Network optimization method and system based on generative correction digital twinning

Abstract

The application discloses a network dynamic optimization method and system based on generative correction digital twinning, and relates to the technical field of digital twinning models. The method comprises the following steps: constructing a unified space-time digital twinning model based on a target physical entity, and obtaining a physical snapshot instance; generating a condition vector of the physical snapshot instance based on a preset operation strategy; a generative large model outputs an ideal state instance according to the condition vector, calculates a weighted difference tensor between the ideal state instance and the physical snapshot instance, and deduces a network correction strategy graph by using a correction strategy; a subsequent simulation instance is generated by using a state evolution predictor, the correction strategy graph is parsed into atomic instructions, and the atomic instructions are executed; an updated physical snapshot instance is obtained, and a prediction deviation is calculated; and model parameters of the state evolution predictor are optimized based on the prediction deviation. The application continuously iterates and optimizes the prediction model by using the deviation between reality and prediction, so that the continuous self-evolution of the entire computing power network is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of digital twin model, and particularly relates to a network optimization method and system based on generative correction digital twin. BACKGROUND

[0002] Currently, from the intelligence calculation center supporting artificial intelligence large model training, to the edge computing node serving automatic driving and industrial internet, the computing power demand is showing the trend of ubiquitous, heterogeneous and real-time. In order to cope with this trend, the industry generally proposes and actively constructs the "1ms latency circle" integrated computing power network covering the city, which is composed of widely distributed computing power resources, ultra-high speed and low latency optical transport network (OTN) and a large number of dynamic user terminals, forming a highly complex giant system. How to realize the global optimal scheduling and management in this complex system has become a key problem in the evolution of computing power network.

[0003] However, the prior art has significant limitations in coping with the above demand. First, in terms of perception and monitoring, traditional network management systems mostly rely on SNMP, NetFlow and other protocols, which can only provide isolated and lagging index monitoring, lack of global view of fusion of computing power, network, storage, business, geographical location and latency, and are difficult to reveal the dynamic law and causal relationship of resources in the city. Secondly, in terms of scheduling and decision-making, the current mostly adopts a static threshold driven response mechanism, such as triggering migration or traffic engineering based on CPU usage or link bandwidth threshold, which often has a response period of minutes, which cannot meet the needs of millisecond-level latency sensitive businesses such as vehicle-road cooperation, remote surgery and holographic communication, and this local greedy strategy often leads to suboptimal overall resource utilization and service quality. Thirdly, in the application of digital twin, existing practices mostly stay at the visualization level, lack of "intelligent kernel" that can simulate nonlinear evolution and generate optimal state, and cannot guide the optimization of physical network from the perspective of prediction and decision-making. Finally, the spatio-temporal information processing of computing power network is still fragmented, and there is a lack of deep integration between geographic information system, network performance monitoring system and cloud management platform, making it difficult to achieve dynamic scheduling considering both spatial location and full link latency.

[0004] In summary, the passive and localized management paradigm of the prior art cannot meet the needs of real-time, global and intelligent optimization of the new generation of urban computing power network. Therefore, a new method and system are needed to upgrade digital twin from a static observation tool to an active intelligent collaborator, which can generate a more optimal operation blueprint and derive correction strategies, so as to realize the continuous self-optimization and evolution of physical network in a dynamic environment. SUMMARY

[0005] In view of the deficiencies in the prior art, the present application provides a network optimization method and system based on generative correction digital twinning, aiming to continuously iterate and optimize the prediction model by using the deviation between reality and prediction, so as to realize the continuous self-evolution and dynamic optimization of the entire computing power network.

[0006] The application discloses a network optimization method based on generative correction digital twinning, comprising:

[0007] S1, constructing a unified space-time digital twinning model based on a target physical entity;

[0008] S2, obtaining a physical snapshot instance of the unified space-time digital twinning model , the physical snapshot instance is used to represent the network running state of the unified space-time digital twinning model at the current time;

[0009] S3, generating a condition vector of the physical snapshot instance based on a preset operation strategy;

[0010] S4, inputting the condition vector into a generative large model, and obtaining an ideal state instance output by the generative large model, the generative large model being a conditional space-time diffusion model based on a U-Net architecture;

[0011] S5, calculating a weighted difference tensor between the ideal state instance and the physical snapshot instance ;

[0012] S6, generating a correction strategy graph based on the physical snapshot instance and the weighted difference tensor by using a correction strategy derivation network;

[0013] S7, generating a subsequent simulation instance based on the physical snapshot instance and the correction strategy graph by using a state evolution predictor; the subsequent simulation instance is used to represent the predicted network running state at the next time step;

[0014] S8, parsing the correction strategy graph into atomic instructions, and executing the atomic instructions to update the unified space-time digital twinning model;

[0015] S9, obtaining an updated physical snapshot instance , and generating a prediction deviation based on the updated physical snapshot instance and the subsequent simulation instance ;

[0016] S10, optimizing the model parameters of the state evolution predictor based on the prediction deviation;

[0017] S11, performing S3 to S10 according to a preset number of cycles.

[0018] Preferably, the generative large model comprises an encoder and a decoder, and the encoder comprises a plurality of down-sampling modules;

[0019] Each down-sampling module comprises two 3D convolution layers;

[0020] After each 3D convolution layer, two multi-head self-attention modules are connected in parallel, including a spatial attention module and a temporal attention module.

[0021] Preferably, the conditional vector is input into the generative large model, and an ideal state instance output by the generative large model is obtained , comprising:

[0022] The conditional vector is converted into an embedding vector of a preset dimension;

[0023] A cross-attention layer is constructed in the up-sampling module of the decoder, and the cross-attention layer comprises a transpose 3D convolution;

[0024] The embedding vector is up-sampled by the transpose 3D convolution, and a preset resolution feature is introduced from the encoder through a skip connection, to obtain the ideal state instance .

[0025] Preferably, a weighted difference tensor between the ideal state instance and the physical snapshot instance is calculated, comprising:

[0026] A weight tensor corresponding to the conditional vector is generated according to a preset weight rule;

[0027] A difference value between the ideal state instance and the physical snapshot instance is calculated; and the weight tensor and the difference value are multiplied element by element to obtain the weighted difference tensor.

[0028] Preferably, the correction strategy derivation network is a graph attention network;

[0029] The correction strategy derivation network is used to generate a correction strategy graph based on the physical snapshot instance and the weighted difference tensor, comprising:

[0030] A multi-modal attribute graph is constructed based on the physical snapshot instance , and parameters of the multi-modal attribute graph include nodes, edges, node features and edge features; the nodes are used to represent computing power centers, and the edges are used to represent network connections; the node features and the edge features are generated based on the physical snapshot instance .

