multi-source data perception and fusion in an embodiment of the wireless network dynamic planning and optimization method based on vehicle positioning information of the present invention; Figure 3 This is a schematic diagram of the network traffic two-layer prediction model in an embodiment of the wireless network dynamic planning and optimization method based on vehicle positioning information of the present invention; Figure 4 This is a simulation verification platform architecture diagram based on digital twins in an embodiment of the wireless network dynamic planning and optimization method based on vehicle positioning information of the present invention. Figure 5 This is a schematic diagram illustrating dynamic network resource optimization and adjustment in an embodiment of the wireless network dynamic planning and optimization method based on vehicle positioning information of the present invention. Detailed Implementation

[0029] The following detailed description provides further details on specific implementation methods.

[0030] A method for dynamic programming and optimization of wireless networks based on vehicle positioning information, such as Figure 1 As shown, it includes the following steps: Step 100, Multi-source data perception and fusion, such as Figure 2 As shown, multi-dimensional data is collected through multiple interfaces and standardized to form a joint spatiotemporal dataset that combines vehicles, networks, and traffic.

[0031] The interfaces for collecting data include: V2X Interface: Collects vehicle data via On-Board Unit (OBU) and Roadside Unit (RSU), including: high-precision GNSS positioning data (latitude, longitude, altitude, positioning error <1m, sampling frequency 10Hz) and motion status data (vehicle speed, acceleration, heading angle, data accuracy ±0.1). ).

[0032] Network-side interface: Collect network performance data through base station (gNodeB) and core network probe, including: real-time cell load rate, number of users, channel quality indication, handover success rate (HOSR), and bit error rate (BLER, threshold ≤10%).

[0033] Traffic Information Platform Interface: Obtain macro-level traffic data through API interface, including: average road speed (accuracy ±1km / h), congestion index (0-10 levels), traffic flow (vehicles / hour), and real-time events (road accidents, major events, temporary traffic control, with an event impact radius of 500m).

[0034] The standardization process is as follows: Data cleaning removes outliers with GNSS positioning errors > 1m and invalid network data with BLER > 10%. Outliers are corrected using differential positioning technology, and missing traffic flow data for ≤ 1 minute is supplemented using linear interpolation. Timestamp alignment: Based on the timestamp of network-side data, V2X interface data, network-side interface data, and traffic information platform interface data are unified to the same time dimension through linear interpolation. That is, the full-dimensional data includes V2X interface data, network-side interface data, and traffic information platform interface data. Data association involves linking vehicle motion data with user-level network data based on the vehicle's unique ID, and linking joint data with road grids based on latitude and longitude coordinates, forming a five-dimensional structured joint spatiotemporal dataset with vehicle ID, time, location, network status, and traffic scene. The road grid size is 500m×500m.

[0035] Step 200: Construct a two-layer prediction model based on the joint spatiotemporal dataset, combining individual trajectories and macroscopic traffic flow. This model outputs a network load heatmap within a preset future timeframe (5-15 minutes). This preset timeframe ensures that the dynamic optimization of the vehicle-to-everything (V2X) scenario is performed within the optimal window, avoiding significant prediction errors beyond the preset timeframe. Figure 3 As shown, the two-layer prediction model includes: Individual trajectory prediction: For key vehicles currently performing high-bandwidth services (bandwidth ≥ 100Mbps, such as in-vehicle high-definition video transmission and real-time interaction of autonomous driving data), trajectory prediction is achieved by using a Long Short-Term Memory (LSTM) network model combined with the Adam optimizer.

[0036] 1) Simplified design of core components LSTM model: Input 30 seconds of historical data, including velocity, acceleration, and heading angle. 64 samples are trained each time. It adopts a 2-layer structure with 256 units per layer and outputs the latitude and longitude sequence for the next 5-15 minutes.

[0037] The Adam optimizer starts with an initial learning rate of 0.001 and optimizes parameters based on momentum and adaptive learning rate to avoid inefficient or oscillating training. The optimized parameters include: 1) weights and biases for the gating mechanisms of each LSTM layer (two layers in total, 256 units each), including parameters related to input gates, forget gates, output gates, and cell state updates, used to capture temporal features of vehicle motion from historical data such as velocity, acceleration, and heading angle; 2) inter-layer connection parameters, i.e., the connection weights from the output of the first LSTM layer to the input of the second LSTM layer, ensuring effective transfer of motion features; and 3) output layer parameters, including weights and biases for converting the output of the second LSTM layer into future latitude and longitude sequences, directly affecting trajectory prediction accuracy. The specific values ​​of these optimized parameters are dynamically adjusted according to actual conditions and will not be elaborated further here.

