Large-range wireless network dynamic deployment system and resource optimization method

Abstract

The invention is applicable to the technical field of wireless networks, and provides a large-range wireless network dynamic deployment system and a resource optimization method, and the system comprises a central control unit, a plurality of intelligent access points, a user density prediction module, a staged ad hoc network protocol module and a reinforcement learning resource scheduling module. Through a staged ad hoc network protocol, a wireless Mesh backhaul topology is quickly constructed and access parameter initialization is completed under the conditions that deployment time is limited and backhaul link resources are limited; a user density prediction module is used for predicting future user density levels of areas covered by all access points, and prediction results, channel utilization rates, return loads and other indexes jointly form a state space of multi-agent reinforcement learning; through an improved MADDPG model, the working channel, the transmitting power, the beam direction and the beacon frame sending interval of each access point are jointly optimized. According to the method, the total throughput of the whole network can be improved, and the congestion occurrence rate of the hotspot area is reduced.

Description

Technical Field

[0001] This invention relates to the field of wireless network technology, specifically to a large-scale wireless network dynamic deployment system and resource optimization method. Background Technology

[0002] With the increasing number of large-scale open-air concerts, marathons, and city festivals, it is becoming increasingly common for tens of thousands or even hundreds of thousands of people to gather in limited spaces within a short period of time. In such scenarios, the following typical requirements are placed on wireless networks: 1. High user density and high bandwidth requirements: The user density per unit area is much higher than that in ordinary shopping malls, parks and other scenarios. The number of terminals online at the same time can reach tens of thousands, and users have extremely high bandwidth requirements for services such as live video streaming, social media, and short video uploads.

[0003] 2. Extremely short deployment time: The network usually needs to be deployed and debugged within hours before the event starts and quickly dismantled after the event ends. It cannot be carried out with long-term, meticulous construction and optimization like macro base stations or campus networks.

[0004] 3. Limited backhaul conditions: Due to terrain, temporary structures and construction limitations at the event site, it is often impossible to lay wired backhaul links. Only a limited number of wireless mesh methods can be used for backhaul, which limits both link capacity and reliability.

[0005] 4. Strong dynamic changes in business load: The distribution of people in the venue changes significantly in time and space during the stages of entering, watching together, resting, and leaving. The network load has strong non-stationary characteristics in both time and space, making it difficult to apply traditional optimization strategies designed for long-term stable scenarios.

[0006] 5. High environmental uncertainty: Temporary outdoor activities are often affected by weather, emergencies, etc., which can lead to sudden changes in population distribution patterns, making it difficult for traditional fixed thresholds and static models to adapt.

[0007] In existing technologies, several independent research and commercial solutions exist for wireless mesh topology optimization, AP transmit power and channel scheduling, beamforming coverage enhancement, and machine learning-based wireless resource management. However, these solutions generally suffer from the following shortcomings: 1. Single optimization objective: Most solutions only perform local optimization on a single indicator such as throughput, coverage, or interference, lacking a comprehensive balance between "overall network capacity + edge experience + configuration stability." 2. Dispersed control dimensions and lack of a unified decision-making framework: Existing solutions often design topology planning, access layer parameter adjustment, and beam direction control independently, operating independently and failing to achieve collaborative optimization under a unified state awareness and decision-making framework. 3. Lack of scenario-specific design for temporary large-scale outdoor events: Existing optimization solutions are mostly designed for long-term deployed macro base stations, campus networks, or indoor enterprise networks, without fully considering special constraints such as short deployment time, limited wireless backhaul, and limited number of devices. 4. Lack of prediction and adaptive mechanisms for highly dynamic and uncertain user behavior: Even when some solutions incorporate machine learning methods, they are mostly used for offline planning or long-term statistical optimization, lacking targeted prediction and online adaptive capabilities for the strong transient impacts caused by the rapid gathering and migration of large numbers of people in a short period. 5. Rigid model combinations and lack of adaptive weight adjustment: Existing prediction models mostly use fixed combinations, unable to dynamically adjust the weights of each sub-model based on real-time prediction errors, leading to a significant drop in prediction accuracy during sudden environmental changes. Therefore, a large-scale wireless network dynamic deployment system and resource optimization method are needed to address the above problems. Summary of the Invention

[0008] In view of the shortcomings of the existing technology, the purpose of this invention is to provide a large-scale wireless network dynamic deployment system and resource optimization method to solve the problems existing in the background technology.