[0031] The multi-modal attribute graph and the weighted difference tensor are input into the correction strategy derivation network, and a structured correction strategy graph output by the correction strategy derivation network is obtained, the correction strategy graph being a data structure including discrete actions and continuous parameters.

[0032] Preferably, a state evolution predictor is utilized to predict a subsequent simulation instance based on the physical snapshot instance and the correction strategy graph , comprising:

[0033] The physical snapshot instance is copied , obtaining a current simulation instance ;

[0034] The correction strategy graph is applied to the current simulation instance by using a preset strategy application operator , obtaining a hypothetical simulation instance ;

[0035] A state evolution predictor pre-trained is utilized to generate network states in a future time window based on the hypothetical simulation instance , obtaining a future state sequence, the future state sequence including the subsequent simulation instance ;

[0036] The future state sequence is checked according to a preset checking rule, and the subsequent simulation instance is obtained when the future state sequence satisfies a predefined constraint condition .

[0037] Preferably, the correction strategy graph is parsed into atomic instructions, and the atomic instructions are executed to update the unified space-time digital twin model, comprising:

[0038] An instruction of the correction strategy graph is obtained;

[0039] The instruction is decomposed into an interface call and / or a command sequence by using an instruction translator library;

[0040] The interface call and / or the command sequence are issued and executed, and a unified space-time digital twin model of a next time step is obtained.

[0041] Preferably, model parameters of the state evolution predictor are optimized based on a prediction bias, comprising:

[0042] The prediction bias is saved in a database;

[0043] The prediction bias is extracted from the database within a preset period, and the model parameters of the state evolution predictor are fine-tuned online by using the prediction bias.

[0044] Preferably, the network optimization method based on the generative correction digital twin further comprises:

[0045] The training state sample is obtained, Gaussian noise is injected into the training state sample through a preset number of time steps, and a noise state sample is obtained;

[0046] The noise state sample is input into the generative large model to train the generative large model to identify Gaussian noise.

[0047] The application discloses a network optimization system based on generative correction digital twinning, which is used for executing a network optimization method based on generative correction digital twinning.

[0048] The model construction module is configured to construct a unified space-time digital twinning model based on a target physical entity, and obtain a physical snapshot instance of the unified space-time digital twinning model , the physical snapshot instance is used to represent the network running state of the unified space-time digital twinning model at the current time;

[0049] The state prediction module is configured to generate a condition vector of the physical snapshot instance based on a preset operation strategy, input the condition vector into a space-time diffusion model, and obtain an ideal state instance output by the space-time diffusion model The space-time diffusion model is a neural network model based on a U-Net architecture;

[0050] The difference acquisition module is configured to calculate a weighted difference tensor between the ideal state instance and the physical snapshot instance ; and derive a correction strategy graph based on the physical snapshot instance and the weighted difference tensor using a correction strategy;

[0051] The simulation generation module is configured to generate a subsequent simulation instance based on the physical snapshot instance and the correction strategy graph using a state evolution predictor , the subsequent simulation instance is used to represent the predicted network running state at the next time step;

[0052] The bias calculation module is configured to parse the correction strategy graph into atomic instructions, execute the atomic instructions to update the unified space-time digital twinning model, obtain an updated physical snapshot instance , and generate a prediction bias based on the updated physical snapshot instance and the subsequent simulation instance

[0053] The model optimization module is configured to optimize the model parameters of the state evolution predictor based on the prediction bias.

[0054] The loop execution module is configured to: repeatedly execute the network optimization method based on generative correction digital twin.

[0055] Compared with existing technologies, the beneficial effects of this invention are as follows: by introducing generative correction digital twin technology, accurate prediction and dynamic optimization of the computing network's operating state are achieved. Compared with traditional methods, this solution can integrate multi-dimensional spatiotemporal information to construct a digital twin with intelligent evolution capabilities. Specifically, through the collaborative work of a generative large model and a graph attention network, the system can automatically identify the deviation between the physical network and the ideal state, and derive a correction strategy that includes discrete actions and continuous parameters. This deviation-driven closed-loop optimization mechanism enables the state evolution predictor to continuously iterate the model parameters, not only confirming the optimization effect, but also using the deviation between reality and prediction to continuously iterate and optimize its internal generation, derivation, and prediction models, thereby achieving continuous self-evolution and dynamic optimization of the entire computing network. Attached Figure Description

[0056] Figure 1 This is a flowchart illustrating the network optimization method based on generative correction digital twin provided by the present invention.

[0057] Figure 2 This is a schematic diagram of the network optimization system based on generative correction digital twin provided by the present invention. Detailed Implementation

[0058] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0059] The present invention will now be described in further detail with reference to the accompanying drawings.

[0060] like Figure 1 As shown in the figure, this embodiment of the invention provides a network optimization method based on generative correction digital twin, which includes the following steps.

[0061] S1. Construct a unified spatiotemporal digital twin model based on the target physical entity.

[0062] In the embodiments of the present application, the generative corrective digital twin (G-CDT) is not a simple mirror image of the physical world, but an intelligent agent with perception, generation, deduction, correction and learning ability, which is deeply coupled with the physical world and co-evolves. It continuously "pushes" the physical network from the current state to an ideal state theoretically optimal dynamically constructed by generative AI through a fast iterative intelligent closed loop, thereby realizing global, real-time and forward-looking autonomous optimization.

[0063] In the embodiments of the present application, the data structure (Unified Spatio-Temporal Twin, USTT) of the digital twin body as the core of system information interaction is defined. The USTT is a high-dimensional, multi-modal and structured spatio-temporal tensor, which is a "universal language" for lossless information exchange between the physical world and the digital world. The USTT is a one-time snapshot of the physical network and records the trajectory record of the continuous evolution of the physical network in the time dimension. Its core data structure is a four-dimensional tensor:

[0064] ,

[0065] Among them, represents the data object at time t, represents the dimension of the tensor, H represents the height, W represents the width, D represents the depth, and R represents the real number, represents the feature vector.

[0066] For example, for a city of 100km x 100km, if it is divided at a resolution of 100m, then H=1000 and W=1000. The D dimension can be used to represent the floor or altitude, thereby accurately encoding the deployment position of the MEC node in the building or the position of the low-altitude computing node such as the unmanned aerial vehicle.

[0067] It should be noted that, represents the feature vector on each three-dimensional grid element , the dimension and content of which are carefully designed to ensure the completeness of the network state description. The design of the USTT is divided into several logical groups to ensure the structured and interpretable information, including static physical attributes (Static Physical Attributes), dynamic node attributes (Dynamic Node Attributes), dynamic link attributes (Dynamic Link Attributes) and business and service quality attributes (Business&QoS Attributes).

