A smart city network resource scheduling method based on digital twinning

Abstract

The application relates to the technical field of network resource scheduling, in particular to a smart city network resource scheduling method based on digital twinning, which comprises the following steps: S1, constructing a city digital twinning model; S2, data acquisition and real-time updating; S3, adaptive learning and prediction analysis; S4, modeling of a multi-objective optimization problem; S5, distributed optimization solution; S6, edge cooperation and local scheduling; S7, dynamic strategy execution and feedback; and S8, evaluation and continuous optimization. Through the introduction of online learning and reinforcement learning mechanisms, the system can dynamically adjust the update frequency and granularity of the digital twinning model, avoid response delays caused by fixed update cycles or thresholds, combine real-time streaming anomaly detection and event-driven strategies, actively predict the diffusion trend of abnormal events such as network congestion or device failure, and trigger the dynamic adjustment of resource scheduling strategies in advance.

Description

Technical Field

[0001] This invention relates to the field of network resource scheduling technology, and in particular to a smart city network resource scheduling method based on digital twins. Background Technology

[0002] With the accelerating pace of global urbanization, cities are expanding in size and becoming highly concentrated in population, presenting unprecedented challenges to urban operation and management. Problems such as traffic congestion, energy waste, environmental pollution, and unequal access to public services are becoming increasingly prominent, and traditional urban management models are no longer sufficient to meet the demands of efficient, intelligent, and sustainable development in modern cities.

[0003] Digital twin technology, as one of the key enabling technologies for smart cities, achieves real-time mapping, simulation, and prediction of the city's operational status by constructing a virtual model highly synchronized with the physical city. Digital twins can not only integrate multi-source heterogeneous data but also support complex decision-making processes through high-fidelity modeling and simulation. In smart city network resource scheduling, digital twins provide managers with a visible, measurable, and controllable virtual operating platform, making the scheduling of urban infrastructure resources more precise and efficient. According to the authorization announcement number "CN117912235A", a method and system for processing planning data in smart cities is disclosed, belonging to the field of data processing technology. The method includes: acquiring historical urban planning data; calculating the vehicle attraction volume of all transportation hubs, and designating stops with vehicle attraction volumes exceeding preset volumes as new transportation hubs; predicting the probability of vehicles shifting to different routes and correcting traffic flow; establishing and training a planning decision model; acquiring real-time urban basic data and data transmission latency; calculating the data processing priority for each area; establishing a network resource allocation model and allocating network resources to real-time urban planning data according to processing priorities; and using the planning decision model to perform decision analysis to obtain new transportation hubs and corresponding traffic flow. This scientifically sets planning data and orderly schedules network resources according to regional levels, improving real-time performance, reducing traffic congestion, and enhancing traffic flow.

[0004] Currently, existing smart city network resource scheduling methods mainly suffer from the following problems: 1. The urban environment is highly uncertain and time-varying. Existing digital twin models mostly adopt periodic updates or adjustment mechanisms based on fixed thresholds, which makes it difficult to achieve rapid response to emergencies. 2. Existing scheduling methods often focus on a single optimization objective, ignoring the conflicts and balances between different objectives; 3. Traditional centralized scheduling methods are ill-suited to handle the massive access demands of smart cities. Large amounts of data need to be remotely transmitted to the central cloud for processing, which not only consumes core network bandwidth but also introduces high latency, failing to meet the millisecond-level response requirements of real-time control applications. Summary of the Invention

[0005] The purpose of this invention is to address the aforementioned shortcomings in the existing technology by proposing a smart city network resource scheduling method based on digital twins.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a smart city network resource scheduling method based on digital twins, wherein the smart city network resource scheduling method is as follows: S1. Construct a digital twin model of the city and establish a high-fidelity, multi-dimensional, and real-time updated digital twin model of the city; S2. Data acquisition and real-time updates: Real-time acquisition of urban operation data through IoT devices, sensor networks, and edge nodes; S3, Adaptive Learning and Predictive Analytics, introduces machine learning and deep learning algorithms to perform real-time analysis and prediction of the collected data; S4. Multi-objective optimization problem modeling: Define the multi-objective optimization problem for resource scheduling; S5. Distributed optimization solution: It adopts a distributed computing architecture, decomposes the global optimization problem into multiple sub-problems, and distributes them to different computing nodes for parallel solution, thereby improving computing efficiency and system robustness. S6, Edge Collaboration and Local Scheduling: Deploy a lightweight scheduler at the edge layer to support real-time local decision-making and resource allocation, reduce cloud communication latency, and improve response speed; S7. Dynamic strategy execution and feedback: The optimized resource scheduling strategy is sent to the physical execution unit. S8. Evaluation and continuous optimization: Regularly evaluate scheduling performance.

[0007] Preferably, in S3, the specific processing steps to address the inability to quickly respond to sudden situations by improving the use of periodic updates or fixed threshold mechanisms include: S3.1 Introduce online learning and reinforcement learning mechanisms, and adopt online learning algorithms or model prediction control to dynamically adjust the update frequency of the digital twin model and automatically trigger model updates based on changes in data flow; S3.2 Integrating real-time anomaly detection and prediction, using streaming processing technology combined with deep learning to detect anomalies in real time and predict the spread and impact of sudden events in advance; S3.3 Event-driven adaptive update strategy: Design an event-based triggering mechanism to dynamically adjust the update granularity of the digital twin model, reduce computational overhead and improve response speed.