[0009] This invention is implemented as follows: a large-scale wireless network dynamic deployment system includes: a central control unit, multiple intelligent access points, a user density prediction module, a phased self-organizing network protocol module, and a reinforcement learning resource scheduling module, wherein: The phased self-organizing network protocol module is configured to control each smart access point to sequentially execute the discovery and registration phase, the topology and channel planning phase, and the access parameter initialization phase after power-on. The topology and channel planning phase adopts an improved minimum spanning tree algorithm based on a preset signal quality threshold for candidate link pre-screening and combined with parallel computation pruning to construct the wireless Mesh backhaul topology. The user density prediction module is configured to: based on the historical user association time series collected by each smart access point, use a combination model of sliding window averaging, exponential smoothing and long short-term memory neural network to predict the user density level and corresponding prediction confidence of the coverage area of ​​each smart access point in the future time period, and output the prediction results to the reinforcement learning resource scheduling module. The reinforcement learning resource scheduling module is configured to: employ an improved multi-agent deep deterministic policy gradient algorithm to construct a state space containing the real-time network status and user density prediction results of each intelligent access point; define an action space containing at least one of the control actions of adjusting the working channel, transmit power, beam direction, and beacon frame transmission interval; and jointly optimize the wireless parameters of each intelligent access point through centralized training and distributed execution, based on a reward function containing the total system throughput, worst-case user latency, configuration handover count, interference suppression index, and prediction confidence index. The central control unit is also configured to: based on the optimization results output by the reinforcement learning resource scheduling module, trigger the update of the wireless Mesh backhaul topology when preset conditions are met through the backhaul load threshold linkage logic, so as to realize cross-layer collaborative optimization of access layer parameters and backhaul network topology.

[0010] Another objective of this invention is to provide a method for dynamic deployment and resource optimization of a large-scale wireless network, the method comprising the following steps: After the multiple smart access points are powered on, they perform discovery and registration and report information to the central control unit. Based on the reported location information and preset link quality conditions, the central control unit executes an improved minimum spanning tree algorithm to perform topology planning and channel allocation, forming an initial wireless mesh backhaul topology. Based on historical user data collected from each smart access point, a combined prediction model is used to generate user density levels and prediction confidence levels for each region in the future time period. Construct a state space and action space for reinforcement learning. The state space includes at least the real-time channel utilization, number of users associated, neighbor channel list, average packet loss rate, backhaul parent node load, user density level, and prediction confidence of each smart access point. The action space includes the adjustment of at least one parameter among the working channel, transmit power, beam direction, and beacon frame transmission interval. Run the improved MADDPG model, select actions based on the current state space, generate wireless parameter configuration results, and determine whether to update the Mesh backhaul topology based on the backhaul load threshold linkage logic, and send the results to each smart access point for execution. Based on network performance feedback, the improved MADDPG model and prediction model are continuously optimized and their parameters are adjusted online.

[0011] As a further aspect of the present invention, in the combined model adopted by the user density prediction module, the average window duration of the sliding window is 60 seconds, the smoothing coefficient of the exponential smoothing is 0.3, the input layer of the LSTM neural network has 10 time steps, and the hidden layer contains 50 neurons.

[0012] As a further aspect of the present invention, the improvements of the improved MADDPG algorithm include: the input layer of the Critic network adds a dimension reflecting the load of the backhaul parent node; and a topology stability penalty term is introduced into the loss function of the Actor network, which is positively correlated with the number of backhaul topology switching times within the statistical period.

[0013] As a further embodiment of the present invention, the smart access point includes an antenna array that supports beamforming; the central control unit is also configured to: determine areas with a density level of or above the third level as potential hotspot areas based on the user density level output by the user density prediction module, and control the corresponding smart access point to point the main lobe of the beam to the area, while narrowing the beamwidth by 20%.

[0014] As a further embodiment of the present invention, the backhaul load threshold linkage logic is configured as follows: when the predicted user density level of the target intelligent access point coverage area exceeds the first preset threshold and the load rate of its current backhaul parent node exceeds the second preset threshold of 85%, a node with a load rate lower than the third preset threshold of 45% is selected from the candidate backhaul nodes as the new backhaul parent node to complete the topology switch, and the switch process takes no more than 2 seconds and there is no service interruption.

[0015] As a further aspect of the present invention, the reward function R used by the reinforcement learning resource scheduling module is expressed as: R = α T_sys-β L_min-γ S_cfg+δ I+ε C_pred_avg, where T_sys is the total system throughput, L_min is the average latency of the worst 5% of users, S_cfg is the number of configuration switching times, I is the interference suppression index, C_pred_avg is the average prediction confidence of the entire network, and α, β, γ, δ, and ε are weighting coefficients.

[0016] As a further aspect of the present invention, when the average prediction confidence C_pred_avg of the entire network is lower than 0.7 and lasts for 3 consecutive statistical periods, a strategy update is triggered, and the weight coefficient α is adjusted to 1.2 and ε is adjusted to 0.02.

[0017] As a further aspect of the present invention, the improved MADDPG model adopts a centralized training and distributed execution architecture. During the training phase, the central control unit centrally updates the policy network and value network based on the overall network status information. During the execution phase, each intelligent access point autonomously selects actions based on its local observation status and the policy network parameters obtained from the central control unit, and receives policy parameter updates every 2 minutes.

