An optimal path planning method, system, and device for SD-WAN networks

By integrating AI Agent and reinforcement learning model into SD-WAN network, loop-free paths are generated and future quality is predicted, solving the path planning problem of existing SD-WAN network in complex environments and realizing intelligent, real-time and highly available path optimization.

CN123123239APending Publication Date: 2026-11-17GENERATION TIMES TECHNOLOGY (SHANGHAI) CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202610879535.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-06-17
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing SD-WAN networks, when faced with the increasing complexity of global business, dynamic changes in network environment, and multi-vendor, multi-cloud architectures, suffer from rigid path calculation algorithms, insufficient multi-path support capabilities, frequent loop problems, and a lack of predictive optimization, making it difficult to meet the demands of modern enterprises for intelligent, real-time, and highly available global networks.

Method used

By acquiring real-time operational metrics and external network data through the SD-WAN controller, combining it with an AI Agent for path planning, using a reinforcement learning model to generate loop-free paths, predicting future quality and making dynamic adjustments, and supporting real-time evaluation and inclusion of emerging links.

Benefits of technology

It enables intelligent and real-time path planning for SD-WAN networks, improving network flexibility and reliability, reducing the risk of service interruption, and supporting the rapid access and optimization of emerging technologies.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123123239A_ABST
    Figure CN123123239A_ABST
Patent Text Reader

Abstract

This invention discloses an optimal path planning method, system, and device for SD-WAN networks, effectively solving the problem that existing SD-WAN networks cannot meet the core requirements of modern enterprises for intelligent, real-time, and highly available global networks. The method provided by this invention includes: acquiring real-time operational indicator data of the SD-WAN network and external global network data; integrating the real-time operational indicator data and external global network data to obtain a unified feature vector; parsing the user's input natural language routing strategy according to a pre-set AI Agent to obtain a target semantic graph; mapping the target semantic graph to obtain the objective function and constraint set for user path optimization; using the unified feature vector, objective function, and constraint set, the AI ​​Agent generates multiple candidate loop-free paths with the network topology as the state space and path decision as the action space; selecting the optimal path from the multiple candidate loop-free paths and executing the optimal path to complete the optimal path planning for the SD-WAN network.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of SD-WAN technology, and in particular to an optimal path planning method, system and device for SD-WAN networks. Background Technology

[0002] SD-WAN technology, with its core advantage of enabling dynamic path selection over the public internet, has been widely adopted in global enterprise network deployments, becoming a key technology supporting cross-border and cross-regional business data transmission. Current mainstream SD-WAN solutions primarily use path planning algorithms based on single network quality metrics such as end-to-end latency, packet loss rate, and jitter, or employ weighted normalization to transform multiple factors like bandwidth utilization, cost, and security into composite metrics for calculation. Examples include Cisco ViPtela SD-WAN's BGP policy routing and VMware VeloCloud's application-aware routing. In stable network environments, these algorithms can generally meet transmission requirements.

[0003] However, with the increasing complexity of global business layouts, dynamic changes in network environments, and the widespread adoption of multi-vendor, multi-cloud architectures, existing technologies have gradually exposed numerous bottlenecks that are difficult to overcome. First, the problem of rigid path calculation algorithms is prominent, and routing strategies rely on hard-coded scripts. When enterprises need to switch core requirements based on business changes (such as shifting from "delay priority" to "cost priority") or to cope with emergencies such as geopolitical events or submarine cable interruptions, a large amount of code refactoring, simulation verification, and field testing are required, with a cycle of up to several weeks. In addition, there is the risk of controller firmware upgrades, which can easily lead to business interruptions, resulting in a significant increase in the enterprise's SLA default rate.

[0004] Secondly, multi-path support is insufficient. The classic Dijkstra algorithm can only output a single path, lacking redundancy and backup, making it susceptible to single points of failure. Multi-path algorithms such as K-ShortestPaths suffer from high computational complexity and high parameter tuning barriers, resulting in excessive time consumption in large global topologies with over 1000 nodes, failing to meet real-time scheduling requirements. Simultaneously, loop problems are frequent in multi-path environments. Traditional static solutions such as TTL counters and BGPAS-Path filtering have high failure rates in dynamic SD-WAN architectures, while loop detection tools suitable for LANs are slow to respond at the WAN scale, with an average detection time exceeding 10 seconds, easily leading to resource exhaustion or even network paralysis.

[0005] Furthermore, existing technologies lack predictive optimization capabilities, relying solely on real-time network snapshots for path selection. They cannot anticipate potential problems such as peak-hour congestion and link attenuation based on historical trends, and can only passively respond to network deterioration. In addition, the fragmented global network environment, inconsistent QoS standards among different operators, and the difficulty of quickly evaluating and integrating emerging technologies like satellite links further limit the resilience and reliability of SD-WAN networks. These shortcomings make existing solutions inadequate for meeting the core needs of modern enterprises for intelligent, real-time, and highly available global networks, necessitating an innovative technological solution that breaks through the limitations of traditional algorithms. Summary of the Invention

[0006] In view of the shortcomings of existing technologies, the purpose of this invention is to provide an optimal path planning method, system and device for SD-WAN networks, which effectively solves the problem that existing SD-WAN networks are unable to meet the core needs of modern enterprises for intelligent, real-time and highly available global networks.

[0007] The technical solution provided in this application is: an optimal path planning method for SD-WAN networks, the method comprising:

[0008] The real-time operation index data of the SD-WAN network and external global network data are obtained through the acquisition interface of the SD-WAN controller, and the real-time operation index data and external global network data are integrated to obtain a unified feature vector.

[0009] The target semantic graph is obtained by parsing the natural language routing strategy input by the user based on the pre-set AI Agent, and the objective function and constraint set for user path optimization are obtained by mapping the target semantic graph.

[0010] Using the unified feature vector, objective function, and constraint set, the AI ​​Agent generates multiple candidate loop-free paths with network topology as the state space and path decision as the action space.

