Network architecture intelligent planning method and system

Abstract

Embodiments of the present disclosure provide a network architecture intelligent planning method and system, which comprises: acquiring network data and constructing a network digital twin, establishing a multi-dimensional network state matrix containing node state, link state and traffic state, designing a local matrix exponential function to construct a discrete dynamic model and a continuous dynamic model, and unifying the two models to the same mathematical framework through a unified transformation operator to form a unified modeling framework. Subsequently, based on the framework, the state space and the action space are defined, a composite reward function is constructed, a deep reinforcement learning model with a fusion Transformer structure is designed, and a multi-agent collaborative learning framework is constructed, and a candidate network architecture scheme set is generated through multi-objective optimization technology. Finally, simulation testing and iterative optimization are carried out in the network digital twin environment, and the optimal network architecture planning scheme is output. Embodiments of the present disclosure realize intelligent planning and continuous optimization of network architecture.

Description

Technical Field

[0001] This disclosure relates to the field of network architecture design, and in particular to a method and system for intelligent network architecture planning. Background Technology

[0002] Network architecture planning is a crucial foundational task in fields such as the internet and data centers. Its core mission is to design the network topology, determine the location of network nodes and link capacity based on business needs and operational objectives to meet multi-dimensional requirements such as bandwidth, latency, and reliability. With the rapid development of 5G / 6G, edge computing, and cloud-native applications, network architectures are becoming increasingly complex, posing significant challenges to traditional planning methods.

[0003] Current mainstream network architecture planning techniques primarily rely on empirical models and mathematical optimization methods. For example, graph-based topology optimization algorithms construct network connections using minimum spanning trees or shortest path algorithms; while resource allocation models based on mixed-integer programming (MIP) solve for the optimal deployment scheme by defining constraints and objective functions. These methods have achieved certain results in static network environments.

[0004] In recent years, some research has begun to combine machine learning with network planning, using historical network operation data to train predictive models and assist in planning decisions. This type of technology typically employs supervised learning methods to build traffic prediction models, combines them with heuristic algorithms to generate network topology schemes, and verifies the feasibility of the schemes through simulation. This approach can improve planning efficiency to some extent, but due to the lack of environmental feedback mechanisms and global optimization capabilities, it struggles to cope with complex and ever-changing network environments. However, existing technologies suffer from several key problems: First, traditional methods mostly employ static planning models, which cannot adapt to the dynamic changes in network services; second, existing algorithms are often designed for a single optimization objective, making it difficult to achieve a balance between multiple dimensions such as bandwidth, latency, energy consumption, and cost; third, the planning process lacks a unified modeling framework for different network layers, leading to local optimization failing to guarantee global optimality; finally, existing methods struggle to effectively integrate mathematical tools such as semi-Lagrange probabilities and finite difference calculus, failing to accurately characterize the dynamic characteristics of network systems and limiting the accuracy and adaptability of planning results. Summary of the Invention

[0005] The purpose of this disclosure is to provide a method and system for intelligent planning of network architecture. This disclosure achieves intelligent planning and dynamic optimization of complex network architecture by introducing digital twin technology, a unified modeling framework for local matrix exponents, a multi-objective reinforcement learning optimization engine, and a dynamic adaptive adjustment mechanism.

[0006] To achieve the above objectives, embodiments of this disclosure provide a network architecture intelligent planning method, including: Acquire network data, preprocess the network data and establish a digital model to construct a network digital twin; Based on the network state data provided by the network digital twin, a network state matrix is ​​constructed and a local matrix exponential function is designed. Based on the local matrix exponential function, a discrete dynamic model and a continuous dynamic model are constructed. The discrete dynamic model and the continuous dynamic model are unified under the same mathematical framework through a unified transformation operator to obtain a unified modeling framework. Based on the network representation generated by the unified modeling framework, a state space and an action space are defined, a reward function is constructed, a reinforcement learning model is designed based on the state space, the action space, and the reward function, and a collaborative learning framework is constructed. Based on the learning results of the collaborative learning framework, an optimal solution set is generated through multi-objective optimization techniques to obtain a set of candidate network architecture schemes. Based on the candidate network architecture scheme set, a planning constraint model is generated. According to the planning constraint model and the candidate network architecture scheme set, the deployment location of network nodes, link capacity configuration, resource scheduling strategy and fault self-healing path are determined. Simulation test and iterative optimization are carried out in the network digital twin environment to output the optimal network architecture planning scheme. According to the deployed optimal network architecture planning scheme, network operation data is collected in real time. Based on the network operation data, network performance deviations are identified and the network architecture planning model is updated and optimized. Based on the updated network architecture planning model, dynamic adjustment of the network architecture is performed.

[0007] Furthermore, the steps of acquiring network data, preprocessing the network data, establishing a digital model, and constructing a network digital twin include: Based on the network data, which includes network topology data, device parameters, traffic load, and service requirement data, the network data is cleaned, standardized, and structured to obtain a standardized network dataset. Based on the standardized network dataset, a digital model of network nodes, links, and service flows is established to construct a digital mapping body of network elements. Based on the digital mapping body of network elements, network traffic distribution patterns, service fluctuation characteristics, and fault propagation patterns are extracted through time series analysis and pattern recognition algorithms to form a network behavior pattern library. Based on the digital mapping of the network elements and the network behavior pattern library, a network digital twin model containing the coupling relationship between the physical layer, logical layer and business layer is constructed. By comparing and verifying the simulation data output by the network digital twin model with the actual network data, the parameters of the network digital twin model are continuously adjusted and optimized to obtain the network digital twin.

[0008] Furthermore, based on the network state data provided by the network digital twin, a network state matrix is ​​constructed and a local matrix exponential function is designed. Based on the local matrix exponential function, discrete and continuous dynamic models are constructed. A unified transformation operator unifies the discrete and continuous dynamic models under the same mathematical framework, resulting in a unified modeling framework, including: Based on the network status data provided by the network digital twin, a multi-dimensional network status matrix is ​​constructed, which includes node status, link status, and traffic status. Based on the multidimensional network state matrix, a local matrix exponential function incorporating adaptive parameters and a region segmentation mechanism is designed. Based on the local matrix exponential function, the network dynamic change process is spatiotemporally discretized. A high-order finite difference scheme is constructed to establish a discrete mathematical model of network state transition, and a finite difference network dynamic model is generated as the discrete dynamic model. Based on the local matrix exponential function, a semi-Lagrange framework is constructed to establish the evolution equation of network traffic and service demand, generating a continuous semi-Lagrange network model as the continuous dynamic model. Based on the local matrix exponential function, a unified transformation operator is designed to unify the discrete representation of the dynamic model of the finite difference network and the continuous representation of the continuous model of the semi-Lagrange network under the same mathematical framework, thereby constructing a unified representation system for the network system and obtaining the unified modeling framework.

[0009] Further, based on the network representation generated by the unified modeling framework, a state space and action space are defined, a reward function is constructed, a reinforcement learning model is designed based on the state space, the action space, and the reward function, and a collaborative learning framework is constructed. Based on the learning results of the collaborative learning framework, an optimal solution set is generated through multi-objective optimization techniques to obtain a set of candidate network architecture schemes, including: Based on the network representation generated by the unified modeling framework, a network state space containing node configuration, link parameters, and resource allocation strategies is defined as the state space, and an action space including node deployment, link planning, and resource scheduling operations is designed. Based on the state space and the action space, a composite reward function with a weight adaptive mechanism is constructed as the reward function to achieve a dynamic balance between bandwidth cost, latency performance, energy efficiency and reliability, thus forming a multi-dimensional reward evaluation mechanism for network planning. Based on the state space, the action space, and the multidimensional reward evaluation mechanism, a deep reinforcement learning model integrating the Transformer structure is designed as the reinforcement learning model. A mapping relationship between network state and optimal action is established, and a decision network with long-term memory and attention mechanism is constructed. Based on the decision network, a multi-agent collaborative learning framework is constructed as the collaborative learning framework to realize collaborative optimization among planning tasks of different network domains and different levels, and to establish a multi-agent system based on shared experience and distributed decision-making. Based on the collaborative learning results of the multi-agent system, a Pareto optimal solution set that satisfies the bandwidth cost, latency performance, energy efficiency, and reliability is generated through multi-objective optimization techniques such as non-dominated sorting and congestion distance calculation, thus obtaining the candidate network architecture scheme set.

[0010] Further, the step of generating a planning constraint model based on the candidate network architecture scheme set, and determining the network node deployment location, link capacity configuration, resource scheduling strategy, and fault self-healing path according to the planning constraint model and the candidate network architecture scheme set, includes: Based on the candidate network architecture scheme set and business scenario requirements, analyze business characteristics, security level, available resources and deployment limitations, and generate a scenario-based planning constraint model; Based on the planning constraint model, the candidate network architecture scheme set is screened, and combined with geographic information, resource distribution and connectivity requirements, the optimal deployment location of network nodes is determined by the location optimization algorithm to generate a network node deployment scheme. Based on the network node deployment scheme, the network traffic distribution pattern and service fluctuation characteristics in the network digital twin are utilized to construct a traffic prediction model through a time series prediction algorithm. This model analyzes service growth trends and fluctuation characteristics, determines the capacity configuration and expansion strategy for each link, and generates a link capacity planning scheme that includes redundancy design.

[0011] Furthermore, the step of determining the network node deployment location, link capacity configuration, resource scheduling strategy, and fault self-healing path based on the planning constraint model and the candidate network architecture scheme set also includes: Based on the network node deployment scheme and the link capacity planning scheme, a resource allocation algorithm and a dynamic scheduling mechanism are designed to determine the allocation strategies for computing resources, storage resources and bandwidth resources, and to generate a resource scheduling strategy scheme. Based on the network node deployment scheme and the link capacity planning scheme, the dynamic characteristics of the network are characterized by the local matrix exponential function and the unified modeling framework, potential fault points and fault propagation paths are analyzed, and fault risk assessment results are generated. Based on the fault risk assessment results, a multi-level backup strategy and a reconstruction strategy are designed. The multi-level backup strategy includes primary and backup link configuration and redundant node deployment. The reconstruction strategy includes traffic rerouting mechanism and resource reallocation mechanism to build the network's fault self-healing capability and obtain a fault self-healing path planning scheme.

[0012] Furthermore, the step of collecting network operation data in real time according to the deployed optimal network architecture planning scheme, identifying network performance deviations based on the network operation data and updating and optimizing the network architecture planning model, and performing dynamic adjustments to the network architecture based on the updated network architecture planning model includes: Based on the optimal network architecture plan, network operation status data, resource usage and abnormal events are collected in real time to obtain network operation data stream; Based on the network operation data stream and expected performance indicators, network performance deviations and potential problems are identified by constructing an anomaly detection model and trend analysis algorithm, and a performance deviation analysis report is generated. Based on the performance deviation analysis report, information on network performance index deviations and resource utilization anomalies from actual operation feedback is extracted. The parameters and strategies of the reinforcement learning model in the network architecture planning model are updated using incremental learning and transfer learning techniques. The constraints of the planning constraint model are adjusted according to the network performance index deviations, constructing a self-evolving continuous learning mechanism to obtain the updated network architecture planning model. This updated model is built upon the reinforcement learning model, the planning constraint model, and the unified modeling framework. Based on the updated network architecture planning model, a dynamic network architecture adjustment scheme is generated according to the current deployment status of the optimal network architecture planning scheme. The dynamic network architecture adjustment scheme includes node reconfiguration instructions, link adjustment instructions, and resource reallocation instructions, and performs dynamic adjustments to the optimal network architecture planning scheme.

[0013] Furthermore, the step of performing simulation testing and iterative optimization in the network digital twin environment to output the optimal network architecture planning scheme includes: Based on the overall network architecture planning scheme, by constructing simulation scenarios and conducting simulation tests in the network digital twin environment, the performance indicators of the network architecture planning scheme, including throughput, latency, packet loss rate and energy consumption, under different load conditions, traffic patterns and fault scenarios are evaluated, and a performance evaluation report is generated. Based on the performance evaluation report and preset performance targets, the network architecture planning scheme is iteratively optimized by adjusting the deployment location of network nodes, link capacity configuration, and resource scheduling strategy until the performance indicators including throughput, latency, packet loss rate, and energy consumption all meet the preset performance target requirements, and the optimal network architecture planning scheme is output.

[0014] Furthermore, the step of unifying the discrete dynamic model and the continuous dynamic model under the same mathematical framework through a unified transformation operator includes: Based on the local matrix exponential function, when performing spatiotemporal discretization processing on the dynamic change process of the network, a high-order finite difference scheme is constructed using the central difference scheme or the upwind difference scheme to establish a discrete mathematical model of the network state transition, thereby obtaining the dynamic model of the finite difference network. Based on the local matrix exponential function, by constructing a semi-Lagrange framework, tracing the traffic evolution trajectory along the feature lines, establishing the partial differential evolution equations of network traffic and service demand, and obtaining the semi-Lagrange network continuous model; Based on the finite difference network dynamic model and the semi-Lagrange network continuous model, a unified transformation operator is defined to map the discrete time step of the finite difference network dynamic model to the continuous time manifold of the semi-Lagrange network continuous model, thereby achieving mathematical unification of discrete and continuous representations and obtaining the unified modeling framework.

[0015] This disclosure also provides an intelligent network architecture planning system, including: A digital twin construction module is used to acquire network data, preprocess the network data and establish a digital model to construct a network digital twin. The unified modeling module is used to construct a network state matrix and design a local matrix exponential function based on the network state data provided by the network digital twin. Based on the local matrix exponential function, a discrete dynamic model and a continuous dynamic model are constructed. The discrete dynamic model and the continuous dynamic model are unified under the same mathematical framework through a unified transformation operator to obtain a unified modeling framework. The multi-objective optimization module is used to define the state space and action space based on the network representation generated by the unified modeling framework, construct a reward function, design a reinforcement learning model and construct a collaborative learning framework based on the state space, the action space and the reward function, and generate an optimal solution set based on the learning results of the collaborative learning framework through multi-objective optimization techniques to obtain a set of candidate network architecture schemes. The intelligent planning and dynamic adjustment module is used to generate a planning constraint model based on the candidate network architecture scheme set, determine the network node deployment location, link capacity configuration, resource scheduling strategy, and fault self-healing path according to the planning constraint model and the candidate network architecture scheme set, conduct simulation testing and iterative optimization in the network digital twin environment, output the optimal network architecture planning scheme, collect network operation data in real time according to the deployed optimal network architecture planning scheme, identify network performance deviations based on the network operation data and update and optimize the network architecture planning model, and perform dynamic adjustment of the network architecture based on the updated network architecture planning model.

[0016] Compared with the prior art, the beneficial effects of the embodiments of this disclosure include: 1. This embodiment achieves a unified mathematical expression for discrete and continuous characteristics in network systems through a local matrix exponential unified modeling framework, solving the problem of the separation between discrete and continuous representations in traditional network modeling and improving model accuracy. At the same time, it adopts a multi-objective reinforcement learning Pareto optimal engine to achieve balanced optimization among multiple dimensions such as bandwidth cost, latency performance, energy efficiency and reliability, overcoming the limitations of single-objective optimization.

[0017] 2. This embodiment combines digital twins and reinforcement learning into a dual-driven optimization mechanism, forming an optimization system encompassing simulation, learning, verification, and deployment, significantly improving planning efficiency and accuracy. Simultaneously, through dynamic adaptive network adjustment and knowledge accumulation mechanisms, the planning strategy is continuously optimized based on actual operational data, achieving continuous improvement in system capabilities. Furthermore, this embodiment utilizes a local matrix exponentially driven fault self-healing path construction to accurately identify potential fault points and predict fault propagation paths, constructing an efficient fault self-healing mechanism and enhancing network reliability. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating the intelligent network architecture planning method of this embodiment; Figure 2 This is a schematic diagram of the network architecture intelligent planning system in this embodiment. Detailed Implementation

[0020] The preferred embodiments of the present disclosure are described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the embodiments of the present disclosure.

[0021] Example 1: Figure 1 A flowchart illustrating the intelligent network architecture planning method of this embodiment is shown, as follows: Figure 1 As shown, this embodiment provides a network architecture intelligent planning method, which includes: Step S1: Acquire network data, preprocess the network data and establish a digital model to construct a network digital twin.