[0038] 2) Collaborative training (1000 iterations in total) The optimization is performed using a 4-step loop, processing 64 samples each time, specifically as follows: Forward propagation: LSTM uses historical data to calculate predicted latitude and longitude; Loss calculation: Compare the predicted latitude and longitude with the actual latitude and longitude using the mean square error (MSE) to calculate the error; Backpropagation: Based on the calculated error, determine the adjustment direction (gradient) of the LSTM parameters (weights, biases); Parameter updates: Adam uses gradients, combined with momentum and adaptive learning rate, to adjust LSTM parameters and reduce error.

[0039] Simultaneously, Dropout (dropout rate 0.2) is added to prevent overfitting.

[0040] After 1000 iterations of optimization, the LSTM model outputs the latitude and longitude sequence of the target vehicle for the next 5-15 minutes, with the prediction error stabilizing within 5m, providing high-precision individual trajectory data for subsequent network load heatmap generation and dynamic resource scheduling.

[0041] Macro-level traffic prediction: The coverage area is divided into grids of a preset size, 500m × 500m. Uniform network coverage is maintained within each grid. Different models are configured for macro-level traffic prediction based on different road scenarios, generating vehicle density (vehicles / grid) and network service demand (Mbps / grid) for each grid, presented as a color heatmap. Road scenarios include highways and urban arterial roads.

[0042] When dealing with highway scenarios, a combined LSTM and Graph Neural Network (GNN) model is employed. The model leverages road topology (interchanges, service areas) to capture vehicle movement patterns and predict vehicle density within a grid cell. The GNN assigns "influence weights" to different road elements, with the core rules being: service areas / toll stations: weight 0.7-0.8; interchanges: weight 0.5-0.6; number of main lanes: weight 0.3-0.4. These influence weights are derived from historical traffic flow data. For example, if the vehicle density around a service area is high, the corresponding weight is automatically increased. The LSTM first outputs the trajectory of a single vehicle for the next 5-15 minutes. The GNN binds the "road topology weights" to a 500m×500m grid cell. Combining the individual trajectories from the LSTM, the number of vehicles entering each grid cell is calculated and then adjusted according to the GNN weights, which are the influence weights. The final grid cell vehicle density is then output.

[0043] When dealing with main road scenarios, a Spatiotemporal Convolutional Network (STCNN) combined with an attention mechanism is used to assign weights (event impact weight ≥ 0.6) to real-time traffic events such as traffic accidents and traffic lights. The attention mechanism assigns "impact weights" to different events, ensuring an impact weight ≥ 0.6: Traffic accidents: weight 0.8-0.9; Traffic lights (morning rush hour): weight 0.6-0.7; Temporary traffic control (such as construction): weight 0.7-0.8. The impact weight is calculated as "event impact duration × coverage area". For example, an accident with an impact of 20 minutes has a higher weight than a traffic light with an impact of 10 minutes. Input historical data (grid traffic flow and real-time events over the past 30 minutes); STCNN extracts "spatiotemporal features" (such as "morning rush hour intersection traffic flow patterns"); The attention mechanism allows STCNN to focus on "high-weight events", such as "accident grids" which are calculated first; The prediction results are corrected based on the event weights; The final grid network service requirements are output.

[0044] Taking a certain intersection (grid A) during the morning rush hour in the urban area as an example: Input: Traffic flow in grid A over the past 30 minutes (120-180 vehicles / 5min), real-time events (traffic lights at intersection, construction in adjacent grid B); Weighting calculation: Based on "duration × range", traffic lights (15 minutes × 1 grid, weight 0.65) and construction (20 minutes × 2 grids, weight 0.78) prioritize attention to construction. STCNN extracts the following pattern: During the morning rush hour, traffic flow increases by 10% every 5 minutes, and construction will divert 20% of the traffic flow to grid A. Corrected forecast: The initial forecast for service demand in grid A was 240Mbps. After weighting adjustments (7.8Mbps for traffic lights and 28.1Mbps for construction), the final forecast is 276Mbps. Output: Operators allocate resources in advance according to the network service demand of 276Mbps to avoid congestion.

