Smart city resource allocation method and system based on end-cloud collaboration

Abstract

The invention discloses a smart city resource allocation method and system based on end-cloud collaboration. The method comprises the steps of collecting urban multi-dimensional resource state data in real time through an Internet of Things sensor, and performing standardization processing; edge calculation is executed on the end side, noise is filtered, and a resource load feature vector is extracted; uploading the data to a cloud for multi-source data fusion and association analysis to obtain a resource interaction relationship quantitative index; a resource multi-dimensional situation map model is constructed, dynamic adjustment is triggered according to the change rate, and resource availability distribution is judged; simulating a resource flow path by using a graph neural network, and generating an optimal allocation scheme draft; issuing a control instruction to an end side for execution, and determining a final plan according to feedback iterative updating; and continuous correction is performed through a monitoring feedback mechanism. The system comprises corresponding modules and realizes intelligent resource allocation. According to the method, the real-time performance and the self-adaptability of resource allocation are improved, and the city management efficiency is optimized.

Description

Technical Field

[0001] This application relates to the field of urban resource allocation technology, and in particular to a smart city resource allocation method and system based on edge-cloud collaboration. Background Technology

[0002] Smart city resource allocation is a crucial area of ​​urban management, aiming to optimize the distribution and use of resources such as energy, transportation, healthcare, and education through intelligent means to improve urban operational efficiency and residents' quality of life. With accelerating urbanization and increasingly complex resource demands, efficient allocation has become a vital guarantee for sustainable urban development.

[0003] Chinese patent, publication number: CN120579803A, publication date: September 2, 2025, discloses a resilient urban resource allocation method and system based on blockchain and edge computing, involving the field of urban resource allocation technology. It utilizes a federated protocol to exchange sheared gradient summaries to synchronize model versions, then writes the predicted summary containing fingerprints into a permissioned chain and solidifies the timing through Byzantine consensus. On-chain smart contracts generate allocation intentions with verifiable paths based on sliding window prediction and adaptive threshold evaluation. These intentions are delivered to the execution subsystem with low latency via a sidechain message bus, and cross-departmental resource rearrangement is completed by an adaptive decision driver through rolling optimization. Execution receipts are mapped as incremental patches and written back to the local state tensor, dynamically adjusting node trust weights. When the coverage and weight increase meet the trigger strategy, differential shearing training is initiated, generating a new structural fingerprint and model signature. This ensures efficient, reliable, dynamic allocation of urban public resources and sustainable learning and evolution in the event of emergencies.

[0004] The shortcomings of the above technical solutions are: 1. They rely on complex mechanisms such as blockchain consensus (e.g., Byzantine fault tolerance), federated learning gradient convergence, and differential privacy masks. While these technologies improve the system's credibility and resilience, they introduce significant computational, communication, and management overhead. 2. They emphasize ensuring credibility through blockchain consensus and federated learning, but multi-step on-chain operations and gradient synchronization may introduce high latency, affecting the real-time performance of resource allocation. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this application aims to provide a smart city resource allocation method and system based on edge-cloud collaboration. This method enables efficient, real-time allocation and adaptive optimization of smart city resources. Through edge-side data preprocessing, cloud-based collaborative analysis, and iterative feedback mechanisms, it reduces system complexity and deployment costs while improving response speed, scalability, and overall management efficiency.