[0022] In this step, comprehensive network data is first collected, including network topology data (node ​​locations, link connections, device types and configurations), historical traffic data (service traffic distribution, fluctuation characteristics, peak period features), performance indicator data (latency, packet loss rate, throughput, energy consumption), and service requirement data (service level agreements, key service characteristics, user distribution). After collection, this raw data undergoes systematic preprocessing: the data cleaning stage removes outliers, supplements missing data, and corrects erroneous records; the data standardization stage unifies the format, units, and timestamps of data from different sources; the data aggregation stage performs statistical aggregation according to time granularity and spatial scope, extracting key features; the feature extraction stage applies methods such as principal component analysis and autoencoders to extract the essential features of the data, reduce dimensionality, and retain key information. Based on the preprocessed data, a high-fidelity network digital twin model is constructed, achieving an accurate digital mirror of the physical network.

[0023] Based on this, the twin comprises a network topology layer (precisely representing the physical and logical connections of the network), a device layer (simulating the functional characteristics and performance parameters of various network devices), a service layer (simulating the traffic characteristics and service requirements of various services), and an environment layer (simulating the impact of external conditions on the network). Through this tightly coupled structure of four layers, a complete network digital twin is formed, which is beneficial for subsequent modeling and optimization.

[0024] Step S2: Based on the network state data provided by the network digital twin, construct a network state matrix and design a local matrix exponential function. Based on the local matrix exponential function, construct a discrete dynamic model and a continuous dynamic model. Through a unified transformation operator, unify the discrete dynamic model and the continuous dynamic model under the same mathematical framework to obtain a unified modeling framework.

[0025] In step S2, based on the real-time and historical state data provided by the network digital twin in step S1, a multi-dimensional network state matrix is ​​constructed to comprehensively characterize the dynamic characteristics of the network system. This state matrix includes several core dimensions: node state dimension (capturing key indicators such as processing capacity, cache status, and load level of each network node), link state dimension (recording transmission characteristics such as bandwidth utilization, congestion level, and latency changes of each communication link), and traffic state dimension (describing the distribution, direction, and changing trends of service flows in the network).

[0026] Next, a local matrix exponential function is designed as a mathematical tool to characterize the dynamic properties of the network system. This function can efficiently express the evolution of the state matrix in a local region, forming a compact mathematical description of the system's changes. Based on this function, a discrete dynamic model is constructed on the one hand, using the central difference scheme or the upwind difference scheme for spatiotemporal discretization to accurately capture discrete events and abrupt changes in the network; on the other hand, a continuous dynamic model is constructed, using a semi-Lagrange framework to trace the flow evolution trajectory along the characteristic lines, accurately describing the continuous evolution process in the network.

[0027] Finally, a unified transformation operator is defined to map discrete and continuous dynamic models under the same mathematical framework, thereby unifying discrete and continuous representations and forming a complete unified modeling framework. This lays the mathematical foundation for the comprehensive analysis and optimization of network systems.

[0028] Step S3: Based on the network representation generated by the unified modeling framework, define the state space and action space, construct a reward function, design a reinforcement learning model and construct a collaborative learning framework based on the state space, action space and reward function, and generate an optimal solution set based on the learning results of the collaborative learning framework through multi-objective optimization techniques to obtain a candidate network architecture scheme set.

[0029] In this step, based on the unified modeling framework, the state space of the network architecture planning problem is first defined, including the network topology state (node ​​distribution, link connection), resource configuration state (computing, storage, and bandwidth resource allocation), and service distribution state (traffic patterns, user distribution), comprehensively representing the current state of the network system. At the same time, the action space is defined, including topology adjustment actions (adding / deleting nodes, modifying links), resource allocation actions (adjusting resource configuration ratios, optimizing scheduling strategies), and traffic engineering actions (rerouting, load balancing), covering all possible operations in network architecture planning.

[0030] Next, a composite reward function is constructed, comprehensively considering performance rewards (increased throughput, reduced latency, etc.), cost rewards (equipment investment, operating costs, etc.), reliability rewards (fault tolerance, recovery speed, etc.), and scalability rewards (the ability to support business growth). A weight balancing mechanism coordinates and optimizes these multiple objectives. Based on the defined state space, action space, and reward function, a deep reinforcement learning model incorporating the Transformer structure is designed. This model fully utilizes self-attention mechanisms to capture long-range dependencies between network elements and employs multi-head attention mechanisms to process different types of network information. Simultaneously, a multi-agent collaborative learning framework is constructed, decomposing the network architecture planning problem into multiple sub-problems, which are then collaboratively solved by specialized agents. Finally, multi-objective optimization techniques, such as the improved NSGA-III algorithm and multi-objective particle swarm optimization, are applied to process the output of the learning framework, generating a Pareto optimal solution set and forming a diverse set of candidate network architecture schemes.

[0031] Step S4: Based on the candidate network architecture scheme set, generate a planning constraint model. Determine the network node deployment location, link capacity configuration, resource scheduling strategy, and fault self-healing path according to the planning constraint model and the candidate network architecture scheme set. Perform simulation testing and iterative optimization in the network digital twin environment, output the optimal network architecture planning scheme, collect network operation data in real time according to the deployed optimal network architecture planning scheme, identify network performance deviations based on the network operation data, update and optimize the network architecture planning model, and perform dynamic adjustment of the network architecture based on the updated network architecture planning model.

[0032] In this step, based on the candidate network architecture scheme set, a scenario-based planning constraint model is first generated. This model analyzes business characteristics (traffic model, service response requirements), security levels (data protection requirements, access control policies), available resources (budget constraints, equipment resource limits), and deployment constraints (geographical constraints, regulatory requirements), constructing a complete constraint system. Secondly, a detailed design is performed based on the planning constraint model: using location optimization algorithms combined with geographic information, resource distribution, and connectivity requirements, the optimal deployment locations of network nodes are determined; based on the traffic prediction model in the network digital twin, business growth trends and fluctuation characteristics are analyzed to determine the capacity configuration and expansion strategies for each link; resource allocation algorithms and dynamic scheduling mechanisms are designed to determine the allocation strategies for computing, storage, and bandwidth resources; potential failure points and propagation paths are analyzed, and multi-level backup and reconstruction strategies are designed to build the network's fault self-healing capability.

[0033] Understandably, after the design is completed, various simulation scenarios are built in the network digital twin environment, including different load conditions, traffic patterns and fault scenarios, to comprehensively evaluate the performance indicators of the planning scheme, such as throughput, latency, packet loss rate and energy consumption, and generate a performance evaluation report. Based on the evaluation results, the planning scheme is iteratively optimized by adjusting node positions, link capacity and resource policies until all performance indicators meet the preset targets, and the optimal network architecture planning scheme is output.

[0034] Specifically, after the solution is deployed, network operation status data, resource usage and abnormal events are continuously collected. Performance deviations are identified through anomaly detection models and trend analysis algorithms, and analysis reports are generated. Based on the analysis results, the parameters of the reinforcement learning model are updated through incremental learning and transfer learning techniques, and the planning constraints are adjusted to build a self-evolving continuous learning mechanism.

[0035] In summary, based on the updated planning model, this embodiment generates a dynamic adjustment scheme that includes instructions for node reconfiguration, link adjustment, and resource reallocation, thereby achieving continuous optimization of the network architecture and ensuring that the network always maintains an efficient and reliable operating state.

[0036] In this embodiment, step S1, which involves acquiring network data, preprocessing the network data, establishing a digital model, and constructing a network digital twin, includes: Step S11: Based on the network data, which includes network topology data, device parameters, traffic load, and service requirement data, the network data is cleaned, standardized, and structured to obtain a standardized network dataset.

[0037] Understandably, network data acquisition is a fundamental step in building a network digital twin. This embodiment designs a multi-source heterogeneous data acquisition framework, collecting comprehensive network data through four types of interfaces. The network management system API interface provides access to network devices and services, collecting static and dynamic data such as network topology change records, device configuration history, and operational status logs. The SNMP protocol interface, targeting network devices supporting the protocol, collects real-time performance metrics such as CPU utilization, memory usage, interface status, and traffic throughput through standard MIB object access methods. The Netflow / sFlow data flow interface collects traffic characteristic data, including source and destination addresses, ports, protocol types, and traffic volume, by deploying probes at key network nodes, providing raw data for traffic analysis. The configuration file analysis interface extracts key configuration information such as business policies, security rules, and quality of service settings from structured and semi-structured texts such as device configuration files, business planning documents, and resource lists through a parsing engine.

[0038] These interfaces together constitute a complete data acquisition system, ensuring the comprehensiveness, timeliness, and accuracy of the collected data, and providing high-quality raw materials for subsequent data processing.

[0039] In this step, data preprocessing is a crucial step in transforming raw network data into a standardized dataset. For the collected multi-source heterogeneous data, this embodiment designs a three-stage data processing flow.

[0040] First, the data cleaning stage mainly addresses data quality issues, including outlier detection and handling, using the statistical analysis-based Z-Score method and the density-based Local Outlier Factor (LOF) method to identify outlier data points; missing value completion, using methods such as mean / median imputation, time series interpolation, or machine learning prediction based on data characteristics; and duplicate data removal, identifying and merging duplicate records through hash indexing and timestamp comparison.

[0041] Secondly, the data standardization stage addresses the issue of inconsistent data formats from different sources, including unifying units of measurement by converting data from different units to a unified measurement standard; normalizing numerical values ​​by using Min-Max or Z-score standardization to map data to a specific interval for easier subsequent analysis; and aligning time by using interpolation or downsampling methods to unify data from different sampling frequencies onto the same time axis.

[0042] Furthermore, the data structuring stage organizes the cleaned and standardized data into a structured data model, including relational model construction to establish relationships between various types of data; data classification and labeling to add semantic tags to support multidimensional analysis; and time series organization to organize dynamic data in chronological order to support time series analysis.

[0043] Understandably, through these multiple stages of processing, the raw network data is transformed into a structured and standardized network dataset, providing a foundation for subsequent modeling.

[0044] Step S12: Based on the standardized network dataset, establish digital models of network nodes, links, and service flows to construct a digital mapping body of network elements. Based on the digital mapping body of network elements, extract network traffic distribution patterns, service fluctuation characteristics, and fault propagation patterns through time series analysis and pattern recognition algorithms to form a network behavior pattern library.

[0045] In this step, the network element digital mapping is a digital representation of the physical and logical components of the real network. Based on a standardized network dataset, this embodiment constructs three core digital models.

[0046] Among them, the network node digital model describes the characteristics and status of various devices in the network, including static attributes (such as device type, manufacturer, model, function set and geographical location) and dynamic status (such as computing resource utilization, forwarding performance, cache status and queue length), and accurately characterizes node characteristics through a multi-dimensional parameter set.

[0047] The link digital model expresses the physical and logical characteristics of the connection in the network, including link type (copper cable, fiber optic, wireless, etc.), topology (end-to-end connection information), performance parameters (bandwidth, latency, packet loss rate, and jitter, etc.), and operating status (active, degraded, or failed), and fully describes the link characteristics.

[0048] The business flow digital model captures the characteristics of business transmission in the network, including business type (real-time streaming media, batch file transfer, interactive applications, etc.), traffic characteristics (packet size distribution, burstiness, periodicity, etc.), quality of service requirements (latency sensitivity, packet loss tolerance, etc.), and transmission path, comprehensively expressing the characteristics of business flows. These three types of digital models are combined through correlation to form a complete digital mapping of network elements, laying the foundation for network behavior pattern extraction and digital twin construction.

[0049] Understandably, the network behavior pattern library is a core intelligent component of the digital twin, used to capture and represent the dynamic behavioral characteristics of the network system. Based on the digital mapping of network elements, two types of advanced analysis techniques are applied to extract network behavior patterns.

[0050] Among them, time series analysis technology is used to discover patterns in the time dimension of the network, mainly using three methods: time series decomposition, which decomposes traffic data into trend, seasonal and residual terms to identify long-term growth trends and periodic changes; autoregressive integral moving average (ARIMA) model, which captures the autocorrelation and periodicity of traffic time series to identify daily fluctuation patterns; and wavelet transform analysis, which discovers characteristic patterns at different time scales through multi-scale decomposition to identify multi-level periodic patterns.

[0051] Pattern recognition technology is used to discover structured features in network behavior, primarily employing four methods: unsupervised clustering algorithms (K-means and DBSCAN) to classify services based on traffic characteristics and identify typical service patterns; association rule mining (Apriori algorithm) to discover causal relationships between different network events and identify fault propagation chains; anomaly detection algorithms (isolation forest and single-class SVM) to identify abnormal behavioral patterns in the network; and graph pattern mining to discover spatial patterns of traffic distribution and congestion propagation through graph structure analysis of the network topology. These analytical results collectively form a network behavior pattern library, containing traffic distribution patterns (such as temporal and spatial distribution characteristics), service fluctuation characteristics (such as suddenness and periodicity), and fault propagation patterns (such as cascading fault paths and critical failure points), providing dynamic behavioral models for network digital twins.

[0052] Step S13: Based on the digital mapping body of the network elements and the network behavior pattern library, construct a network digital twin model that includes the coupling relationship between the physical layer, logical layer and business layer. By comparing and verifying the simulation data output by the network digital twin model with the actual network data, continuously adjust and optimize the parameters of the network digital twin model to obtain the network digital twin.

[0053] Understandably, a network digital twin model is a highly realistic virtual network representation that accurately reflects the structure, characteristics, and behavior of a real network. The network digital twin model constructed in this embodiment adopts a three-layer coupled architecture to comprehensively capture the multi-layered characteristics of the network system.

[0054] Specifically, the physical layer model expresses the physical entities and connections of the network, including the physical characteristics of devices (computing power, storage capacity, energy consumption characteristics, etc.), the physical characteristics of links (transmission medium, physical topology, transmission delay, etc.), and the geographical distribution characteristics (location coordinates, spatial distribution, environmental factors, etc.), thus constructing a physical digital twin of the network.

[0055] The logical layer model describes the functional logic and protocol characteristics of the network, including the network protocol stack implementation (routing protocols, transport protocols, application protocols, etc.), service function chains (VPN, firewall, load balancing, and other service links), and logical topology (VLAN, VxLAN, and other overlay networks), constructing a logical digital twin of the network.

[0056] The business layer model expresses the business characteristics and user experience carried by the network, including business models (business type, requirement characteristics, priority, etc.), user behavior models (access patterns, operation sequences, etc.) and experience quality models (response time, interaction smoothness, etc.), and constructs a business digital twin of the network.

[0057] In summary, these three models in this step are tightly coupled through a cross-layer mapping mechanism to form a unified network digital twin model, which can accurately reflect the inter-layer influence and overall behavior.

[0058] Model parameter optimization is a crucial step in ensuring the high fidelity of network digital twins. This embodiment employs an iterative verification and optimization method to continuously improve the model's accuracy.

[0059] First, network simulation is performed using a twin model to generate various metrics data that simulate network operation, including device status metrics (such as CPU utilization, memory usage, etc.), link performance metrics (such as bandwidth utilization, latency, packet loss rate, etc.), and service experience metrics (such as response time, throughput, etc.).

[0060] Next, these simulation data are compared and analyzed with similar indicator data collected from actual networks. The error rate and similarity of various indicators are calculated, including evaluation indicators such as root mean square error (RMSE), mean absolute percentage error (MAPE), and Pearson correlation coefficient.

[0061] Based on the comparative analysis results, for the parts with larger errors, parameter optimization algorithms are applied to adjust the model parameters and improve the model accuracy. Parameter optimization employs multiple methods, including gradient descent to adjust continuous parameters, genetic algorithms to optimize complex multi-parameter combinations, and Bayesian optimization to find a better solution within a fewer iterations. This embodiment, through continuous iteration of this verification and optimization process, gradually improves the model accuracy until it reaches a predetermined accuracy threshold, ultimately obtaining a high-fidelity network digital twin that accurately reflects the structural characteristics and dynamic behavior of the actual network, providing a reliable virtual environment for subsequent network architecture planning.

[0062] In this embodiment, in step S2, the process of constructing a network state matrix and designing a local matrix exponential function based on the network state data provided by the network digital twin, constructing a discrete dynamic model and a continuous dynamic model based on the local matrix exponential function, and unifying the discrete dynamic model and the continuous dynamic model under the same mathematical framework through a unified transformation operator to obtain a unified modeling framework includes: Step S21: Based on the network status data provided by the network digital twin, construct a multi-dimensional network status matrix that includes node status, link status, and traffic status.