[0045] To adapt to changes in traffic scenarios and continuously ensure prediction accuracy, a dual mechanism of "short-cycle parameter iteration + long-cycle structural optimization" is adopted to simultaneously update the models for two types of scenarios: highways (LSTM+GNN) and main roads (STCNN+attention mechanism).

[0046] Step 300, as follows Figure 4 As shown, a network digital twin system covering the entire target area is constructed. The execution effect of the optimization strategy within a preset time range is verified through virtual simulation. The optimization strategy is based on the congested areas identified in step 200 based on the two-layer prediction model, the important network services in each grid, and the flow direction of vehicles. Multiple candidate strategies are generated based on network resource means (parameters / PRB / slicing / switching). The optimization strategy that meets the set conditions is output as the optimal optimization strategy. The set conditions are: latency ≤ 20ms, disconnection rate ≤ 0.5%, and SINR ≥ 10dB. The effectiveness of the optimization strategy is verified through virtual simulation, the optimal solution is selected, and the risk of trial and error in the live network is avoided.

[0047] Step 301, the construction process of the network digital twin system is as follows: Import electronic maps, base station engineering parameters, and wireless propagation models. The electronic maps include road topology, building height, and vegetation distribution. The base station engineering parameters include location, transmit power of 43dBm, azimuth angle, and downtilt angle. The wireless propagation models include the COST-Hata model used in highway scenarios and the Okumura-Hata model used in urban scenarios. Import the output network load heat map and real-time traffic event data into the twin system to construct a virtual network environment that maps 1:1 to the physical world.

[0048] The core of building a network digital twin system is to create a 1:1 virtual environment based on the "geography, network, and dynamic events" of the physical world, which can be achieved in five simple steps: 1) Establish a basic geographical framework Using high-precision electronic maps (accuracy ≤ 1 meter), geographical information such as roads (number of lanes, curves), buildings (height, material), and vegetation is made into a three-dimensional grid model, which serves as the "base map" of the virtual environment.

[0049] 2) Build a virtual base station cluster The parameters of the real base stations (location, transmit power 43dBm, azimuth angle, downtilt angle) are "transferred" into the virtual base map according to the actual coordinates. Each real base station corresponds to one virtual base station with completely identical parameters.

[0050] 3) Define signal propagation rules Select model based on scenario: In high-speed scenarios, the COST-Hata model is used to simulate signal attenuation in open areas with a small amount of vegetation. In urban scenes, the Okumura-Hata model is used to simulate signal loss due to tall buildings; this ensures that the signal propagation in the virtual environment follows the same laws as in reality (for example, tall buildings will weaken the signal).

[0051] 4) Integrating dynamic data Put the result of step 200 into the virtual environment: Load heatmap: Load is marked by grid (red = high load, green = low load); Traffic incidents: Mark the impact range and duration at the accident location and update the status in real time.

[0052] 5) Virtual / Real Calibration By comparing virtual and real data (such as signal strength and load rate), fine-tuning parameters (for example, if the virtual signal is 5dB weaker, add 5dB of loss) to ensure that the deviation is ≤5% and achieve a 1:1 accurate mapping.

[0053] Step 302: Monte Carlo simulation is used, with ≥1000 iterations to simulate the execution effect of different optimization strategies within a preset time range in the future. For potential congested areas, at least two optimization strategies are designed. The simulation results are verified according to KPI evaluation indicators, which are the signal-to-interference-plus-noise ratio (SINR) distribution, handover failure risk rate, network throughput, end-to-end latency, and disconnection rate in the congested area.

[0054] Step 400, such as Figure 5 As shown, based on the optimal optimization strategy, standardized optimization instructions are issued to perform multi-dimensional dynamic resource scheduling. This multi-dimensional dynamic resource scheduling follows the priority order of "first solidifying the basic guarantee, then performing fine-grained optimization" to avoid network fluctuations caused by operational conflicts and ensure stable operation. First, coverage optimization adjustments are performed; then load balancing adjustments are performed; next, network slice reconfiguration is performed; and finally, mobility management enhancements are implemented. This requires stable configurations of coverage parameters and CIO after the previous adjustments to accurately predict the handover timing and avoid service interruptions caused by handovers that are too early or too late. Through this layered sequence, the stability of the network is guaranteed step by step from basic operation to fine-grained optimization.