[0006] To achieve the above objectives, this application adopts the following technical solution: This application provides a smart city resource allocation method based on edge-cloud collaboration, the method comprising: S101 collects resource status data in real time from scattered nodes in the city by deploying IoT sensors. The resource status data includes multi-dimensional heterogeneous format data of energy, transportation, medical care and education. The collected resource status data is initially labeled and converted in format using a standardized protocol to obtain a standardized resource status dataset. S102, execute edge computing tasks on the edge device, filter noise in the standardized resource status dataset, calculate the data dimension of the standardized resource status data, and if the data dimension of the standardized resource status data is greater than the preset threshold of data dimension, apply the K-means clustering algorithm to group the standardized resource status data and determine the preliminary extracted resource load feature vector. S103, the initially extracted resource load feature vector is uploaded to the cloud, and the resource load feature vectors from different end devices are fused through a multi-source data fusion mechanism to obtain a fused vector. A deep learning model is used to perform correlation analysis on the fused vector to obtain a set of quantitative indicators of the interaction relationship between resources. S104: For the set of quantitative indicators, construct a multi-dimensional resource situation map model, perform real-time update operation on the multi-dimensional resource situation map model in the cloud, and calculate the change rate of interaction relationship indicators. If the change rate of interaction relationship indicators is higher than the preset threshold of change rate, a dynamic adjustment mechanism is triggered to determine the resource availability distribution pattern. S105: Extract the status of key nodes from the resource availability distribution pattern, simulate the resource flow path through a graph neural network algorithm model, and obtain an optimized resource allocation scheme draft. S106. Based on the optimized resource allocation scheme draft, a sequence of control commands is generated in the cloud, sent to the end device to perform resource reconfiguration and obtain execution feedback data. If the execution feedback data shows a deviation, the resource allocation scheme draft is iteratively updated to determine the final resource allocation execution plan. S107. Based on the final resource allocation execution plan, the monitoring module continuously tracks the implementation effect and uses a feedback loop mechanism to correct the deviations in the final resource allocation execution plan, thereby obtaining a real-time optimized urban resource management status.

[0007] As a preferred technical solution, in S101, the preliminary labeling and format conversion of the collected resource status data using a standardized protocol includes: converting unstructured or semi-structured data into unified structured format data; mapping and converting fields with the same name but different meanings from different sources, or fields with different names but the same meaning; and adding metadata tags such as source, timestamp, and device ID to the data.

[0008] As a preferred technical solution, in S102, noise filtering of the standardized resource status dataset includes: for numerical data, using statistical filtering or simple digital filters for noise filtering; for image and video data, using Gaussian filtering or median filtering to perform convolution operations on the image to remove noise; setting a preset threshold for data quality confidence, and triggering noise filtering of the standardized resource status dataset only when the data quality confidence is lower than the preset threshold; The standardized resource status data is grouped and processed using the K-means clustering algorithm. The initial extracted resource load feature vector is determined by: randomly selecting K data points as initial cluster centers and calculating the distance from each data point to the K centers; recalculating the mean of all data points in each cluster and using this mean as the new cluster center; repeating the calculation until the change in the center point is less than a preset threshold or the maximum number of iterations is reached; after grouping, each cluster represents a data pattern with similar resource load status, and key information is extracted from each cluster to form a resource load feature vector, which includes: cluster center, number of data points in the cluster, and cluster radius or variance.

[0009] As a preferred technical solution, in S103, uploading the initially extracted resource load feature vector to the cloud includes: the end device uploading the preprocessed resource load feature vector to the cloud through a secure communication protocol; the cloud platform establishing a unified data receiving interface with authentication, traffic control, and data verification capabilities; the data receiving interface performing preliminary format verification and timestamp alignment on the resource load feature vector to ensure data consistency during subsequent processing; and the resource load feature vector being buffered in a message queue to decouple the data receiving and processing processes and cope with traffic peaks. The fusion vector is obtained by fusing resource load feature vectors from different end devices through a multi-source data fusion mechanism. The fusion vector is obtained by: calculating the weight of the resource load feature vectors uploaded by different end devices for the current analysis task based on the attention mechanism method, and performing weighted averaging or weighted concatenation to obtain the fusion vector. A deep learning model is used to perform correlation analysis on the fused vectors to obtain a set of quantitative indicators for the interaction relationships between resources. This includes: using a graph neural network-based deep learning model to abstract the feature vectors of various resource loads as nodes in a graph and the potential dependencies between resources as edges, thereby directly building a deep learning model for the resource network; converting the fused vectors into a tensor format recognizable by the deep learning model; performing nonlinear combinations and transformations of features of different resource types in the hidden layers of the deep learning model; outputting a resource correlation matrix, where each element represents the degree of mutual influence between two resource nodes in the current state; and deriving the indirect coupling degree or risk transmission coefficient between resources by analyzing the internal state of the deep learning model, thus obtaining a set of quantitative indicators for the interaction relationships between resources.