[0063] In this step, the multidimensional network state matrix is ​​a comprehensive mathematical expression of the current state of the network system. Based on the rich state data provided by the network digital twin, this embodiment constructs a high-dimensional matrix structure to comprehensively capture the state characteristics of the network system. This matrix can be represented as a multidimensional tensor, containing a spatial dimension (distribution of network nodes and links), a state dimension (different types of state parameters), and a time dimension (the evolution of the state over time).

[0064] Specifically, the matrix contains three types of core state information: the node state matrix captures the key state parameters of all nodes in the network, including computing resource status (CPU utilization, memory usage, processing queue length, etc.), storage resource status (storage capacity, I / O performance, cache hit rate, etc.), and device health status (temperature, power status, hardware alarms, etc.), forming a complete set of indicators characterizing node status; the link state matrix describes the performance and state characteristics of all connections in the network, including transmission performance (bandwidth utilization, remaining capacity, congestion level, etc.), transmission quality (latency, jitter, packet loss rate, etc.), and link health status (error packet rate, signal-to-noise ratio, link availability, etc.), comprehensively characterizing the link operating status; and the traffic state matrix expresses the distribution and characteristics of service flows in the network, including traffic distribution (the size and direction of traffic on each node and link), traffic composition (the proportion of traffic of different service types), and traffic change characteristics (growth rate, fluctuation range, periodicity, etc.).

[0065] Understandably, these state matrices are organized according to time series, forming a complete record containing historical and current states, providing a data foundation for subsequent dynamic modeling.

[0066] Step S22: Based on the multidimensional network state matrix, design a local matrix exponential function that includes adaptive parameters and a region segmentation mechanism.

[0067] Understandably, the local matrix exponential function is used to describe the temporal evolution characteristics of network states. Traditional matrix exponential functions suffer from problems such as excessive parameters, high computational complexity, and poor global adaptability when applied to large-scale networks. To address these challenges, this embodiment designs a local matrix exponential function with adaptive parameters and a region segmentation mechanism.

[0068] Specifically, the adaptive parameter mechanism refers to the automatic adjustment of function parameters based on local network characteristics. This is achieved by monitoring the changing trends and fluctuations in network conditions, and calculating optimal parameter values ​​in real time to match the function characteristics with network behavior. For example, in areas of stable network traffic, parameters are adjusted to settings suitable for slow changes; while in areas of drastic traffic fluctuations, parameters automatically switch to configurations suitable for rapid changes.

[0069] Region segmentation divides the network space into multiple sub-regions, each applying an independent set of parameters to more accurately describe the differentiated characteristics of different areas within the network. Segmentation methods are primarily based on network topology and functional partitioning, such as using different parameter configurations for access areas, aggregation areas, and core areas. This localization significantly reduces computational complexity while improving the model's adaptability to the heterogeneity of the network space.

[0070] Based on this, the mathematical expression of the local matrix exponential function adopts the matrix exponential expansion, where the matrix parameters are constructed according to the characteristic spectrum of the network state matrix, ensuring that the function can accurately express the evolution characteristics of the network system in different regions and under different states.

[0071] Step S23: Based on the local matrix exponential function, the network dynamic change process is spatiotemporally discretized. A high-order finite difference scheme is constructed to establish a discrete mathematical model of network state transition, and a finite difference network dynamic model is generated as the discrete dynamic model.

[0072] In this step, the finite difference network dynamic model focuses on capturing the discrete change characteristics of the network state. Based on the local matrix exponential function, this embodiment performs spatiotemporal discretization processing on the network dynamic change process, discretizing the continuous time and spatial dimensions into finite time steps and spatial grids, and establishing a discrete mathematical model of network state transition.

[0073] In constructing high-order finite difference schemes, appropriate difference schemes are selected based on the variation characteristics of different network parameters: for parameters with smooth changes (such as most traffic load changes), the central difference scheme is adopted. This scheme samples symmetrically in the time dimension and has high accuracy (the truncation error is on the order of the square of the time step), which can accurately capture smooth change trends; for parameters with steep changes or oscillations (such as sudden changes in traffic or fault states), the upwind difference scheme is adopted. This scheme samples upstream when calculating the rate of change of the current state, has good numerical stability, and can effectively handle abrupt changes in the network without generating numerical oscillations.

[0074] In critical network areas where high precision is required, this embodiment also applies a fourth-order precision compact difference scheme to achieve high-precision difference approximation with fewer grid points. Through this high-order finite difference scheme, the discrete-time evolution of the network state is precisely mathematically represented, forming a complete finite difference network dynamic model, effectively capturing discrete events in the network (such as device state changes, link failures, and service switching) and their evolution.

[0075] Step S24: Based on the local matrix exponential function, a semi-Lagrange framework is constructed to establish the evolution equation of network traffic and service demand, generating a semi-Lagrange network continuous model as the continuous dynamic model.

[0076] It's worth noting that the semi-Lagrangian network continuous model focuses on describing the continuous changes in network traffic and service demands. Unlike the traditional Eulerian perspective, the semi-Lagrangian method adopts a matter-tracking approach, tracing system evolution along the characteristic lines of traffic transmission, making it suitable for handling continuous processes such as traffic transmission and changes in service demands within a network.

[0077] The semi-Lagrange framework constructed in this embodiment mainly includes several core components: the feature line calculation module calculates the trajectory (feature line) of flow particles in the network based on the network topology and flow transmission path, capturing the migration law of flow in space and time; the matter derivative calculation module calculates the total derivative of the network state quantity along the feature line, accurately describing the rate of change of flow particles with time, avoiding the numerical difficulties of nonlinear convection term processing in the Euler perspective; the grid projection module maps the results calculated along the feature line back to a fixed computational grid, realizing the combination of the feature line method and the traditional grid method.

[0078] The greatest advantage of this semi-Lagrange framework lies in its ability to accurately handle high-speed flow phenomena in networks, effectively avoiding the numerical instability of traditional Eulerian methods when dealing with high Reynolds number flows. It is suitable for describing continuous but rapidly changing phenomena in networks, such as burst traffic, congestion wave propagation, and fluctuations in service demand. This embodiment establishes a system of partial differential equations for the evolution of network traffic and service demand using this method, forming a continuous semi-Lagrange network model, providing a high-precision mathematical framework for describing continuous changes in networks.

[0079] Step S25: Based on the local matrix exponential function, design a unified transformation operator to unify the discrete representation of the dynamic model of the finite difference network and the continuous representation of the continuous model of the semi-Lagrange network under the same mathematical framework, construct a unified representation system for the network system, and obtain the unified modeling framework.

[0080] In this step, the unified transformation operator is a key mathematical tool for fusing discrete and continuous models. Traditional network modeling methods often treat discrete events and continuous processes separately, making it difficult for the model to comprehensively and accurately describe the complex dynamic characteristics of the network system.

[0081] To effectively address this issue, this embodiment designs a unified transformation operator that unifies the discrete representation of the dynamic model of the finite difference network and the continuous representation of the continuous model of the semi-Lagrange network under the same mathematical framework. The core of the unified transformation operator is to construct a special mapping function that maps the discrete time step to the continuous time manifold, achieving seamless unification of the two representation methods.

[0082] The operator design is based on several key principles: the point value consistency principle, which ensures that the transformed continuous representation matches the original discrete value precisely at discrete time nodes; the derivative consistency principle, which guarantees that the derivative value of the continuous representation matches the difference approximation value of the discrete model at discrete time nodes; and the energy conservation principle, which ensures that the transformation process does not introduce additional energy and maintains the physical properties of the system.

[0083] Understandably, in its implementation, the unified transformation operator employs a piecewise high-order polynomial interpolation method to construct smooth connection functions between discrete time nodes, while simultaneously satisfying the aforementioned principles. For highly nonlinear regions, the operator adopts an adaptive subdivision strategy to increase the density of local interpolation points and improve transformation accuracy. Through this unified transformation operator, the abrupt change characteristics captured by the discrete model and the evolutionary trends described by the continuous model are organically integrated, forming a unified mathematical expression. This provides a more accurate and comprehensive theoretical foundation for the integrated analysis and optimization of network systems.

[0084] Specifically, the unified modeling framework is a comprehensive mathematical system that integrates discrete and continuous dynamic characteristics. By applying a unified transformation operator, this embodiment unifies the dynamic model of finite difference networks and the continuous model of semi-Lagrange networks into the same mathematical expression system, forming a complete unified modeling framework.

[0085] Understandably, this framework possesses several key characteristics: integrated expressive capability, enabling it to simultaneously describe discrete events (such as fault occurrence and state transition) and continuous processes (such as traffic fluctuations and performance evolution) in the network, avoiding the problem of separation between discrete and continuous models in traditional modeling; multi-scale analysis capability, enabling it to analyze network behavior at different time scales, from microsecond-level data packet transmission to day-level service fluctuations, achieving consistent analysis across scales; and causal relationship tracing capability, through unified mathematical expression, enabling it to accurately trace the causal links of state changes in the network and identify the root causes of complex problems.

[0086] In practical applications, the unified modeling framework provides several functions for network planning and optimization: state prediction function, which predicts the future state evolution of the network under given inputs and configurations through forward model calculation; backward analysis function, which determines the input conditions and configuration parameters required to achieve the desired network state through backward model derivation; and sensitivity analysis function, which evaluates the degree of influence and sensitivity of different factors on network performance through parameter perturbation experiments.

[0087] In summary, the implementation of this unified modeling framework enables the dynamic characteristics of network systems to be fully and accurately mathematically represented, providing a solid theoretical foundation for subsequent optimization decisions.

[0088] In this embodiment, in step S3, the process involves defining a state space and action space based on the network representation generated by the unified modeling framework, constructing a reward function, designing a reinforcement learning model and constructing a collaborative learning framework based on the state space, action space, and reward function, and generating an optimal solution set using multi-objective optimization techniques based on the learning results of the collaborative learning framework, thereby obtaining a candidate network architecture scheme set, including: Step S31: Based on the network representation generated by the unified modeling framework, define a network state space containing node configuration, link parameters and resource allocation strategies as the state space, and design an action space including node deployment, link planning and resource scheduling operations.

[0089] Specifically, the precise definition of state space and action space is fundamental to the construction of reinforcement learning models. Based on the network representations generated by the unified modeling framework, state space and action space with clear physical meaning are constructed.

[0090] The network state space is a comprehensive description of the current state of the network system, containing the following core information: Node configuration information describes the characteristics and state of network nodes, including node type (such as core router, edge switch, access point, etc.), performance parameters (such as processing capacity, cache size, forwarding rate, etc.), functional characteristics (such as supported protocols, security features, service capabilities, etc.), location information (such as geographical coordinates, topology location, etc.), and operating status (such as utilization, load level, health status, etc.).

[0091] Link parameter information expresses the characteristics of network connection, including link type (such as fiber optic, copper cable, wireless, etc.), performance indicators (such as bandwidth capacity, transmission delay, packet loss rate, etc.), utilization (such as current traffic, remaining capacity, etc.), and reliability characteristics (such as redundancy, backup status, etc.).

[0092] Resource allocation strategy information describes the allocation status of network resources, including computing resource allocation (such as CPU allocation, memory allocation, etc.), bandwidth resource allocation (such as QoS policies, traffic engineering configurations, etc.), and storage resource allocation (such as caching policies, storage allocation, etc.). This information together constitutes a high-dimensional vector space, comprehensively representing the current state of the network system.

[0093] The action space defines the set of operations that the planning system can execute, including the following core operations: node deployment operations control the configuration and deployment of network nodes, including node addition (deploying a new node at a specified location), node removal (removing an existing node), node upgrade (improving node performance or functionality), node downgrade (reducing node specifications to save costs), and node function adjustment (modifying the functional characteristics of a node).

[0094] Link planning operations manage the establishment and configuration of network connections, including link establishment (establishing a new connection between specified nodes), link removal (revoking an existing connection), bandwidth adjustment (increasing or decreasing link capacity), link type change (modifying the physical or logical type of a link), and routing policy adjustment (modifying traffic forwarding rules).

[0095] Resource scheduling operations control the allocation and use of network resources, including computing resource allocation (adjusting node computing power allocation), bandwidth resource allocation (adjusting link bandwidth allocation strategies), storage resource allocation (adjusting cache and storage allocation), and quality of service (QoS) adjustments (modifying service priorities and resource guarantees). These operations involve both discrete decision-making and continuous parameter adjustments, forming a hybrid action space that provides comprehensive operational capabilities for network planning.

[0096] Step S32: Based on the state space and the action space, construct a composite reward function with a weight adaptive mechanism as the reward function to achieve a dynamic balance between bandwidth cost, latency performance, energy efficiency and reliability, and form a multi-dimensional reward evaluation mechanism for network planning.

[0097] In this step, the composite reward function is a key mechanism guiding the reinforcement learning algorithm to achieve a balance among multiple objectives. The composite reward function constructed in this embodiment includes multiple core dimensions to comprehensively evaluate the performance and cost of the network architecture.

[0098] Specifically, the bandwidth cost incentive measures the economic benefits of a network and mainly considers two aspects: the degree of matching between bandwidth configuration and demand, which rewards schemes with reasonable configuration (neither excessive waste nor serious shortage) by calculating the gap between configured bandwidth and actual demand; and the link construction cost, which rewards schemes with lower costs while meeting performance requirements by calculating construction investment based on link type, length and capacity.

[0099] Latency performance incentives evaluate network response speed, primarily including: end-to-end transmission latency, which measures the transmission time of data packets in the network, and rewards architectures that provide low-latency transmission; and service response time, which evaluates the complete processing time of user requests, including transmission latency and processing latency, and rewards architectures that provide fast response.

[0100] Energy efficiency rewards focus on the network's energy consumption, including: total energy consumption level, which calculates the energy consumption of the entire network system and rewards solutions with lower energy consumption; and energy efficiency ratio, which measures the energy required per unit of data processing and rewards solutions with lower energy consumption or higher throughput for the same throughput.

[0101] Reliability rewards assess the stability and resilience of a network, including: service availability, which calculates the probability that the system can maintain service under various failure scenarios and rewards solutions with high availability; and fault recovery capability, which assesses the speed at which the network can restore service after a failure and rewards architectural designs with rapid recovery capabilities.

[0102] It's important to note that the adaptive weight mechanism is the core innovation of the composite reward function. It enables the reward function to dynamically adjust the importance of each dimension based on the network state and optimization stage. This mechanism mainly comprises several key components: a state-aware module, which monitors the network's current state and performance metrics, identifying the most prominent problems and needs; a stage-identification module, which determines the stage of the optimization process (e.g., initial exploration, direction determination, fine-tuning), employing different weight strategies for different stages; and a weight adjuster, which dynamically calculates the weight coefficients of each dimension's reward based on the results of state awareness and stage identification.

[0103] For example, in the initial exploration phase, the weights of multiple dimensions may be relatively balanced; when latency performance is identified as a bottleneck, the weight of latency performance rewards will automatically increase; in the fine-tuning phase, more emphasis may be placed on balancing energy efficiency and reliability. This dynamic adjustment mechanism allows the reward function to more accurately guide the reinforcement learning algorithm, finding the optimal balance among multiple objectives and generating a network architecture solution that meets practical needs.

[0104] Step S33: Based on the state space, the action space, and the multi-dimensional reward evaluation mechanism, design a deep reinforcement learning model that integrates the Transformer structure as the reinforcement learning model, establish the mapping relationship between the network state and the optimal action, and construct a decision network with long-term memory and attention mechanisms.

[0105] It should be noted that the deep reinforcement learning model integrating the Transformer architecture is the core intelligent decision engine in this embodiment. Traditional reinforcement learning models often struggle to capture long-term dependencies and complex interactions between key elements when dealing with complex decision problems such as network planning. To address this challenge, this embodiment integrates the Transformer architecture with deep reinforcement learning, designing a decision network with long-term memory and attention mechanisms.

[0106] This model is built on the Deep Q-Network (DQN) framework, but introduces key mechanisms of the Transformer structure in the state encoding and action generation stages. In state encoding, the multi-head self-attention mechanism enables the model to simultaneously focus on different aspects of the network state, calculate the correlation between different network elements, and automatically identify the state features most critical to the current decision.