[0011] The optimal path is selected from the multiple candidate loop-free paths and the optimal path is executed to complete the optimal path planning for the SD-WAN network.

[0012] Furthermore, selecting the optimal path from the multiple candidate acyclic paths includes:

[0013] The weighted scores for the multiple candidate acyclic paths are calculated based on the objective function, and future quality predictions are performed to obtain the quality prediction results.

[0014] The weighted score is dynamically adjusted based on the quality prediction results to select the optimal path from the multiple candidate acyclic paths.

[0015] Furthermore, the process of obtaining quality prediction results through future quality prediction includes:

[0016] Based on path quality data over historical time periods and external influencing factors, the AI ​​Agent predicts the quality change trend of the optimal path in future time periods.

[0017] Determine whether the quality change trend meets the preset quality deterioration conditions to generate the quality prediction result of the optimal path.

[0018] Further, after generating the quality prediction result of the optimal path, the process includes:

[0019] If the quality prediction result is satisfactory, then the backup path of the candidate loop-free path is generated in advance and injected into the SD-WAN controller queue.

[0020] The SD-WAN controller is controlled to perform a forward-looking switch of the optimal path based on the backup path.

[0021] Furthermore, selecting the optimal path from the multiple candidate acyclic paths also includes:

[0022] Acquire and push the multiple candidate loop-free paths and their corresponding indicator data to a preset visualization panel;

[0023] Based on the aforementioned indicator data and various unquantified factors, the optimal path is selected from the multiple candidate acyclic paths.

[0024] Furthermore, the AI ​​Agent generates multiple candidate acyclic paths using network topology as the state space and path decision as the action space, including:

[0025] The state space and action space are constructed based on the network topology and path decision sequence of the SD-WAN network, respectively.

[0026] The multiple candidate acyclic paths are generated by training a reinforcement learning model to interact between the state space and the action space.

[0027] Further, generating the multiple candidate loop-free paths includes:

[0028] Verify whether there are loop-containing paths among the multiple candidate paths generated by the AI ​​Agent, and obtain the verification results;

[0029] Based on the verification results, the loop-containing paths are filtered, and alternative paths are generated until the multiple candidate loop-free paths are obtained.

[0030] Furthermore, the method also includes:

[0031] Real-time data of emerging network providers is collected in real time to obtain evaluation results by real-time assessment of the service quality of the emerging network providers based on the real-time data;

[0032] Based on the evaluation results, emerging links based on the emerging network providers are constructed and dynamically incorporated into the path planning scope.

[0033] This application provides another solution: an optimal path planning system for SD-WAN networks, the system comprising:

[0034] The acquisition module is used to acquire real-time operating indicator data of the SD-WAN network and external global network data through the acquisition interface of the SD-WAN controller, and integrate the real-time operating indicator data and external global network data to obtain a unified feature vector;

[0035] The parsing module is used to parse the natural language routing strategy input by the user according to the preset AI Agent, obtain the target semantic graph, and map the target semantic graph to obtain the objective function and constraint set for user path optimization;

[0036] The generation module is used to generate multiple candidate acyclic paths using the unified feature vector, objective function, and constraint set, with the AI ​​Agent using network topology as the state space and path decision as the action space.

[0037] The execution module is used to select the optimal path from the multiple candidate loop-free paths and execute the optimal path to complete the optimal path planning of the SD-WAN network.

[0038] This application also provides a solution: an electronic device comprising: a processor, a memory, and a bus, wherein the memory stores machine-readable instructions executable by the processor, and when the electronic device is running, the processor communicates with the memory via the bus, and when the machine-readable instructions are executed by the processor, the steps of any one of the optimal path planning methods for an SD-WAN network are performed.

[0039] This application provides an optimal path planning method for SD-WAN networks. The method first acquires real-time operational metrics data and external global network data of the SD-WAN network through the acquisition interface of the SD-WAN controller, and integrates the real-time operational metrics data and external global network data to obtain a unified feature vector. Second, it parses the user's input natural language routing strategy according to a pre-set AI Agent to obtain a target semantic graph, and maps the target semantic graph to obtain the objective function and constraint set for user path optimization. Then, using the unified feature vector, objective function, and constraint set, the AI ​​Agent generates multiple candidate loop-free paths with the network topology as the state space and path decision as the action space. Finally, it selects the optimal path from the multiple candidate loop-free paths and executes the optimal path to complete the optimal path planning for the SD-WAN network. Attached Figure Description

[0040] Figure 1 This is a flowchart illustrating an optimal path planning method for an SD-WAN network provided in an embodiment of this application.

[0041] Figure 2 This is a schematic diagram illustrating the interaction between a smart device and a cloud system, as provided in an embodiment of this application.

[0042] Figure 3 This is a schematic diagram of the overall framework of the AI ​​Agent provided in the embodiments of this application.

[0043] Figure 4 This is a structural block diagram of an optimal path planning system for an SD-WAN network provided in an embodiment of this application.

[0044] Figure 5 A structural block diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0045] For the purposes of this invention, the foregoing and other technical contents, features and effects are described in conjunction with the appendix below. Figure 1-5 The detailed description of the embodiments will make this clear. All structural details mentioned in the following embodiments are based on the accompanying drawings.

[0046] Existing SD-WAN networks are insufficient to meet the core needs of modern enterprises for intelligent, real-time, and highly available global networks, and there is an urgent need for an innovative technical solution that breaks through the limitations of traditional algorithms.

[0047] Based on this, this application provides an optimal path planning method, system, and device for SD-WAN networks, which are described below with reference to embodiments and accompanying drawings.