[0010] As a preferred technical solution, in S104, the construction of a multi-dimensional resource situation map model for the set of quantitative indicators includes: abstracting various resource entities in the city as nodes in the graph, abstracting the interaction relationship between resource entities as edges, and pre-defining the types, attributes and possible relationships between various resource entities through the ontology model to obtain the multi-dimensional resource situation map model. Performing real-time updates to the resource multidimensional situation map model in the cloud includes: establishing a stable connection between the cloud and the edge device to continuously receive the latest resource load feature vectors and quantitative indicator sets; when new data streams arrive, the cloud automatically identifies the changed parts and updates only these parts. The rate of change of the interaction relationship index is calculated. If the rate of change of the interaction relationship index is higher than the preset threshold, a dynamic adjustment mechanism is triggered. The determination of the resource availability distribution pattern includes: starting from the resource node where drastic changes occur, traversing and analyzing the relationship edges in the multi-dimensional situation map model, simulating the scope of impact, in order to determine which areas in the city are currently in a state of tension, abundance or bottleneck, and thus obtaining the resource availability distribution pattern.

[0011] As a preferred technical solution, in S105, extracting the key node status from the resource availability distribution pattern includes: extracting key nodes from the resource availability distribution pattern, where key nodes include one or more features such as resource bottlenecks, connection hubs, and vulnerability points; quantifying the criticality of nodes through several centrality indicators, including degree centrality, betweenness centrality, and proximity centrality, to obtain the key node status. The optimized resource allocation scheme draft is obtained by simulating resource flow paths using a graph neural network algorithm model. This includes: transforming the multi-dimensional resource situation map model into a graph neural network algorithm model that can be processed by the graph neural network; simulating resource flow through a message passing mechanism, where each node aggregates neighbor node messages and updates its own state representation at each layer, and through multi-layer stacking, ensuring that a node can obtain neighbor node messages beyond multiple hops; modeling resource allocation decisions as actions on the edges of the graph neural network algorithm model, outputting an action probability for each possible allocation edge, and finally selecting the path with the highest probability to form the resource allocation scheme draft.

[0012] As a preferred technical solution, in S106, the generation of a control command sequence in the cloud based on the optimized resource allocation scheme draft includes: the cloud will translate the resource allocation scheme draft into a series of ordered commands according to the preset device control protocol library to obtain the control command sequence; The process of sending control command sequences to the edge devices to perform resource reconfiguration and obtain execution feedback data includes: the cloud sending control command sequences to the corresponding edge devices; after receiving the control command sequences, the edge devices perform identity authentication and command verification to prevent unauthorized operation; after successful verification, the actuators of the edge devices begin to operate, completing the actual reconfiguration of resources; while the edge devices are performing resource configuration, the sensors built into the edge devices collect the status data after execution in real time to obtain execution feedback data. If the execution feedback data shows a deviation, the resource allocation plan draft is iteratively updated to determine the final resource allocation execution plan. This includes: the cloud comparing the execution feedback data with the expected target of the resource allocation plan draft and calculating the execution deviation; if the execution deviation is greater than the preset deviation threshold, a dynamic adjustment mechanism is triggered; the execution feedback data is used as new input to the resource multidimensional situation map model and graph neural network algorithm model to iteratively update and re-optimize the resource allocation plan draft, generate an updated resource allocation plan draft, and generate control command sequences again, which are sent to the end device to execute resource reconfiguration and obtain execution feedback data and calculate the execution deviation until the execution deviation is less than the preset deviation threshold or the maximum number of iterations is reached, thus obtaining the final resource allocation execution plan.

[0013] As a preferred technical solution, in S107, based on the final resource allocation execution plan, the monitoring module continuously tracks the implementation effect, including: the monitoring module accesses the execution feedback data of IoT sensors, business systems, and edge devices in the city's distributed nodes; and calculates resource utilization, response speed, and business continuity to obtain the implementation effect. The feedback loop mechanism is used to correct deviations in the final resource allocation execution plan, resulting in a real-time optimized urban resource management status. This includes: comparing the implementation effect with the expected goals of the final resource allocation execution plan to calculate the implementation effect deviation; activating the feedback loop mechanism when the implementation effect deviation exceeds a preset threshold; performing root cause analysis on the implementation effect deviation, correlating abnormal data, locating the source of the deviation, and obtaining the root cause analysis results; and initiating different levels of correction strategies based on the nature of the implementation effect deviation and the root cause analysis results: for local, small-scale implementation effect deviations, step S105 is re-executed to generate a local optimization plan and issue it for execution; for global, large-scale deviations, the nature of the implementation effect deviation and the root cause analysis results are fed back to the resource multidimensional situation map model or graph neural network algorithm model, and steps S104-S106 are re-executed to obtain a global optimization plan and issue it for execution; finally, a real-time optimized urban resource management status is obtained.