[0018] As a further aspect of the present invention, when it is detected that the regional user density level continues to rise within two consecutive statistical periods, the transmission power level of the corresponding smart access point in that region is increased and switched to a less busy working channel 1 to 2 statistical periods in advance. At the same time, the resource allocation of smart access points in non-hotspot areas is reduced, thereby reducing the access congestion rate in hotspot areas.

[0019] Compared with the prior art, the beneficial effects of the present invention are: By incorporating user density prediction results into the state space of the improved MADDPG and applying its decision results to both access layer parameters and Mesh backhaul topology through backhaul load threshold linkage logic, resource pre-configuration based on future population trends is achieved.

[0020] When a continuous upward trend in user density is detected in a certain area, the corresponding AP's transmit power and available channel resources can be increased 1-2 statistical periods in advance, thereby reducing the probability of access congestion in hotspot areas.

[0021] An usable network is quickly formed within a 2-hour deployment constraint using an improved self-organizing network protocol, and then converged to a high-performance configuration within 20 minutes of the start of the activity using an improved MADDPG.

[0022] The improved MADDPG architecture, which adopts centralized training and distributed execution, can continuously accumulate experience through multiple deployments in different activities and venues, and gradually form an experience strategy library adapted to such temporary large-scale outdoor events.

[0023] With the newly added adaptive weight adjustment mechanism, the system can dynamically optimize the weight allocation of each prediction model based on the real-time prediction error, and maintain prediction accuracy when the environment changes abruptly. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of the overall architecture of the system in an embodiment of the present invention.

[0025] Figure 2 This is a flowchart of the three-stage ad hoc network protocol in an embodiment of the present invention.

[0026] Figure 3 This is a schematic diagram of the structure of the improved multi-agent reinforcement learning model in an embodiment of the present invention.

[0027] Figure 4 This is a schematic diagram of dynamic beamforming and hotspot enhancement in an embodiment of the present invention.

[0028] Figure 5 This is a schematic diagram of AP coverage area division based on Thiessen polygons in an embodiment of the present invention.

[0029] Figure 6 This is a schematic diagram of a heatmap for predicting user density in an embodiment of the present invention.

[0030] Figure 7 This is a flowchart of the whole network resource collaborative scheduling process in an embodiment of the present invention. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0032] The specific implementation of the present invention will be described in detail below with reference to specific embodiments.

[0033] like Figure 1 and Figure 2 As shown, this embodiment of the invention provides a large-scale wireless network dynamic deployment system, including: a central control unit, multiple intelligent access points, a user density prediction module, a phased self-organizing network protocol module, and a reinforcement learning resource scheduling module, wherein: The phased self-organizing network protocol module is configured to control each smart access point to sequentially execute the discovery and registration phase (time ≤ 10 minutes), the topology and channel planning phase (30 AP scenarios ≤ 8 minutes, 50 AP scenarios ≤ 15 minutes), and the access parameter initialization phase (time ≤ 5 minutes) after power-on. The topology and channel planning phase adopts an improved minimum spanning tree algorithm based on a preset signal quality threshold for candidate link pre-screening (signal strength ≥ -75dBm, packet loss rate ≤ 1%) and combined with parallel computation pruning to form a backhaul topology and coverage access network based on wireless mesh.

[0034] The user density prediction module is configured to: based on the historical user association time series collected by each smart access point within a predetermined historical time window (≥60 seconds), use a combination model of sliding window averaging (60-second window), exponential smoothing (coefficient 0.3) and long short-term memory neural network (10 time steps in the input layer and 50 neurons in the hidden layer) to predict the user density level (1-5 levels) and prediction confidence (0-1) of the coverage area of ​​each smart access point within a predetermined time window (600 seconds) in the future. The prediction delay is ≤10 seconds, and the MAE is reduced by ≥38% compared with a single LSTM. The user density level and prediction confidence are output as density prediction parameters.

[0035] The reinforcement learning resource scheduling module is configured to: employ an improved multi-agent deep deterministic policy gradient algorithm to construct a state space containing the real-time network state and user density prediction results of each intelligent access point; define an action space containing at least one of the control actions of adjusting the working channel, transmit power, beam direction, and beacon frame transmission interval; and jointly optimize the wireless parameters of each intelligent access point through centralized training and distributed execution, based on a reward function containing the total system throughput, worst-case user latency, configuration handover count, interference suppression index, and prediction confidence index. Specifically, the state space is composed of the channel utilization (U_i), user association number (N_i), neighboring access point channel list (C_i), average packet loss rate (L_i), backhaul parent node load (B_i), and the density prediction parameters of each smart access point. The action space is composed of joint adjustments to the working channel (switching between 2-3 candidate channels), transmit power (4 preset switching levels), beam direction (±10° / ±20° offset), and beacon frame transmission interval (3 switching levels) of the smart access points. The reward function R is constructed based on the total system throughput (T_sys), worst 5% user latency (L_min), configuration handover count (S_cfg), interference suppression index (I), and the average prediction confidence of the entire network (C_pred_avg). An improved MADDPG algorithm (Critic network includes backhaul load dimension, Actor network includes topology stability penalty term) is used to run a multi-agent reinforcement learning model in the central control unit. The model selects actions according to the state space and jointly optimizes the wireless parameters of each smart access point. The model converges within 20 minutes after the start of the activity.