[0048] Example 1

[0049] To facilitate understanding of this embodiment, a detailed description of an optimal path planning method for an SD-WAN network disclosed in this application embodiment will be provided first. For example... Figure 1 The diagram illustrates an optimal path planning method for an SD-WAN network. The optimal path planning method for an SD-WAN network provided in this application includes:

[0050] S101. Obtain real-time operation index data of SD-WAN network and external global network data through the acquisition interface of SD-WAN controller, and integrate the real-time operation index data and external global network data to obtain a unified feature vector;

[0051] S102. Based on the preset AI Agent, the user input natural language routing strategy is parsed to obtain the target semantic graph. The target semantic graph is then mapped to obtain the objective function and constraint set for user path optimization.

[0052] S103. Using the unified feature vector, objective function and constraint set, the AI ​​Agent generates multiple candidate loop-free paths with network topology as the state space and path decision as the action space.

[0053] S104. Select the optimal path from the multiple candidate loop-free paths and execute the optimal path to complete the optimal path planning of the SD-WAN network.

[0054] The method described in this application can be configured to execute in a smart device, such as... Figure 2 As shown, during execution, data interaction can be conducted with the cloud to ensure the implementation of the methods described in this application.

[0055] In step S101, the SD-WAN controller uses its built-in standardized data acquisition interface and a dual-protocol (SNMP and NetFlow) collaborative acquisition mode to capture real-time core network performance data at a preset interval of 5 seconds per acquisition, ensuring data timeliness and completeness. The real-time core performance data covers transmission quality, service level, and security dimensions: latency (RTT) is accurate to the millisecond level and calculated using bidirectional data packet round-trip time; bandwidth utilization is dynamically calculated based on the ratio of actual link throughput to rated bandwidth, with a 70% threshold triggering an alert; packet loss rate and jitter are quantified by statistically analyzing the proportion of lost packets and the amplitude of transmission latency fluctuations, respectively; SLA compliance is verified in real-time based on service level agreement terms, and the adequacy of latency, availability, and other indicators is quantified and scored; the security score is generated by weighting together dimensions such as link encryption level, node protection capabilities, and access control policy compliance. Simultaneously, a multi-source external global network data access channel is established through a RESTful API interface to achieve real-time synchronization of global network environment information. This external global network data covers four key categories: BGP routing tables from authoritative network data institutions such as RIPE NCC, providing global routing topology and path reachability information; real-time SLA data from operators, obtained through interfaces with mainstream operators such as AT&T and China Telecom, providing real-time service quality data such as link latency and availability; a geopolitical risk index integrating information on international situation dynamics and regional network policy changes to quantify the geopolitical risks of path transmission; and emerging network technology access information captured through an industry technology monitoring platform, including the deployment scope and performance parameters of new technologies such as 5G base station coverage and Starlink satellite links. Furthermore, the collected internal real-time indicator data and external network data undergo standardization processing: normalization transformation is performed for indicators of different magnitudes to eliminate dimensional differences; missing data is filled using a time-series interpolation algorithm, and abnormal data is identified and filtered using the 3σ principle; finally, according to a preset data structure, multi-dimensional data is integrated into a unified feature vector of 50+ dimensions, including static topology features, dynamic operational features, and external environment features, providing a foundation for AI. The agent's path planning and prediction optimization are supported by comprehensive and standardized input data.

[0056] The AI ​​Agent described in this application is built for SD-WAN networks. This Agent employs an Agentic AI architecture, including a perception layer for data acquisition, a decision layer for path generation, and an execution layer for controller integration. Agentic AI is a highly autonomous AI architecture that mimics human cognitive processes: perceiving the environment, making decisions, executing actions, and continuously optimizing through feedback. The optimization approach shifts from algorithm parameter tuning to Agent capability enhancement: using an Actor-Critic reinforcement learning framework, the network topology is used as the state space, path decisions as the action space, and the reward function R = w1*(-latency) + w2bandwidth + w3(-cost) + w4*security. Through pre-training in simulated environments (e.g., Mininet simulation), the Agent learns complex decisions, such as automatically injecting virtual tunnels during congestion. This design ensures that the Agent generalizes well under SD-WAN-specific constraints (such as link symmetry), avoiding the illusion problem of general-purpose models. The decision layer is a parallel module, including an NLP policy parser (parsing user text input), a GNN loop validator (graph embedding for loop detection), an RL path optimizer (state-action reward loop graph), and an LSTM prediction module (time series prediction curves), such as... Figure 3 As shown.

[0057] In step S102, the user inputs a natural language strategy via the mobile app or dashboard, such as: "Calculate 3 paths for video conferencing traffic, prioritizing low-latency paths (<60ms), bandwidth >200Mbps, avoiding loops, supporting costs <0.05$ / GB, and considering green routing (low-carbon links), based on pre-set AI..." The agent parses the natural language routing strategy input by the user to obtain a target semantic graph. Specifically, the NLP module uses zero-shot learning to parse it into JSON rules. First, it performs word segmentation, entity recognition, and semantic association analysis on the natural language text based on an LLM model to extract core elements: business type (cross-border video conferencing), number of paths (3), and performance requirements (latency <60ms, bandwidth utilization <70 / GB). Based on the extracted core elements, a target semantic graph is constructed, representing key information such as target dimensions, constraint types, business attributes, and quantification thresholds in the form of nodes. Edge connections represent the logical relationships between elements, such as a strong association between video conferencing and low latency, and a hierarchical association between low-carbon links and sustainability goals. The semantic graph is stored using a standardized data structure, supports dynamic expansion, and can adapt to complex multi-condition combination strategies. Subsequently, through a preset semantic mapping rule base, the target semantic graph is mapped to obtain the objective function and constraint set for user path optimization that the AIAgent can execute; the objective function is the AIAgent's... The agent's reward function is constructed based on multi-dimensional weighting, with the formula: F = w1 × (-latency) + w2 × (1 - bandwidth utilization) + w3 × (-cost) + w4 × security score + w5 × green coefficient. The weights w1-w5 are dynamically allocated based on the business type and target priority in the semantic graph (e.g., w1's weight increases to 0.4 in a video conferencing scenario). The constraint set is presented in structured JSON format, clearly defining hard constraints (e.g., "latency < 60ms" "region ≠ EU") and soft constraints (e.g., "prioritize low-carbon links"). Hard constraints directly serve as necessary conditions for path selection, while soft constraints reflect priority through weighted objective function calculations.