[0068] Specifically, static physical attributes are loaded at system initialization and are almost constant. They include: grid geographic coordinates (lat, lon, alt); whether the grid is a computing node is_node (Boolean); node type node_type (enumeration: large data center, regional data center, edge node, MEC, RSU, etc.); PUE value of the machine room design; total capacity of the server rack; physical coordinates of the server in the machine room; accurate routing ID of the optical cable on the GIS map; total length and core type of the optical cable (such as G.652.D, G.654.E).

[0069] Dynamic node attributes include, but are not limited to: total number of CPU cores and used cores; total computing power and used computing power of various GPUs (such as NVIDIA A100, H100); computing power and usage of special AI accelerators (NPU / TPU); total memory capacity (GB) and used capacity; total storage capacity (TB), used capacity, read / write IOPS, and average delay (ms); real-time total power consumption (kW) of the node; real-time PUE value reported by the DCIM system. Dynamic node attributes are updated at a high frequency (seconds or sub-seconds).

[0070] Dynamic link attributes describe the connection state between nodes, which is the core of network state. For each node, its feature vector contains connection information to all its physical / logical neighbor nodes, forming an implicit graph structure. For example, for the link from node to neighbor , the features include: end-to-end latency latency_to_j (μs), latency jitter jitter_to_j (μs), total bandwidth bandwidth_to_j (Gbps), used bandwidth used_bw_to_j (Gbps), packet loss rate packet_loss_to_j (%), optical path signal quality (OSNR), etc.

[0071] Business and service quality attributes describe the business load situation running on the computing node. They include: the number of virtual machine / container instances carried; the number of various types of services (such as service_count_V2X, service_count_CloudGaming); service level agreement (SLA) of each type of service, such as required latency upper bound required_latency_k and required bandwidth lower bound required_bw_k; the number of tasks currently queued for processing and the average waiting time.

[0072] S2, obtaining a physical snapshot instance of the unified space-time digital twin model.

[0073] In this embodiment, the system maintains three different semantic USTT instances: The network running state of the unified spatio-temporal digital twin model at the current time is used to represent the network running state, The most real snapshot of the physical world is formed by real-time aggregation through the data collection adapter. : The state copy used for simulation, deduction and operation inside the digital twin system, and Maintain high-frequency synchronization. : A theoretically optimal network running state dynamically generated by the generative model of the present application according to the current global optimization target.

[0074] S3, generate a condition vector of a physical snapshot instance based on a preset operation strategy.

[0075] In an embodiment of the present application, the condition vector C is determined according to the current operation strategy (which may be triggered automatically by the upper business intent system or set by the operation and maintenance personnel) The dimension and content of the condition vector C are customized according to the needs of the specific operation strategy, for example, but not limited to, the following elements: business priority (such as emergency business, high-value business, etc.), resource allocation ratio (such as CPU, memory, storage, etc. Resource allocation ratio), network performance index requirement (such as delay, bandwidth, packet loss rate, etc.), computing power demand of specific area, etc. The condition vector As an input parameter of the generative model, it is used to guide the model to generate a theoretically optimal network running state instance .

[0076] S4, input the condition vector into the generative large model, and obtain the ideal state instance output by the generative large model.

[0077] In an embodiment of the present application, the generative large model is a conditional spatio-temporal diffusion model (Conditional Spatio-Temporal Diffusion Model, C-STDM) based on U-Net architecture. The generative large model does not generate conventional images or text, but is specially used to generate high-dimensional and complex USTT tensors.

[0078] The generative large model based on U-Net architecture includes an encoder and a decoder. The encoder is composed of multiple down-sampling modules, and the core of each down-sampling module is two 3D convolution layers (Conv3D), whose kernel_size=(3,3,3) and padding='same' enable it to capture features in the local spatial neighborhood. After each convolution block, two multi-head self-attention (Multi-Head Self-Attention, MHSA) modules are connected in parallel to capture non-local dependencies.

[0079] Specifically, the parallel multi-head self-attention module is a spatial attention module and a temporal attention module. The spatial attention module is used to calculate the self-attention of the input feature map dimension is flattened into a sequence, and then the self-attention is calculated on the sequence. This enables the generative large model to understand the "remote association between the algorithm congestion at one end of the city and the network delay at the other end". The temporal attention module acts on the historical state sequence of the input of the generative large model, so that the model can review and refer to the past evolution trend when generating the current ideal state, and understand the "tide effect" of the network. The outputs of the spatial attention module and the temporal attention module are fused through a learnable gated recurrent unit (GRU), so that the generative large model can dynamically decide whether to rely more on spatial information or temporal information at the current step.

[0080] In the process of generating an ideal state instance using the generative large model, the conditional vector is first converted into an embedding vector of a predetermined dimension. Then, a cross-attention layer is constructed in the upsampling module of the decoder, and the cross-attention layer includes a transpose 3D convolution. The embedding vector is upsampled by the transpose 3D convolution, and the predetermined resolution feature is introduced from the encoder through a skip connection to form a feature reconstruction path, and finally a denoising step is iteratively performed to obtain an ideal state instance .

[0081] Specifically, the conditional vector C is converted into an embedding vector of a fixed dimension using an embedding layer and a small Transformer encoder. The Query (Q) of the cross-attention layer comes from the feature map of the U-Net itself, and the Key (K) and Value (V) come from the global embedding vector . This enables the conditional vector to influence the reconstruction of features at each step during the entire generation process from noise to clear state, ensuring that the finally generated strictly serves the given conditional vector .

[0082] In the embodiment of the present application, the training process of the generative large model includes: obtaining a training state sample, injecting Gaussian noise into the training state sample through a predetermined number of time steps to obtain a noise state sample, and inputting the noise state sample into the generative large model to train the generative large model to identify the Gaussian noise.

[0083] Specifically, the training and generation of the generative large model follow the standard diffusion model paradigm, and first a large number of training state samples are obtained, and then Step (e.g.) Gradually inject Gaussian noise into it. At any step ,state The formula for generating it is:

[0084] ,

[0085] in, It is a pre-set, gradually increasing noise scheduling table. After... Step, the original clear state It will become a pure isotropic Gaussian noise. . This represents a conditional probability distribution, i.e., given the previous state. At that time, the current state The probability distribution. This represents a multivariate normal distribution (Gaussian distribution). This represents the mean of a normal distribution. Let I be the covariance matrix of the normal distribution, and I be the identity matrix. This represents a parameter that controls the magnitude of the covariance.

[0086] Next, using reparameterization techniques based on Get any :

[0087] ,

[0088] in, , , This indicates the injected noise.