[0008] Preferably, the method for improving the fidelity decay of the perceptible digital twin model by dynamically adjusting the update frequency according to S3.1 includes: S3.1.1 Define the fidelity decay function F(t). For each key component of the model, including the traffic flow model of a certain intersection and the network load model of a certain area, define one or more metrics, including prediction error, distribution offset, and uncertainty estimation. S3.1.2 Construct a fidelity decay map, calculate F(t) for different regions and subsystems in real time, and form a dynamic, fine-grained heat map that clearly identifies which parts are most in need of updating.

[0009] Preferably, the process of enhancing the reinforcement learning-driven processing according to 3.1 for dynamically adjusting the update frequency includes: S3.2.1 The environmental state perceived by the agent includes the global and local fidelity decay spectrum F(t), the status of available computing resources, the throughput and value density of the real-time data stream, and the feedback on the effect of historical update strategies.

[0010] S3.2.2 The decision made by the intelligent agent is a complex action, which includes updating the target selection, updating the granularity control, and allocating the resource budget; S3.2.3 Design a multi-objective reward function to guide the agent's learning, including positive rewards and negative rewards.

[0011] Preferably, the optimization method that ensures the trade-offs between objectives and improves the applicability of the adjustment strategy in real-world scenarios in S4 includes: S4.1 Dynamic Weights and Preference Learning: Introduces a weight adjustment mechanism based on user preferences or scenario awareness, and dynamically adjusts the weights of multiple objectives according to real-time needs. S4.2. Employ meta-learning or transfer learning optimization algorithms. Use meta-learning to quickly adapt to new scenarios, or use transfer learning to transfer historical optimization experience to new problems, thereby improving the convergence speed and Pareto solution quality of the algorithm. S4.3 Introduce a human-machine collaborative decision-making mechanism, combining expert knowledge base and reinforcement learning, allowing human intervention to adjust and optimize objectives, and achieving a flexible balance among multiple objectives.

[0012] Preferably, the processing steps based on the enhanced capture of complex and diverse requirements in S4.1 include: S4.1.1 Multi-source spatiotemporal scene perception and feature extraction: Utilizing a digital twin model, massive amounts of data from time, space, and event dimensions are fused in real time. A deep learning model is used to extract and fuse features from these multimodal data, outputting a dynamic scene embedding vector that can comprehensively represent the current urban state. S4.1.2 Weighted decision maker based on deep reinforcement learning: The dynamic scene embedding vector mentioned above outputs a continuous action vector, which directly corresponds to the weight values ​​of each objective in the multi-objective optimization problem. The reward function guides the DRL agent to learn how to make trade-offs. S4.1.3 Multi-agent consensus mechanism for conflict resolution: In the digital twin environment, multiple agent agents are deployed, each representing a stakeholder and having its own independent reward function. A consensus module is introduced to merge these different proposals into a final, consensus-based weight vector, which is then input into the global multi-objective optimizer. S4.1.4 Visual Analysis and Human Feedback Loop: Develop a weighted decision dashboard to visually display to administrators the current scene characteristics, the weights of DRL agent decisions, the consensus process, and the expected performance of each target under that weight. This allows administrators to manually override or adjust decisions when they deem them inappropriate.

[0013] Preferably, the method for further optimizing the convergence speed and Pareto front quality based on S4.2 includes: The S4.2.1 hypernetwork-driven meta-optimizer no longer trains or fine-tunes a single optimization algorithm model for each new scenario, but instead trains a hypernetwork. S4.2.2 Scene Memory Pool and Experience Reuse: Construct a scene memory pool to store historical optimization cases; S4.2.3 Closed-loop meta-reinforcement learning training models the entire multi-objective optimization process as a meta-reinforcement learning problem.

[0014] Preferably, the method for achieving millisecond-level real-time requirements in S5 and S6 includes: M1, hierarchical federated learning and edge intelligence, introduce federated learning into a distributed architecture, allowing edge nodes to train models locally and only upload model parameters, reducing cloud load while protecting privacy; M2, a decentralized collaborative scheduling mechanism, uses blockchain smart contracts or distributed consensus algorithms to achieve autonomous negotiation and resource scheduling among edge nodes, reducing cloud center intervention; M3, dynamic task offloading and compute migration, designs a lightweight decision-maker that dynamically determines the task allocation path between cloud, edge, and device based on network status and compute load, achieving low-latency response.

[0015] Preferably, the optimization steps to further improve the response speed of federated learning in M1 include: M1.1 Asynchronous Federated Learning: The global model in the cloud center serves as a continuously updated model pool. After any edge node completes its local training, it immediately uploads its model incrementally to the cloud without waiting for other nodes. Upon receiving an update from any node, the cloud immediately merges it asynchronously into the global model and quickly distributes the updated global model to idle or relevant edge nodes. M1.2 Incremental learning and personalized fine-tuning: After receiving the global model, each edge node will continuously use the latest local real-time data stream for incremental learning, so that the model can quickly adapt to the unique local environment. The node locally stores a very small personalized adaptation layer to fine-tune for local features, forming a hybrid model of global knowledge + local characteristics. M1.3, cross-layer knowledge distillation: The cloud center uses its powerful computing capabilities and global data to train a large, high-precision teacher model. Through knowledge distillation technology, the knowledge of the teacher model is compressed and transferred to a small student model, which is then directly deployed on edge devices or even terminal IoT devices. M1.4, Feedback-based elastic aggregation weight adjustment: After the cloud distributes a new model, it will monitor the feedback performance of each edge node after applying the model. For nodes with significant performance improvement, it is considered that their local data quality is high and representative, and they will be given higher weights in the next federated aggregation. For nodes with poor data quality or encountering abnormal situations, their weights will be automatically reduced, or they may even be temporarily excluded from aggregation.