[0058] The entire process requires no manual coding or configuration adjustments, achieves over 95% accuracy in parsing, supports rapid policy iteration and A / B testing, and reduces the traditional weekly cycle of policy changes to the hourly cycle, significantly improving network operation and maintenance efficiency.

[0059] In step S103, after obtaining the unified feature vector, objective function, and constraint set, the reinforcement learning RL path optimizer of the AI ​​Agent uses the network topology as the state space and path decision as the action space, and iteratively generates multiple candidate paths using the Proximal Policy Optimization (PPO) algorithm. The number of candidate paths is preset by the user or adaptively configured by the system. The AI ​​Agent first transforms the network data in the unified feature vector into specific parameters in the state space, such as labeling each node and link with real-time delay, cost, and security score. At the same time, it loads the objective function (scoring rules) and constraint set (taboo list) to clarify the decision boundary. The RL path optimizer uses the current topology as the state S_t, generates action A_t (path sequence), normalizes the function R = -0.4delay-0.3loss+0.2bw-0.1cost, and iterates for 500 steps using the Proximal Policy Optimization (PPO) algorithm to output K=3 paths. Each path, including a node sequence, expected metrics, and confidence interval, explores different path decisions in the state space. For example, starting from the source node, according to the scoring logic of the objective function (i.e., the reward function), links with high scores (such as links with low latency and low cost) are prioritized and combined to form an initial path. To avoid the risk of a single path, the RL path optimizer explores multiple different node combinations to generate multiple initial candidate paths (generally 3-10, which can be preset by the user). For the generated initial paths, the AI ​​performs loop detection through a graph neural network (GNN), such as checking whether there are duplicate nodes or loops in the path, and performs secondary screening according to the constraint set, such as eliminating paths with latency ≥60ms, passing through EU nodes, or cost >0.05$ / GB, and only retaining paths that meet all hard requirements to form the final multiple candidate loop-free paths.

[0060] In a specific implementation of step S103, one embodiment is as follows: the AI ​​Agent generates multiple candidate loop-free paths using network topology as the state space and path decision as the action space, including:

[0061] S1031. Construct the state space and action space based on the network topology and path decision sequence of the SD-WAN network, respectively;

[0062] S1032. The reinforcement learning model is trained interactively between the state space and the action space to generate the multiple candidate acyclic paths.

[0063] In steps S1031-S1032, this application uses the actual SD-WAN network topology as a blueprint, mapping all nodes, links, and real-time attributes of links / nodes (retrieved from a unified feature vector) into digital features to construct a state space, clarifying the decision-making environment boundary of the AI ​​Agent. Based on this topology, it sorts out all legal path decision sequences (node ​​jump order) from the source node to the destination node, constructs the action space, and clarifies the AI... The agent's selectable path range; basic acyclic rules (such as prohibiting repeated traversal of the same node) are embedded during action space initialization to reduce loop paths from the source and lower subsequent training costs. The AI ​​Agent continuously interacts between the state space and action space through a cycle of state awareness, action selection, result feedback, and policy adjustment, gradually learning to select high-quality paths. Specifically, it selects a path from the action space as a candidate path for this decision, and performs a double evaluation on the selected path, giving rewards / penalties (the core driving force of reinforcement learning): positive rewards: if the path meets the acyclic requirement, fits the objective function (such as low latency and low cost), and meets the constraints (such as avoiding designated areas), a high reward score is given; negative penalties: if the path contains loops, violates constraints, or has poor performance (such as excessively high packet loss rate), a low reward score or even a penalty score is given. The AI ​​Agent automatically adjusts its internal decision-making strategy based on the reward / penalty results, reducing the probability of selecting penalized paths and increasing the probability of selecting rewarded paths, making the next path selection more in line with the requirements, thereby generating the multiple candidate acyclic paths.

[0064] In a specific implementation of step S1032, one embodiment involves generating the multiple candidate loop-free paths, including:

[0065] S10321. Verify whether there are loop-containing paths among the multiple candidate paths generated by the AI ​​Agent, and obtain the verification results;

[0066] S10322. Based on the verification results, filter the loop-containing paths and trigger the generation of alternative paths until the multiple candidate loop-free paths are obtained.

[0067] In steps S10321-S10322, after the AI ​​Agent generates candidate paths based on the action space, this application verifies whether there are loop-containing paths among the multiple candidate paths generated by the AI ​​Agent, and obtains the verification results. Specifically, the AI ​​Agent integrates a Graph Neural Network (GNN) module to perform topological reasoning and loop detection on the pre-planned paths. Since the current topology is generated by reasoning, due to the illusion of a large model, it cannot guarantee 100% that the path is loop-free. Other mathematical algorithms can guarantee loop-free paths, but they lack flexibility and are not suitable for complex real-world scenarios. GNN represents the network topology with node embeddings, runs a message passing algorithm to simulate packet propagation, and detects negative weighted loops or simple cycles (such as using a variant of Bellman-Ford). If a loop exists and the loop detection score is >0.5, resampling (adding noise perturbation) is triggered until all paths are loop-free, and the verification log records potential risk nodes. If the verification result shows that the multiple candidate paths include loop-containing paths, the loop-containing paths are filtered, and alternative paths are generated. That is, the AI ​​Agent generates alternative paths through Monte Carlo Tree Search (MCTS), injects "bridging nodes" or rerouting strategies, and ensures 100% loop-free output. Compared with traditional DFS / BFS rules, this method adapts to dynamic topology changes and supports mesh, hub-spoke, and full mesh structures. In the test, the loop detection accuracy reached 99.5%, and the response time was <50ms, until the multiple candidate loop-free paths were obtained.