[0055] This includes wireless parameter self-optimization, dynamic allocation of physical resource blocks, network slice reconfiguration, and enhanced mobility management. Among these: Step 401, wireless parameter self-optimization includes: Coverage optimization and adjustments: Core objective: To create a strip-shaped coverage area that conforms to the road, with an overlap area of ​​≤100m and a blind spot coverage rate of <1%; Contextual parameters: Highway corridors: Prioritize adjusting the base station downtilt angle by 1-3° (2-3° for high base stations above 30m, and 1-2° for low base stations 10-20m), with an azimuth deviation of ≤3° (straight road sections) or ≤5° (curved road sections) to avoid signal deviation from the lane; For congested urban areas: focus on optimizing base stations at intersections, with a downtilt angle of 1-2° (to reduce the impact of tall buildings blocking the view), and an azimuth deviation of ≤5°, to ensure continuous coverage between intersections and main roads (e.g., intersections in commercial areas need to have no blind spots within a 300m radius). Verification logic: After adjustment, signals are collected through road test equipment to ensure that the SINR of the target road section is ≥10dB and the proportion of blind spots (SINR<5dB) is reduced to less than 1%.

[0056] Load balancing adjustments: Triggering condition: Real-time load rate of the cell ≥ 85% (based on heatmap prediction in step 200) Parameter configuration and scenario adaptation: Highway scenario: CIO sets 2-3dB (to avoid a sudden increase in load to >80% in adjacent cells), and a handover threshold of -95 to -92dBm (to stabilize the signal and reduce frequent handovers), guiding 20% ​​of vehicles to handover to adjacent cells with a load rate ≤60%; In urban areas: CIOs set the threshold to 4-5 dB (for rapid traffic flow relief), with a switching threshold of -92 to -90 dBm (to withstand signal fluctuations), guiding 25%-30% of vehicles to switch. Anomaly Handling: If the load rate of the congested cell is still >75% after adjustment, adjust the transmit power of the adjacent cell by ±2dB to avoid failure of single parameter adjustment.

[0057] Step 402: Dynamic allocation of Physical Resource Blocks (PRBs) Based on the "grid vehicle density + service type" predicted in step 200, PRBs are allocated according to "hotspot reservation and priority scheduling" to ensure resource utilization ≥ 80% and that critical services are not congested. PRB reservations in hotspot areas: Reserve basis: Calculated at 1.2-1.5 times the predicted vehicle density (e.g., if a highway service area predicts 20 vehicles / 5min, reserve 25% PRB).

[0058] Scenario-based reserve ratio: Highway service areas / interchanges: Reserve 20%-30% of PRB (high probability of sudden traffic congestion); For urban commercial areas / school surroundings: reserve 25%-30% of PRB during morning peak hours (7:00-9:00) and 15%-20% of PRB during off-peak hours (12:00-14:00); Release mechanism: When the actual vehicle density of the grid is less than 60% of the predicted value, the reserved PRB will be released to other high-demand areas within 10 minutes.

[0059] QoS priority scheduling: Service priority ranking: autonomous driving data transmission (latency ≤ 20ms) > in-vehicle high-definition video (bandwidth ≥ 100Mbps) > general data service (latency ≤ 100ms). Resource guarantee rules: Allocate an independent PRB pool (accounting for 15%-20% of the total PRB in the cell) for autonomous driving services to ensure that its PRB utilization rate is not less than 90%; high-definition video services are allocated on demand to avoid competing for autonomous driving resources.

[0060] Step 403: Network slice reconfiguration (dynamic adaptation to services and time periods) Based on the "change in business share" predicted in step 200, the slices are reconfigured according to the principle of "time period adjustment and low latency response" to ensure that the delay in issuing slice instructions is ≤500ms; Time-dimension slice bandwidth adjustment: Time-based configuration: During the morning rush hour (7:00-9:00): "High-definition map download slices" account for 40%-50% of the total bandwidth (to meet the map update needs when vehicles start up), and "sensor data upload slices" account for 20%-30%. During off-peak hours (10:00-17:00): "Sensor data upload slices" will increase to 30%-40% (as the demand for real-time perception in autonomous driving remains stable), while "high-definition map download slices" will decrease to 20%-25%. Evening rush hour (17:00-19:00): "High-definition map download slices" will be restored to 35%-45% (road conditions will be updated for return vehicles); Adaptation logic: Fine-tune based on real-time business volume every 30 minutes, and trigger reconfiguration when the deviation exceeds 10%.