[0016] Preferably, in M2, to further meet the millisecond-level real-time requirements, the improved edge collaboration mechanism processing steps include: M2.1, Local consensus domain based on resource groups: The edge nodes are dynamically divided into multiple resource groups according to the geographical or functional relevance of smart city services. M2.2 Within the resource group, a fast voting mechanism based on token ring or time slice is adopted; M2.3. For critical requests at the millisecond or even microsecond level, establish a high-priority channel; M2.4. Use blockchain only as an audit and arbitration layer, rather than a real-time decision-making layer.

[0017] The design scheme proposed in this invention has the following beneficial effects during application: 1. By introducing online learning and reinforcement learning mechanisms, this solution enables the system to dynamically adjust the update frequency and granularity of the digital twin model, avoiding response delays caused by fixed update cycles or thresholds. Combined with real-time streaming anomaly detection and event-driven strategies, the system can proactively predict the spread of abnormal events such as network congestion or equipment failure, and trigger dynamic adjustments to resource scheduling strategies in advance. This allows urban infrastructure to quickly adapt to changes in scenarios such as large-scale events or extreme weather, effectively reducing the risk of service interruption and ensuring continuous and stable system operation. 2. The system adopts a dynamic weight adjustment mechanism and deep reinforcement learning technology to automatically generate weight values ​​for each optimization objective based on real-time scene characteristics. This allows the system to seek the best balance between conflicting objectives such as minimizing energy consumption and maximizing service quality. By introducing a multi-agent consensus mechanism and a human-machine collaborative decision-making loop, the system can integrate the preferences of different stakeholders and also allows administrators to intervene and adjust the system. This enhances the applicability and interpretability of the scheduling strategy in real-world complex environments and makes the decision-making process more in line with diverse needs. 3. By adopting a distributed computing model, the global optimization problem is decomposed into parallel solutions on cloud centers and edge servers, reducing the load on central nodes. A lightweight scheduler is deployed at the edge layer to support local real-time decision-making such as traffic signal control, reducing cloud communication overhead. Combined with federated learning and knowledge distillation techniques, local training and personalized fine-tuning of models are achieved at edge nodes, further improving response speed while protecting data privacy and meeting millisecond-level real-time scheduling requirements. 4. By monitoring the execution effect of resource scheduling strategies in real time and re-inputting performance feedback data into the digital twin model, the system can periodically evaluate indicators such as energy efficiency, latency, and reliability. By utilizing historical optimization experience bases and meta-learning technology, the system continuously adjusts algorithm parameters and model structure, enabling the scheduling method to have the ability to learn online and continuously improve. Ultimately, this achieves performance evolution and stability improvement of the smart city system in long-term operation. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the smart city network resource scheduling method of the present invention; Figure 2 This is a flowchart of the agent update decision process based on S3.1 according to the present invention. Detailed Implementation

[0019] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Example

[0020] Reference Figures 1-2 A smart city network resource scheduling method based on digital twins is proposed, as follows: S1. Construct a digital twin model of the city and establish a high-fidelity, multi-dimensional, and real-time updated digital twin model of the city; Integrate real-time and historical data from urban infrastructure such as transportation, energy, communication, and environment to form a virtual mapping system; S2. Data acquisition and real-time updates: Real-time acquisition of urban operation data through IoT devices, sensor networks, and edge nodes; The data includes traffic flow, equipment status, network load, and environmental information, and the digital twin model is dynamically updated to ensure it is synchronized with the physical city. S3, Adaptive Learning and Predictive Analytics, introduces machine learning and deep learning algorithms to perform real-time analysis and prediction of the collected data; The predictions include: predicting abnormal situations such as network congestion and equipment failure; Dynamically adjust resource allocation strategies to respond to urban events, such as large-scale events and extreme weather; S4. Multi-objective optimization problem modeling: Define the multi-objective optimization problem for resource scheduling; The objectives include: minimizing energy consumption, minimizing latency, maximizing service quality, and maximizing resource utilization. Modeling is performed using multi-objective optimization algorithms, such as NSGA-II or MOEA / D. S5. Distributed optimization solution: It adopts a distributed computing architecture, decomposes the global optimization problem into multiple sub-problems, and distributes them to different computing nodes for parallel solution, thereby improving computing efficiency and system robustness. The nodes include cloud centers and edge servers; S6, Edge Collaboration and Local Scheduling: Deploy a lightweight scheduler at the edge layer to support real-time local decision-making and resource allocation, reduce cloud communication latency, and improve response speed; The allocation types include traffic signal control and street light regulation; S7. Dynamic strategy execution and feedback: The optimized resource scheduling strategy is sent to the physical execution unit. The execution unit includes servers, base stations, and smart devices, and monitors the execution effect in real time to form a closed-loop feedback and continuously adjust the strategy. S8. Evaluation and continuous optimization: Regularly evaluate scheduling performance; Performance includes energy efficiency, latency, and reliability. By combining digital twin models for simulation testing and strategy iteration, algorithm parameters and model structure are continuously optimized to achieve system self-evolution.