[0068] In step S104, this application selects the optimal path from multiple candidate loop-free paths based on the comprehensive decision-making of the AI ​​Agent, which is based on multi-dimensional evaluation, dynamic adaptation, and predictive verification. This ensures that the selected optimal path meets the service requirements of the SD-WAN network. After selecting the optimal path, the AI ​​Agent converts the node sequence and link configuration requirements (such as encryption level and bandwidth reservation) of the optimal path into standardized instructions that can be recognized by the SD-WAN controller of the SD-WAN network. After receiving the instructions, the SD-WAN controller first performs local verification to confirm that the nodes and links in the optimal path are in an available state, without hardware failures or configuration conflicts, to avoid deployment failure. After the verification is passed, the controller performs a hot switch (switching latency <100ms), smoothly migrating the service traffic from the current path to the optimal path. The entire process does not interrupt data transmission (such as video conferencing and file transfer are not affected), thereby enabling the SD-WAN controller to execute the optimal path to complete the optimal path planning of the SD-WAN network. During the execution process, the SD-WAN controller collects the actual operating indicators (latency, packet loss rate, cost, etc.) of the optimal path every 5 seconds and synchronously feeds them back to the AI ​​Agent for subsequent strategy optimization.

[0069] In a specific implementation of step S104, one embodiment involves selecting the optimal path from the plurality of candidate loop-free paths, including:

[0070] S10411. Calculate the weighted score and perform future quality prediction on the multiple candidate acyclic paths based on the objective function to obtain the quality prediction result.

[0071] S10412. Based on the quality prediction results, dynamically adjust the weighted score to select the optimal path from the multiple candidate acyclic paths.

[0072] In steps S10411-S10412, the AI ​​Agent re-verifies whether all candidate loop-free paths 100% meet the constraint set. If some paths no longer meet the constraints due to real-time changes in the SD-WAN network (such as a sudden increase in latency on a certain link), they are directly eliminated to avoid invalid decisions. Using the previously generated objective function, i.e., the reward function (e.g., latency weight 40%, cost 30%, security 20%, green 10%), the paths that pass the re-examination are quantified and scored one by one. The higher the score, the more closely they match user preferences. The LSTM prediction module is then called to predict the quality changes of each high-scoring path in the next 1-24 hours (e.g., whether there will be congestion, whether the latency will increase). If a path currently has the highest score, but its predicted deterioration probability exceeds 70% after 1 hour, its priority is reduced. Conversely, if a path currently has the second highest score, but its prediction quality is stable, its priority can be increased. If a path has a significantly higher score (≥10% of the second-place path) and its prediction is stable, it is directly designated as the optimal path. If multiple paths have similar scores, they are all Pareto optimal solutions, each with its own advantages: for example, A has low latency and B has low cost. The optimal path is obtained by following the user's preset priority, such as prioritizing low latency for video conferencing, or by supporting manual selection (displaying indicator details for administrator confirmation).

[0073] In a specific implementation of step S10411, one embodiment is as follows: obtaining the quality prediction result by performing future quality prediction includes:

[0074] A1. Based on path quality data over historical time periods and external influencing factors, the AI ​​Agent predicts the quality change trend of the optimal path over future time periods.

[0075] A2. Determine whether the quality change trend meets the preset quality deterioration conditions, so as to generate the quality prediction result of the optimal path.

[0076] In steps A1-A2, this application pre-collects real-time operational metrics of the optimal path over the past 7 days (configurable), including latency, bandwidth utilization, packet loss rate, jitter, SLA compliance rate, and fault frequency, and stores them at a sampling rate of 1 minute / sample to form a complete time-series dataset. For example, the latency of a certain path previously increased from 40ms to 80ms every Monday morning at 10:00 AM (peak business hours). This historical pattern is the core basis for prediction. The external influencing factors are factors outside the SD-WAN network itself that indirectly affect path transmission, such as time-related factors: weekdays / weekends, peak hours (e.g., 9:00-12:00), holidays; environmental factors: weather such as heavy rain and typhoons that may affect submarine cables, natural disaster warnings such as earthquakes and floods; external network factors: operator maintenance announcements, BGP route adjustments, new supplier access; geopolitical policy factors: changes in regional network policies, escalation of geopolitical risks (e.g., routing restrictions in a certain area). AI The agent uses temporal deep learning models (such as LSTM and Transformer) to extract hidden temporal patterns from historical path quality data. For example, it identifies patterns like increased latency during peak hours and slightly higher packet loss rates during operator maintenance on Wednesdays. It also analyzes the correlation between external influencing factors and path quality, such as an 80% correlation between heavy rain and increased latency in submarine optical cables, and a 3-fold increase in the probability of path interruption after escalating geopolitical risks. Finally, it combines historical patterns and external factors to construct a prediction model, outputting the path quality change trend over a future time period (1-24 hours, configurable), including predicted index values ​​and the probability of deterioration. Based on the deterioration probability, it determines whether the quality change trend meets a preset quality deterioration condition, i.e., whether the deterioration probability has reached a quality deterioration condition, which can be a pre-set threshold such as 0.7. This generates the optimal path quality prediction result, which includes whether deterioration has occurred, the corresponding deterioration probability, and predicted index values.

[0077] In a specific implementation of step A2, one embodiment includes: after generating the quality prediction result of the optimal path, the following steps are taken:

[0078] A21. If the quality prediction result is satisfactory, then the backup path of the candidate loop-free path is generated in advance and injected into the SD-WAN controller queue.

[0079] A22. Control the SD-WAN controller to perform forward switching of the optimal path based on the backup path.