[0014] This application also provides a smart city resource allocation system based on edge-cloud collaboration, the system comprising: The data acquisition and edge computing module is used to collect multi-dimensional and heterogeneous resource status data from scattered nodes in the city in real time by deploying IoT sensors. The collected data is initially labeled and converted using a standardized protocol to obtain a standardized resource status dataset. Edge computing tasks are executed on the edge device to filter noise from the standardized resource status dataset, calculate the data dimension, and if the data dimension is greater than a preset threshold, the K-means clustering algorithm is applied to group the data and extract the resource load feature vector. The cloud-based data fusion and analysis module receives resource load feature vectors from edge devices, fuses feature vectors from different edge devices through a multi-source data fusion mechanism to obtain a fused vector, and uses a deep learning model to perform correlation analysis on the fused vector to obtain a set of quantitative indicators of the interaction relationship between resources. The resource situation modeling and dynamic adjustment module is used to construct a multi-dimensional resource situation map model for a set of quantitative indicators. The model is updated in real time in the cloud and the rate of change of interactive relationship indicators is calculated. If the rate of change is higher than a preset threshold, a dynamic adjustment mechanism is triggered to determine the resource availability distribution pattern. The resource allocation optimization and execution module is used to extract the status of key nodes from the resource availability distribution pattern, simulate the resource flow path through a graph neural network algorithm model, and obtain an optimized resource allocation scheme draft. Based on the draft, a control command sequence is generated and sent to the end device to execute resource reconfiguration. The scheme is iteratively updated based on execution feedback data to determine the final resource allocation execution plan. The monitoring and feedback module is used to continuously track the implementation effect based on the final resource allocation execution plan, and to correct deviations through a feedback loop mechanism to obtain a real-time optimized status of urban resource management.

[0015] Compared with the prior art, the beneficial effects of this application are as follows: This application achieves rapid standardization and feature extraction of resource status data through edge computing tasks on the edge device, significantly improving data processing efficiency and reducing system complexity. Utilizing cloud-based multi-source data fusion and dynamic adjustment mechanisms, combined with real-time edge acquisition, this application achieves low-latency response and high adaptability in resource allocation. By introducing graph neural networks and feedback loop mechanisms, this application simulates resource flow paths and corrects deviations, thereby optimizing allocation accuracy. The edge-cloud collaborative architecture of this application simplifies system components, and through standardized and modular design, enhances scalability and practicality. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the steps of a smart city resource allocation method based on edge-cloud collaboration proposed in this application. Detailed Implementation

[0017] To enable those skilled in the art to better understand the present application, the technical solutions in specific embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings.

[0018] like Figure 1 As shown, this application provides a smart city resource allocation method based on edge-cloud collaboration, the method comprising: S101 collects resource status data in real time from scattered nodes in the city by deploying IoT sensors. The resource status data includes multi-dimensional heterogeneous data of energy, transportation, medical care and education. The collected resource status data is initially labeled and converted in format using a standardized protocol to obtain a standardized resource status dataset.

[0019] S102, execute edge computing tasks on the edge device, filter noise in the standardized resource status dataset, calculate the data dimension of the standardized resource status data, and if the data dimension of the standardized resource status data is greater than the preset threshold, apply the K-means clustering algorithm to group the standardized resource status data and determine the preliminary extracted resource load feature vector.

[0020] S103, the initially extracted resource load feature vectors are uploaded to the cloud, and the resource load feature vectors from different end devices are fused through a multi-source data fusion mechanism to obtain a fused vector. A deep learning model is used to perform correlation analysis on the fused vector to obtain a set of quantitative indicators of the interaction relationship between resources.