[0021] In S3, the specific handling of the inability to quickly respond to sudden situations is addressed by improving the use of periodic updates or fixed threshold mechanisms as follows: S3.1 Introduce online learning and reinforcement learning mechanisms, and adopt online learning algorithms or model prediction control to dynamically adjust the update frequency of the digital twin model and automatically trigger model updates based on changes in data flow; 1) Specifically, the trigger condition is U, defining a data change indicator. Examples of parameters include KL divergence, mean shift, and prediction error variation, with thresholds set. , in: This is an indicator of the degree of data change at time t; The absolute change threshold; The relative change threshold; 2) Specifically, for a precision of A, reinforcement learning is used to learn the optimal update interval. In other words, how often to update to maximize long-term benefits: , in: For the benefit function, a trade-off is made between accuracy improvement and update costs; λ is the cost weighting coefficient; α is the learning rate; β is the penalty term; 3) At each time step, optimize the update strategy for the next N steps: , , in: =1 indicates that the model is updated at time t; This represents the prediction error; λ is the update cost weight.

[0022] S3.2 Integrating real-time anomaly detection and prediction, using streaming processing technology combined with deep learning to detect anomalies in real time and predict the spread and impact of sudden events in advance; S3.3 Event-driven adaptive update strategy: Design an event-based triggering mechanism to dynamically adjust the update granularity of the digital twin model, reduce computational overhead and improve response speed.

[0023] According to S3.1, the method for dynamically adjusting the update frequency to improve the perceived digital twin model fidelity degradation is as follows: S3.1.1 Define the fidelity decay function F(t). For each key component of the model, including the traffic flow model of a certain intersection and the network load model of a certain area, define one or more metrics, including prediction error, distribution offset, and uncertainty estimation. Specifically: traffic is m, network load is i, and fidelity decay function: , in: The model fidelity index at time t typically decays from 1 (perfectly faithful to reality) to 0 (completely unrealistic, completely unrelated to reality). The higher the value, the more reliable the model. 1) Traffic flow model fidelity components: ; Traffic flow attenuation coefficient, a constant greater than 0, represents the rate of distortion of the traffic flow model at this intersection. The larger the coefficient, the faster the attenuation. This coefficient depends on the complexity of the model and its dependence on real-time data. Traffic flow weights represent the importance of the traffic flow model in the overall system fidelity. ; 2) Network load model fidelity components: ; This is the network load attenuation coefficient, a constant greater than 0, representing the distortion rate of the regional network load model. Typically, the network state changes faster, so this coefficient may be larger. The network load weight represents the importance of the network load model in the overall system fidelity. 3) The accelerating effect of external environmental changes on model decay rate: ; This is the environmental disturbance index at time t. This can be a composite index, such as sudden changes in traffic demand, abrupt changes in road capacity, or abnormal peaks in network traffic. For environmental baseline or disturbance threshold, a constant used for normalization, representing the normal fluctuation range of the system; α is the perturbation sensitivity coefficient, a constant greater than 0, which represents the sensitivity of the entire model to external environmental perturbations. The larger α is, the greater the destructive impact of external events on the model's fidelity.

[0024] S3.1.2 Construct a fidelity attenuation map and calculate the fidelity attenuation of models in different regions and subsystems in real time. This creates a dynamic, fine-grained heatmap that clearly identifies which parts most urgently need updating.

[0025] Based on the dynamic adjustment of the update frequency in S3.1, the reinforcement learning-driven processing of the augmentation meta-method is as follows: S3.2.1 The environmental state perceived by the agent, including global and local fidelity degradation maps. Available computing resource status, throughput and value density of real-time data streams, and feedback on the effectiveness of historical update strategies.

[0026] S3.2.2 The decision made by the intelligent agent is a complex action, which includes updating the target selection, updating the granularity control, and allocating the resource budget; S3.2.3 Design a multi-objective reward function to guide the agent's learning, including positive rewards and negative rewards; Positive rewards include improved prediction accuracy of the updated model, successful prediction of sudden events, and improved service quality. Negative rewards: The computational resources and communication bandwidth consumed by update operations, as well as the potential risks of model unavailability during the update period.

[0027] In S4, the following optimizations are made to ensure the trade-offs between objectives and improve the applicability of the adjustment strategy in real-world scenarios: S4.1 Dynamic Weights and Preference Learning: Introduces a weight adjustment mechanism based on user preferences or scenario awareness, and dynamically adjusts the weights of multiple objectives according to real-time needs. Preference score Calculation: ; in: It is the preference score of target i at time t, and the softmax function is used to ensure weight normalization; This refers to the current system state, such as sensor and environmental variables; This is a user preference vector, obtained through interaction, historical behavior, or explicit feedback; For scene context, such as time, location, and device status; It is a function used to calculate the preference score for target i, designed as a linear model, neural network or other learning model.

[0028] S4.2. Employ meta-learning or transfer learning optimization algorithms. Use meta-learning to quickly adapt to new scenarios, or use transfer learning to transfer historical optimization experience to new problems, thereby improving the convergence speed and Pareto solution quality of the algorithm. S4.3 Introduce a human-machine collaborative decision-making mechanism, combining expert knowledge base and reinforcement learning, allowing human intervention to adjust and optimize objectives, and achieving a flexible balance among multiple objectives.