[0080] In steps A21-A22, if the quality prediction result is satisfactory, the probability of deterioration is >0.7. The AI ​​Agent then uses the objective function and constraint set of the optimal path as a benchmark to select the Top 3 Pareto optimal solutions from the candidate acyclic path pool as backup paths. These are prioritized according to prediction stability: the first backup path has the most stable predicted quality over the next 24 hours; the second backup path has the best cost-performance ratio; and the third backup path has the strongest risk resistance, such as those containing multiple redundant nodes. If no path meets the requirements in the candidate pool, the AI ​​Agent will immediately re-trigger the path generation process, quickly generating 2-3 acyclic paths that meet the constraints based on the latest network feature vectors to ensure a sufficient supply of backup paths. After generating backup paths, a secondary acyclic verification is performed using the GNN loop validator, and the LSTM prediction module is called to evaluate their future quality stability (the predicted probability of deterioration must be <30%). Paths with loop risks or unstable predictions are eliminated and replaced with newly generated backup paths to ensure that the backup paths are available, acyclic, and stable. The Agent converts the node sequence, link configuration parameters (encryption level, bandwidth reservation threshold, etc.), and priority identifiers of the backup paths into standardized instructions (encapsulated based on the gRPC protocol) that the SD-WAN controller can recognize. These instructions include path IDs, switching trigger conditions (such as optimal path latency ≥ 60ms, degradation probability ≥ 70%), and switching timeout. Upon receiving the instructions, the SD-WAN controller injects the backup paths into a dedicated path queue according to priority, employing a first-in-first-out (FIFO) and priority-based queuing mechanism: when the first backup path no longer meets requirements due to network changes, it is automatically removed, and the second backup path is upgraded to first priority. Simultaneously, the AI ​​Agent updates the predicted quality data of the backup paths every 30 seconds, and the controller adjusts the queue order in real time to ensure that the queue always contains the optimal combination of backup paths. The SD-WAN controller monitors the actual operating metrics of the optimal path in real time and, combined with the AI ​​Agent's real-time prediction results, initiates a proactive switchover when any of the following trigger conditions are met. The switchover uses hot-swapping technology, completing traffic migration within 100ms.

[0081] In the specific implementation of step S104, another embodiment further includes: selecting the optimal path from the multiple candidate loop-free paths, which also includes:

[0082] S10421. Obtain and push the multiple candidate loop-free paths and their corresponding indicator data to the preset visualization panel;

[0083] S10422. Based on the index data and various unquantified factors, select the optimal path from the multiple candidate loop-free paths.

[0084] In steps S10421-S10422, to balance the efficiency of AI intelligent decision-making with the flexibility of human intervention, this application uses a model of data visualization, multi-dimensional reference assistance, and final human decision-making to allow network administrators to accurately select the optimal path based on quantified indicators and unquantified factors. This application uses an AI Agent to extract the core quantified indicators of each path from the candidate loop-free path set, forming a structured data set, including basic attributes (path ID, node sequence, link type combination), performance indicators (latency, bandwidth utilization, packet loss rate, jitter), cost indicators (link lease fee, data transmission cost, total cost), compliance and security indicators (compliance score, security score, encryption level), reliability indicators (availability in the past 30 days, fault recovery time), sustainability indicators (carbon emission intensity, green coefficient), and quality trend data predicted by the AI ​​Agent for the next 24 hours (probability of deterioration, indicator fluctuation range). The standardized data is pushed to a pre-set visualization panel in real time through a RESTful API, with the push frequency consistent with the network data collection cycle (synchronization every 5 seconds) to ensure that the panel data is synchronized with the actual network status. Meanwhile, when candidate paths are updated due to network changes (such as adding / removing paths), the panel refreshes the path list and indicator data in real time to avoid network administrators making decisions based on outdated information. The network administrator selects the optimal path from the multiple candidate loop-free paths based on the indicator data and various unquantified factors to ensure that the optimal path meets the actual needs. The network administrator submits the selection through the "Confirm Optimal" button on the visual panel, and the instruction is immediately synchronized to the SD-WAN controller and AI Agent to trigger the path deployment process.

[0085] In a specific implementation of step S104, another embodiment exists where the method further includes:

[0086] S10431. Collect real-time data of emerging network providers in real time, and obtain evaluation results by real-time evaluation of the service quality of the emerging network providers based on the real-time data.

[0087] S10432. Based on the evaluation results, construct emerging links based on the emerging network providers and dynamically incorporate the emerging links into the path planning scope.