[0021] S104: For the set of quantitative indicators, construct a multi-dimensional resource situation map model, perform real-time update operation on the multi-dimensional resource situation map model in the cloud, and calculate the change rate of interaction relationship indicators. If the change rate of interaction relationship indicators is higher than the preset threshold, a dynamic adjustment mechanism is triggered to determine the resource availability distribution pattern.

[0022] S105: Extract the status of key nodes from the resource availability distribution pattern, simulate the resource flow path through a graph neural network algorithm model, and obtain an optimized resource allocation scheme draft.

[0023] S106. Based on the optimized resource allocation scheme draft, a sequence of control commands is generated in the cloud, sent to the end device to perform resource reconfiguration and obtain execution feedback data. If the execution feedback data shows a deviation, the resource allocation scheme draft is iteratively updated to determine the final resource allocation execution plan.

[0024] S107. Based on the final resource allocation execution plan, the monitoring module continuously tracks the implementation effect and uses a feedback loop mechanism to correct the deviations in the final resource allocation execution plan, thereby obtaining a real-time optimized urban resource management status.

[0025] This application achieves rapid standardization and feature extraction of resource status data through edge computing tasks (S101-S102) on the edge device, significantly improving data processing efficiency and reducing system complexity. Specifically: In S101, IoT sensors collect multi-dimensional heterogeneous data such as energy, transportation, and healthcare data, and perform preliminary annotation and format conversion using standardized protocols to generate a standardized dataset; in S102, the edge device performs noise filtering and data dimension calculation. If the data dimension exceeds a threshold, K-means clustering algorithm is applied for group processing to extract resource load feature vectors. This avoids complex data alignment and privacy masking mechanisms, simplifying the process. Edge preprocessing reduces the burden on the cloud, lowers computational and communication overhead, and makes the system lighter and easier to deploy. This application avoids high-cost decentralized operations, improving scalability and cost-effectiveness.

[0026] In S101, specialized IoT sensors are deployed at key nodes of critical infrastructure in sectors such as energy, transportation, healthcare, and education. For example, smart meters are installed in the power grid; traffic flow monitoring sensors or cameras are deployed along major transportation routes; real-time data on bed availability and registration is acquired through device interfaces in medical institutions; and information on classroom energy consumption and equipment utilization is collected in educational institutions. These sensors form the hardware foundation for data acquisition, responsible for generating initial, raw resource status data from the physical world.

[0027] These resource status data are inherently heterogeneous, manifesting in several ways: Diverse formats: They may include real-time streaming data from sensors, structured tables in databases, log files, and even image and video streams. Inconsistent protocols: The data acquisition devices may use different communication protocols, such as low-power wide-area network technologies like ZigBee, LoRaWAN, and NB-IoT, or wired networks. Differences in frequency and granularity: Traffic data may be in the millisecond range, while energy consumption data may be in the minute or even hour range, resulting in different data granularity and acquisition frequencies.

[0028] This application employs standardized communication protocols suitable for IoT scenarios, such as MQTT and CoAP, to uniformly transmit data collected by distributed nodes to a data aggregation point. These protocols typically feature low overhead and suitability for unstable networks, effectively ensuring data transmission efficiency and reliability. At the data entry point, data acquisition adapters or IoT gateways are used. These components act as bridges, receiving information from different protocols and data formats, and performing preliminary protocol parsing and format conversion to prepare for subsequent in-depth standardization. They also act as buffers to handle peak data inflow pressures.

[0029] Furthermore, in S101, the preliminary labeling and format conversion of the collected resource status data using standardized protocols includes: Unstructured or semi-structured data is converted into a unified structured format (such as JSON, XML, or a specific columnar format). Fields with the same name but different meanings from different sources, or fields with the same name but different meanings, are mapped and transformed. For example, the "energy consumption value" field from different systems is uniformly named "energy_consumption," and the unit is uniformly "kilowatt-hours." Metadata tags such as source (e.g., "traffic-intersection A"), timestamp, and device ID are added to the data to enhance its traceability. Through the above processing, a high-quality, standardized resource status dataset is finally generated. This standardized resource status dataset is the sole and reliable data source for subsequent edge computing and cloud analytics.