[0107] For example, when deciding on the deployment location of a node, the model can automatically consider key information such as traffic distribution in the relevant area, existing node load, and link status. In terms of action generation, the self-attention mechanism can evaluate the mutual influence between different possible actions, generating coordinated action strategies and avoiding global suboptimal problems caused by local optimization. For instance, when planning multiple network regions simultaneously, the model can consider the interaction effects between regions, ensuring that the planning of each region is coordinated.

[0108] Furthermore, the Transformer's positional encoding mechanism is adapted to network topology encoding, enabling the model to understand the positional relationships of network elements within the topology, thus providing a basis for spatially-aware decision-making.

[0109] Specifically, long-term memory and attention mechanisms are two core advantages of reinforcement learning models that incorporate the Transformer structure. The long-term memory mechanism is achieved through the Transformer's self-attention structure, enabling the model to capture long-term dependencies in the network state sequence.

[0110] Traditional reinforcement learning models are often plagued by short-sightedness, struggling to effectively correlate current decisions with long-term rewards, leading to local optima in policy optimization. In contrast, the Transformer architecture can directly compute the correlation between any two time steps, unrestricted by sequence distance. This allows the model to learn the impact of current decisions on long-term network performance, enabling more forward-looking planning strategies.

[0111] For example, when planning network expansion, the model can foresee the long-term impact of current deployment decisions on network performance over the next few years, avoiding long-term bottlenecks caused by short-term optimizations. Attention mechanisms enable the model to intelligently allocate computing resources, focusing on the state information most critical to the decision. In complex network systems, the state space has extremely high dimensionality, but not all state variables are equally important for a particular decision. Through attention weight calculation, the model can automatically identify key factors and ignore secondary information, improving decision accuracy and computational efficiency. For example, when planning link capacity in a certain area, the model will focus on the traffic patterns and the status of surrounding nodes in that area, rather than the network conditions far from that area.

[0112] It can be understood that these mechanisms work together to enable the model to exhibit superior performance in complex decision-making tasks such as network planning, and to generate network architecture schemes with long-term optimality.

[0113] Step S34: Based on the decision network, construct a multi-agent collaborative learning framework as the collaborative learning framework to realize collaborative optimization between different network domains and different levels of planning tasks, and establish a multi-agent system based on shared experience and distributed decision-making.

[0114] In this step, a multi-agent collaborative learning framework is an effective strategy for addressing the complexity of large-scale network planning. Considering the large scale and complex structure of modern network systems, this embodiment decomposes the planning task into multiple sub-tasks, which are completed collaboratively by different agents.

[0115] The framework comprises three types of specialized agents: domain agents are responsible for planning decisions in specific network areas, such as access network planning, aggregation network planning, and core network planning. These agents possess expertise and specific constraints for their respective areas. Layer agents focus on planning tasks at specific network layers, such as physical layer planning, protocol layer planning, and service layer planning. These agents have specialized decision-making capabilities in their respective domains. Functional agents are responsible for decisions regarding specific functions, such as bandwidth optimization, latency optimization, energy consumption optimization, and reliability assurance. These agents perform in-depth optimization for specific objectives.

[0116] Specifically, the aforementioned intelligent agents achieve overall optimization through three collaborative mechanisms: an experience-sharing mechanism, where each agent stores its own learning experience in a shared experience pool, which other agents can access and learn from, accelerating the learning process; a distributed decision-making mechanism, where each agent makes independent decisions based on local observations, and then integrates them through negotiation to form a global solution, avoiding the computational bottleneck of centralized decision-making; and an ensemble learning mechanism, which integrates the decision results of multiple agents to obtain a more robust and comprehensive planning scheme, reducing the bias and uncertainty of single agent decisions.

[0117] In summary, this multi-agent collaborative learning framework can effectively handle the planning complexity of large-scale networks, and achieve efficient solutions to complex planning tasks through specialized division of labor and collaborative cooperation.

[0118] Step S35: Based on the collaborative learning results of the multi-agent system, a Pareto optimal solution set that satisfies the bandwidth cost, latency performance, energy efficiency, and reliability is generated as the optimal solution set through multi-objective optimization techniques such as non-dominated sorting and congestion distance calculation, thereby obtaining the candidate network architecture scheme set.

[0119] In this step, the generation of the candidate network architecture scheme set adopts multi-objective optimization techniques, aiming to find the optimal solution set that satisfies multiple competing objectives.

[0120] Understandably, in network planning, objectives such as bandwidth cost, latency performance, energy efficiency, and reliability are often mutually constrained and cannot reach their respective extreme values ​​simultaneously. It is necessary to find multiple balance points to form a Pareto optimal solution set.

[0121] This embodiment employs the Non-Dominated Sorting Genetic Algorithm II (NSGA-II) to achieve multi-objective optimization. This algorithm comprises two core mechanisms: a non-dominated sorting mechanism, which layers all possible network architecture schemes according to their dominance relationships (defined as scheme A being superior to scheme B on at least one objective and not inferior to scheme B on other objectives), and a crowding distance calculation. Within the same non-dominated layer, the algorithm calculates the distance between each solution and its surrounding solutions in the objective space, prioritizing solutions with larger distances to ensure a uniform distribution of the final solution set in the objective space and providing diverse choices.

[0122] Specifically, the algorithm continuously explores the solution space through genetic operations such as selection, crossover, and mutation, and the iterative optimization process continues until convergence or the maximum number of iterations is reached. The final Pareto optimal solution set contains a series of network architecture schemes that achieve different balances on multiple objectives such as bandwidth cost, latency performance, energy efficiency, and reliability, forming a candidate network architecture scheme set, providing a rich selection space for subsequent adaptation to specific scenarios.

[0123] In this embodiment, step S4, which involves generating a planning constraint model based on the candidate network architecture scheme set, and determining the network node deployment location, link capacity configuration, resource scheduling strategy, and fault self-healing path based on the planning constraint model and the candidate network architecture scheme set, includes: Step S411: Based on the candidate network architecture scheme set and business scenario requirements, analyze business characteristics, security level, available resources and deployment limitations, and generate a scenario-based planning constraint model.

[0124] In this step, the scenario-based planning constraint model is a key tool for achieving precise matching between network architecture and specific business scenarios. Based on a set of candidate network architecture solutions and business scenario requirements, this model constructs a constraint system from multiple dimensions: business characteristic constraints, including business traffic models, service response time requirements, and peak processing capacity requirements, to ensure the network can meet business service quality requirements; security level constraints, including data protection levels, access control policies, and audit trail requirements, to ensure network security; available resource constraints, including budget limits, equipment resource limits, and implementation time windows, to ensure the feasibility of the planned solutions; and deployment constraint constraints, including geographical location limitations, equipment compatibility requirements, and regulatory compliance requirements, to address various limitations of the actual deployment environment.

[0125] The constraints are mathematically expressed in the form of mixed-integer linear programming (MILP), using binary variables to represent discrete choices and continuous variables to represent parameter adjustments, thus forming a formalized constrained optimization problem. Scenario-based processing is a core feature of this model. By setting specific sets of constraint parameters for different scenarios (such as finance, healthcare, education, or manufacturing), the constrained model achieves scenario adaptability, ensuring that the final planning solution accurately meets the needs of specific business scenarios.

[0126] Specifically, business characteristic analysis is the first step in constructing a scenario-based planning constraint model. This embodiment adopts a hierarchical business modeling approach to comprehensively capture business requirement characteristics. In application layer analysis, the focus is on the functional type of the business (such as transaction processing, content distribution, real-time communication, etc.), user experience requirements (such as response time, interaction smoothness, etc.), and business criticality (such as core business, supporting business, general business, etc.). In traffic layer analysis, the focus is on data traffic characteristics (such as packet size distribution, burstiness, periodicity, etc.), bandwidth requirements (such as average bandwidth, peak bandwidth, etc.), and transmission quality requirements (such as packet loss tolerance, latency sensitivity, etc.). In security layer analysis, the focus is on data sensitivity (such as public data, internal data, sensitive data, etc.), protection requirements (such as encryption strength, access control level, etc.), and compliance requirements (such as data localization, audit trail, etc.).

[0127] These analytical results are organized in a structured manner to form a business requirement characteristic map, providing input for the construction of constraint models. The business characteristic analysis adopts a hybrid approach, combining expert knowledge and data-driven analysis. It utilizes both the experience and judgment of domain experts and the statistical analysis of historical data to discover potential patterns, ensuring the comprehensiveness and accuracy of the analysis results.

[0128] Security level assessment is a crucial dimension in constraint model construction. This embodiment designs a risk-based security grading method, quantifying network security requirements into measurable constraints. This method first identifies potential security risks through threat modeling, comprehensively assessing the risk level based on asset value, threat probability, and vulnerability severity. Then, based on the risk assessment results, it determines corresponding security control measures and protection levels, including authentication strength (e.g., single-factor, two-factor, or multi-factor authentication), access control granularity (e.g., network-level, application-level, or data-level control), data protection methods (e.g., transmission encryption, storage encryption, or end-to-end encryption), network isolation requirements (e.g., physical isolation, logical isolation, or VLAN separation), and monitoring and auditing depth (e.g., basic logs, detailed auditing, or behavioral analysis).

[0129] These security requirements are translated into constraints that affect multiple aspects of network architecture design, such as node deployment (security zone division), link configuration (encrypted channel requirements), and resource allocation (security function resource reservation). Security constraints are set using a tiered principle, with different levels of constraints applied to different security applications to ensure that security investment matches the risk level and avoid over-protection or under-protection.

[0130] Available resource assessment and deployment constraint analysis together constitute the implementation boundary of the constraint model. Available resource assessment is mainly conducted from several aspects: budgetary resource assessment, which determines the upper limit of available financial investment, including the initial investment budget and operating cost budget, and translates it into constraints on the number of equipment, performance level, and function selection; physical resource assessment, which inventoryes available space resources (such as data center area, rack capacity, etc.), power supply resources (such as power capacity, backup power, etc.), and heat dissipation resources (such as cooling capacity, heat dissipation space, etc.), forming the constraint boundary of physical deployment; and time resource assessment, which clarifies the project's time window requirements, including the implementation period, phased goals, and business launch time, and translates it into constraints on solution complexity and implementation schedule.

[0131] Deployment constraint analysis focuses on environmental and strategic constraints: geographical constraints analysis examines the geographical characteristics of the deployment environment (such as terrain conditions and climate characteristics) and regional policies (such as land use restrictions and environmental protection requirements); technical compatibility analysis assesses compatibility requirements with existing systems (such as equipment interoperability and protocol compatibility) and technical route limitations (such as vendor strategies and technology standard selection); and regulatory compliance analysis identifies applicable industry regulations (such as financial regulation and medical data protection) and regional laws (such as privacy laws and data localization requirements). These assessment results collectively determine the feasible areas for the planning scheme, ensuring the feasibility of the final solution.

[0132] Step S412: Based on the planning constraint model, the candidate network architecture scheme set is screened, and combined with geographical information, resource distribution and connectivity requirements, the optimal deployment location of network nodes is determined by the location optimization algorithm to generate a network node deployment scheme.

[0133] In this step, the location optimization algorithm is the core technology for determining the optimal deployment location of network nodes. After preliminary screening of the candidate network architecture scheme set based on the planning constraint model, this embodiment applies the location optimization algorithm to determine the node deployment location by combining multiple key information: geographic information, including the geographic coordinates of available deployment points, terrain features, traffic conditions, and natural environmental factors, used to assess the suitability of the deployment environment; resource distribution information, including power resource distribution, cooling resource availability, maintenance support radius, and cost map, used to assess operating cost factors; and connectivity requirements, including user distribution, access point distribution, backbone network connection requirements, and interconnection strategies, used to ensure the rationality of the network topology.

[0134] Understandably, the location optimization algorithm combines a modified weighted centroid method with simulated annealing. The weighted centroid method determines the initial location of nodes based on user distribution and business importance, with the calculation formula considering the spatial distribution of user locations and the business weight of each location. Simulated annealing, on the other hand, continuously attempts small-scale location adjustments to evaluate changes in total cost (including construction cost, operating cost, and service quality cost), achieving a balance between global search and local optimization to avoid getting trapped in local optima. During algorithm iteration, the annealing temperature parameter is dynamically adjusted according to the optimization stage; a high initial temperature allows for large-scale exploration, while a low initial temperature focuses on fine-tuning, forming an efficient search strategy. The final network node deployment scheme includes the precise location coordinates, functional positioning, capacity specifications, and connectivity relationships of each node, providing a foundation for subsequent link planning.

[0135] Step S413: Based on the network node deployment scheme, utilize the network traffic distribution pattern and service fluctuation characteristics in the network digital twin, construct a traffic prediction model through a time series prediction algorithm, analyze the service growth trend and fluctuation characteristics, determine the capacity configuration and expansion strategy of each link, and generate a link capacity planning scheme that includes redundancy design.

[0136] In this step, the traffic prediction model serves as the basis for link capacity planning decisions. Based on the network node deployment scheme, this embodiment utilizes historical data accumulated in the network digital twin to construct an accurate traffic prediction model.

[0137] The model employs a hierarchical time series analysis framework to capture traffic change characteristics at different time scales: long-term trend analysis uses a Long Short-Term Memory (LSTM) network to capture the overall trend of business growth and predict traffic growth curves for the next 1-3 years, providing long-term guidance for capacity planning; seasonality analysis uses a seasonal autoregressive integral moving average (SARIMA) model to identify periodic traffic change patterns, including intraday, weekly, and seasonal variations, providing a reference for capacity demand at different times; and burst prediction analysis combines extreme value theory and convolutional neural networks (CNN) to predict burst traffic behavior and abnormal peaks, providing a basis for capacity redundancy design.

[0138] These three layers of prediction results are integrated using ensemble learning methods (such as Stacking) to form the final traffic prediction model. This model can not only predict the average traffic level but also estimate the variance and peak value of the traffic distribution, providing comprehensive information for risk-aware capacity planning. Prediction accuracy is evaluated through cross-validation with historical data, and an appropriate safety margin is set based on the prediction error to ensure the reliability of capacity planning.

[0139] Specifically, the link capacity planning scheme is a systematic decision-making process based on traffic prediction results. This embodiment adopts a hierarchical link capacity configuration strategy, designing differentiated capacity planning methods for different network layers: the core layer link capacity is designed at 1.5-2 times the predicted peak value to ensure that critical network backbones have sufficient redundancy to cope with sudden traffic surges and future growth; the aggregation layer link capacity is configured according to the regional traffic aggregation characteristics, typically at 1.3-1.5 times the predicted peak value, while considering the balancing problem of uneven traffic distribution; the access layer link capacity is determined based on the number of end users and service types, prioritizing the quality of user experience. The capacity configuration adopts a tiered growth strategy, pre-planning multiple expansion trigger points and corresponding expansion schemes to achieve smooth evolution of network capacity.

[0140] Redundancy design is a crucial component of link capacity planning, primarily encompassing two redundancy mechanisms: bandwidth redundancy, which reserves extra capacity to address traffic fluctuations and prediction errors, with the redundancy ratio adjusted based on business criticality and traffic predictability; and path redundancy, which ensures service continuity in the event of a primary link failure through the configuration of backup links and multi-path transmission. The output of the link capacity planning scheme includes the initial capacity configuration for each link, expansion trigger conditions, maximum expansion limit, and redundancy safeguards, forming a complete capacity evolution blueprint. The planning process considers both technical feasibility and economic rationality, optimizing investment returns while meeting performance requirements to ensure the implementation value of the scheme.

[0141] In this embodiment, step S4, which involves determining the network node deployment location, link capacity configuration, resource scheduling strategy, and fault self-healing path based on the planning constraint model and the candidate network architecture scheme set, further includes: Based on the network node deployment scheme and the link capacity planning scheme, a resource allocation algorithm and a dynamic scheduling mechanism are designed to determine the allocation strategies for computing resources, storage resources and bandwidth resources, and to generate a resource scheduling strategy scheme.

[0142] In this step, the resource scheduling strategy is a key mechanism for achieving efficient utilization of network resources. Based on the network node deployment scheme and link capacity planning scheme, this embodiment designs an adaptive resource allocation algorithm and a dynamic scheduling mechanism, covering three types of core network resources.