[0036] The central control unit is also configured to: set a scenario adaptation module to dynamically adjust preset thresholds (50,000-person scenario: first preset threshold = AP maximum capacity × 70%; 100,000-person scenario: first preset threshold = AP maximum capacity × 60%); based on the optimization results output by the reinforcement learning resource scheduling module, trigger the update of the wireless Mesh backhaul topology when preset conditions are met through the backhaul load threshold linkage logic, so as to realize cross-layer collaborative optimization of access layer parameters and backhaul network topology.

[0037] It should be noted that the embodiments of this invention involve interdisciplinary technical fields such as wireless communication, self-organizing networks (Mesh), dynamic scheduling of network resources, and reinforcement learning optimization. In particular, they relate to a large-scale, high-capacity wireless network dynamic deployment system and its intelligent resource optimization method for temporary large-scale outdoor events (such as concerts, sporting events, city celebrations, and large gatherings). The embodiments of this invention simultaneously cover wireless Mesh backhaul topology planning, access layer parameter optimization (channel, power, beacon period), beamforming control, and cross-layer collaborative resource scheduling based on multi-agent reinforcement learning. Through innovative designs such as predictive resource pre-optimization, access-backhaul layer linkage decision-making, and adaptive model weight adjustment, it solves the problems of deployment efficiency, dynamic adaptation, and performance bottlenecks in temporary scenarios, and is a typical cross-layer adaptive network system.

[0038] In this embodiment of the invention, the Central Control Unit (CCU) includes: a topology planning module (improved minimum spanning tree), a user density prediction module (combined model), and a multi-agent reinforcement learning resource scheduling module (improved MADDPG). Multiple intelligent access points (APs) include: a wireless access unit, a wireless backhaul unit, a local state acquisition module, and a beam control unit. The phased ad hoc network protocol module includes: Phase 1: Discovery and registration (≤10 minutes), Phase 2: Topology and channel planning (30 APs ≤8 minutes, 50 APs ≤15 minutes), and Phase 3: Access parameter initialization (≤5 minutes). The enhanced user density prediction module includes: sliding window averaging (60-second window), exponential smoothing (coefficient 0.3), an LSTM neural network (10 time steps in the input layer, 50 neurons in the hidden layer), and an adaptive weight adjustment mechanism. The improved MADDPG module includes: a state space, an action space (power, channel, beam, beacon spacing), and a reward function. The dynamic beam control module includes: hotspot area determination: density level ≥ 3; beam adjustment: ±10° / ±20° offset, width reduction of 20%, resulting in a ≥ 25% improvement in signal quality for edge users. The threshold adaptive optimization module includes: Q-learning algorithm to learn the optimal threshold, multi-dimensional threshold matrix, and fuzzy logic control to avoid frequent switching.

[0039] In this embodiment of the invention, for the i-th smart access point, the state space S_i defined by AP_i includes at least the following components: U_i: the current channel utilization rate of the i-th smart access point, ranging from 0 to 1; N_i: the number of user terminals currently associated with the i-th smart access point; C_i: the working channel coding list of neighboring access points within a two-hop range of the i-th smart access point; L_i: the average packet loss rate of the i-th smart access point within a predetermined statistical period (10 seconds); D_i: the predicted user density level (1-5 levels) in the coverage area of ​​the i-th smart access point; B_i: the load rate of the current backhaul parent node of the i-th smart access point (0-1). The state space S_i also includes: C_pred_i: the prediction confidence level of user density in the coverage area of ​​the i-th smart access point (0-1), calculated based on the prediction error of the past 5 statistical periods. When the error rate is ≤5%, the confidence level is ≥0.95, and when the error rate is ≥20%, the confidence level is ≤0.8. For the i-th smart access point, the action space A_i defined by AP_i includes at least one or more of the following control actions: adjusting the transmit power of the i-th smart access point between four preset levels: P1=10dBm, P2=15dBm, P3=20dBm, and P4=25dBm; switching the working channel of the i-th smart access point between two or three pre-allocated candidate channels; adjusting the beam direction of the i-th smart access point within a preset angle offset range of ±10° and ±20° from the current direction; and switching the beacon frame transmission interval of the i-th smart access point between three preset time slot lengths: {40ms, 80ms, and 160ms}.