[0089] In the reverse process, the decoder is trained given a noisy state. Current time step and optimization conditions In this case, it is possible to accurately predict the first... Noise added to the step The loss function is a weighted mean squared error:

[0090] ,

[0091] in, Represents the loss function. This represents a weighting function used to assign different levels of importance to different stages of the denoising process, typically giving higher weights to the earlier stages when noise is greater.

[0092] After the generative large model has been trained, any optimization conditions are given. , the generative large model can generate a new, clear, physically consistent and ideal state satisfying the optimization condition from a randomly sampled pure noise tensor , through iterative reverse denoising steps. The iterative formula of each step is:

[0093] ,

[0094] wherein, represents a random term for increasing the diversity of the generation process.

[0095] S5, calculate the weighted difference tensor between the ideal state instance and the physical snapshot instance.

[0096] In the embodiment of the present application, the weight tensor corresponding to the condition vector is generated according to the preset weight rule, the difference between the ideal state instance and the physical snapshot instance is calculated, and the weight tensor and the difference are multiplied element by element to obtain the weighted difference tensor.

[0097] Specifically, the weighted difference tensor is calculated using the following formula:

[0098] ,

[0099] wherein, represents Hadamard product (element by element multiplication), is a weight tensor of the same dimension as USTT, which is dynamically generated by the condition vector , and represents the weighted difference tensor.

[0100] For example, if emphasizes "minimizing the delay for V2X services", the weight of the delay feature channel corresponding to the geographical location of the V2X service in will be set very high. This enables the difference analysis to automatically focus on the current most critical point of contention.

[0101] S6, using the correction strategy derivation network to generate a correction strategy graph based on the physical snapshot instance and the weighted difference tensor.

[0102] In the embodiment of the present application, the correction strategy derivation network (Policy Derivation Network, PDN) is a graph attention network (Graph Attention Network, GAT) for processing network topology data. Based on the physical snapshot instance A multi-modal attributed graph is constructed, the multi-modal attributed graph and the weighted difference tensor are input into a correction strategy derivation network, and a structured correction strategy graph output by the correction strategy derivation network is acquired. Parameters of the multi-modal attributed graph include nodes, edges, node features and edge features; the nodes are used for representing computing power centers, the edges are used for representing network connections, and the node features and the edge features are generated based on the physical snapshot instance . The correction strategy graph is a data structure including discrete actions and continuous parameters.

[0103] Specifically, the difference tensor points out "where the problem is" and "how big the gap is", but does not explain "how to solve it". The USTT tensor is dynamically constructed into a multi-modal attributed graph at the physical snapshot instance at the time point , wherein the nodes are computing power centers, the edges are network connections, and the features of the nodes and the edges are directly obtained from . The input of the PDN is the graph and the difference tensor (as additional features of the graph nodes), and the output of the PDN is a structured correction strategy graph.

[0104] ,

[0105] wherein, represents the correction strategy graph, ( ) represents a composite function of a probability diffusion network (PDN) and a graph attention network (GAT), and G( ) represents a mapping to a real state.

[0106] The nodes and the edges of the correction strategy graph have specific "action" attributes, which are a complex data structure including discrete actions (such as action_type:'migrate_vm') and continuous parameters (such as cpu_allocation: 8.5 cores). The attention mechanism of the GAT enables it to learn highly complex, topology-dependent strategies, such as "if an edge node has a high load, and its neighbor node A has free resources but a large connection delay, and its neighbor node B has a tight resource but a small connection delay, then prefer to migrate a delay-sensitive task to B rather than A".

[0107] S7, using a state evolution predictor to generate a subsequent simulation instance based on the physical snapshot instance and the correction strategy graph.

[0108] In the embodiment of the present application, the subsequent simulation instance is used to represent the predicted network running state at the next time step. Before the correction strategy graph is applied to the physical world, it is necessary to first simulate the network running state at the next time step​ high-fidelity "digital dress rehearsal".

[0109] In the process of generating the subsequent simulation instance , the physical snapshot instance is copied to obtain the current simulation instance ; the correction strategy graph is applied to the current simulation instance using a preset strategy application operator to obtain the hypothetical simulation instance . The state evolution predictor is pre-trained based on the hypothetical simulation instance to generate network states in a future time window, and a future state sequence is obtained, wherein the subsequent simulation instance is included. According to a preset verification rule, the future state sequence is verified, and the subsequent simulation instance is obtained when the future state sequence meets the predefined constraint condition.

[0110] Specifically, the derived strategy is applied to the twin in a mathematical manner to obtain a hypothetical subsequent state:

[0111] ,

[0112] wherein is a defined strategy application operator.

[0113] A pre-trained state evolution predictor (SEP) is called to predict the chain reaction of the operation in a future time window . The SEP is a temporal graph convolutional network (T-GCN) trained to learn the "physical dynamics" of the network, that is, how an action will trigger a series of nonlinear state evolutions. The future state sequence is obtained by using the following formula:

[0114]

[0115] The preset verification rule is used to verify the future state sequence, to verify whether it meets the predefined hard constraint (for example, the CPU load of any node cannot exceed 98%, and the end-to-end delay of any key service cannot exceed the upper limit of its SLA), and to evaluate whether the strategy can reduce the overall cost function of the system (a weighted function that comprehensively considers the delay, energy consumption, resource cost, and SLA penalty). When the verification is passed, the subsequent simulation instance is obtained.

[0116] S8, parse the correction policy graph into atomic instructions, and execute the atomic instructions to update the unified space-time digital twin model.

[0117] In the embodiment of the present application, the instruction of obtaining the correction policy graph is obtained, and the instruction is decomposed into interface calls and / or command sequences by using the instruction translator library. The interface calls and / or command sequences are issued and executed, and the unified space-time digital twin model of the next time step is obtained.

[0118] It should be noted that the verified correction policy graph is high-level and abstract, and needs to be translated into specific instructions that can be understood by the underlying device. The instruction translator library of the decision and execution driving layer is used to decompose the high-level instructions in the correction policy graph into a series of specific, idempotent, ordered API calls or CLI command sequences. Among them, the instruction translator library includes translation modules for different manufacturer devices (such as Cisco, Huawei, Juniper) and different cloud platforms (such as OpenStack, Kubernetes, VMWare vSphere).

[0119] Further, through the SDN controller, cloud orchestrator, automation configuration tool (such as Ansible, SaltStack), etc., these atomic instructions are accurately and correctly issued to the physical device for execution. After the instruction execution, the state of the physical network changes, and the unified space-time digital twin model of the next time step is obtained.

[0120] S9, obtain the updated physical snapshot instance, and generate a prediction deviation based on the updated physical snapshot instance and the subsequent simulation instance.