[0030] Noise in standardized resource status datasets includes: Random noise: caused by sensor acquisition errors or transmission interference, manifested as irregular fluctuations in data points. Outliers: data that deviates significantly from the normal range due to temporary equipment malfunctions or abnormal events (such as traffic sensors being blocked by vehicles).

[0031] Furthermore, in S102, noise filtering of the standardized resource state dataset includes: For numerical data (such as energy consumption, temperature, and traffic flow), statistical filtering or simple digital filters can be used for noise filtering. For example, median filtering can be used to process time-series data sequences; it effectively smooths impulse noise and is computationally simple, making it suitable for edge devices. For slowly changing parameters, a first-order low-pass filter can also be used to suppress high-frequency noise.

[0032] For image and video data (such as traffic camera footage), Gaussian filtering or median filtering can be used to perform convolution operations on the image to remove noise. For example, optimization libraries such as OpenCV can be used to remove noise.

[0033] A preset threshold for data quality confidence is set. Noise filtering of the standardized resource status dataset is triggered only when the data quality confidence falls below the preset threshold. While ensuring no key information is lost, using half-precision (FP16) or fixed-point arithmetic for floating-point numbers can significantly improve computation speed and reduce memory usage.

[0034] The number of feature fields is read from the standardized resource status dataset to obtain the data dimension. Calculating the data dimension is to determine whether clustering is necessary, avoiding unnecessary complex operations on simple datasets. For example, the preset threshold for data dimension is set when the number of feature fields exceeds 20. As a more preferred implementation, the preset threshold for data dimension is linked to the real-time computing power of the edge device. The edge device dynamically adjusts the preset threshold for data dimension based on the current CPU / memory load; when the load is high, the preset threshold for data dimension is increased to reduce the clustering task.

[0035] When the data dimension is high, K-means is used for grouping, which aims to summarize massive, high-dimensional data into several representative categories, thereby significantly reducing the data volume and providing a foundation for subsequent analysis.

[0036] Specifically, the K-means clustering algorithm is applied to group standardized resource status data to determine the initial extracted resource load feature vectors. This includes: randomly selecting K data points as initial cluster centers and calculating the distance from each data point to the K centers. The value of K can be set based on historical experience or estimated using a simplified version of the Elbow Method. The mean of all data points in each cluster is recalculated, and this mean is used as the new cluster center. This calculation is repeated until the change in center points is less than a preset threshold or the maximum number of iterations is reached.

[0037] More specifically, approximate algorithms or sampling of standardized resource state data are employed, with clustering performed only on a subset of the standardized resource state data to accelerate the process. The maximum number of iterations is limited to prevent excessive time consumption on edge devices. If the data is collected continuously, the cluster centers of the previous batch are used as a reference for the current initialization, speeding up convergence.

[0038] After grouping, each cluster represents a data pattern with similar resource load states. Key information is extracted from each cluster to form a resource load feature vector, which includes: cluster center, number of data points within the cluster, and cluster radius or variance. Specifically, the cluster center represents the average characteristic of all data points in the cluster, indicating the typical state of this pattern. The number of data points within the cluster reflects the frequency or prevalence of the pattern. The cluster radius or variance describes the dispersion and stability of the data within this pattern. Finally, this information is concatenated to form a preliminary resource load feature vector, ready for uploading to the cloud.

[0039] This application utilizes a cloud-based multi-source data fusion and dynamic adjustment mechanism (S103-S104), combined with real-time edge data acquisition, to achieve low-latency response and high adaptability in resource allocation. Specifically: In S103, after the resource load feature vector is uploaded to the cloud, correlation analysis is performed through a multi-source fusion mechanism and a deep learning model to obtain a set of quantitative indicators; in S104, a multi-dimensional resource situation map model is constructed, and the rate of change of interaction relationship indicators is calculated in real time. If the rate of change exceeds a threshold, a dynamic adjustment mechanism is triggered to determine the resource availability distribution. This ensures that the system can quickly adapt to dynamic changes in the urban environment. Centralized cloud analysis reduces on-chain consensus latency (such as the Byzantine fault-tolerant process of comparison files), making resource allocation decisions more agile. For example, the dynamic threshold triggering mechanism avoids the complex temporal dependencies of sliding window risk assessment in comparison files, improving the response speed in sudden scenarios (such as traffic congestion or energy peaks). At the same time, real-time model updates enhance the system's adaptability and reduce false positives and false negatives.