[0040] In this embodiment of the invention, the reward function R used by the reinforcement learning resource scheduling module is expressed as: R = α T_sys-β L_min-γ S_cfg+δ I+ε C_pred_avg, where T_sys is the total system throughput, L_min is the average latency of the worst 5% of users, S_cfg is the number of configuration handovers, I is the interference suppression index, C_pred_avg is the network-wide average prediction confidence, and α, β, γ, δ, and ε are weighting coefficients. The initial values ​​are set to α=1.0, β=0.5, γ=0.1, δ=0.2, and ε=0.05. When the network-wide average prediction confidence C_pred_avg is below 0.7 for three consecutive statistical periods, a policy update is triggered, adjusting the weighting coefficients α to 1.2 and ε to 0.02.

[0041] In this embodiment of the invention, the user density prediction module adopts a combination of sliding window averaging and exponential smoothing. For the time series of user association numbers within a preset time period (≥60 seconds) of each smart access point, an LSTM neural network model is further used to extrapolate the trend to generate the density prediction parameters. Among them, sliding window averaging is used to smooth short-term noise, exponential smoothing is used to quickly track trends, and LSTM is used to fit nonlinear changes. The three work together to make the prediction delay ≤10 seconds and reduce the MAE by ≥38% compared with the single LSTM model.

[0042] In this embodiment of the invention, when performing topology and channel planning, the phased self-organizing network protocol module uses the minimum spanning tree algorithm to select backbone links from candidate backhaul links with signal strength ≥ -75dBm and packet loss rate ≤ 1% to minimize the average number of hops across the entire network (controlled at 2-3 hops). Based on the tree structure, interference constraints on inter-link channel reuse are considered to allocate backhaul link channels. At the same time, parallel computing and greedy pruning strategies are introduced to ensure that the topology planning time is ≤ 8 minutes in a 30 AP scenario and ≤ 15 minutes in a 50 AP scenario.

[0043] In this embodiment of the invention, each of the intelligent access points includes an antenna array capable of beamforming (supporting ±10° / ±20° angle offset). The central control unit, based on the future user distribution trend output by the user density prediction module (areas with density levels ≥3 are determined as potential hotspots), prioritizes pointing the main lobe of the beam to areas with higher predicted user density, while simultaneously reducing the beamwidth of hotspot areas by 20% to enhance the received signal quality of edge users (improvement ≥25%).

[0044] In this embodiment of the invention, the multi-agent reinforcement learning model adopts a centralized training and distributed execution architecture, and employs an improved multi-agent deep deterministic policy gradient (MADDPG) algorithm. Improvements include: the input layer of the Critic network adds a dimension reflecting the load of the parent node being fed back, avoiding cross-layer decision disconnect; the loss function of the Actor network introduces a topology stability penalty term, which is positively correlated with the number of topology switching events during the statistical period. During the training phase, the central control unit centrally trains the policy network and value network of each intelligent access point based on the overall network state information (learning rate 0.001, discount factor 0.99). During the execution phase, each intelligent access point autonomously selects actions based on local observations and policy network parameters obtained from the central control unit. The model converges within 20 minutes after the activity begins.

[0045] In this embodiment of the invention, the central control unit is further configured to: while generating the wireless parameter configuration result of the smart access point based on the multi-agent reinforcement learning model, trigger the "backhaul load threshold linkage logic": when the predicted user density level of the coverage area of ​​a target smart access point exceeds the first preset threshold (50,000 users scenario: AP maximum capacity × 70%; 100,000 users scenario: AP maximum capacity × 60%), and the load of the backhaul link where the current backhaul parent node of the target smart access point is located exceeds the second preset threshold (85%), select a candidate backhaul node with a load lower than the third preset threshold (45%) from multiple candidate backhaul nodes, and switch the backhaul parent node of the target smart access point to the candidate backhaul node. The switching time is ≤2 seconds, with no service interruption, so as to avoid the backhaul link becoming a performance bottleneck in the hotspot area.