[0121] In the embodiment of the present application, the prediction deviation is calculated by using the following formula:

[0122] ,

[0123] Among them, represents the prediction deviation, which is used to represent the deficiency of the state evolution predictor.

[0124] S10, optimize the model parameters of the state evolution predictor based on the prediction deviation.

[0125] In the embodiment of the present application, the prediction deviation is saved to the database; and in a preset period, the prediction deviation is extracted from the database, and the model parameters of the state evolution predictor are fine-tuned online by using the prediction deviation.

[0126] S11, according to a preset number of times, the S3 to S10 are executed cyclically.

[0127] In the embodiment of the present application, the new The self-improving optimization engine is formed as the starting point of the next optimization cycle.

[0128] The network optimization method based on generative correction digital twin provided by the application realizes global, real-time and forward-looking autonomous optimization by continuously "pushing" the physical network from the current state to an ideal state that is theoretically optimal and dynamically constructed by generative AI through a fast iterative intelligent closed loop.

[0129] As shown in Figure 2 , the application also provides a network optimization system based on generative correction digital twin, which is used to execute the network optimization method based on generative correction digital twin described above, and includes a model construction module, a state prediction module, a difference acquisition module, a simulation generation module, a deviation calculation module, a model optimization module and a loop execution module.

[0130] The model construction module is configured to construct a unified space-time digital twin model based on a target physical entity, and acquire a physical snapshot instance of the unified space-time digital twin model , the physical snapshot instance is used to represent the network running state of the unified space-time digital twin model at the current time.

[0131] The state prediction module is configured to generate a conditional vector of the physical snapshot instance based on a preset operation strategy; input the conditional vector into a space-time diffusion model, and acquire an ideal state instance output by the space-time diffusion model , the space-time diffusion model is a neural network model based on U-Net architecture;

[0132] The difference acquisition module is configured to calculate a weighted difference tensor between the ideal state instance and the physical snapshot instance ; and derive a correction strategy graph based on the physical snapshot instance and the weighted difference tensor using a correction strategy.

[0133] The simulation generation module is configured to generate a subsequent simulation instance based on the physical snapshot instance and the correction strategy graph using a state evolution predictor, the subsequent simulation instance is used to represent the predicted network running state at the next time step.

[0134] The deviation calculation module is configured to parse the correction strategy graph into atomic instructions, and execute the atomic instructions to update the unified space-time digital twin model; acquire an updated physical snapshot instance , and generate a prediction deviation based on the updated physical snapshot instance and the subsequent simulation instance .

[0135] The model optimization module is configured to optimize the model parameters of the state evolution predictor based on the prediction bias.

[0136] The loop execution module is configured to loop the network optimization method based on the generative correction digital twin.

[0137] From the above technical solution, the present application comprehensively and in real time perceives the complex heterogeneous computing power network, collects multi-source heterogeneous data, and constructs a digital twin state that is accurately synchronized with the physical world. Subsequently, the core generative AI engine intervenes, which creatively "generates" an "ideal state" that the current physical network should reach at the next moment, which is theoretically optimal according to the preset global optimization goal (such as the lowest latency, the lowest energy consumption or the highest resource utilization). Then, the ideal state is compared with the real-time perceived "real state" in depth, and the multi-dimensional "correction gap" between the two is quantified. Based on this gap, an intelligent strategy derivation module generates a specific and executable correction action sequence (for example, virtual machine migration, traffic path re-routing, network slice dynamic adjustment, etc.). To ensure absolute safety, this strategy will be first simulated in the digital twin environment before being issued to the physical network, predicting all possible chain reactions it may cause and verifying its effectiveness and safety. Once verified, the high-level strategy will be translated into atomic instructions for the underlying devices and executed accurately through the controller and orchestrator. Finally, through continuous monitoring of the post-execution state of the physical network, a feedback loop is formed, not only confirming the optimization effect, but also using the deviation between reality and prediction to continuously iterate and optimize the internal generation, derivation and prediction models, thereby realizing the continuous self-evolution and dynamic optimization of the entire computing power network.

[0138] To make the technical solutions, objectives and advantages of the present application clearer, the following will elaborate on the actual operation of the network optimization method based on the generative correction digital twin of the present application through three examples covering different application fields.

[0139] The first embodiment scenario describes a "Century Intersection" in the CBD area of a certain megacity, which is the intersection of multiple trunk roads, with huge traffic flow during peak hours and complex wireless environment. A fleet of 50 autonomous driving net taxis needs to provide services in this area. These vehicles rely on V2X (vehicle-road cooperation) technology to offload perception fusion and collaborative decision-making computing tasks that are extremely sensitive to latency to roadside units (RSUs) and edge computing (MEC) nodes. According to the agreement between the vehicle manufacturer and the operator, the p99 latency of the end-to-end (OBU to MEC and back to OBU) must be strictly controlled within 10ms.

[0140] The network optimization method and system based on generative correction digital twinning provided by the application can provide millisecond-level computing power guarantee for the automatic driving vehicle fleet of the urban core transportation hub.

[0141] Firstly, the system continuously collects data from probes deployed in RSUs, 5G base stations, MEC servers and on-board units (OBU) around the intersection, constructs and updates At T0 moment, The relevant features show that the user density (user_density) feature in the grid unit of the CBD area rises sharply, causing the 5G channel congestion (channel_congestion) to rise to 70%. The GPU utilization of the MEC-01 node (located at (h1, w1, d1)) closest to the intersection reaches 85%, and its average processing delay feature rises to 9ms. The V2X service end-to-end delay P99 value reported by the associated vehicle terminal appears 12-15ms jitter, and the SLA is about to be breached.

[0142] The system's built-in SLA monitoring module detects the risk of breach, automatically constructs and activates a high-priority conditional vector C:

[0143] {

[0144] "objective": "minimize_latency_p99",

[0145] "target_service_type": "V2X_collaborative_perception",

[0146] "target_geo_area": {"h": [h1-5, h1+5], "w": [w1-5, w1+5], "d": [d1,d1]},

[0147] "sla_constraints":{"latency_p99_upper_bound":"10ms", "jitter_upper_bound": "2ms"},

[0148] "priority": "critical"

[0149] };

[0150] The conditional vector C and the current (as historical context) are input into the generative model C-STDM. The model generates Within In the tensor, the following significant changes can be observed: MEC-01 at the (h1, w1, d1) position has its GPU utilization increased to 95%, and its internal service queue-related features show that V2X services are given the absolute highest priority. And, An innovative task splitting and distributed execution scheme is proposed: the "high-precision map matching and updating" sub-task (about 30% of the computing power) of the V2X task, which originally belonged to MEC-01 and had a slightly higher tolerance to latency (15ms acceptable), is scheduled to the 1.5km away MEC-02 node (located at (h2, w2, d2)) with a current GPU utilization of only 40%. At the same time, on the link characteristics connecting MEC-01, MEC-02 and the area where the vehicle fleet is located, A dedicated 5G network slice with guaranteed bandwidth (guaranteed_bw) and latency budget (latency_budget) features is planned, with ideal values set.