[0143] The computing resource allocation adopts a multi-tiered scheduling strategy based on Quality of Service (QoS) levels. Services are divided into multiple priority levels, with high-priority services receiving resource reservation guarantees and low-priority services using an elastic allocation mechanism. The scheduling algorithm combines static resource allocation and dynamic adjustment. The initial configuration sets a baseline resource allocation ratio based on the expected service load. During operation, resource allocation is dynamically adjusted by monitoring system load changes in real time to ensure maximum resource utilization while meeting QoS commitments.

[0144] Storage resource allocation employs a data access frequency-aware tiered storage strategy, establishing a tiered storage system comprising a cache layer, a standard storage layer, and an archive storage layer. Through a data lifecycle management mechanism, data is automatically migrated between different storage layers based on access frequency and importance, optimizing storage resource utilization efficiency. Simultaneously, Content Delivery Network (CDN) technology is used to implement distributed caching of frequently accessed data, reducing access latency and alleviating pressure on the central storage.

[0145] Bandwidth resource allocation adopts a flow-based fine-grained control mechanism. The software-defined networking (SDN) controller monitors network traffic in real time and dynamically adjusts flow tables and queue weights based on service priority, traffic characteristics, and network status to achieve precise allocation and on-demand adjustment of bandwidth resources.

[0146] Specifically, the dynamic scheduling mechanism is the core of resource management, ensuring that resource allocation can adaptively adjust to changes in the environment. The dynamic scheduling mechanism designed in this embodiment includes several key components: the resource monitoring module continuously collects the utilization rate and performance indicators of various resources, and monitors the resource status in real time through distributed monitoring probes and a centralized data analysis platform, providing a basis for scheduling decisions; the demand forecasting module analyzes business load patterns based on historical data, and uses time series forecasting algorithms to predict short-term (minute-level) to medium-term (hour-level) changes in resource demand, enabling the system to prepare for resource adjustments in advance and avoid passive responses; the scheduling decision module calculates the optimal resource allocation scheme through a multi-objective optimization algorithm based on monitoring data and forecast results, and executes corresponding configuration changes through the network controller.

[0147] The scheduling decision-making adopts a hierarchical control strategy. Emergency adjustments (such as fault response) are handled by the rapid response controller, while routine optimization is handled by the periodic planning controller. These two controllers work together to ensure both response speed and maintain global optimality. The adaptability of the resource scheduling strategy is reflected in several aspects: load adaptability, automatically adjusting resource allocation according to changes in business load; scenario adaptability, enabling preset resource strategies for different scenarios (such as daily operation, promotional activities, emergency states, etc.); and evolution adaptability, continuously learning and optimizing scheduling algorithm parameters to adapt to long-term changes in the network environment. These mechanisms together constitute a complete resource scheduling strategy scheme, ensuring efficient utilization of network resources and continuous guarantee of service quality.

[0148] Based on the network node deployment scheme and the link capacity planning scheme, the dynamic characteristics of the network are characterized by the local matrix exponential function and the unified modeling framework, potential fault points and fault propagation paths are analyzed, and fault risk assessment results are generated.

[0149] In this step, fault risk assessment is a prerequisite for building a highly reliable network, used to systematically identify and quantify potential risk points in the network. This embodiment utilizes local matrix exponential functions and a unified modeling framework to accurately characterize the dynamic properties of the network, providing a mathematical foundation for comprehensive risk assessment.

[0150] Specifically, the assessment process mainly includes several stages: potential failure point identification involves analyzing the complex network topology to calculate the centrality indicators of each node (such as degree centrality, betweenness centrality, and proximity centrality) and identify the critical nodes that have the greatest impact on network stability. Simultaneously, combined with equipment reliability data (such as mean time between failures (MTBF), annual failure rate, etc.), the failure probability of each node is assessed, and a failure risk heatmap is generated.

[0151] Fault propagation analysis utilizes the state transition matrix in a unified modeling framework to simulate the propagation process of a fault from its initial point to other parts of the network. The analysis employs Monte Carlo simulation, employing extensive testing under numerous random fault scenarios to calculate the fault's impact range and propagation speed, identify fault cascading effects and amplification mechanisms, and construct a fault propagation model.

[0152] The risk quantification assessment comprehensively considers multiple dimensions, including failure probability, scope of impact, and business importance, to calculate a comprehensive risk index for each potential failure scenario. The assessment employs a risk matrix method, with the horizontal axis representing failure probability and the vertical axis representing the severity of impact. The position in the matrix corresponds to the risk level, facilitating intuitive understanding and prioritization.

[0153] Understandably, the assessment results of this step are output in the form of a risk report, which includes high-risk area identification, key vulnerability analysis and risk mitigation recommendations, providing a scientific basis for the design of subsequent fault self-healing strategies.

[0154] Based on the fault risk assessment results, a multi-level backup strategy and a reconstruction strategy are designed. The multi-level backup strategy includes primary and backup link configuration and redundant node deployment. The reconstruction strategy includes traffic rerouting mechanism and resource reallocation mechanism to build the network's fault self-healing capability and obtain a fault self-healing path planning scheme.

[0155] In this step, the multi-level backup strategy is a fundamental component of fault self-healing path planning, providing a preventative protection mechanism. Based on the fault risk assessment results, this embodiment designs a hierarchical, multi-dimensional backup strategy system.

[0156] The primary and backup link configuration adopts a three-level protection mechanism based on the importance and reliability requirements of the business: Level 1 protection (1+1 full protection) is suitable for critical businesses, where data is transmitted simultaneously on the primary and backup paths, achieving seamless switching at the millisecond level, but the resource cost is relatively high; Level 2 protection (1:1 backup) is suitable for important businesses, where the backup path is reserved but does not transmit data under normal circumstances, and a switching process is required in case of failure, with recovery time at the second level, and the resource efficiency is relatively high; Level 3 protection (N:M shared backup) is suitable for ordinary businesses, where multiple primary paths share backup resources, which is economical and efficient but the recovery time is relatively long.

[0157] Redundant node deployment provides protection from a spatial perspective, primarily employing two deployment modes: geographically distributed deployment, which deploys backup nodes with identical functions in different geographical locations to mitigate the impact of regional disasters; and clustered deployment, which deploys multiple nodes with identical functions within the same region to form a cluster, distributing traffic through load balancing mechanisms, ensuring that the failure of any single node does not affect the overall service. The two modes are used in combination to form multi-layered node redundancy protection.

[0158] The key technology of the backup system is the state synchronization mechanism, which ensures data consistency between the primary and backup nodes. This embodiment adopts incremental synchronization and checkpoint technology to minimize synchronization overhead while ensuring consistency, so that the backup system can seamlessly take over when a failure occurs and ensure business continuity.

[0159] The reconstruction strategy is a proactive repair mechanism that automatically restores services after a network failure. The reconstruction strategy designed in this embodiment includes two core components that work together to achieve the network's self-healing capability. The traffic rerouting mechanism is a network-level fault response mechanism. Based on software-defined networking technology, it uses a central controller to calculate new optimal paths in real time and quickly issues flow table updates, redirecting affected service traffic to alternative paths.

[0160] Specifically, the mechanism includes several key steps: fault detection, which quickly identifies link or node faults through a combination of active detection and passive monitoring; path calculation, which uses an improved Dijkstra algorithm or constrained shortest path algorithm to calculate the optimal alternative path, taking into account multi-dimensional constraints such as bandwidth, latency, and reliability; and flow table update, which uses a batch-based, priority-ordered approach to efficiently issue flow table change commands and minimize traffic interruption time.

[0161] The resource reallocation mechanism is a system-level fault response mechanism. After a fault occurs, it reassesses available resources and service priorities, and reallocates computing, storage, and bandwidth resources according to preset strategies. This mechanism includes priority management, which classifies affected services based on their importance to ensure that critical services receive priority recovery; resource negotiation, which determines resource allocation schemes through preset rules or real-time negotiation when resources are insufficient; and service degradation, which provides basic functionality to non-critical services according to acceptable service degradation strategies when resources are severely insufficient, ensuring overall system availability. These mechanisms together constitute the network's fault self-healing capability framework, enabling the network to autonomously recover in the event of a fault and minimize the impact of service interruptions.

[0162] The fault self-healing path planning scheme is a systematic integration of the aforementioned backup and reconstruction strategies, forming a holistic solution for network resilience. This embodiment adopts a scenario-based design approach, pre-planning the optimal recovery path for different fault scenarios and constructing a complete self-healing path map.

[0163] The planning process first identifies high-risk failure scenarios based on the failure risk assessment results, including typical scenarios such as single-node failure, single-link failure, regional failure, and cascading failure. Then, a dedicated recovery plan is designed for each scenario, clarifying the triggering conditions, execution steps, and expected results, forming a structured decision tree. Finally, the plan is verified through a network digital twin to test the effectiveness and performance characteristics of the self-healing path and optimize and adjust the plan parameters.

[0164] The core of self-healing path planning is the path optimization algorithm, which seeks a balance among multiple objectives: recovery speed (aiming for the shortest service interruption time); resource efficiency (optimizing resource utilization during the recovery process); performance assurance (ensuring that the service quality after recovery is comparable to that before the failure); and cascading effects (avoiding new network problems caused by the recovery process). The planning results include detailed recovery path definitions, decision logic, and execution scripts, which can be directly called by automated systems to achieve automated recovery when a failure occurs.

[0165] Understandably, a complete fault self-healing path planning scheme significantly improves the reliability and resilience of the network, enabling the network to maintain service continuity in the face of various faults and attacks, and providing stable and reliable infrastructure support for business applications.

[0166] Deployment evolution strategy is a crucial element in ensuring the smooth implementation of network architecture planning and is integral to the entire planning and implementation process. Based on network node deployment schemes, link capacity planning schemes, resource scheduling strategies, and fault self-healing path planning schemes, this embodiment designs a phased deployment evolution strategy, comprising several key components: The phased deployment plan decomposes the overall plan into multiple independently implementable phases, each achieving specific functional objectives and ensuring the continuity of network services. Typical phased strategies include core-first (building the backbone network first, then expanding the access network), regional advancement (implementing gradually by region), or functional progression (prioritizing basic functions, adding advanced functions later), etc.; Critical path analysis identifies the critical task chains and dependencies during the implementation process, and optimizes the implementation sequence through project management techniques (such as Critical Path Method (CPM) and Project Review Technique (PERT)) to minimize implementation risks and timelines.

[0167] The switching window planning design the optimal execution time for each switching operation that requires service interruption, taking into account off-peak business periods, user impact, and resource availability, and preparing detailed rollback plans to ensure rapid recovery in case of problems; the capacity evolution roadmap is based on business growth forecasts, planning capacity expansion nodes and triggering conditions for the next 3-5 years, forming a long-term evolution strategy that enables the network to expand smoothly with business development.

[0168] Understandably, the core value of deployment evolution strategies lies in transforming static planning into dynamic implementation paths, ensuring that network architecture plans can be smoothly implemented in real-world environments and possess future-oriented evolution capabilities. Through scientific deployment evolution strategies, not only can implementation risks be reduced, but investment pace can also be optimized, resource utilization efficiency improved, and the network architecture consistently maintained at its best.

[0169] In this embodiment, step S4, which involves collecting network operation data in real time according to the deployed optimal network architecture planning scheme, identifying network performance deviations based on the network operation data and updating and optimizing the network architecture planning model, and performing dynamic adjustments to the network architecture based on the updated network architecture planning model, includes: Step S431: Based on the optimal network architecture planning scheme, collect network operation status data, resource usage and abnormal events in real time to obtain network operation data stream.

[0170] In this step, network operation data streams serve as the fundamental data source for network performance evaluation and optimization. This embodiment designs a distributed data acquisition system to collect three types of key operational data in real time, based on the optimal network architecture deployment plan.

[0171] Network operation status data includes network performance indicators such as device health status (e.g., CPU temperature, fan status, power supply status), link quality parameters (e.g., bit error rate, signal quality, stability), transmission latency (e.g., end-to-end latency, processing latency, queuing latency), packet loss rate (e.g., interface packet loss, buffer overflow, congestion packet loss), and throughput (e.g., interface traffic, application throughput, number of sessions). These data are obtained through network monitoring protocols (e.g., SNMP, RMON) and active probing technologies (e.g., ICMP probes, TCP probes). Resource usage data includes resource consumption indicators such as CPU utilization, memory usage, storage space utilization, queue depth, and bandwidth utilization. These data are obtained through device agent programs and resource monitoring interfaces.

[0172] Anomaly event data includes records of abnormal situations such as device alarms (e.g., hardware failures, performance alarms), performance threshold exceedances (e.g., excessive latency, high packet loss rate), security events (e.g., abnormal access, attack attempts), and connection interruptions (e.g., link failures, session interruptions). These are captured through log analysis and event correlation techniques. Data collection frequency is differentiated based on data type and importance. Key performance indicators are sampled at high frequency (every 5-30 seconds), resource usage is sampled at medium frequency (every 1-5 minutes), and anomaly events are collected in real-time.

[0173] The data processing pipeline is a crucial step in transforming raw collected data into a structured network operation data stream. This embodiment designs a multi-stage data processing flow to ensure data quality and availability.

[0174] The data cleaning stage addresses outliers, missing values, and duplicate records in the raw data. It employs a combination of statistical filtering and business rules to identify anomalous data points and supplements missing values ​​using interpolation algorithms or preset rules, ensuring data integrity and accuracy. The data standardization stage converts data from different sources and formats into a unified standard, including unit conversion, numerical normalization, and timestamp alignment, enabling consistent data processing and comparison. The data aggregation stage aggregates data according to time windows (e.g., minutes, hours, days) and spatial ranges (e.g., device-level, region-level, network-level), generating statistical indicators of different granularities to meet multi-level analysis needs. The feature extraction stage extracts key features from the processed data, including statistical features (e.g., mean, variance, quantiles), time-series features (e.g., periodicity, trend, abrupt changes), and correlation features (e.g., correlation between indicators, event correlations), providing high-dimensional feature representations for subsequent analysis.

[0175] Understandably, after this series of processes, the raw collected data is transformed into a structured and standardized network operation data stream, which has temporal continuity, spatial consistency and feature richness, providing a high-quality data foundation for network performance analysis.

[0176] Step S432: Based on the network operation data flow and expected performance indicators, identify network performance deviations and potential problems by constructing an anomaly detection model and trend analysis algorithm, and generate a performance deviation analysis report.

[0177] In this step, anomaly detection models and trend analysis algorithms are the core technologies for identifying network performance deviations. Based on network operation data flow and preset expected performance indicators, this embodiment constructs two types of analysis models.

[0178] Anomaly detection models focus on identifying anomalous states and performance deviations in networks, and mainly include several detection mechanisms: statistical anomaly detection, which uses statistical methods such as Z-scores and moving average control charts to identify situations where indicator values ​​exceed the normal range of variation; clustering-based anomaly detection, which uses clustering algorithms such as K-means and DBSCAN to identify outliers that are significantly separated from the main data group; isolation forest-based anomaly detection, which quickly identifies anomalous samples with short paths by constructing a set of random decision trees; and deep learning-based anomaly detection using autoencoders, which identifies complex anomaly patterns that are difficult to express using conventional models by reconstructing the error magnitude.

[0179] Trend analysis algorithms focus on identifying long-term trends in network performance indicators and predicting potential problems. They mainly include three types of techniques: time series decomposition, which decomposes the indicator series into trend, seasonal, and residual terms to extract long-term growth or decline trends; autoregressive moving average (ARIMA) model, which constructs a time series forecasting model with autocorrelation to predict future trends in indicators; and linear and nonlinear regression analysis, which establishes correlation models between indicators to analyze the driving factors and influencing mechanisms of key performance indicators.

[0180] In this step, the performance deviation analysis report is a comprehensive presentation of the results of anomaly detection and trend analysis, providing a basis for subsequent optimization decisions. The report includes several core parts: performance indicator deviation analysis, comparing the gap between actual performance and expected targets, calculating deviation rates and stability indicators, and identifying key indicators of substandard performance; anomaly event analysis, statistically analyzing the frequency, duration, and impact of various anomalies, and identifying high-frequency and severe anomalies; trend early warning analysis, based on trend prediction results, identifying indicators and time points where performance bottlenecks or capacity shortages may occur in the future; and root cause analysis, exploring the potential causes of performance deviations and anomalies through correlation analysis and causal inference, distinguishing between systemic and incidental problems.