[0046] Please see Figures 3 to 7 As shown, this embodiment of the invention also provides a method for dynamic deployment and resource optimization of a large-scale wireless network, applied to the aforementioned dynamic deployment system for a large-scale wireless network. The method includes the following steps: Step S1: After each smart access point is powered on, discovery and registration are performed through a phased self-organizing network protocol. Each smart access point reports its geographical location (pre-configured / simple mapping and data entry) and device capability information (whether it supports beamforming and transmit power range) to the central control unit. This stage takes ≤10 minutes. Step S2: Based on the location information and candidate backhaul links with signal strength ≥ -75dBm and packet loss rate ≤ 1%, the central control unit performs improved minimum spanning tree topology planning (parallel computation + greedy pruning) and backhaul link channel allocation (using non-overlapping channels) to form a wireless mesh backhaul topology; this stage takes ≤ 8 minutes in a 30 AP scenario and ≤ 15 minutes in a 50 AP scenario. Step S3: Collect the time series of user associations for each smart access point within a predetermined historical time window (≥60 seconds). Use the user density prediction module (sliding window averaging + exponential smoothing + LSTM combination) to generate the user density level (1-5 levels) and prediction confidence (0-1) of the coverage area of ​​each smart access point within a future predetermined time window (600 seconds). The prediction delay is ≤10 seconds, and the MAE is reduced by ≥38% compared to LSTM alone. Step S4: Construct the state space by considering the channel utilization (U_i), number of user associations (N_i), channel list of neighboring access points (C_i), average packet loss rate (L_i), backhaul parent node load (B_i), and corresponding user density level (D_i) and prediction confidence (C_pred_i) of each smart access point; construct the action space by adjusting the working channel (2-3 candidate switching), transmit power (4-level switching), beam direction (±10° / ±20° offset), and beacon frame transmission interval (3-level switching) of the smart access point; and construct the reinforcement learning reward function R based on the total system throughput (T_sys), worst 5% user latency (L_min), configuration switching count (S_cfg), interference suppression index (I), and the network-wide average prediction confidence (C_pred_avg). Step S5: Run the improved MADDPG model (Critic includes backhaul load dimension, Actor includes topology stability penalty term) in the central control unit, select the corresponding action according to the current state space, and generate wireless parameter configuration results for each smart access point; at the same time, trigger the "backhaul load threshold linkage logic": if the target AP prediction density level is greater than the first preset threshold and the current parent node load is greater than the second preset threshold (85%), then select the candidate node with load less than the third preset threshold (45%) as the new parent node and update the Mesh topology; send the configuration results to each smart access point for execution, and the model converges within 20 minutes after the start of the activity; Step S6: Based on the wireless parameter configuration results, the central control unit updates the wireless mesh backhaul topology and access layer coverage parameters as needed; if the average prediction confidence is <0.7 and lasts for 3 statistical periods (30 seconds), the model is triggered to update online and the reward function weights are adjusted (α=1.2, ε=0.02); in subsequent statistical periods, the multi-agent reinforcement learning model is trained based on actual network performance feedback (collected in 10-second cycles) to achieve continuous optimization of the entire network resources in temporary large-scale outdoor event scenarios; at the same time, the preset threshold is dynamically adjusted through the "scene adaptation module" to adapt to events of different scales such as 50,000 people / 100,000 people.

[0047] In this embodiment of the invention, the user density prediction in step S3 includes: performing sliding window averaging on the number of user associations at each smart access point over the past 60 seconds (window 60 seconds) to smooth short-term noise; using an exponential smoothing algorithm (coefficient 0.3) to extrapolate the time series to quickly track population changes; inputting the processed sequence into an LSTM model (10 time steps in the input layer and 50 neurons in the hidden layer) to fit a nonlinear trend; comparing the predicted number of users with the maximum capacity of the AP to classify user density levels (Level 1: ≤20% capacity, Level 5: ≥80% capacity), and calculating the prediction confidence based on the prediction error over the past 5 statistical periods.

[0048] In this embodiment of the invention, in step S5, the multi-agent reinforcement learning model adopts the improved MADDPG method based on policy gradient; during the training phase, the entire network state information (including the backhaul load of each AP) is jointly updated in a centralized manner, the Critic network calculates the global Q value, and the Actor network updates the policy through gradient descent; during the execution phase, each intelligent access point selects actions according to its local state (U_i, N_i, B_i, etc.) using the trained policy, and receives policy parameter updates from the CCU every 2 minutes, reducing the signaling volume by 60%.

[0049] In this embodiment of the invention, the method is applied to an open-air concert with no fewer than 50,000 people. When it is detected that the user density level in a certain area continuously increases (from level 2 to level 4) within two consecutive statistical periods (20 seconds), the transmission power of the corresponding smart access point in that area is increased by at least one level (e.g., from P2 to P3) 1-2 statistical periods in advance (10-20 seconds), the available channel resources are switched to idle channels, and the resource configuration of smart access points in non-hotspot areas is reduced (e.g., the power is reduced from P3 to P2), so that the congestion rate in hotspot areas is reduced by ≥38%, and sudden congestion in hotspot areas is avoided.

[0050] In this embodiment of the invention, step S6 further includes: when the average prediction confidence output by the user density prediction module in the current statistical period is lower than the preset confidence threshold (0.7), the central control unit increases the weight of the real-time network state feedback term (α is adjusted from 1.0 to 1.2) and decreases the weight of the prediction confidence term (ε is adjusted from 0.05 to 0.02) in the reinforcement learning reward function, and triggers the online update process of the multi-agent reinforcement learning model; at the same time, a time lag effect is introduced, and the update is only triggered when the average confidence is continuously lower than 0.7 for 3 statistical periods (30 seconds), so as to reduce the negative impact of prediction error on resource scheduling decision and improve the robustness of the system in scenarios with high prediction uncertainty.