[0151] Then, calculate = ⊙ ( - ). Since the weights are concentrated on the latency and V2X service-related features, the GPU usage, service allocation, and related network links of MEC-01 and MEC-02 show significant positive or negative values, clearly indicating the direction and amplitude of the adjustments needed.

[0152] Further, input the dynamic graph G(T0) constructed from and to the PDN (GAT model). The PDN outputs a structured correction strategy graph, which after analysis contains the following specific atomic action instructions:

[0153] {"target_node": "MEC-01", "action": "reprioritize_gpu_queue", "params": {"service_type":"V2X_collaborative_perception","priority_level": "realtime_critical"}}

[0154] {"target_node": "MEC-02", "action": "instantiate_service", "params": {"service_name": "V2X_hdmap_update", "docker_image": "...", "resources": {"gpu_cores": 2, "memory_gb": 64}}}

[0155] {"target_link": ["AV_Fleet_Group_01", "MEC-02"], "action": "establish_5g_slice", "params": {"slice_id": "SLA_AV_01", "qos_profile": "URLLC", "guaranteed_bw_mbps": 500}}

[0156] {"target_app": "AV_Fleet_Brain_Controller", "action": "update_service_endpoint", "params": {"sub_task": "hdmap_update", "new_endpoint_ip":"IP_of_service_on_MEC-02"}}

[0157] Further, the correction policy graph is applied to , resulting in a hypothetical successor state . The SEP (T-GCN model) is used to predict the state evolution for the next 30 seconds, outputting . .

[0158] Then, check whether the CPU utilization and GPU utilization of MEC-01 and MEC-02 in each frame of the sequence are both <98%. Check whether the newly created 5G slice causes the adjacent base stations to be overloaded. Check whether the end-to-end delay p99 of the V2X service in the sequence is steadily decreasing and remains around 8.5 ms. And calculate the change of the overall cost function , confirming that J predicted < J current due to the avoidance of SLA penalties and more efficient resource utilization. If the check passes, the policy is confirmed to be safe and efficient.

[0159] Further, the high-level instructions in the correction strategy graph are translated into specific API calls to the underlying systems, including: sending a RESTful API request to the Kubernetes cluster of MEC-01, PATCHing the Pod object of the V2X perception service, modifying its priorityClassName. Sending a template to the OpenStack Heat orchestration engine of MEC-02, creating a new compute instance stack containing GPU resources and the specified image. Sending instructions to the operator's 5G core network NEF (Network Exposure Function) and SDN controller, configuring a new PDU session and network slice, and issuing corresponding flow table rules. Sending an API request to the service discovery module (such as Consul) of the cloud control center of the autonomous vehicle fleet, updating the address record of the "hdmap_update" sub-service.

[0160] After the instructions are executed, at T0+5 seconds, the perception layer collects new The system compares and finds that the key indicators (latency, load) of the physical network are highly consistent with the predictions in the twin deduction, and the model prediction error is recorded as <5%.

[0161] As can be seen, the SLA of the vehicle fleet is fully met, successfully passing through the business peak period. The system seamlessly enters the next round of continuous monitoring and fine-tuning cycle, always keeping the network running at the edge of the optimal state.

[0162] Second embodiment scenario description: The national gene sequencing center (GNC) located in city A needs to transmit 50TB of raw sequencing data generated in a day to the national supercomputing center (NSCC) located in city B for large-scale parallel analysis. This task is not sensitive to latency, but requires completion during the night low electricity price period (01:00-05:00), and the total cost (including network bandwidth cost, energy consumption during data transmission, and computing and storage cost of the supercomputing center) is the lowest.

[0163] Using the network optimization method and system based on generative correction digital twin provided by the application, the cost and energy efficiency optimal scheduling of cross-regional scientific research data "computing power express" can be generated.

[0164] At 00:55 in the morning, The city network load is low, and the default public network BGP path between the GNC in A and the NSCC in B passes through multiple aggregation layer devices, with an end-to-end available bandwidth available_bw of only 20Gbps, which cannot complete the 50TB transmission within 4 hours. At the same time, the electricity price characteristic electricity_price enters the night low value. The operation and maintenance personnel set or the system automatically triggers the condition vector C according to the plan on the GUI:

[0165] {

[0166] "objective": "minimize_total_cost",

[0167] "cost_components": ["network_transit_fee", "energy_consumption", "compute_instance_price", "storage_price"],

[0168] "task_id": "GeneData_Transfer_GNC_NSCC_20231027",

[0169] "constraints": { "time_window": ["01:00", "05:00"], "data_volume_tb": 50, "source_node_id": "GNC_A", "dest_node_id": "NSCC_B"}

[0170] };

[0171] After the large generative model model receives C, considering the cost optimization objective, it generates . This ideal state presents a completely different solution from the default state: at the network level, it does not choose to simply widen the public network path, but suggests using the OTN (Optical Transport Network) underlying network of the operator to temporarily establish a 100Gbps wavelength-level direct optical path from the A zone to the B zone switching node. In S_ideal, the link characteristics connecting the two nodes are rewritten as {link_type:'OTN_Private', bandwidth: 100000, transit_cost_factor: 0.2}. At the computing power level, it instructs the supercomputing center NSCC to pre-start a batch of CPU-type virtual machines with the highest energy efficiency but the lowest unit price, which are used for data decompression and preprocessing, rather than directly writing raw data to expensive high-performance parallel file systems. In S_ideal, the cpu_instance_count and cpu_utilization of the NSCC node are increased accordingly.

[0172] Further, there is a huge positive difference in the bandwidth and type of the link connecting the A and B zones, and a positive difference in the CPU resource usage of the NSCC, clearly indicating the need for resource creation and configuration.

[0173] PDN analysis After that, a correction policy graph is generated, which contains the following cross-domain coordination policies:

[0174] {"target_system": "Optical_Network_Controller", "action": "provision_otn_path", "params": {"source_port": "GNC_OTN_Port_1", "dest_port": "NSCC_OTN_Port_3", "bandwidth_gbps": 100, "protection_level": "none", "lease_duration_hours": 4}}

[0175] {"target_system": "NSCC_Cloud_Platform", "action": "scale_up_vm_group", "params": {"group_id": "cpu_preprocess_pool_low_cost", "instance_type": "c6g.8xlarge", "count": 100}}

[0176] {"target_system": "Data_Transfer_Manager", "action": "start_transfer_job", "params": {"protocol": "GridFTP", "source_path": "...", "dest_path":"...", "parallel_streams": 64, "network_interface": "otn_bond0"}};

[0177] Further, the security inference is performed using the state evolution predictor, and the inference shows that after the 100G optical path is established, the data transmission rate is stable at about 98Gbps, and it is expected to be completed within 3.5 hours, meeting the time window requirement. This operation has almost zero impact on the network quality of other night services in the city. The total cost of starting 100 CPU virtual machines is much lower than the cost of directly using high-performance computing nodes for preprocessing. The estimated total cost of the entire task Compared with the default scheme, it is reduced by 40%, and the verification is passed.