[0181] Understandably, the report employs a multi-layered structure, providing both a high-level summary overview to help decision-makers quickly grasp the overall situation and detailed analytical data and charts to support technical personnel in delving into specific issues. During report generation, key findings are automatically identified and sorted by severity, highlighting issues requiring priority attention and providing initial improvement suggestions to offer clear direction for subsequent model updates and network adjustments.

[0182] Step S433: Based on the performance deviation analysis report, extract the deviation of network performance indicators and abnormal information of resource utilization from the actual operation feedback. Update the parameters and strategies of the reinforcement learning model in the network architecture planning model using incremental learning and transfer learning techniques. Adjust the constraints of the planning constraint model according to the deviation of network performance indicators to construct a continuous learning mechanism with self-evolution capabilities, and obtain the updated network architecture planning model. The network architecture planning model is constructed based on the reinforcement learning model, the planning constraint model, and the unified modeling framework.

[0183] In step S433, the continuous learning mechanism is a key technology for maintaining the advancement and adaptability of the network architecture planning model. Based on the performance deviation analysis report, this embodiment constructs a self-evolving continuous learning system, which mainly updates the network architecture planning model through two technical paths.

[0184] The first path involves updating the parameters and policies of the reinforcement learning model, primarily employing incremental learning and transfer learning techniques. Incremental learning allows the model to update its parameters using only new data while retaining existing knowledge, avoiding the high computational cost of training from scratch. This technique is mainly implemented through Elastic Weight Consolidation, which adds constraints to parameters in the model that are related to previous experience, preventing catastrophic forgetting. Transfer learning, on the other hand, utilizes experiential knowledge gained in similar network environments to quickly adapt to the specific needs of the current network. This technique is mainly implemented through feature mapping and domain adaptation methods, transferring knowledge from the source domain to the target domain, accelerating the learning process and improving the model's generalization ability. The combined use of these two techniques maintains the model's stability while enhancing its adaptability to new situations.

[0185] The second approach involves adjusting the constraints of the planning constraint model. Based on deviations in network performance indicators, the constraint parameters are automatically adjusted: for indicators with insufficient performance, the relevant constraints are strengthened to ensure that the updated plan meets performance requirements; for overly conservative indicators, the constraints are appropriately relaxed to improve resource utilization efficiency and reduce unnecessary costs.

[0186] The core of the continuous learning mechanism is a closed-loop feedback system. In this embodiment, model performance is continuously optimized through actual operation results. This system comprises several key components: an experience replay buffer, which stores historical decisions and their results, providing training samples for the learning algorithm, with particular attention to key decisions that lead to significant performance improvements or declines; a parameter update engine, which calculates the gradient of model parameters and performs updates based on new observations and feedback information, employing a priority sampling strategy to focus on high-value experiences; a policy evaluation module, which evaluates the performance difference between the new and old policies through A / B testing or control experiments to ensure that updates bring substantial improvements; and a knowledge extraction module, which extracts key knowledge from complex models, transforming it into interpretable rules and insights to support human-machine collaborative improvement.

[0187] Understandably, these components work together to form a self-improving learning system that continuously accumulates experience from network operation practices, optimizing the performance and adaptability of the planning model. Through this continuous learning mechanism, the network architecture planning model can automatically evolve with changes in the network environment and business needs, maintaining long-term effectiveness and providing intelligent support for continuous optimization of the network architecture.

[0188] Step S434: Based on the updated network architecture planning model, generate a dynamic network architecture adjustment scheme according to the current deployment status of the optimal network architecture planning scheme. The dynamic network architecture adjustment scheme includes node reconfiguration instructions, link adjustment instructions, and resource reallocation instructions, and executes dynamic adjustments to the optimal network architecture planning scheme.

[0189] The dynamic network architecture adjustment scheme is a set of specific optimization measures generated based on the updated planning model and the current network state. This embodiment designs multiple adjustment instructions to comprehensively cover the optimization needs of all layers of the network.

[0190] Among them, the node reconfiguration command is responsible for adjustments at the network node level, including operations such as expanding processing capacity (upgrading CPU, adding memory or expanding processing units), starting and stopping functional modules (enabling or disabling specific network functions such as QoS, security policies, etc.), adjusting node roles (such as promoting backup nodes to primary nodes, changing the role of nodes in the topology), and optimizing locations (adjusting node positions in physical or logical topologies), thereby enabling dynamic adjustments to network node resources and functions.

[0191] Link adjustment commands are responsible for adjustments at the network connection layer, including bandwidth capacity adjustment (increasing or decreasing link bandwidth), link priority change (adjusting the priority of links in load balancing or backup systems), transmission path replanning (optimizing traffic routing paths), and backup link activation (activating redundant links to share the load or improve reliability), thereby optimizing network connectivity and transmission efficiency.

[0192] Resource reallocation instructions are responsible for adjusting network resources, including computing resource reallocation (adjusting CPU time and memory allocation for each service or function), storage space adjustment (redistributing cache and persistent storage space), cache strategy optimization (adjusting cache algorithm parameters and strategies), and queue weight adjustment (optimizing the processing priority of different service flows), to achieve optimal allocation of network resources.

[0193] Specifically, the dynamic adjustments are implemented using a secure and controllable strategy to ensure that the adjustment process does not affect network stability. This embodiment designs a tiered implementation framework, employing different execution strategies based on the scope of the adjustment's impact and the level of risk.

[0194] Low-risk adjustments (such as optimization of non-critical node parameters and fine-tuning of resources) adopt a direct execution strategy, which is implemented directly through automated scripts to simplify the process and improve efficiency. Medium-risk adjustments (such as changes in link capacity and reconfiguration of secondary nodes) adopt a phased execution strategy, which first tests on a small scale to verify the effect before expanding the implementation scope to control the risk range. High-risk adjustments (such as core node reconstruction and changes to the backbone link) adopt a manual review strategy. After the adjustment plan is generated, it must be manually reviewed and approved before it can be implemented, and it must be implemented during the maintenance window to minimize the risk.

[0195] It's important to note that regardless of the strategy employed, multiple rollback checkpoints are implemented to continuously monitor network status during execution. Upon detecting any anomalies, the rollback mechanism is immediately triggered, restoring the network to a safe state. After the adjustment is implemented, network performance changes continue to be monitored, and the effectiveness of the adjustments is evaluated, forming a complete closed-loop optimization system. This ensures that the network architecture can continuously evolve and optimize, always maintaining its optimal state and providing stable and efficient infrastructure support for business applications. Through this dynamic adjustment mechanism, this embodiment enables the network architecture to adapt to constantly changing business needs and the technological environment, achieving long-term sustainable development.

[0196] In this embodiment, step S4, which involves performing simulation testing and iterative optimization in the network digital twin environment to output the optimal network architecture planning scheme, includes: Step S421: Based on the overall network architecture planning scheme, by constructing a simulation scenario and conducting simulation tests in the network digital twin environment, evaluate the performance indicators of the network architecture planning scheme, including throughput, latency, packet loss rate and energy consumption, under different load conditions, traffic patterns and fault scenarios, and generate a performance evaluation report.

[0197] In this step, network digital twin simulation testing is a crucial step in verifying the effectiveness of the network architecture planning scheme. Based on the overall network architecture planning scheme, this embodiment constructs a series of high-fidelity simulation scenarios in the digital twin environment. These scenarios cover various typical and extreme cases of network operation.

[0198] The load conditions scenarios include three basic conditions: normal load (average utilization 30-50%), high load (average utilization 70-90%), and overload (exceeding design capacity), as well as dynamic scenarios with rapid load changes (load increases 3-5 times in a short period). Traffic pattern scenarios cover typical distribution patterns such as uniformly distributed traffic, centrally concentrated traffic, edge-intensive traffic, and multi-center traffic, as well as time-varying characteristics such as daily fluctuation patterns, burst traffic patterns, and periodic change patterns. Fault scenarios include various fault types such as complete node failure, link interruption, partial performance degradation, and cascading failures to test the network's resilience and recovery capabilities.

[0199] In these scenarios, the simulation system employs discrete event simulation technology and traffic mathematical models to accurately simulate the transmission process of network data packets and the processing flow of service flows, comprehensively evaluating four key performance indicators: throughput measures the network's data transmission capacity, including overall throughput, node processing capacity, and link utilization; latency measures the time consumption of data transmission, including end-to-end latency, processing latency, and queuing latency; packet loss rate measures the reliability of data transmission, including congestion packet loss, error packet loss, and buffer overflow; and energy consumption measures the network's energy efficiency, including total energy consumption, energy consumption per unit of data, and peak power. These performance indicators are derived through simulation testing, generating a comprehensive performance evaluation report to provide a basis for subsequent optimization.

[0200] Simulation scenario construction is a fundamental step in simulation testing, determining the comprehensiveness and reliability of the test results. This embodiment employs a hierarchical scenario construction method, building the system layer by layer from the basic environment to specific test cases. At the basic environment layer, a virtual network topology is created based on the network architecture planning scheme, including node layout, link connections, and resource configuration, ensuring that the virtual environment is completely consistent with the planning scheme.

[0201] At the traffic model layer, multiple service traffic models are constructed, including synthetic traffic based on statistical distributions (such as Poisson and Pareto distributions) and replay traffic captured from actual networks, ensuring the authenticity of traffic characteristics. At the scenario configuration layer, specific network conditions and events are set according to the test objectives, such as background traffic intensity, interference factors, and fault triggers. Particularly in fault scenario construction, this embodiment employs a fault injection framework, which can inject various fault events into the system at precisely controlled points in time, including hardware faults (such as device downtime and link disconnection), software faults (such as protocol errors and configuration failures), and resource exhaustion (such as buffer overflows and full processing queues), testing the network's fault tolerance and recovery mechanisms. During scenario construction, the configuration parameters are kept precisely controllable to ensure the reproducibility of test results and support comparative evaluation of different planning schemes.

[0202] Performance evaluation is the core output of simulation testing. In this embodiment, a multi-level indicator system is used to comprehensively evaluate the network architecture's performance from different dimensions during the evaluation process.

[0203] Throughput metrics not only focus on the overall data processing capacity of the network (such as Gbps throughput and transactions per second), but also analyze spatial distribution characteristics (load distribution balance in different regions) and temporal variation characteristics (peak processing capacity and fluctuation coefficient).

[0204] Latency metrics assess network response speed from a user experience perspective, including end-to-end average latency, latency distribution (95th and 99th percentiles), latency jitter, and round-trip latency for interactive services. Packet loss rate metrics focus on network transmission reliability, including long-term average packet loss rate, packet loss rate during congestion periods, spatial distribution of packet loss, and packet retransmission rate. Energy consumption metrics assess network efficiency from a green energy-saving perspective, including total power consumption, power consumption per unit of data processing, power efficiency ratio (the ratio of energy consumption to throughput), and peak load power consumption.

[0205] It can be understood that these metrics are automatically collected, statistically analyzed, and processed by the simulation system to form a structured performance evaluation report, which intuitively shows the performance and potential bottlenecks of the network architecture in various scenarios, providing a clear direction for subsequent optimization.

[0206] Step S422: Based on the performance evaluation report and preset performance targets, the network architecture planning scheme is iteratively optimized by adjusting the deployment location of network nodes, link capacity configuration, and resource scheduling strategy until the performance indicators including throughput, latency, packet loss rate, and energy consumption all meet the preset performance target requirements, and the optimal network architecture planning scheme is output.

[0207] In step S422, the performance evaluation report is a comprehensive document integrating simulation test results, providing a scientific basis for iterative optimization. The report format designed in this embodiment includes several core parts, comprehensively presenting the network performance analysis results. The overall performance overview section provides a macro-level assessment, including key indicator achievement rates, performance radar charts, and comparative analysis with benchmark solutions, helping to quickly grasp the overall performance of the solution; the bottleneck analysis section identifies performance bottlenecks and potential risk areas through hotspot maps and correlation analysis, including high-load nodes, congested links, and resource contention points, pointing the way for optimization; the scenario adaptability analysis section evaluates the stability of the solution's performance under various scenarios, quantifying the network's robustness and resilience through performance fluctuation range and adaptability index, highlighting the solution's advantages and disadvantages in specific scenarios; the improvement suggestion section, based on empirical data, proposes specific optimization directions and parameter adjustment suggestions, including node relocation, link capacity reconfiguration, and resource scheduling strategy optimization.

[0208] The report generation process employs automated analysis technology, combined with an expert rule system, to not only present data but also provide intelligent interpretation, enabling technical personnel to quickly understand complex test results and provide clear guidance for the next round of optimization.

[0209] In this step, the iterative optimization mechanism is a key technology for achieving continuous improvement of the network architecture. Based on the performance evaluation report and preset performance targets, this embodiment constructs a systematic iterative optimization process. This process first analyzes the gap between performance indicators and targets, identifies key improvement points, and uses sensitivity analysis to calculate the impact of different adjustment measures on performance indicators, determining the most effective optimization direction. Then, corresponding optimization measures are taken for different performance bottlenecks.

[0210] To address insufficient throughput, network processing capacity is primarily improved by adjusting node processing capacity configurations, increasing critical path bandwidth, and optimizing traffic allocation strategies. For high latency, transmission delays are reduced by optimizing node deployment locations (reducing physical transmission distance), adjusting routing strategies (selecting the lowest latency path), and improving queue scheduling algorithms (prioritizing latency-sensitive services). For excessive packet loss, transmission reliability is improved by increasing buffer capacity, adjusting congestion control parameters, and optimizing traffic shaping strategies. For high energy consumption, energy consumption is reduced by optimizing resource utilization efficiency, adjusting energy-saving strategies (such as shutting down some resources under low load), and improving heat dissipation design. These optimization measures are prioritized based on their impact and implementation difficulty, with measures that have a significant impact and are easy to implement implemented first, forming an efficient optimization strategy.

[0211] Node deployment location optimization is a crucial step in iterative optimization, directly impacting network geographic coverage and service quality. This embodiment employs a multi-objective location optimization algorithm to minimize latency and resource consumption while meeting coverage requirements. The algorithm comprehensively considers several key factors: service distribution factors, including user density distribution, traffic hotspots, and service importance weights, ensuring node deployments are close to key service areas; physical constraints, including limitations on available deployment points, power and cooling resource availability, and physical security conditions, ensuring deployment feasibility; and connectivity requirements, including connection requirements with other nodes, backup path redundancy, and interconnection structure, ensuring a rational network topology. The algorithm uses a hybrid strategy: first, a weighted centroid method is used to determine initial locations, then simulated annealing is applied for fine-tuning. By continuously attempting small-scale location adjustments and evaluating changes in the total cost function, the global optimum is sought. The location optimization process is conducted in conjunction with other optimization measures, particularly link capacity planning, to ensure consistency between node locations and link configurations, collectively improving network performance.

[0212] In this step, link capacity configuration optimization is the core means to solve bandwidth bottlenecks and improve network throughput. Based on traffic distribution and service demand analysis, this embodiment designs an adaptive link capacity planning method to allocate the most suitable bandwidth resources to different links. This method first estimates the future traffic demand of each link using a traffic prediction model, including average traffic, peak traffic, and burst coefficient, creating a traffic demand map. Then, based on service importance and Service Level Agreement (SLA) requirements, it sets service quality objectives for different links, such as bandwidth guarantee rate and maximum utilization limit. Finally, it applies a capacity optimization algorithm to calculate the optimal bandwidth configuration for each link while meeting the service quality objectives. During the optimization process, special attention is paid to potential bottleneck links in the network. Traffic stress testing identifies links that may become congested under high load conditions, and the capacity configuration of these critical paths is strengthened.

[0213] Meanwhile, considering the economics and feasibility of capacity expansion, the algorithm evaluates the cost-effectiveness ratio of different bandwidth levels and selects the most cost-effective configuration scheme to ensure maximum performance improvement within a limited budget. In this way, link capacity configuration achieves demand-driven and cost-controlled precise planning, providing the network with sufficient and non-wasteful transmission resources.

[0214] In this step, optimizing resource scheduling strategies is a key technology for improving resource utilization efficiency and service quality. Addressing the issues of uneven resource allocation and competition identified in simulation tests, this embodiment designs an intelligent resource scheduling optimization mechanism to improve resource management from multiple dimensions.