[0029] The processing of enhanced capture of complex and diverse requirements in S4.1 is as follows: S4.1.1 Multi-source spatiotemporal scene perception and feature extraction: Utilizing a digital twin model, massive amounts of data from time, space, and event dimensions are fused in real time. A deep learning model is used to extract and fuse features from these multimodal data, outputting a dynamic scene embedding vector that can comprehensively represent the current urban state. S4.1.2 Weighted decision maker based on deep reinforcement learning: The dynamic scene embedding vector mentioned above outputs a continuous action vector, which directly corresponds to the weight values ​​of each objective in the multi-objective optimization problem. The reward function guides the DRL agent to learn how to make trade-offs. The reward consists of two parts: Global utility reward: Substitute the weights output by DRL into the multi-objective optimization algorithm of S4 to solve for a set of solutions on the Pareto front. Select one of them and calculate the comprehensive performance score of the scheme after simulation in a physical city or digital twin model. Constraint, penalty, and reward: Significantly negatively reward behaviors that violate hard constraints such as SLAs and resource capacity limits; S4.1.3 Multi-agent consensus mechanism for conflict resolution: In the digital twin environment, multiple agent agents are deployed, each representing a stakeholder and having its own independent reward function. A consensus module is introduced to merge these different proposals into a final, consensus-based weight vector, which is then input into the global multi-objective optimizer. S4.1.4 Visual Analysis and Human Feedback Loop: Develop a weighted decision dashboard to visually display to administrators the current scene characteristics, the weights of DRL agent decisions, the consensus process, and the expected performance of each target under that weight. This allows administrators to manually override or adjust decisions when they deem them inappropriate.

[0030] The following method further optimizes the convergence speed and Pareto front quality based on S4.2: The S4.2.1 hypernetwork-driven meta-optimizer no longer trains or fine-tunes a single optimization algorithm model for each new scenario, but instead trains a hypernetwork. The input to the hypernetwork is the real-time feature vector of the current scene, such as traffic flow patterns, network load distribution, event types, and environmental data. The output is a set of dynamic and optimal algorithm parameters generated for the target multi-objective optimization algorithm, such as crossover rate, mutation rate, and selection pressure. S4.2.2 Scene Memory Pool and Experience Reuse: Construct a scene memory pool to store historical optimization cases; Each case is a key-value pair: Key: Encoded scene feature vector; Value: The high-quality Pareto solution set obtained in this scenario and the algorithm parameter configuration used when the solution was successfully obtained; Operation process: Retrieval: When a new scene appears, the system calculates its scene feature vector in real time and retrieves the K most similar historical scenes from the memory pool; Migration: Inject the Pareto solution sets corresponding to these similar scenarios into the current optimization algorithm as a high-quality initial population, instead of randomly initializing from scratch; Configuration: Use the successful algorithm parameter configurations in similar scenarios as a reference for the hypernetwork initialization or directly as the baseline configuration; S4.2.3 Closed-loop meta-reinforcement learning training models the entire multi-objective optimization process as a meta-reinforcement learning problem; Intelligent agents: Hypernetworks; Action: Parameter configuration generated for optimizing the algorithm; Status: The current optimization process status fed back by the digital twin model, such as population diversity, generational improvement rate, and degree of constraint violation.

[0031] Rewards: Based on a comprehensive calculation of the optimization results, including indicators such as hypervolume, spacing, and distribution, actions that can quickly generate high-quality, highly diverse Pareto fronts are rewarded.

[0032] The following methods can be used to achieve millisecond-level real-time requirements in S5 and S6: M1, hierarchical federated learning and edge intelligence, introduce federated learning into a distributed architecture, allowing edge nodes to train models locally and only upload model parameters, reducing cloud load while protecting privacy; M2, a decentralized collaborative scheduling mechanism, uses blockchain smart contracts or distributed consensus algorithms to achieve autonomous negotiation and resource scheduling among edge nodes, reducing cloud center intervention; , in: The number of scheduled tasks completed through direct negotiation between local edge nodes; The number of scheduling tasks completed through blockchain smart contracts; This represents the total number of scheduled tasks. The self-regulation rate represents the proportion of scheduling tasks that do not require intervention from the cloud center. The higher the rate, the higher the degree of decentralization of the system and the less intervention from the cloud center. The range of the self-regulation rate is [0,1]. When it is 1, all scheduling is completed autonomously by the edge nodes. When it is 0, all scheduling depends on the cloud center.

[0033] M3, dynamic task offloading and compute migration, designs a lightweight decision-maker that dynamically determines the task allocation path between cloud, edge, and device based on network status and compute load, achieving low-latency response.

[0034] The optimization steps to further improve the response speed of federated learning in M1 are as follows: M1.1 Asynchronous Federated Learning: The global model in the cloud center serves as a continuously updated model pool. After any edge node completes its local training, it immediately uploads its model incrementally to the cloud without waiting for other nodes. Upon receiving an update from any node, the cloud immediately merges it asynchronously into the global model and quickly distributes the updated global model to idle or relevant edge nodes. M1.2 Incremental learning and personalized fine-tuning: After receiving the global model, each edge node will continuously use the latest local real-time data stream for incremental learning, so that the model can quickly adapt to the unique local environment. The node locally stores a very small personalized adaptation layer to fine-tune for local features, forming a hybrid model of global knowledge + local characteristics. M1.3, cross-layer knowledge distillation: The cloud center uses its powerful computing capabilities and global data to train a large, high-precision teacher model. Through knowledge distillation technology, the knowledge of the teacher model is compressed and transferred to a small student model, which is then directly deployed on edge devices or even terminal IoT devices. M1.4, Feedback-based elastic aggregation weight adjustment: After the cloud distributes a new model, it will monitor the feedback performance of each edge node after applying the model. For nodes with significant performance improvement, it is considered that their local data quality is high and representative, and they will be given higher weights in the next federated aggregation. For nodes with poor data quality or encountering abnormal situations, their weights will be automatically reduced, or they may even be temporarily excluded from aggregation.