[0051] In summary, the execution steps of the method are as follows: Step S1: AP power-on and registration (≤10 minutes), data flow: AP location and capability information → CCU registration. Step S2: Mesh topology construction (30 APs ≤8 minutes, 50 APs ≤15 minutes), data flow: link detection → candidate screening → improved minimum spanning tree → channel allocation. Step S3: User density prediction (latency ≤10 seconds), data flow: historical user data → three-model parallel prediction → weight adjustment → density level output. Step S4: Reinforcement learning state and action construction, data flow: real-time state + prediction results → state space → reward function construction. Step S5: MADDPG decision and topology update, data flow: state input → improved MADDPG → action output + threshold linkage → parameter distribution. Step S6: Closed-loop optimization and model update, data flow: performance feedback → confidence detection → weight adjustment → continuous optimization. Step S7: Threshold adaptive optimization, data flow: historical performance data → Q-learning → optimal threshold learning.

[0052] In this embodiment of the invention, the combined model used by the user density prediction module has an average sliding window duration of 60 seconds, an exponential smoothing coefficient of 0.3, an input layer of the LSTM neural network with 10 time steps, and a hidden layer containing 50 neurons.

[0053] In this embodiment of the invention, the smart access point includes an antenna array that supports beamforming; the central control unit is further configured to: determine areas with a density level of or above the third level as potential hotspot areas based on the user density level output by the user density prediction module, and control the corresponding smart access point to point the main lobe of the beam to the area, while narrowing the beamwidth by 20%.

[0054] In this embodiment of the invention, the improved MADDPG model adopts a centralized training and distributed execution architecture. During the training phase, the central control unit centrally updates the policy network and value network based on the overall network status information. During the execution phase, each intelligent access point autonomously selects actions based on its local observation status and the policy network parameters obtained from the central control unit, and receives policy parameter updates every 2 minutes.

[0055] Taking a 50,000-person open-air concert as an example. Equipment configuration: Number of APs: 30 (15 supporting beamforming); Frequency band: 5GHz access and backhaul; Venue area: 1-2 square kilometers. Self-organizing network process: Phase 1: Discovery and registration (8 minutes), Phase 2: Topology planning (7 minutes), Phase 3: Parameter initialization (4 minutes), Total time: 19 minutes. Performance optimization results are shown in the table below: Taking a 100,000-seat sports event as an example. Scenario adaptation: Number of APs: 50 (30 supporting beamforming), venue area: 3 square kilometers, threshold configuration: density trigger value = AP maximum capacity × 60%. Cross-layer linkage example: Prediction: The density of the east grandstand will increase from level 2 to level 5 in 15 minutes. Access layer optimization: power increase, channel switching, beam adjustment; backhaul layer optimization: parent node switching (load 92% → 42%). Result: throughput 3.2Gbps, no congestion.

[0056] Taking a scenario of sudden weather changes as an example. Adaptive performance: Change detection: Confidence drops sharply from 0.85 to 0.65; Weight adjustment: Sliding window weight changes from 0.3 to 0.5; Threshold optimization: Backpropagation threshold changes from 85% to 80%; Results: Prediction error fluctuation ≤12%, throughput fluctuation <15%.

[0057] The above description only details the preferred embodiments of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0058] Other embodiments of this disclosure will readily occur to those skilled in the art upon consideration of the disclosure in the specification and embodiments. This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not disclosed herein. The specification and embodiments are to be considered exemplary only, and the true scope and spirit of this disclosure are indicated by the claims.

Claims

1. A large-area wireless network dynamic deployment system, characterized in that, include: The system comprises a central control unit, multiple intelligent access points, a user density prediction module, a phased self-organizing network protocol module, and a reinforcement learning resource scheduling module, among which: The phased self-organizing network protocol module is configured to control each smart access point to sequentially execute the discovery and registration phase, the topology and channel planning phase, and the access parameter initialization phase after power-on. The topology and channel planning phase adopts an improved minimum spanning tree algorithm based on a preset signal quality threshold for candidate link pre-screening and combined with parallel computation pruning to construct the wireless Mesh backhaul topology. The user density prediction module is configured to: based on the historical user association time series collected by each smart access point, use a combination model of sliding window averaging, exponential smoothing and long short-term memory neural network to predict the user density level and corresponding prediction confidence of the coverage area of ​​each smart access point in the future time period, and output the prediction results to the reinforcement learning resource scheduling module. The reinforcement learning resource scheduling module is configured to: employ an improved multi-agent deep deterministic policy gradient algorithm to construct a state space containing the real-time network status and user density prediction results of each intelligent access point; define an action space containing at least one of the control actions of adjusting the working channel, transmit power, beam direction, and beacon frame transmission interval; and jointly optimize the wireless parameters of each intelligent access point through centralized training and distributed execution, based on a reward function containing the total system throughput, worst-case user latency, configuration handover count, interference suppression index, and prediction confidence index. The central control unit is also configured to: based on the optimization results output by the reinforcement learning resource scheduling module, trigger the update of the wireless Mesh backhaul topology when preset conditions are met through the backhaul load threshold linkage logic, so as to realize cross-layer collaborative optimization of access layer parameters and backhaul network topology.