[0040] Furthermore, in S103, the initially extracted resource load feature vector is uploaded to the cloud, including: The edge devices upload the preprocessed resource load feature vectors to the cloud via secure communication protocols (such as HTTPS, MQTT over TLS). The cloud platform establishes a unified data receiving interface with authentication, flow control, and data verification capabilities. The received feature vectors may come from different types of edge devices (such as traffic sensors and smart meters). The data receiving interface performs preliminary format verification and timestamp alignment on the resource load feature vectors to ensure data consistency during subsequent processing. The resource load feature vectors are buffered in a message queue (such as Kafka, RabbitMQ) to decouple the data receiving and processing processes and handle traffic spikes.

[0041] The resource load feature vectors from different endpoint devices are fused using a multi-source data fusion mechanism to obtain a fused vector. This fused vector includes feature vectors from different endpoint devices, whose dimensions, units, and physical meanings may differ. First, feature alignment is required. For example, traffic flow feature vectors may have inconsistent time granularity with energy consumption feature vectors, requiring alignment through interpolation or aggregation methods. Next, resource load feature vectors from different sources are associated based on business logic (such as spatial location association and time window matching) to prepare for subsequent fusion. Using an attention mechanism, the weights of resource load feature vectors uploaded from different endpoint devices for the current analysis task are calculated, and a weighted average or weighted concatenation is performed to obtain the fused vector. This allows the deep learning model to dynamically focus on more critical information sources; for example, during morning rush hour, traffic flow data features are given higher weights.

[0042] A deep learning model is used to perform correlation analysis on the fused vectors, obtaining a set of quantitative indicators of the interaction relationships between resources, including: A deep learning model based on graph neural networks is employed to abstract the feature vectors of various resource loads as nodes in a graph, and the potential dependencies between resources as edges, thereby directly building a deep learning model for the resource network. If the time-series characteristics of the data are significant, a Long Short-Term Memory (LSTM) network or a Transformer model is used to effectively capture its dynamic evolution patterns. For more complex spatiotemporal relationships, hybrid architectures such as CNN-LSTM or ST-GNN (Spatiotemporal Graph Neural Network) can be used.

[0043] The fused vectors are transformed into a tensor format recognizable by the deep learning model. For example, a fused vector sample can be constructed for each analysis period (e.g., the past hour). Simultaneously, explicit prior knowledge needs to be provided to the model, such as initializing the weights of edges in the graph using the inverse of spatial distance or business logic-based association strength (e.g., "relevance weights between the school and surrounding bus stops").

[0044] In the hidden layers of deep learning models, features of different resource types are combined and transformed non-linearly. For example, a neuron in the model might learn a high-order interaction pattern between "peak energy load in a commercial area on a weekday evening" and "traffic flow on a main road." Attention mechanisms are particularly important in this process, as they can dynamically calculate the importance weights of different resource features in a specific context, thus clarifying the most critical associations for "attention." For instance, when predicting regional congestion, the model might automatically assign higher attention weights to the vacancy rate of surrounding parking lots and subway passenger flow.

[0045] The deep learning model outputs a resource association matrix. Each element in the resource association matrix represents the degree of mutual influence between two resource nodes in the current state. By analyzing the internal state of the deep learning model, the indirect coupling degree or risk transmission coefficient between resources is derived, resulting in a set of quantitative indicators of the interaction relationship between resources.

[0046] Furthermore, in S104, a multi-dimensional resource situation map model is constructed for the set of quantitative indicators, including: Various resource entities in the city are abstracted as nodes in a graph (e.g., a traffic intersection, a hospital, a school, a substation, etc.). The interactions between resource entities are abstracted as edges (e.g., "transportation-energy coupling"). Edge types can represent different relationships, such as "physical connection," "functional dependency," and "spatial proximity." By predefining the types, attributes, and possible relationships between various resource entities through an ontology model, a multi-dimensional resource situational graph model is obtained. For example, a "dependency / support" relationship can be defined to indicate that a business application depends on a certain middleware service.