[0061] Slice isolation and fault tolerance: The core slice (autonomous driving) adopts a hard isolation mechanism to prevent other slice resources from crowding out the resources; If a slice fails (such as a sudden drop in bandwidth), a backup slice switch will be triggered within 100ms to ensure that the service interruption time is less than 1s.

[0062] Step 404: Enhanced Mobility Management (Dynamic Scenario Switching Assurance): Based on the vehicle trajectory prediction results from step 200, the handover failure rate is reduced to ≤1% through "predictive handover + conflict avoidance," adapting to the vehicle movement characteristics of highways and urban areas. Predictive switching: Triggering timing: Based on vehicle trajectory prediction (error < 5m), the switching command is issued 1-2 seconds in advance (1.5-2 seconds for highway scenarios, 1-1.5 seconds for urban scenarios, matching vehicle speed differences). Resource preparation: Before the handover, pre-allocate PRBs to the target cell (accounting for 10%-15% of the target cell's idle PRBs) to ensure that there are no resource waiting times when the vehicle connects; Scene optimization: Trigger handover 300m in advance at highway tunnel entrances (due to rapid signal attenuation), and trigger handover 100m in advance at urban intersections (due to vehicle deceleration and longer handover window).

[0063] Switching conflict avoidance: Multi-vehicle concurrent handover processing: A "time-division handover" mechanism is adopted, which segments handover requests by vehicle ID (processing 10 vehicles every 50ms) to avoid the instantaneous load of the target cell exceeding 80%; Conflict rollback strategy: If the target cell load suddenly increases to >85%, the alternative cell handover will be triggered immediately, with a rollback delay of <50ms, to avoid handover failure.

[0064] Each scheduling stage needs to collect execution data (such as switchover success rate and PRB utilization rate) synchronously and feed it back to the closed-loop monitoring in step 500 to ensure that parameter adjustments are continuously adapted to scenario requirements.

[0065] Step 500: Monitor the optimized network performance KPIs and vehicle location data in real time. The network performance KPIs include cell load rate, end-to-end latency, handover success rate, and drop rate. Through simulation verification in Step 300 and data comparison with the actual optimized configuration in Step 400, calculate the deviation rate between the optimized network performance and the predicted simulation performance, and determine whether the deviation rate reaches the iteration trigger condition. If yes, proceed with the iterative operation according to the preset steps; otherwise, continue monitoring. The iteration trigger condition is: a single deviation rate > a preset threshold, or the deviation rate for three consecutive time slices is greater than the preset threshold.

[0066] Deviation rate = |actual value (B) - predicted value (A)| / predicted value (A) × 100%.

[0067] The actual value (B) is the "real KPI data within the same time slice" collected by the base station monitoring system and vehicle terminal after the optimization is performed in step 404; The predicted value (A) is the "KPI prediction result within a preset time period" output when the optimal strategy is verified in step 300 simulation.

[0068] The preset steps are as follows: Retrain the two-layer prediction model: Based on the latest full-dimensional data collected before the iteration (including vehicle location data, network performance KPIs, and traffic event records for the past 7 days), the two-layer prediction model is fully retrained. For the individual trajectory prediction module (LSTM), the number of hidden layer nodes is adjusted to enhance the ability to capture temporal patterns. For the macro traffic prediction module (LSTM+GNN for highway scenarios and STCNN+attention mechanism for urban scenarios), the road topology weights and traffic event weights are optimized to ensure that the "vehicle trajectory and network load heatmap" output by the model accurately matches the current scene features.