[0215] In terms of computing resource scheduling, a multi-level queue scheduling algorithm is adopted. Based on business priority and resource demand characteristics, reasonable processing time and priority are allocated to different types of business flows to ensure that critical businesses are processed in a timely manner, while avoiding long-term starvation of low-priority businesses. In terms of storage resource scheduling, a data access frequency-aware caching strategy is implemented. By analyzing data access patterns, hot data is kept in high-speed cache, while cold data is moved to low-speed storage to maximize cache hit rate and reduce data access latency. In terms of bandwidth resource scheduling, software-defined networking (SDN) technology and advanced queue management (AQM) algorithm are applied to achieve fine-grained bandwidth control and congestion management at the flow level, ensuring fair allocation of bandwidth resources among different business flows and dynamically adjusting the allocation strategy according to business needs.

[0216] Understandably, these resource scheduling strategies dynamically sense changes in network status and business needs, adjust resource allocation in real time, improve resource utilization efficiency while ensuring service quality, and provide intelligent software-level support for network performance optimization.

[0217] The optimization-evaluation closed loop is the core mechanism of iterative optimization. This embodiment adopts a cyclical iterative approach, where a comprehensive simulation test is performed again in the digital twin environment after each round of optimization to evaluate the actual effect of the optimization measures. The evaluation results are compared with the preset performance targets. If there is a gap, the next round of optimization is initiated; if the target is met, the iteration process is completed.

[0218] This closed-loop process employs a heuristic search strategy, automatically adjusting the optimization direction and step size based on the results of previous optimization rounds to accelerate the convergence process. Simultaneously, a multi-objective balancing mechanism is used during optimization to ensure that optimizing one metric does not excessively sacrifice other metrics, maintaining overall performance balance. After multiple rounds of iterative optimization, the final network architecture plan meets the preset target requirements in key performance indicators such as throughput, latency, packet loss rate, and energy consumption, achieving comprehensive optimal network performance.

[0219] Understandably, this iterative optimization method based on digital twins avoids the trial-and-error costs of traditional planning, significantly improves planning efficiency and accuracy, and ensures that the final solution achieves the expected results in actual deployment. This embodiment, through comprehensive verification in a digital twin environment, demonstrates that the optimal network architecture planning scheme has obtained reliable performance guarantees, providing a solid foundation for actual deployment and implementation.

[0220] In this embodiment, step S2, which unifies the discrete dynamic model and the continuous dynamic model under the same mathematical framework using a unified transformation operator, includes: Based on the local matrix exponential function, when performing spatiotemporal discretization processing on the dynamic change process of the network, a higher-order finite difference scheme is constructed using the central difference scheme or the upwind difference scheme to establish a discrete mathematical model of the network state transition, thereby obtaining the dynamic model of the finite difference network.

[0221] In this step, the high-order finite difference scheme is the core mathematical tool for constructing the discrete dynamic model. When performing spatiotemporal discretization of the network's dynamic changes based on the local matrix exponential function, this embodiment selects an appropriate difference scheme to construct a high-precision discrete model according to the changing characteristics of the network parameters.

[0222] Among them, the central difference scheme is suitable for smoothly changing network parameters. It calculates the rate of change by symmetrically sampling the state values ​​before and after the time point, and has high accuracy (the truncation error is on the order of the square of the time step). It can accurately capture the smooth changing trend of network parameters. The upwind difference scheme is suitable for network parameters with steep changes or oscillations. It calculates the rate of change by sampling the state values ​​in the upstream direction. It has good numerical stability and can effectively handle abrupt changes in the network without producing numerical oscillations.

[0223] In practical applications, this embodiment adopts an adaptive selection mechanism, dynamically switching different difference formats according to the local state change rate, and using high-order differences (such as fourth-order precision compact differences) to improve accuracy in critical regions, while using low-order differences to reduce computational burden in secondary regions, thus forming a difference strategy that balances computational efficiency and accuracy.

[0224] This embodiment uses this high-order finite difference scheme to precisely mathematically represent the discrete-time evolution of network states, forming a complete finite difference network dynamic model, which provides an accurate mathematical expression for capturing discrete events in the network (such as device state changes, link failures, and service switching).

[0225] In this step, discrete-time mapping is a fundamental operation of higher-order finite-difference schemes, converting continuous time and spatial dimensions into discrete grid points. This embodiment designs an adaptive grid generation technique that automatically adjusts the grid density based on the network's dynamic characteristics, increasing the grid density in regions of rapid change and decreasing it in regions of gradual change, thus balancing computational accuracy and efficiency.

[0226] In the time dimension, a variable step size strategy is adopted, which automatically adjusts the time step size according to the rate of change of the state. Smaller time steps are used around critical events to improve accuracy, while larger time steps are used in the stable phase to improve efficiency. In the spatial dimension, a multi-level grid technology is adopted, which applies spatial discretization of different scales to different regions according to the network topology and functional partitions. For example, fine-grained grids are used in the core network region and coarse-grained grids are used in the edge region, balancing accuracy and computational cost.

[0227] In particular, the discretization process places special emphasis on boundary condition handling. For network boundary nodes, virtual node technology and extrapolation of boundary conditions are applied to ensure the stability and accuracy of boundary region calculations. This adaptive discretization strategy not only improves modeling accuracy but also significantly reduces computational complexity, making discrete dynamic models of complex network systems feasible in practical applications.

[0228] In this embodiment, the selection of the difference scheme is a key decision to ensure the numerical stability and accuracy of the model. Based on network state characteristics and numerical analysis theory, this embodiment constructs an intelligent difference scheme selection system to dynamically select the most suitable difference scheme for different types of network parameters.

[0229] For parameters such as bandwidth utilization and service load, which typically exhibit smooth changes, the central difference scheme is preferred to accurately capture gradual trends by leveraging its high-order accuracy. For parameters that may exhibit discontinuous changes, such as device state switching and traffic bursts, the upwind difference scheme is used to prevent numerical oscillations and ensure computational stability by utilizing its upwind characteristics. For volatile phenomena such as congestion propagation and fault cascading, the weighted intrinsically non-oscillating scheme (WENO) is adopted to balance numerical stability and shock wave capture capabilities, accurately representing the characteristics of wave propagation.

[0230] To further improve computational accuracy, this embodiment applies a compact high-order difference scheme to key network nodes and performance indicators. This scheme uses a wider range of computational templates to achieve fourth-order or even sixth-order accuracy, significantly reducing numerical errors and providing more accurate evolutionary predictions. This intelligent difference scheme selection system automatically adapts to the dynamic characteristics of the network, ensuring both computational stability and model accuracy, thus providing a solid mathematical foundation for the construction of discrete dynamic models.

[0231] Based on the local matrix exponential function, a semi-Lagrange framework is constructed, and the traffic evolution trajectory is traced along the feature lines to establish the partial differential evolution equations of network traffic and service demand, thereby obtaining the semi-Lagrange network continuous model.

[0232] Understandably, the semi-Lagrangian framework is an advanced method for describing the continuous dynamic characteristics of networks. Unlike the traditional Eulerian perspective, the semi-Lagrangian method describes system evolution by tracing along characteristic lines, making it particularly suitable for handling continuous processes such as traffic transmission and changes in service demands in networks.

[0233] The semi-Lagrange framework constructed in this embodiment mainly includes several core components: a feature line calculation module, which calculates the trajectory (feature line) of flow particles in the network based on the network topology and flow transmission path, capturing the migration law of flow in space and time; a matter derivative calculation module, which calculates the total derivative of the network state variables along the feature line, accurately describing the rate of change of flow particles over time, avoiding the problem of handling nonlinear convection terms in the Euler perspective; and a grid projection module, which maps the results calculated along the feature line back to a fixed computational grid, realizing the combination of the feature line method and the traditional grid method.

[0234] The greatest advantage of this semi-Lagrange framework lies in its ability to accurately handle high-speed flow phenomena in networks, effectively avoiding the numerical instability of traditional Eulerian methods when dealing with high Reynolds number flows. It is suitable for describing continuous but rapidly changing phenomena in networks, such as burst traffic, congestion wave propagation, and fluctuations in service demand. Using this method, this embodiment establishes a system of partial differential equations for the evolution of network traffic and service demand, forming a continuous semi-Lagrange network model, providing a high-precision mathematical framework for describing continuous changes in networks.

[0235] In this step, feature line tracing is the core mechanism of the semi-Lagrange framework, allowing the model to capture dynamic changes along the traffic transmission path. This embodiment designs an adaptive feature line generation algorithm that can intelligently construct a set of feature lines representing the traffic transmission path based on the network topology and traffic characteristics. The algorithm first determines the possible transmission paths of traffic particles based on network routing strategies and topological connections, forming an initial feature line skeleton. Then, it adjusts the feature line density distribution according to link load status and congestion conditions, increasing the feature line density in high-load areas and congestion bottlenecks to improve the computational accuracy in these critical areas. Finally, considering the time-varying characteristics of traffic, it dynamically adjusts the time sampling interval of the feature lines according to the frequency and amplitude of traffic fluctuations to ensure the capture of key time change points.

[0236] The mathematical representation of the feature lines employs parametric curve equations. Each feature line is defined by a series of control points, and smooth curves are generated between these control points through cubic spline interpolation, ensuring the continuity and differentiability of the feature lines. When network topology changes or routing strategies are adjusted, the set of feature lines is automatically updated to maintain consistency with the actual traffic transmission paths, providing a dynamically adaptable computational framework for the continuous model.

[0237] The calculation of the mass derivative is the theoretical core of the semi-Lagrange method. By calculating the total derivative along the characteristic line, it avoids the difficulties in handling nonlinear convection terms in the Euler method. This embodiment implements a high-precision method for calculating the mass derivative, ensuring accurate and reliable calculation of the rate of change of state variables along the characteristic line.

[0238] Specifically, firstly, at discrete time points, the state quantity values ​​at each control point on the characteristic line are calculated to form a spatiotemporal distribution sample of the state quantity; then, higher-order difference or spectral methods are used to calculate the spatial and temporal derivatives at these sample points; finally, combined with the velocity field information of the characteristic line, the material derivative (total derivative) of the state quantity is calculated to obtain the true rate of change along the flow path.

[0239] To improve computational accuracy, this embodiment increases the density of control points in key change regions of the feature lines and applies an adaptive time step strategy to ensure accurate capture of important changes. The material derivative calculation results directly reflect the changing patterns of network traffic and service demands during transmission, providing core data support for constructing accurate partial differential evolution equations and forming the foundational computational link for the continuous model of semi-Lagrange networks.

[0240] Mesh projection is a key technique that combines the feature line method with traditional mesh methods, ensuring that computational results can be presented and analyzed on a fixed mesh. This embodiment develops a high-precision mesh projection algorithm that accurately maps the results calculated along feature lines onto a regular computational mesh, achieving a seamless connection between continuous models and discrete analysis.

[0241] During the projection process, the position of each grid point on the feature line is first determined, and the mapping relationship between the grid point and the feature line is established. Then, based on the calculation results on the feature line, the state quantity value at the grid point is calculated by using conformal interpolation or physical quantity conservation interpolation methods to ensure that the interpolation process does not introduce non-physical oscillations. Finally, boundary processing and consistency verification are performed to ensure the physical rationality and numerical stability of the grid value distribution.

[0242] To improve projection accuracy, high-order local interpolation techniques were applied in key areas, using more feature line information to improve interpolation accuracy, especially in areas with abrupt changes in flow and congestion boundaries, to ensure that these important features can be accurately preserved in the grid representation.

[0243] This embodiment effectively transforms the advantages of the semi-Lagrange method into computational results on a regular grid through this high-precision grid projection, facilitating subsequent analysis and visualization, while preserving the accuracy of tracing along feature lines, providing a practical computational output interface for continuous models.

[0244] Based on the finite difference network dynamic model and the semi-Lagrange network continuous model, a unified transformation operator is defined to map the discrete time step of the finite difference network dynamic model to the continuous time manifold of the semi-Lagrange network continuous model, thereby achieving mathematical unification of discrete and continuous representations and obtaining the unified modeling framework.

[0245] Understandably, the unified transformation operator addresses the key mathematical challenge of fusing discrete and continuous models. This operator, by constructing a special mapping function, maps the discrete-time representation of a finite-difference network dynamic model to the continuous-time manifold of a semi-Lagrange network continuous model, achieving a seamless unification of the two representations.

[0246] The design of the unified transformation operator is based on several key principles: the point value consistency principle, which ensures that the transformed continuous representation matches the original discrete value precisely at discrete time nodes; the derivative consistency principle, which ensures that the derivative value of the continuous representation matches the difference approximation value of the discrete model at discrete time nodes; and the energy conservation principle, which ensures that the transformation process does not introduce additional energy and maintains the physical characteristics of the system.

[0247] Specifically, the unified transformation operator employs a piecewise high-order polynomial interpolation method to construct smooth connection functions between discrete time nodes, while simultaneously satisfying the aforementioned principles. For highly nonlinear regions, the operator adopts an adaptive subdivision strategy to increase the density of local interpolation points and improve transformation accuracy. This design enables the unified transformation operator to accurately capture the complex dynamic characteristics of network systems, which exhibit both discrete jumps and continuous evolution, achieving a true mathematical unification of discrete and continuous representations.

[0248] In this step, polynomial interpolation mapping is the core technology of the unified transformation operator, providing an accurate transformation from discrete points to continuous curves. This embodiment designs a piecewise adaptive polynomial interpolation method, which can accurately preserve the characteristics of the discrete model while ensuring mathematical continuity.

[0249] This method automatically selects appropriate polynomial order and interpolation node distribution based on the local characteristics of discrete data: in regions of gradual change, low-order polynomials (such as cubic Hermitian polynomials) are used to achieve smooth interpolation and reduce computational complexity; in regions of rapid change, high-order polynomials (such as 5th-7th order spline functions) or subdivided piecewise low-order polynomials are used to improve local fitting accuracy; near discontinuities, special transition functions and piecewise strategies are applied to ensure that the smoothness of the overall function is maintained while preserving the discontinuities. During the interpolation process, strict boundary condition matching is performed to ensure that at discrete time points, the values ​​of the interpolation function and their derivatives precisely match the corresponding quantities of the discrete model, achieving seamless connection.

[0250] Understandably, this specially designed polynomial interpolation mapping provides a mathematical foundation for a unified transformation operator, enabling discrete and continuous models to achieve a unified expression while preserving their respective characteristics, thus providing a powerful tool for the comprehensive analysis of network systems.

[0251] Among them, the time manifold mapping is an advanced feature of the unified transformation operator, which maps the discrete time step to a continuous time manifold, achieving complete unification of the time domain. In this embodiment, by introducing differential geometry, the network state evolution is regarded as a manifold structure in a high-dimensional state space. By constructing continuous-time parameterization, the mapping from discrete time points to continuous time flow is realized.

[0252] This method first constructs a local coordinate system on a discrete time series, with each discrete time point corresponding to a local coordinate region; then, it connects these local coordinates into a continuous time manifold through a smooth transition function of the overlapping regions; finally, it defines the state variables as smooth functions on the time manifold to realize the transformation from a discrete state series to a continuous state flow.

[0253] The key technique of time-manifold mapping is the construction of the tangent space. By calculating the rate of change of the state at each discrete time point (corresponding to the tangent vector), the tangent space structure is constructed, and the elements of the tangent space are mapped onto the manifold through exponential mapping to form continuous trajectories. This time-manifold mapping based on differential geometry provides a theoretically complete mathematical framework for unified transformation operators, capable of handling complex network systems with multi-scale time characteristics, realizing a unified expression of discrete dynamic features and continuous evolution laws, and providing a new perspective for network modeling and analysis.

[0254] In this step, the unified modeling framework is a comprehensive mathematical system that integrates discrete and continuous dynamic characteristics. By applying the unified transformation operator, this embodiment successfully unifies the dynamic model of finite difference networks and the continuous model of semi-Lagrange networks into the same mathematical expression system, forming a complete unified modeling framework.

[0255] It should be noted that this framework has several key features: integrated expressive capability, which can simultaneously describe discrete events (such as fault occurrence and state transition) and continuous processes (such as traffic fluctuations and performance evolution) in the network, avoiding the problem of separation between discrete and continuous models in traditional modeling; multi-scale analysis capability, which can analyze network behavior at different time scales, from microsecond-level data packet transmission to day-level service fluctuations, achieving consistent analysis across scales; and causal relationship tracing capability, which, through unified mathematical expression, can accurately trace the causal links of state changes in the network and identify the root causes of complex problems.