[0178] Further, the execution layer sends a standardized T-API call to the operator's optical network controller (such as an ASON / GMPLS controller or a NetConf-based controller) to request configuration and establishment of the temporary OTN connection. It also sends API requests to the Nova and Neutron services of the cloud platform (such as OpenStack) of the NSCC to create virtual machines in bulk and configure their networks. A script is executed on the D-T-N (Data Transport Node) of the GNC to start a GridFTP transfer process with multi-stream parallelism and specified network interfaces.

[0179] Finally, the system continuously monitors the OTN optical path quality and actual transmission rate during the transmission process and compares it with the progress planned in advance. If there is a rate drop (for example, due to a storage write bottleneck at the NSCC end), the G-CDT will immediately start a new round of fine-tuning, which may generate a strategy to dynamically increase the number of pre-processing virtual machines or adjust the write strategy of the parallel file system. After the task is completed, the system will automatically generate and execute the strategy, dismantle the temporary optical path, and release all virtual machines, achieving full life cycle "zero waste" automatic management of resources.

[0180] Third embodiment scenario description: An earthquake occurs in city C area, causing partial communication base stations and optical cables in the region to be interrupted, and an important edge data center MEC-C1 is completely out of contact. The emergency command department needs to immediately establish a stable emergency communication network and urgently mobilize multiple drones for high-resolution video reconnaissance of the disaster area. The video stream needs to be transmitted back in real time and analyzed by AI to quickly evaluate building damage and search for signs of survivors.

[0181] The network optimization method and system based on generative correction digital twinning provided by the application can cope with emergency communication and computing power reconstruction in urban-level sudden disasters.

[0182] First, the system acquires the current state of the network: , The tensor undergoes a dramatic, structural change: the characteristics of multiple network links in C area become {latency: infinity, bandwidth: 0, status: 'down'}, and all dynamic attributes of MEC-C1 node become offline. A large number of users' communication requests fail, and the service_success_rate feature drops sharply.

[0183] Next, determine the optimization target, the system is manually or automatically (by monitoring a large-scale network interruption) switched to "emergency response mode", immediately loads the preset highest priority condition vector C:

[0184] {

[0185] ​"objective": "maximize_service_availability_and_resilience",

[0186] "priority_services": ["emergency_calls_911", "first_responder_pTT","drone_video_analysis_for_rescue"],

[0187] "disaster_geo_area": "Area_C",

[0188] "policy_mode": "forceful_preemption"

[0189] };

[0190] Further, after receiving this extreme condition vector C, the generative large model quickly generates a that aims to "restore critical services at all costs". In this ideal state: all non-critical services (such as video on demand, online games, and advertisement pushing) are cleared or reduced to the minimum guarantee. Through the redundant routing in the network topology, completely bypass the C area breakpoint, plan a new end-to-end channel with high bandwidth and priority for emergency communication services. The off-line MEC-C1 originally responsible for unmanned aerial vehicle video AI analysis tasks is dynamically and distributedly migrated to adjacent MEC-D1 and MEC-D2 which are still working normally. In this way, the GPU resources of the two nodes are integrated to form a temporary "distributed AI analysis cluster".

[0191] Further, the difference quantization, At this moment, a large range and severe difference is presented, reflecting the huge gap from the regular operation mode to the extreme emergency mode. A large-scale reconstruction strategy set is generated by PDN, including a series of concurrent instructions: the QoS policy of the whole network is forced to update, and the DSCP mark of emergency services is raised to the highest. Large-scale network routing re-convergence instructions are issued to the core network and SDN controller to force traffic to avoid the failure area. On MEC-D1 and D2, all low-priority containers are forcibly terminated or suspended, releasing their occupied GPU and CPU resources, and immediately deploying AI analysis application containers. Instructions are sent to nearby low-altitude communication platforms (such as tethered unmanned aerial vehicle base stations) or emergency communication vehicles to start and move to the designated location to provide temporary wireless coverage.

[0192] Further, the security inference is performed by using the state evolution predictor. In this emergency scenario, the inference target is not to pursue long-term cost optimization, but to "ensure that the core function is quickly restored and the system does not crash". The inference verifies that the new routing will not cause secondary congestion in other areas, and the integrated resources of MEC-D1 and D2 are sufficient to undertake the migrated key AI tasks without overloading.

[0193] After verification, the execution layer concurrently issues atomic instructions of this batch of reconstruction plan to the SDN controller, cloud platform and wireless network management platform of the whole network with the highest priority. The instructions are marked as "emergency" to bypass the regular change window and process.

[0194] During the emergency response, the frequency of the G-CDT cycle is automatically increased to the highest (such as once per second). It continuously adjusts the computing power and network resources according to the changing disaster situation (such as new interruption reports, position movement of rescue teams, power return of drones) to provide a tough, intelligent and adaptive neural network hub for the whole city's emergency command. For example, when a drone needs to return due to low power, the system will automatically migrate its video analysis task to the MEC node connected to the drone that replaces it. The whole process is transparent to the upper layer application and has no interruption.

[0195] Through the above three examples covering daily optimization, planned tasks and emergency, it can be seen that the system and method proposed by the present application has strong ability to handle complex optimization problems in different scenarios with its unique generative correction digital twin core. The network optimization method and system based on generative correction digital twin provided by the present application ensures the intelligence, foresight, safety and adaptability of the decision, and provides a solid, complete and highly innovative technical solution for realizing the autonomous intelligent operation of the next generation computing power network.