[0035] In M2, to further meet the millisecond-level real-time requirements, the edge collaboration mechanism is improved as follows: M2.1, Local consensus domain based on resource groups: The edge nodes are dynamically divided into multiple resource groups according to the geographical or functional relevance of smart city services. M2.2 Within the resource group, a fast voting mechanism based on token ring or time slice is adopted; M2.3. For critical requests at the millisecond or even microsecond level, establish a high-priority channel; M2.4. Use blockchain only as an audit and arbitration layer, rather than a real-time decision-making layer.

[0036] Example 2 Project: Intelligent transportation and communication coordination and scheduling during large-scale events Application Background: A large concert is scheduled to be held at a stadium in the city center, with an estimated 50,000 people gathering. Historical data shows that such events can cause severe traffic congestion in the surrounding area, overload of communication base stations, and a decline in the quality of public network services. Implementation process: S1 & S2: Integrate traffic cameras, base station load sensors, ticketing system data, and real-time crowd monitoring equipment data to construct a digital twin of the area surrounding the concert, including traffic flow model, communication load model, and environmental model; S3: Use LSTM networks to predict traffic congestion points and base station overload risks in the next hour (S3.2). The model update frequency is dynamically adjusted through meta-reinforcement learning (S3.2.1–S3.2.3): when a sudden increase in pedestrian flow is detected, the update granularity of the traffic and communication model is automatically increased from once every 5 minutes to once every 30 seconds; S4: Minimize traffic delay, maximize base station throughput, and minimize energy consumption; The NSGA-II algorithm is used, and the target weights are dynamically adjusted through a deep reinforcement learning weight decision-maker (S4.1.2): communication quality is prioritized at the start of the event, and traffic management is prioritized at the end. S5 & S6: The cloud center coordinates the overall resource allocation, and the edge nodes perform local scheduling: real-time green wave control of traffic signals (S6), dynamic bandwidth allocation of base stations (S5), and asynchronous aggregation of models from each edge node through federated learning (M1.1) to ensure the real-time nature of local decision-making; S7 & S8: The strategy is distributed to smart traffic lights and base station controllers to monitor traffic flow and network throughput in real time, adjust signal timing and bandwidth allocation in a closed loop, evaluate overall performance after the activity, and iteratively optimize model parameters. Effect: Traffic congestion was reduced by 35%, network latency was reduced by 40%, and base station downtime was reduced by 50%. By dynamically adjusting weights and using edge collaboration, the system effectively addresses sudden surges in crowds during the event's conclusion.

[0037] Example 3 Project: Resilient Dispatch of Urban Energy and Communications under Extreme Weather Application Background: A typhoon warning has been issued, predicting that the city will be affected within the next 6 hours. Strong winds and heavy rain may cause local power outages, communication disruptions, and traffic paralysis. Implementation process S1 & S2: Integrate real-time typhoon path, wind force, and rainfall data from the meteorological bureau, power grid load data, base station backup power status, and traffic monitoring data to construct a digital twin for disaster emergency response; S3: Use streaming anomaly detection (S3.2) to identify power grid fault risk points and base station power outage risk; Based on event-driven updates (S3.3), each update to the typhoon path triggers a global model update. S4: To maximize power grid coverage, minimize communication interruption time, and ensure smooth emergency transportation, the MOEA / D algorithm is used to quickly generate Pareto solutions by transferring and optimizing parameters from historical typhoon data through meta-learning (S4.2). S5 & S6: The cloud center coordinates resources across the city, edge nodes make autonomous decisions, microgrids dynamically switch loads (S6), base stations switch to backup power or satellite links (M2.2), and trusted collaboration between edge nodes is achieved through blockchain smart contracts (M2). S7 & S8: Distribute scheduling strategies to smart grid switches and base station power controllers, monitor power supply or communication status in real time, dynamically adjust resource allocation, assess system resilience after disasters, and optimize model and algorithm parameters. Effect The area affected by power outages has been reduced by 20%, communication in the core area remains uninterrupted, and emergency access channels remain open. Through edge collaboration and blockchain mechanisms, local autonomous decision-making can be maintained even when the cloud center is disconnected.