[0088] In steps S10431-S10432, this application also collects real-time data from emerging network providers. This real-time data includes basic provider information (service coverage area, link type, bandwidth limit, encryption level), real-time Quality of Service (QoS) data (latency, packet loss rate, jitter, link availability), cost data (link rental unit price, data transmission rate, package discount policies), compliance qualifications (data privacy compliance certification, regional operation license), and operation and maintenance response capabilities (fault reporting response time, average repair time). The collection method involves three channels working together: Provider Open API: connecting to the real-time monitoring interface of emerging providers to pull QoS and cost data at a frequency of 3 seconds per instance; Third-party monitoring platforms (such as ThousandEyes, Akamai EdgeScope): independently collecting cross-regional transmission quality data of provider links through globally distributed probes to verify data authenticity; Industry dynamic crawler: crawling industry news and technology release information to obtain dynamics such as service expansion and technology upgrades of emerging providers.After standardized processing, the collected data forms a unified supplier data archive. Based on the collected real-time data, a multi-dimensional evaluation model is established. The indicators corresponding to this multi-dimensional evaluation model include latency compliance rate (target value <80ms), packet loss rate control (target value <0.5%), link availability (target value ≥99.9%), jitter stability (target value <20ms), and indicators covering cost competitiveness (price advantage compared to mature suppliers of the same type), compliance adaptability (whether it meets the regional compliance requirements for enterprise data transmission), operation and maintenance support capability (fault response time ≤4 hours), and technological foresight (whether it supports advanced functions such as dynamic bandwidth adjustment and green low-carbon transmission). The service quality of emerging suppliers is dynamically scored and graded to obtain the evaluation result. The dynamic scoring network is: Comprehensive Score = 0.3 × Latency Compliance Rate + 0.2 × Packet Loss Rate Control + 0.2 × Link Availability + The evaluation score is calculated using the formula: 0.1 × jitter stability + 0.1 × cost competitiveness + 0.05 × compliance adaptability + 0.05 × operation and maintenance support capability. A three-tier evaluation standard is set: Level A (score ≥ 85): Excellent service quality, directly included in path planning; Level B (60-84): Basically meets transmission requirements, requires small-volume testing for verification; Level C (< 60): Not yet meets requirements, continuous monitoring and optimization are needed. Based on the evaluation results, new links based on the emerging network providers are constructed and dynamically included in the path planning scope. Specifically: For providers with Level A evaluation results: The AI ​​Agent automatically generates link construction instructions, including node access configuration (interconnection parameters with existing SD-WAN edge nodes and core nodes), link transmission protocols (such as IPsec encrypted tunnel configuration), bandwidth reservation thresholds, etc., and synchronizes them to the SD-WAN controller via the gRPC interface; For providers with scores of 60-84, a temporary test link is first constructed, injecting 5%-10% non-core business traffic (such as background data synchronization), and the transmission quality is continuously monitored for 24 hours. Once the formal link is built, AIAgent automatically integrates the feature data of the emerging links (latency, cost, bandwidth, coverage area, compliance attributes, etc.) into a unified feature vector, updating the state space of the network topology. In the subsequent path generation process, the reinforcement learning model treats the emerging links as a component of optional actions (path decision-making), participating in multi-objective optimization and constraint selection on an equal footing with existing links. For example, for cross-border paths, a combination of existing links and emerging satellite links can be prioritized to achieve latency or cost optimization. If there are no serious faults during the testing period (packet loss rate consistently <1%, latency fluctuation <30ms), the evaluation level will be upgraded to Level A, and a commercial link will be formally built. If quality fluctuations occur, feedback will be provided to the supplier for optimization or continuous monitoring.The AI ​​Agent synchronizes real-time quality data of emerging links every 30 seconds. If the evaluation score drops to C level (e.g., link availability is below 99% or packet loss rate suddenly increases), it will be temporarily removed from the path planning scope, triggering an alarm to notify the administrator. Once the supplier fixes the issue and the evaluation rises back to A level, it will be reinstated.

[0089] The following are the effects of the method described in this application in practical use:

[0090] For multinational e-commerce companies, this AI Agent detected the launch of an emerging supplier (such as Telkom's 5G link in Indonesia): its SLA (latency = 40ms, covering Southeast Asia) was queried via API, topology compatibility was assessed using a GNN, and path optimization was implemented through RL simulation. Results: The latency of the China-Indonesia path decreased from 180ms to 90ms, cost was reduced by 25%, and monthly savings exceeded $100,000. Simultaneously, the AI ​​Agent generated a report recommending migrating 20% ​​of traffic to the new link, with an estimated annual ROI of 150%.

[0091] In SaaS mode, the AI ​​Agent runs offline training weekly: using federated learning to aggregate anonymized data from 10+ tenants (differential privacy ε=1.0), updating the global model, and evolving from an initial delay-first strategy to an adaptive hybrid strategy (dynamically adjusted weights) after 3 iterations. In testing, path satisfaction increased from 82% to 97% after iterations, supporting a scale of 1000+ nodes.

[0092] Deployed in remote branches (such as mining sites in Africa), the AI ​​Agent runs at the edge (TensorFlow Lite), combining local sensor data to predict the impact links of dust storms and pre-plan satellite backup paths. Recovery time has been reduced from 5 minutes to 30 seconds.

[0093] This invention is simple to implement, compatible with existing systems (such as integrating Cisco SD-WAN API), highly scalable, and can be quickly deployed through containerization (Docker).

[0094] The SD-WAN controller described in this application includes a control system, a management system, a visualization front-end, and a historical monitoring system.

[0095] Example 2

[0096] This application also provides an optimal path planning system for SD-WAN networks, such as... Figure 4The diagram shows a block diagram of an optimal path planning system for an SD-WAN network. This system implements functions corresponding to the steps described above in executing an optimal path planning method for an SD-WAN network on a terminal device. The device can be understood as a server component including a processor. The optimal path planning system for an SD-WAN network described in this application includes:

[0097] The acquisition module 401 is used to acquire real-time operating indicator data of the SD-WAN network and external global network data through the acquisition interface of the SD-WAN controller, and integrate the real-time operating indicator data and external global network data to obtain a unified feature vector;

[0098] The parsing module 402 is used to parse the natural language routing strategy input by the user according to the preset AI Agent, obtain the target semantic graph, and map the target semantic graph to obtain the objective function and constraint set for user path optimization;

[0099] The generation module 403 is used to generate multiple candidate acyclic paths by using the unified feature vector, objective function and constraint set, with the AIAgent using network topology as the state space and path decision as the action space;

[0100] The execution module 404 is used to select the optimal path from the multiple candidate loop-free paths and execute the optimal path to complete the optimal path planning of the SD-WAN network.

[0101] In one feasible implementation, the execution module includes:

[0102] The weighted scores for the multiple candidate acyclic paths are calculated based on the objective function, and future quality predictions are performed to obtain the quality prediction results.

[0103] The weighted score is dynamically adjusted based on the quality prediction results to select the optimal path from the multiple candidate acyclic paths.

[0104] In one feasible implementation, the execution module further includes:

[0105] Based on path quality data over historical time periods and external influencing factors, the AI ​​Agent predicts the quality change trend of the optimal path in future time periods.

[0106] Determine whether the quality change trend meets the preset quality deterioration conditions to generate the quality prediction result of the optimal path.