[0196] The above is only a preferred embodiment of the present application and is not used to limit the present application. For those skilled in the art, the present application can have various modifications and changes. Any modification, equivalent replacement, improvement, etc. within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A network optimization method based on generative corrective digital twins, characterized in that, include: S1. Construct a unified spatiotemporal digital twin model based on the target physical entity; S2. Obtain a physical snapshot instance of the unified spatiotemporal digital twin model. The physical snapshot instance Used to characterize the network operating state of the unified spatiotemporal digital twin model at the current moment; S3. Generate the physical snapshot instance based on the preset operation strategy. The condition vector; S4. Input the conditional vector into the generative large model and obtain the ideal state instance output by the generative large model. The generative large model is a conditional spatiotemporal diffusion model based on the U-Net architecture; S5. Calculate the ideal state instance. and the physical snapshot instance The weighted difference tensor between them; S6. Derive the network based on the physical snapshot instance using the correction strategy. The weighted difference tensor generates a correction strategy graph, including: based on the physical snapshot instance. A multimodal attribute graph is constructed, wherein the parameters of the multimodal attribute graph include nodes, edges, node features, and edge features; the nodes are used to represent computing power centers, the edges are used to represent network connections, and the node features and edge features are based on the physical snapshot instance. Generate; input the multimodal attribute graph and the weighted difference tensor into the correction policy derivation network, the correction policy derivation network being a graph attention network, and obtain the structured correction policy graph output by the correction policy derivation network, the correction policy graph being a data structure including discrete actions and continuous parameters; S7. Utilize the state evolution predictor based on the physical snapshot instance. The correction strategy diagram is used to generate subsequent simulation examples. The subsequent simulation examples Used to characterize the predictive network's operating state at the next time step; S8. Parse the correction strategy diagram into atomic instructions and execute the atomic instructions to update the unified spatiotemporal digital twin model; S9. Obtain the updated physical snapshot instance. Based on the updated physical snapshot instance With the subsequent simulation examples Generate prediction bias; S10. Optimize the model parameters of the state evolution predictor based on the prediction bias; S11. Repeat S3 to S10 a preset number of times.

2. The network optimization method based on generative corrective digital twins according to claim 1, characterized in that, The generative large model includes an encoder and a decoder, and the encoder includes multiple downsampling modules; Each of the downsampling modules includes two 3D convolutional layers; After each of the 3D convolutional layers, two multi-head self-attention modules are connected in parallel, including a spatial attention module and a temporal attention module.

3. The network optimization method based on generative corrective digital twins according to claim 2, characterized in that, The conditional vector is input into the generative large model, and the ideal state instance output by the generative large model is obtained. ,include: The conditional vector is converted into an embedding vector of a preset dimension; A cross-attention layer is constructed in the upsampling module of the decoder, the cross-attention layer comprising transposed 3D convolution; The embedded vector is upsampled by the transposed 3D convolution, and features of a preset resolution are introduced from the encoder through skip connections to obtain the ideal state instance. .

4. The network optimization method based on generative corrective digital twins according to claim 1, characterized in that, The calculation of the ideal state instance and the physical snapshot instance The weighted difference tensor between them includes: Generate the weight tensor corresponding to the condition vector according to the preset weighting rules; Calculate the ideal state instance With the physical snapshot instance The difference between them; the weight tensor is multiplied element by element by the difference to obtain the weighted difference tensor.

5. The network optimization method based on generative corrective digital twin according to claim 4, characterized in that, The state evolution predictor is based on the physical snapshot instance. The correction strategy diagram is used to generate subsequent simulation examples. ,include: Copy the physical snapshot instance To obtain the current simulation instance ; The correction strategy graph is applied to the current simulation instance using a preset strategy application operator. To obtain hypothetical simulation examples ; Simulation example using the pre-trained state evolution predictor based on the aforementioned assumptions Generate the network states for a future time window to obtain a future state sequence, which includes the subsequent simulation instances. ; The future state sequence is validated according to preset validation rules. If the future state sequence satisfies the predefined constraint pairs, the subsequent simulation instance is obtained. .

6. The network optimization method based on generative corrective digital twin according to claim 5, characterized in that, The step of parsing the correction strategy diagram into atomic instructions and executing the atomic instructions to update the unified spatiotemporal digital twin model includes: Instructions for obtaining the correction strategy diagram; The instructions are broken down into interface calls and / or command sequences using an instruction translator library; Issue and execute the interface calls and / or the command sequence, and obtain the unified spatiotemporal digital twin model for the next time step.

7. The network optimization method based on generative corrective digital twins according to claim 6, characterized in that, The optimization of the model parameters of the state evolution predictor based on the prediction bias includes: The prediction deviation is saved to the database; Within a preset period, the prediction deviation is extracted from the database, and the model parameters of the state evolution predictor are fine-tuned online using the prediction deviation.

8. The network optimization method based on generative corrective digital twin according to claim 3, characterized in that, Also includes: Acquire training state samples, and inject Gaussian noise into the training state samples through a preset number of time steps to obtain noise state samples; The noise state samples are input into the generative large model to train the generative large model to recognize the Gaussian noise.

9. A network optimization system based on generative correction digital twin, characterized in that, The network optimization method based on generative corrective digital twins as described in any one of claims 1 to 8 is characterized by comprising: The model building module is configured to: construct a unified spatiotemporal digital twin model based on the target physical entity, and obtain a physical snapshot instance of the unified spatiotemporal digital twin model. The physical snapshot instance Used to characterize the network operating state of the unified spatiotemporal digital twin model at the current moment; The state prediction module is configured to: generate a condition vector for the physical snapshot instance based on a preset operation strategy; input the condition vector into a spatiotemporal diffusion model; and obtain the ideal state instance output by the spatiotemporal diffusion model. The spatiotemporal diffusion model is a neural network model based on the U-Net architecture; The difference acquisition module is configured to: calculate the ideal state instance. and the physical snapshot instance The weighted difference tensor between them; the network is derived based on the physical snapshot instance using a correction strategy. The weighted difference tensor generates a correction strategy graph, including: based on the physical snapshot instance. A multimodal attribute graph is constructed, wherein the parameters of the multimodal attribute graph include nodes, edges, node features, and edge features; the nodes are used to represent computing power centers, the edges are used to represent network connections, and the node features and edge features are based on the physical snapshot instance. Generate; input the multimodal attribute graph and the weighted difference tensor into the correction policy derivation network, the correction policy derivation network being a graph attention network, and obtain the structured correction policy graph output by the correction policy derivation network, the correction policy graph being a data structure including discrete actions and continuous parameters; The simulation generation module is configured to: utilize the state evolution predictor based on the physical snapshot instance. The correction strategy diagram is used to generate subsequent simulation examples. The subsequent simulation examples Used to characterize the predictive network's operating state at the next time step; The deviation calculation module is configured to: parse the correction strategy diagram into atomic instructions, execute the atomic instructions to update the unified spatiotemporal digital twin model, and obtain the updated physical snapshot instance. Based on the updated physical snapshot instance With the subsequent simulation examples Generate prediction bias; The model optimization module is configured to optimize the model parameters of the state evolution predictor based on the prediction bias. The loop execution module is configured to repeatedly execute the network optimization method based on generative correction digital twins.

Images