[0256] The implementation of this unified modeling framework in this embodiment enables the complete and accurate mathematical representation of the dynamic characteristics of network systems, providing a solid theoretical foundation for subsequent analysis, optimization, and prediction. Compared with traditional discrete event simulation or continuous system modeling, the unified modeling framework has significant advantages in describing the dynamic characteristics of complex network systems, especially in scenarios where discrete events are closely coupled with continuous changes. It can provide more accurate predictions of system behavior, offering more reliable theoretical support for network planning and optimization decisions.

[0257] It should be noted that the practical application of the framework is the final step in verifying the effectiveness of the unified modeling framework. This embodiment successfully applies the unified modeling framework to the analysis of various network scenarios, demonstrating its powerful modeling capabilities and practical value. In network capacity planning, the framework can simultaneously consider the mutual influence of device expansion (discrete event) and traffic growth (continuous process), accurately predicting the long-term performance evolution under different expansion strategies; in fault impact analysis, the framework can accurately simulate the link response of congestion propagation (continuous process) triggered by device failure (discrete event), predicting the cascading effects and scope of impact of the failure; in quality of service optimization, the framework can evaluate the impact of queuing policy adjustments (discrete event) on end-to-end delay distribution (continuous characteristic), guiding QoS parameter optimization.

[0258] These application examples demonstrate that the unified modeling framework not only theoretically unifies discrete and continuous models but also provides a powerful tool for solving complex network problems in practice. Compared with traditional single modeling methods, the unified modeling framework exhibits significant advantages in analytical accuracy, computational efficiency, and the breadth of application scenarios. Particularly in complex network systems that require simultaneous consideration of multiple time scales and dynamic characteristics, it can provide more comprehensive and accurate analytical results, offering a scientific basis for network architecture planning and optimization.

[0259] Example 2: Figure 2 A schematic diagram of the network architecture intelligent planning system of this embodiment is shown; as follows: Figure 2 As shown, this embodiment also provides a network architecture intelligent planning system, which includes: The digital twin construction module 201 is used to acquire network data, preprocess the network data and establish a digital model to construct a network digital twin; The unified modeling module 202 is used to construct a network state matrix and design a local matrix exponential function based on the network state data provided by the network digital twin. Based on the local matrix exponential function, a discrete dynamic model and a continuous dynamic model are constructed. The discrete dynamic model and the continuous dynamic model are unified under the same mathematical framework through a unified transformation operator to obtain a unified modeling framework. The multi-objective optimization module 203 is used to define the state space and action space based on the network representation generated by the unified modeling framework, construct a reward function, design a reinforcement learning model and construct a collaborative learning framework based on the state space, the action space and the reward function, and generate an optimal solution set based on the learning results of the collaborative learning framework through multi-objective optimization technology to obtain a candidate network architecture scheme set. The intelligent planning and dynamic adjustment module 204 is used to generate a planning constraint model based on the candidate network architecture scheme set, determine the network node deployment location, link capacity configuration, resource scheduling strategy and fault self-healing path according to the planning constraint model and the candidate network architecture scheme set, perform simulation testing and iterative optimization in the network digital twin environment, output the optimal network architecture planning scheme, collect network operation data in real time according to the deployed optimal network architecture planning scheme, identify network performance deviations and update and optimize the network architecture planning model based on the network operation data, and perform dynamic adjustment of the network architecture based on the updated network architecture planning model.

[0260] It should be noted that this system implements all the functions described above and consists of four main modules: The digital twin construction module 201 is responsible for acquiring network data, preprocessing it, establishing a digital model, constructing a network digital twin, and providing a basic network environment model for the system.

[0261] The unified modeling module 202 constructs a network state matrix based on data provided by the network digital twin, designs a local matrix exponential function, constructs discrete dynamic models and continuous dynamic models, and forms a unified modeling framework through a unified transformation operator.

[0262] The multi-objective optimization module 203 defines the state space and action space based on the network representation generated by the unified modeling framework, constructs the reward function, designs the reinforcement learning model and collaborative learning framework, and generates a set of candidate network architecture schemes through multi-objective optimization.

[0263] The intelligent planning and dynamic adjustment module 204 generates a planning constraint model based on the candidate scheme set, determines the deployment location of network nodes, link capacity configuration, etc., performs simulation testing and iterative optimization, outputs the optimal network architecture planning scheme, and makes dynamic adjustments based on real-time operation data.

[0264] It should be noted that each functional module of the system corresponds one-to-one with the corresponding steps in the aforementioned method embodiment 1. Through the collaborative work of the above modules, intelligent planning and dynamic optimization of complex network architecture can be achieved, adapting to the continuous changes in network services, balancing multi-dimensional optimization objectives, and improving planning efficiency and accuracy.

[0265] The above description is only a preferred embodiment of the present disclosure and is not intended to limit the present disclosure. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present disclosure should be included within the protection scope of the present disclosure.

Claims

1. A network architecture intelligent planning method, characterized in that, include: Acquire network data, preprocess the network data and establish a digital model to construct a network digital twin; Based on the network state data provided by the network digital twin, a network state matrix is ​​constructed and a local matrix exponential function is designed. Based on the local matrix exponential function, a discrete dynamic model and a continuous dynamic model are constructed. The discrete dynamic model and the continuous dynamic model are unified under the same mathematical framework through a unified transformation operator to obtain a unified modeling framework. Based on the network representation generated by the unified modeling framework, a state space and an action space are defined, a reward function is constructed, a reinforcement learning model is designed based on the state space, the action space, and the reward function, and a collaborative learning framework is constructed. Based on the learning results of the collaborative learning framework, an optimal solution set is generated through multi-objective optimization techniques to obtain a set of candidate network architecture schemes. Based on the candidate network architecture scheme set, a planning constraint model is generated. According to the planning constraint model and the candidate network architecture scheme set, the deployment location of network nodes, link capacity configuration, resource scheduling strategy and fault self-healing path are determined. Simulation test and iterative optimization are carried out in the network digital twin environment to output the optimal network architecture planning scheme. According to the deployed optimal network architecture planning scheme, network operation data is collected in real time. Based on the network operation data, network performance deviations are identified and the network architecture planning model is updated and optimized. Based on the updated network architecture planning model, dynamic adjustment of the network architecture is performed.

2. The method according to claim 1, characterized in that, The process of acquiring network data, preprocessing the network data, establishing a digital model, and constructing a network digital twin includes: Based on the network data, which includes network topology data, device parameters, traffic load, and service requirement data, the network data is cleaned, standardized, and structured to obtain a standardized network dataset. Based on the standardized network dataset, a digital model of network nodes, links, and service flows is established to construct a digital mapping body of network elements. Based on the digital mapping body of network elements, network traffic distribution patterns, service fluctuation characteristics, and fault propagation patterns are extracted through time series analysis and pattern recognition algorithms to form a network behavior pattern library. Based on the digital mapping of the network elements and the network behavior pattern library, a network digital twin model containing the coupling relationship between the physical layer, logical layer and business layer is constructed. By comparing and verifying the simulation data output by the network digital twin model with the actual network data, the parameters of the network digital twin model are continuously adjusted and optimized to obtain the network digital twin.

3. The method according to claim 1, characterized in that, The process involves constructing a network state matrix and designing a local matrix exponential function based on the network state data provided by the network digital twin. A discrete dynamic model and a continuous dynamic model are then constructed based on the local matrix exponential function. A unified transformation operator is used to unify the discrete dynamic model and the continuous dynamic model under the same mathematical framework, resulting in a unified modeling framework, including: Based on the network status data provided by the network digital twin, a multi-dimensional network status matrix is ​​constructed, which includes node status, link status, and traffic status. Based on the multidimensional network state matrix, a local matrix exponential function incorporating adaptive parameters and a region segmentation mechanism is designed. Based on the local matrix exponential function, the network dynamic change process is spatiotemporally discretized. A high-order finite difference scheme is constructed to establish a discrete mathematical model of network state transition, and a finite difference network dynamic model is generated as the discrete dynamic model. Based on the local matrix exponential function, a semi-Lagrange framework is constructed to establish the evolution equation of network traffic and service demand, generating a continuous semi-Lagrange network model as the continuous dynamic model. Based on the local matrix exponential function, a unified transformation operator is designed to unify the discrete representation of the dynamic model of the finite difference network and the continuous representation of the continuous model of the semi-Lagrange network under the same mathematical framework, thereby constructing a unified representation system for the network system and obtaining the unified modeling framework.

4. The method according to claim 1, characterized in that, The process involves defining a state space and an action space based on the network representation generated by the unified modeling framework, constructing a reward function, designing a reinforcement learning model and constructing a collaborative learning framework based on the state space, action space, and reward function, and generating an optimal solution set through multi-objective optimization techniques based on the learning results of the collaborative learning framework, thereby obtaining a set of candidate network architecture schemes, including: Based on the network representation generated by the unified modeling framework, a network state space containing node configuration, link parameters, and resource allocation strategies is defined as the state space, and an action space including node deployment, link planning, and resource scheduling operations is designed. Based on the state space and the action space, a composite reward function with a weight adaptive mechanism is constructed as the reward function to achieve a dynamic balance between bandwidth cost, latency performance, energy efficiency and reliability, thus forming a multi-dimensional reward evaluation mechanism for network planning. Based on the state space, the action space, and the multidimensional reward evaluation mechanism, a deep reinforcement learning model integrating the Transformer structure is designed as the reinforcement learning model. A mapping relationship between network state and optimal action is established, and a decision network with long-term memory and attention mechanism is constructed. Based on the decision network, a multi-agent collaborative learning framework is constructed as the collaborative learning framework to realize collaborative optimization among planning tasks of different network domains and different levels, and to establish a multi-agent system based on shared experience and distributed decision-making. Based on the collaborative learning results of the multi-agent system, a Pareto optimal solution set that satisfies the bandwidth cost, latency performance, energy efficiency, and reliability is generated through multi-objective optimization techniques such as non-dominated sorting and congestion distance calculation, thus obtaining the candidate network architecture scheme set.

5. The method according to claim 1, characterized in that, The step of generating a planning constraint model based on the candidate network architecture scheme set, and determining the network node deployment location, link capacity configuration, resource scheduling strategy, and fault self-healing path according to the planning constraint model and the candidate network architecture scheme set, includes: Based on the candidate network architecture scheme set and business scenario requirements, analyze business characteristics, security level, available resources and deployment limitations, and generate a scenario-based planning constraint model; Based on the planning constraint model, the candidate network architecture scheme set is screened, and combined with geographic information, resource distribution and connectivity requirements, the optimal deployment location of network nodes is determined by the location optimization algorithm to generate a network node deployment scheme. Based on the network node deployment scheme, the network traffic distribution pattern and service fluctuation characteristics in the network digital twin are utilized to construct a traffic prediction model through a time series prediction algorithm. This model analyzes service growth trends and fluctuation characteristics, determines the capacity configuration and expansion strategy for each link, and generates a link capacity planning scheme that includes redundancy design.

6. The method according to claim 5, characterized in that, The step of determining the network node deployment location, link capacity configuration, resource scheduling strategy, and fault self-healing path based on the planning constraint model and the candidate network architecture scheme set also includes: Based on the network node deployment scheme and the link capacity planning scheme, a resource allocation algorithm and a dynamic scheduling mechanism are designed to determine the allocation strategies for computing resources, storage resources and bandwidth resources, and to generate a resource scheduling strategy scheme. Based on the network node deployment scheme and the link capacity planning scheme, the dynamic characteristics of the network are characterized by the local matrix exponential function and the unified modeling framework, potential fault points and fault propagation paths are analyzed, and fault risk assessment results are generated. Based on the fault risk assessment results, a multi-level backup strategy and a reconstruction strategy are designed. The multi-level backup strategy includes primary and backup link configuration and redundant node deployment. The reconstruction strategy includes traffic rerouting mechanism and resource reallocation mechanism to build the network's fault self-healing capability and obtain a fault self-healing path planning scheme.

7. The method according to claim 1, characterized in that, The process of collecting network operation data in real time according to the deployed optimal network architecture planning scheme, identifying network performance deviations based on the network operation data and updating and optimizing the network architecture planning model, and performing dynamic adjustments to the network architecture based on the updated network architecture planning model includes: Based on the optimal network architecture plan, network operation status data, resource usage and abnormal events are collected in real time to obtain network operation data stream; Based on the network operation data stream and expected performance indicators, network performance deviations and potential problems are identified by constructing an anomaly detection model and trend analysis algorithm, and a performance deviation analysis report is generated. Based on the performance deviation analysis report, information on network performance index deviations and resource utilization anomalies from actual operation feedback is extracted. The parameters and strategies of the reinforcement learning model in the network architecture planning model are updated using incremental learning and transfer learning techniques. The constraints of the planning constraint model are adjusted according to the network performance index deviations, constructing a self-evolving continuous learning mechanism to obtain the updated network architecture planning model. This updated model is built upon the reinforcement learning model, the planning constraint model, and the unified modeling framework. Based on the updated network architecture planning model, a dynamic network architecture adjustment scheme is generated according to the current deployment status of the optimal network architecture planning scheme. The dynamic network architecture adjustment scheme includes node reconfiguration instructions, link adjustment instructions, and resource reallocation instructions, and performs dynamic adjustments to the optimal network architecture planning scheme.

8. The method according to claim 1, characterized in that, The process of conducting simulation tests and iterative optimizations in the network digital twin environment to output the optimal network architecture planning scheme includes: Based on the overall network architecture planning scheme, by constructing simulation scenarios and conducting simulation tests in the network digital twin environment, the performance indicators of the network architecture planning scheme, including throughput, latency, packet loss rate and energy consumption, under different load conditions, traffic patterns and fault scenarios are evaluated, and a performance evaluation report is generated. Based on the performance evaluation report and preset performance targets, the network architecture planning scheme is iteratively optimized by adjusting the deployment location of network nodes, link capacity configuration, and resource scheduling strategy until the performance indicators including throughput, latency, packet loss rate, and energy consumption all meet the preset performance target requirements, and the optimal network architecture planning scheme is output.

9. The method according to claim 3, characterized in that, The process of unifying the discrete dynamic model and the continuous dynamic model under the same mathematical framework through a unified transformation operator includes: Based on the local matrix exponential function, when performing spatiotemporal discretization processing on the dynamic change process of the network, a high-order finite difference scheme is constructed using the central difference scheme or the upwind difference scheme to establish a discrete mathematical model of the network state transition, thereby obtaining the dynamic model of the finite difference network. Based on the local matrix exponential function, by constructing a semi-Lagrange framework, tracing the traffic evolution trajectory along the feature lines, establishing the partial differential evolution equations of network traffic and service demand, and obtaining the semi-Lagrange network continuous model; Based on the finite difference network dynamic model and the semi-Lagrange network continuous model, a unified transformation operator is defined to map the discrete time step of the finite difference network dynamic model to the continuous time manifold of the semi-Lagrange network continuous model, thereby achieving mathematical unification of discrete and continuous representations and obtaining the unified modeling framework.

10. A network architecture intelligent planning system, characterized in that, include: A digital twin construction module is used to acquire network data, preprocess the network data and establish a digital model to construct a network digital twin. The unified modeling module is used to construct a network state matrix and design a local matrix exponential function based on the network state data provided by the network digital twin. Based on the local matrix exponential function, a discrete dynamic model and a continuous dynamic model are constructed. The discrete dynamic model and the continuous dynamic model are unified under the same mathematical framework through a unified transformation operator to obtain a unified modeling framework. The multi-objective optimization module is used to define the state space and action space based on the network representation generated by the unified modeling framework, construct a reward function, design a reinforcement learning model and construct a collaborative learning framework based on the state space, the action space and the reward function, and generate an optimal solution set based on the learning results of the collaborative learning framework through multi-objective optimization techniques to obtain a set of candidate network architecture schemes. The intelligent planning and dynamic adjustment module is used to generate a planning constraint model based on the candidate network architecture scheme set, determine the network node deployment location, link capacity configuration, resource scheduling strategy, and fault self-healing path according to the planning constraint model and the candidate network architecture scheme set, conduct simulation testing and iterative optimization in the network digital twin environment, output the optimal network architecture planning scheme, collect network operation data in real time according to the deployed optimal network architecture planning scheme, identify network performance deviations based on the network operation data and update and optimize the network architecture planning model, and perform dynamic adjustment of the network architecture based on the updated network architecture planning model.

Images