[0069] Adjusting Wireless Parameter Self-Optimization Parameters: Based on the root cause analysis of deviations before iteration, the core configuration of wireless parameter self-optimization in step 400 is fine-tuned according to specific scenarios. If the deviation is due to "unrelieved congestion": In highway scenarios, the CIO will be increased from 2-3dB to 3-4dB and the switching threshold will be increased from -95 to -92dBm to -94 to -91dBm. In urban scenarios, the CIO will be increased from 4-5dB to 5-6dB to enhance the vehicle switching guidance effect. If the deviation originates from "coverage blind spots": the downtilt angle of base stations in highway tunnel sections will be reduced from 2°-3° to 1°-2° to expand coverage, and the azimuth deviation of base stations at urban intersections will be reduced from 5° to 4° to ensure that the signal matches the road direction; Iteration effect verification: After each iteration operation, two time slices (5 minutes each, 10 minutes in total) are continuously monitored, and "actual network performance KPIs (cell load rate, end-to-end latency, handover success rate, drop rate)" and "vehicle location data" are collected simultaneously, and the simulation performance values ​​output by the retrained two-layer prediction model are compared. If the deviation rate between the actual performance and the simulation value is less than or equal to the preset threshold (e.g., load rate deviation ≤ 10%, latency deviation ≤ 15%), and the core KPIs are met, then the verification is successful and the normal monitoring process is restored. If the deviation rate still exceeds the preset threshold or the KPI fails to meet the target, repeat the process of retraining the two-layer prediction model, adjusting the wireless parameter self-optimization parameter, and verifying the effect after iteration until the effect meets the target, ultimately ensuring that the system adapts to the dynamic characteristics of scenarios such as traffic flow changes and business demand fluctuations in the long term.

[0070] In the preset steps, the adjustment rules for the same type of deviation root cause are fixed. For example, if the congestion is not relieved, adjust the CIO; if the timing is insufficient, adjust the LSTM node. However, the specific parameter value of each adjustment will change dynamically with the current parameter baseline, the degree of deviation, and the physical / performance upper limit of the parameters. The adjustment amount will not be mechanically repeated with a fixed amplitude. Each adjustment amount is continuously adapted based on the result of the previous iteration to ensure that the adjustment is effective and does not cause new problems, and ultimately achieves the gradual optimization of model and network parameters.

[0071] Compared with existing technologies, the method in this embodiment first increases the diversity of data dimensions through multi-source data fusion and a two-layer prediction model, providing a solid foundation for improving subsequent prediction accuracy and reducing network traffic prediction errors. This allows for the early identification of micro-mobile hotspots, achieving "proactive optimization" instead of "passive response." Next, dynamic resource allocation allows optimization strategies to be dynamically adapted to different application scenarios, optimizing load spatiotemporal balance, improving base station PRB utilization, and significantly reducing operator expenses. Furthermore, predictive handover and coverage optimization improve the success rate of high-speed scenario handover, and corresponding simulation verification is conducted to minimize trial-and-error costs and reduce end-to-end latency for autonomous driving services, meeting the critical business needs of vehicle-to-everything (V2X) and ensuring the continuity of critical services. Finally, a self-evolving closed-loop optimization system is constructed to adapt to long-term changes in traffic scenarios and business upgrades.

[0072] The above descriptions are merely embodiments of the present invention, and common knowledge regarding specific structures and characteristics is not elaborated upon here. It should be noted that those skilled in the art can make various modifications and improvements without departing from the structure of the present invention, and these should also be considered within the scope of protection of the present invention. These modifications and improvements will not affect the effectiveness of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. A method for dynamic planning and optimization of wireless networks based on vehicle positioning information, characterized in that, The method comprises the following steps: Step 100, collecting full-dimensional data through multiple interfaces, and forming a joint space-time data set of vehicles, networks and traffic through standardization processing; Step 200, constructing a double-layer prediction model of individual trajectory and macro flow based on the joint space-time data set, and outputting a network load heat map in a future preset time range from the double-layer prediction model; Step 300, constructing a network digital twin system covering the entire target area, verifying the execution effect of the optimization strategy in the future preset time range through virtual simulation, and outputting an optimization strategy meeting the set conditions as an optimal optimization strategy; Step 400, issuing a standardized optimization instruction based on the optimal optimization strategy, and performing multi-dimensional resource dynamic scheduling, including wireless parameter self-optimization, physical resource block dynamic allocation, network slice reconfiguration, and mobility management enhancement; Step 500, monitoring the network performance and vehicle positioning data after optimization in real time, comparing the data of simulation verification in step 300 and actual optimization configuration in step 400, calculating the deviation rate of the network performance after optimization and the predicted simulation performance, and judging whether the deviation rate reaches an iteration triggering condition; if yes, performing iteration operation according to the preset steps, and if no, continuing to monitor.