2. A method for dynamic deployment and resource optimization of a large-scale wireless network, applied to the large-scale wireless network dynamic deployment system described in claim 1, characterized in that, The method includes the following steps: After the multiple smart access points are powered on, they perform discovery and registration and report information to the central control unit. Based on the reported location information and preset link quality conditions, the central control unit executes an improved minimum spanning tree algorithm to perform topology planning and channel allocation, forming an initial wireless mesh backhaul topology. Based on historical user data collected from each smart access point, a combined prediction model is used to generate user density levels and prediction confidence levels for each region in the future time period. Construct a state space and action space for reinforcement learning. The state space includes at least the real-time channel utilization, number of users associated, neighbor channel list, average packet loss rate, backhaul parent node load, user density level, and prediction confidence of each smart access point. The action space includes the adjustment of at least one parameter among the working channel, transmit power, beam direction, and beacon frame transmission interval. Run the improved MADDPG model, select actions based on the current state space, generate wireless parameter configuration results, and determine whether to update the Mesh backhaul topology based on the backhaul load threshold linkage logic, and send the results to each smart access point for execution. Based on network performance feedback, the improved MADDPG model and prediction model are continuously optimized and their parameters are adjusted online.

3. The large-area wireless network dynamic deployment system according to claim 1, characterized in that, In the combined model used by the user density prediction module, the average window duration of the sliding window is 60 seconds, the smoothing coefficient of the exponential smoothing is 0.3, the input layer of the LSTM neural network has 10 time steps, and the hidden layer contains 50 neurons.

4. The method for dynamic deployment and resource optimization of large-scale wireless networks according to claim 2, characterized in that, The improvements of the improved MADDPG algorithm include: the input layer of the Critic network adds a dimension reflecting the load of the backhaul parent node; and the loss function of the Actor network introduces a topology stability penalty term, which is positively correlated with the number of backhaul topology switching times within the statistical period.

5. The large-area wireless network dynamic deployment system according to claim 1, characterized in that, The smart access point includes an antenna array that supports beamforming; the central control unit is also configured to: determine areas with a density level of 3 or higher as potential hotspot areas based on the user density level output by the user density prediction module, and control the corresponding smart access point to point the main lobe of the beam to the area, while narrowing the beamwidth by 20%.

6. The method for dynamic deployment and resource optimization of large-scale wireless networks according to claim 2, characterized in that, The backhaul load threshold linkage logic is configured as follows: when the predicted user density level of the target smart access point coverage area exceeds the first preset threshold and the load rate of its current backhaul parent node exceeds the second preset threshold of 85%, a node with a load rate lower than the third preset threshold of 45% is selected from the candidate backhaul nodes as the new backhaul parent node to complete the topology switch, and the switch process takes no more than 2 seconds and there is no service interruption.

7. The method for dynamic deployment and resource optimization of large-scale wireless networks according to claim 2, characterized in that, The reward function R used by the reinforcement learning resource scheduling module is expressed as: R = α T_sys-β L_min-γ S_cfg+δ I+ε C_pred_avg, where T_sys is the total system throughput, L_min is the average latency of the worst 5% of users, S_cfg is the number of configuration switching times, I is the interference suppression index, C_pred_avg is the average prediction confidence of the entire network, and α, β, γ, δ, and ε are weighting coefficients.

8. The method for dynamic deployment and resource optimization of large-scale wireless networks according to claim 7, characterized in that, When the network-wide average prediction confidence C_pred_avg is below 0.7 for three consecutive statistical periods, a strategy update is triggered, adjusting the weight coefficient α to 1.2 and ε to 0.

02.

9. The method for dynamic deployment and resource optimization of large-scale wireless networks according to claim 2, characterized in that, The improved MADDPG model adopts a centralized training and distributed execution architecture. During the training phase, the central control unit updates the policy network and value network based on the overall network status information. During the execution phase, each intelligent access point autonomously selects actions based on its local observation status and the policy network parameters obtained from the central control unit, and receives policy parameter updates every 2 minutes.

10. The method for dynamic deployment and resource optimization of a large-scale wireless network according to claim 2, characterized in that, When the user density level in a region is detected to be continuously increasing over two consecutive statistical periods, the transmit power level of the corresponding smart access point in that region is increased and switched to a less busy working channel 1 to 2 statistical periods in advance. At the same time, the resource allocation of smart access points in non-hotspot areas is reduced, thereby reducing the access congestion rate in hotspot areas.

Images