[0038] In practice This solution uses a high-fidelity, multi-dimensional, and real-time updated digital twin model of the city as its core foundation to achieve deep mapping and synchronization between the physical city and the virtual space. Its working principle begins with the comprehensive integration of massive real-time and historical data of urban infrastructure such as transportation, energy, communication, and environment to form a virtual system that can accurately reflect the city's operating status. This model is not static, but continuously collects data through IoT devices, sensor networks, and edge nodes throughout the city and dynamically updates itself to ensure that it always remains consistent with the physical world, providing a real and reliable digital environment for all subsequent analysis, prediction, and decision-making. Employing advanced adaptive learning and predictive analysis mechanisms, and building upon these, a complex multi-objective optimization and distributed solution framework is constructed. The system introduces machine learning and deep learning algorithms to analyze real-time data streams, enabling it not only to predict anomalies such as network congestion and equipment failures, but also to pre-adjust strategies for events such as large-scale activities or extreme weather. More importantly, it precisely defines the resource scheduling problem as a multi-objective optimization problem that includes minimizing energy consumption and latency, and maximizing service quality and resource utilization. It employs advanced algorithms such as NSGA-II for solving the problem. To address the massive computational challenges of city-scale systems, it adopts a distributed computing architecture, decomposing the global problem and solving it in parallel on cloud centers and edge servers, greatly improving efficiency and system robustness. At the network edge layer close to the data source, a lightweight scheduler is deployed, enabling it to make local real-time decisions and allocate resources for tasks such as traffic signal control and street light adjustment. This effectively avoids the latency caused by transmitting all data back to the cloud. The optimized strategies are quickly distributed to physical execution units such as base stations and servers. The system also monitors the execution effect in real time and sends the monitoring results back to the decision center as feedback information, thus forming a continuous "perception-decision-execution-feedback" closed loop, which allows the scheduling strategy to be dynamically adjusted and continuously optimized according to the actual effect. It regularly conducts comprehensive evaluations of key performance indicators such as energy efficiency, latency, and reliability, and utilizes the digital twin model as an excellent "testing ground" for simulation testing and strategy iteration. Its evolutionary capabilities are deeply integrated into multiple stages: at the learning and prediction layer, it dynamically adjusts the model update strategy through meta-reinforcement learning; at the multi-objective optimization layer, it uses meta-learning and scenario memory pools to quickly adapt to new situations, improving the quality of solutions and convergence speed; at the distributed architecture layer, it continuously optimizes collaborative efficiency through mechanisms such as asynchronous federated learning, elastic aggregation, and local consensus.

[0039] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A smart city network resource scheduling method based on digital twins, characterized in that: The smart city network resource scheduling method includes: S1. Construct a digital twin model of the city and establish a high-fidelity, multi-dimensional, and real-time updated digital twin model of the city; S2. Data acquisition and real-time updates: Real-time acquisition of urban operation data through IoT devices, sensor networks, and edge nodes; S3, Adaptive Learning and Predictive Analytics, introduces machine learning and deep learning algorithms to perform real-time analysis and prediction of the collected data; S4. Multi-objective optimization problem modeling: Define the multi-objective optimization problem for resource scheduling; S5. Distributed optimization solution: It adopts a distributed computing architecture, decomposes the global optimization problem into multiple sub-problems, and distributes them to different computing nodes for parallel solution, thereby improving computing efficiency and system robustness. S6, Edge Collaboration and Local Scheduling: Deploy a lightweight scheduler at the edge layer to support real-time local decision-making and resource allocation, reduce cloud communication latency, and improve response speed; S7. Dynamic strategy execution and feedback: The optimized resource scheduling strategy is sent to the physical execution unit. S8. Evaluation and continuous optimization: Regularly evaluate scheduling performance.

2. The method for scheduling smart city network resources based on digital twins according to claim 1, characterized in that: In S3, the specific steps for addressing the inability to quickly respond to sudden situations by improving the use of periodic updates or fixed threshold mechanisms include: S3.1 Introduce online learning and reinforcement learning mechanisms, and adopt online learning algorithms or model prediction control to dynamically adjust the update frequency of the digital twin model and automatically trigger model updates based on changes in data flow; S3.2 Integrating real-time anomaly detection and prediction, using streaming processing technology combined with deep learning to detect anomalies in real time and predict the spread and impact of sudden events in advance; S3.3 Event-driven adaptive update strategy: Design an event-based triggering mechanism to dynamically adjust the update granularity of the digital twin model, reduce computational overhead and improve response speed.

3. The method for scheduling smart city network resources based on digital twins according to claim 2, characterized in that: The method described in S3.1 for dynamically adjusting the update frequency to improve the fidelity decay of the perceptible digital twin model includes: S3.1.1 Define the fidelity decay function F(t). For each key component of the model, including the traffic flow model of a certain intersection and the network load model of a certain area, define one or more metrics, including prediction error, distribution offset, and uncertainty estimation. S3.1.2 Construct a fidelity decay map, calculate F(t) for different regions and subsystems in real time, and form a dynamic, fine-grained heat map that clearly identifies which parts are most in need of updating.

4. The smart city network resource scheduling method based on digital twins according to claim 3, characterized in that: The reinforcement learning-driven processing procedure for dynamically adjusting the update frequency, as described in section 3.1, includes: S3.2.1 The environmental state perceived by the agent, including the global and local fidelity decay spectrum F(t), the state of available computing resources, the throughput and value density of the real-time data stream, and the feedback on the effect of historical update strategies; S3.2.2 The decision made by the intelligent agent is a complex action, which includes updating the target selection, updating the granularity control, and allocating the resource budget; S3.2.3 Design a multi-objective reward function to guide the agent's learning, including positive rewards and negative rewards.