2. The method of claim 1, wherein: In the step 100, the standardization processing process is: Data cleaning, excluding abnormal values with GNSS positioning error > 1m and invalid network data with BLER > 10%, and using linear interpolation method to complete traffic flow data missing values ≤ 1 minute; Timestamp alignment, aligning all-dimensional data to the same time dimension through linear interpolation based on network side data timestamp as the reference; Data association, associating vehicle motion data and user-level network data based on vehicle unique ID, and associating joint data and road grid based on latitude and longitude coordinates, forming a five-dimensional structured joint space-time data set of vehicle ID, time, location, network state and traffic scene.

3. The method of claim 1, wherein: In the step 200, the double-layer prediction model comprises: Individual trajectory prediction: for target key vehicles, a long short-term memory network model combined with Adam optimizer is used to predict and output latitude and longitude sequences in a future preset time range; Macro flow prediction: the coverage area is divided into grids of a preset size, the uniformity of network coverage is maintained in each grid, different models are configured according to different road scenes to predict macro flow, vehicle density and network service demand of each grid are generated, and a color heat map is presented.

4. The method of claim 3, wherein: In the step 200, the road scene includes highway scene and urban trunk road scene; when it is highway scene, an LSTM and graph neural network combined model is used to capture vehicle group motion law by using road topology structure to predict vehicle density in the grid; when it is trunk road scene, a spatio-temporal convolution network combined with attention mechanism is used to assign weights to traffic real-time events to predict macro flow.

5. The method of claim 4, wherein: In the step 200, the graph neural network assigns influence weights to different road elements, including service area / toll station, interchange, and number of main lanes. The LSTM first outputs the trajectory of a single vehicle for a preset time period. The graph neural network binds road topology weights to the grid, combines the individual trajectory of the LSTM, counts the number of vehicles entering each grid, and then corrects the influence weights to output the final grid vehicle density. The attention mechanism assigns influence weights to different events and satisfies the condition that the influence weight is greater than or equal to 0.

6.

6. The method of claim 1, wherein: In the step 300, the construction process of the network digital twin system is as follows: Import the electronic map, base station engineering parameters, and wireless propagation model. Import the output network load thermal map and real-time traffic event data into the twin system to construct a virtual network environment that is 1:1 mapped to the physical world.

7. The method of claim 6, wherein: In the step 300, Monte Carlo simulation is used to simulate the execution effect of different optimization strategies in a future preset time period. At least two optimization strategies are designed for potential congestion areas, and the simulation results are verified according to the KPI evaluation index. The set conditions are that the delay is less than or equal to 20 ms, the drop rate is less than or equal to 0.5%, and the SINR is greater than or equal to 10 dB.

8. The method of claim 1, wherein: In the step 400, the multi-dimensional resource dynamic scheduling follows the priority order of first ensuring basic support and then making fine optimization to perform coverage optimization adjustment. Then, load balancing adjustment is performed. Next, network slice reconfiguration is performed. Finally, mobility management enhancement is implemented.

9. The method of claim 1, wherein: In the step 500, the iteration trigger condition is that the single deviation rate is greater than a preset threshold, or the deviation rate of three consecutive time slices is greater than the preset threshold.

10. The method of claim 9, wherein: In the step 500, the preset step is: Retrain the double-layer prediction model based on the latest full-dimensional data collected before iteration. The double-layer prediction model is fully retrained, including adjusting the number of hidden layer nodes for individual trajectory prediction, with the number of hidden layer nodes being 256-512, and optimizing road topology weights and traffic event weights for macro traffic prediction. Adjust the wireless parameter self-optimization parameter. Based on the deviation source positioning before iteration, the core configuration of the wireless parameter self-optimization in step 400 is fine-tuned in the scene. The deviation sources include congestion that is not dredged and coverage blind area. After iteration, the effect is verified. The actual network performance KPI and vehicle positioning data are continuously monitored for two time slices. The simulation performance value output by the retrained double-layer prediction model is compared. If the deviation rate of the actual performance and the simulation value is less than or equal to the preset threshold, and the core KPI meets the standard, the verification is passed, and the normal monitoring process is restored. If the deviation rate is still greater than the preset threshold or the KPI does not meet the standard, the process of retraining the double-layer prediction model, adjusting the wireless parameter self-optimization parameter, and verifying the effect after iteration is repeated until the effect meets the standard.

Images