[0107] In one feasible implementation, the execution module also includes:

[0108] If the quality prediction result is satisfactory, then the backup path of the candidate loop-free path is generated in advance and injected into the SD-WAN controller queue.

[0109] The SD-WAN controller is controlled to perform a forward-looking switch of the optimal path based on the backup path.

[0110] In one feasible implementation, the execution module further includes:

[0111] Acquire and push the multiple candidate loop-free paths and their corresponding indicator data to a preset visualization panel;

[0112] Based on the aforementioned indicator data and various unquantified factors, the optimal path is selected from the multiple candidate acyclic paths.

[0113] In one feasible implementation, the generation module includes:

[0114] The state space and action space are constructed based on the network topology and path decision sequence of the SD-WAN network, respectively.

[0115] The multiple candidate acyclic paths are generated by training a reinforcement learning model to interact between the state space and the action space.

[0116] In one feasible implementation, the generation module further includes:

[0117] Verify whether there are loop-containing paths among the multiple candidate paths generated by the AI ​​Agent, and obtain the verification results;

[0118] Based on the verification results, the loop-containing paths are filtered, and alternative paths are generated until the multiple candidate loop-free paths are obtained.

[0119] The device further includes:

[0120] Real-time data of emerging network providers is collected in real time to obtain evaluation results by real-time assessment of the service quality of the emerging network providers based on the real-time data;

[0121] Based on the evaluation results, emerging links based on the emerging network providers are constructed and dynamically incorporated into the path planning scope.

[0122] Example 3

[0123] This application also provides an electronic device, such as Figure 5As shown, it includes: a processor 501, a memory 502, and a bus 503. The memory 502 stores machine-readable instructions that can be executed by the processor 501. When the electronic device is running, the processor 501 and the memory 502 communicate through the bus 503. When the machine-readable instructions are executed by the processor 501, the steps of any one of the optimal path planning methods for an SD-WAN network are performed.

Claims

1. An optimal path planning method for an SD-WAN network, characterized in that, The method includes: The real-time operation index data of the SD-WAN network and external global network data are obtained through the acquisition interface of the SD-WAN controller, and the real-time operation index data and external global network data are integrated to obtain a unified feature vector. The target semantic graph is obtained by parsing the natural language routing strategy input by the user based on the pre-set AI Agent, and the objective function and constraint set for user path optimization are obtained by mapping the target semantic graph. Using the unified feature vector, objective function, and constraint set, the AI ​​Agent generates multiple candidate loop-free paths with network topology as the state space and path decision as the action space. The optimal path is selected from the multiple candidate loop-free paths and the optimal path is executed to complete the optimal path planning for the SD-WAN network.

2. The method of claim 1, wherein, Selecting the optimal path from the multiple candidate acyclic paths includes: The weighted scores for the multiple candidate acyclic paths are calculated based on the objective function, and future quality predictions are performed to obtain the quality prediction results. The weighted score is dynamically adjusted based on the quality prediction results to select the optimal path from the multiple candidate acyclic paths.

3. The method of claim 2, wherein, The process of predicting future quality to obtain quality prediction results includes: Based on path quality data over historical time periods and external influencing factors, the AI ​​Agent predicts the quality change trend of the optimal path in future time periods. Determine whether the quality change trend meets the preset quality deterioration conditions to generate the quality prediction result of the optimal path.

4. The method of claim 3, wherein, After generating the quality prediction results for the optimal path, the process includes: If the quality prediction result is satisfactory, then the backup path of the candidate loop-free path is generated in advance and injected into the SD-WAN controller queue. The SD-WAN controller is controlled to perform forward switching of the optimal path based on the backup path.

5. The method of claim 1, wherein, Selecting the optimal path from the multiple candidate acyclic paths also includes: Acquire and push the multiple candidate loop-free paths and their corresponding indicator data to a preset visualization panel; Based on the aforementioned indicator data and various unquantified factors, the optimal path is selected from the multiple candidate acyclic paths.

6. The method of claim 1, wherein, The AI ​​Agent generates multiple candidate acyclic paths using network topology as the state space and path decision as the action space, including: The state space and action space are constructed based on the network topology and path decision sequence of the SD-WAN network, respectively. The multiple candidate acyclic paths are generated by training a reinforcement learning model to interact between the state space and the action space.

7. The method as described in claim 6, characterized in that, Generating the multiple candidate acyclic paths includes: Verify whether there are loop-containing paths among the multiple candidate paths generated by the AI ​​Agent, and obtain the verification results; Based on the verification results, the loop-containing paths are filtered, and alternative paths are generated until the multiple candidate loop-free paths are obtained.

8. The method as described in claim 1, characterized in that, The method further includes: Real-time data of emerging network providers is collected in real time to obtain evaluation results by real-time assessment of the service quality of the emerging network providers based on the real-time data; Based on the evaluation results, emerging links based on the emerging network providers are constructed and dynamically incorporated into the path planning scope.

9. An optimal path planning system for an SD-WAN network, characterized in that, The system includes: The acquisition module is used to acquire real-time operating indicator data of the SD-WAN network and external global network data through the acquisition interface of the SD-WAN controller, and integrate the real-time operating indicator data and external global network data to obtain a unified feature vector; The parsing module is used to parse the natural language routing strategy input by the user according to the preset AI Agent, obtain the target semantic graph, and map the target semantic graph to obtain the objective function and constraint set for user path optimization; The generation module is used to generate multiple candidate acyclic paths using the unified feature vector, objective function, and constraint set, with the AI ​​Agent using network topology as the state space and path decision as the action space. The execution module is used to select the optimal path from the multiple candidate loop-free paths and execute the optimal path to complete the optimal path planning of the SD-WAN network.

10. An electronic device, characterized in that, include: The device includes a processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the electronic device is running, the processor communicates with the memory via the bus. When the machine-readable instructions are executed by the processor, they perform the steps of an optimal path planning method for an SD-WAN network as described in any one of claims 1 to 8.