5. A smart city network resource scheduling method based on digital twins according to claim 4, characterized in that: The optimization method described in S4 that ensures the trade-offs between objectives and improves the applicability of the adjustment strategy in real-world scenarios includes: S4.1 Dynamic Weights and Preference Learning: Introduces a weight adjustment mechanism based on user preferences or scenario awareness, and dynamically adjusts the weights of multiple objectives according to real-time needs. S4.

2. Employ meta-learning or transfer learning optimization algorithms. Use meta-learning to quickly adapt to new scenarios, or use transfer learning to transfer historical optimization experience to new problems, thereby improving the convergence speed and Pareto solution quality of the algorithm. S4.3 Introduce a human-machine collaborative decision-making mechanism, combining expert knowledge base and reinforcement learning, allowing human intervention to adjust and optimize objectives, and achieving a flexible balance among multiple objectives.

6. The method for scheduling smart city network resources based on digital twins according to claim 5, characterized in that: The processing steps based on the enhanced capture of complex and diverse requirements in S4.1 include: S4.1.1 Multi-source spatiotemporal scene perception and feature extraction: Utilizing a digital twin model, massive amounts of data from time, space, and event dimensions are fused in real time. A deep learning model is used to extract and fuse features from these multimodal data, outputting a dynamic scene embedding vector that can comprehensively represent the current urban state. S4.1.2 Weighted decision maker based on deep reinforcement learning: The dynamic scene embedding vector mentioned above outputs a continuous action vector, which directly corresponds to the weight values ​​of each objective in the multi-objective optimization problem. The reward function guides the DRL agent to learn how to make trade-offs. S4.1.3 Multi-agent consensus mechanism for conflict resolution: In the digital twin environment, multiple agent agents are deployed, each representing a stakeholder and having its own independent reward function. A consensus module is introduced to merge these different proposals into a final, consensus-based weight vector, which is then input into the global multi-objective optimizer. S4.1.4 Visual Analysis and Human Feedback Loop: Develop a weighted decision dashboard to visually display to administrators the current scene characteristics, the weights of DRL agent decisions, the consensus process, and the expected performance of each target under that weight. This allows administrators to manually override or adjust decisions when they deem them inappropriate.

7. A smart city network resource scheduling method based on digital twins according to claim 6, characterized in that: The method for further optimizing convergence speed and Pareto front quality based on S4.2 includes: The S4.2.1 hypernetwork-driven meta-optimizer no longer trains or fine-tunes a single optimization algorithm model for each new scenario, but instead trains a hypernetwork. S4.2.2 Scene Memory Pool and Experience Reuse: Construct a scene memory pool to store historical optimization cases; S4.2.3 Closed-loop meta-reinforcement learning training models the entire multi-objective optimization process as a meta-reinforcement learning problem.

8. A smart city network resource scheduling method based on digital twins according to claim 7, characterized in that: The methods for achieving millisecond-level real-time requirements in S5 and S6 include: M1, hierarchical federated learning and edge intelligence, introduce federated learning into a distributed architecture, allowing edge nodes to train models locally and only upload model parameters, reducing cloud load while protecting privacy; M2, a decentralized collaborative scheduling mechanism, uses blockchain smart contracts or distributed consensus algorithms to achieve autonomous negotiation and resource scheduling among edge nodes, reducing cloud center intervention; M3, dynamic task offloading and compute migration, designs a lightweight decision-maker that dynamically determines the task allocation path between cloud, edge, and device based on network status and compute load, achieving low-latency response.

9. A smart city network resource scheduling method based on digital twins according to claim 8, characterized in that: The optimization steps to further improve the response speed of federated learning in M1 include: M1.1 Asynchronous Federated Learning: The global model in the cloud center serves as a continuously updated model pool. After any edge node completes its local training, it immediately uploads its model incrementally to the cloud without waiting for other nodes. Upon receiving an update from any node, the cloud immediately merges it asynchronously into the global model and quickly distributes the updated global model to idle or relevant edge nodes. M1.2 Incremental learning and personalized fine-tuning: After receiving the global model, each edge node will continuously use the latest local real-time data stream for incremental learning, so that the model can quickly adapt to the unique local environment. The node locally stores a very small personalized adaptation layer to fine-tune for local features, forming a hybrid model of global knowledge + local characteristics. M1.3, cross-layer knowledge distillation: The cloud center uses its powerful computing capabilities and global data to train a large, high-precision teacher model. Through knowledge distillation technology, the knowledge of the teacher model is compressed and transferred to a small student model, which is then directly deployed on edge devices or even terminal IoT devices. M1.4, Feedback-based elastic aggregation weight adjustment: After the cloud distributes a new model, it will monitor the feedback performance of each edge node after applying the model. For nodes with significant performance improvement, it is considered that their local data quality is high and representative, and they will be given higher weights in the next federated aggregation. For nodes with poor data quality or encountering abnormal situations, their weights will be automatically reduced, or they may even be temporarily excluded from aggregation.

10. A smart city network resource scheduling method based on digital twins according to claim 9, characterized in that: In order to further meet the millisecond-level real-time requirements in M2, the improved edge collaboration mechanism processing steps include: M2.1, Local consensus domain based on resource groups: The edge nodes are dynamically divided into multiple resource groups according to the geographical or functional relevance of smart city services. M2.2 Within the resource group, a fast voting mechanism based on token ring or time slice is adopted; M2.

3. For critical requests at the millisecond or even microsecond level, establish a high-priority channel; M2.

4. Use blockchain only as an audit and arbitration layer, rather than a real-time decision-making layer.

Images