[0047] The construction process of a multi-dimensional resource situation map model can draw on the methods used in knowledge graph construction. It is typically not a one-time process but rather a continuous iterative one. Knowledge extraction and fusion: Structured, semi-structured, or unstructured data from different business systems (such as traffic monitoring systems and power dispatching systems) are integrated. Resource entities and their relationships are extracted using named entity recognition and relation extraction techniques, and then fused with a pre-defined ontology model to resolve issues of semantic redundancy or contradictions among entities. Graph storage and visualization: The constructed multi-dimensional resource situation map model is usually stored in a graph database. On a cloud platform, a corresponding interface needs to be developed to visualize the model, enabling city managers to intuitively understand the complex relationships between resources.

[0048] Real-time updates to the multi-dimensional resource status map model in the cloud include: establishing a stable connection between the cloud and the edge devices to continuously receive the latest resource load feature vectors and quantitative indicator sets. When new data streams arrive, the cloud automatically identifies the changed parts and updates only those parts. Updates can be time-driven (e.g., every 5 minutes) or event-driven (e.g., receiving a status alarm for a critical resource).

[0049] Calculate the rate of change of interaction relationship indicators. If the rate of change of interaction relationship indicators exceeds a preset threshold, a dynamic adjustment mechanism is triggered. The determination of resource availability distribution pattern includes: For a specific interaction indicator I, its rate of change R(t) at time t can be calculated using the following formula: R(t) = |I(t) - I(t-Δt)| / I(t-Δt), where Δt is a preset time window (e.g., 5 minutes or 1 hour). The calculated rate of change R(t) is compared with a preset threshold T. This preset threshold needs to be set based on historical data and business experience; for example, a sudden surge in traffic flow exceeding 50% would be considered abnormal.

[0050] Starting from resource nodes experiencing drastic changes, the system traverses and analyzes the relationships within the multi-dimensional situational mapping model to simulate the impact range. This helps determine which areas in the city are experiencing resource scarcity, abundance, or bottlenecks, thus obtaining the resource availability distribution pattern. For example, if the load on a major substation increases sharply, the system analyzes whether critical facilities within its power supply range, such as hospitals and traffic lights, will be affected. The conclusion might be: "Due to the evening rush hour, traffic resources in the city center are strained, but parking resources in surrounding areas are relatively abundant." This conclusion, the determination of the resource availability distribution pattern, provides a direct basis for generating optimization schemes in step S105.

[0051] This application introduces graph neural networks and a feedback loop mechanism (S105-S107) to simulate resource flow paths and correct deviations, thereby optimizing allocation accuracy. Specifically: In S105, the states of key nodes are extracted from the resource availability distribution pattern, and the resource flow path is simulated using a graph neural network algorithm to generate a draft optimization plan; in S106-S107, a sequence of control commands is generated in the cloud and sent to the edge for execution. The effect is tracked through a monitoring module, and the plan is iteratively updated based on feedback data, forming a closed-loop control. The graph neural network provides scientific spatial path analysis, avoiding the shortcomings of empirical allocation, while the iterative learning mechanism ensures long-term accuracy. Real-time deviation correction (S107) continuously optimizes resource management status and improves the reliability of smart city services.

[0052] Furthermore, in S105, the key node status extracted from the resource availability distribution pattern includes: Key nodes are extracted from the resource availability distribution pattern. Key nodes possess one or more characteristics of resource bottlenecks, connection hubs, and vulnerabilities. Resource bottlenecks have extremely high resource loads, and their congestion hinders resource flow throughout the network. Connection hubs have a large number of connections (i.e., high centrality) and are essential pathways for multiple resource flows. Vulnerabilities have a significant impact on network connectivity due to their failure or deterioration. The criticality of key nodes is quantified using several centrality indicators, including degree centrality, betweenness centrality, and proximity centrality, to determine the key node status. Degree centrality directly measures the number of connections a node has. The more connections, the more critical the node. Betweenness centrality measures how many pairs of shortest paths a node is on. A higher value indicates that the node is an essential pathway for more resource flows and is a key hub. Proximity centrality measures the average distance from a node to all other nodes in the network. A higher value indicates that the node is closer to the network center topologically and has greater influence. The extracted key node states (such as node ID, resource load value, and centrality index) will serve as important input features for subsequent graph neural network algorithm models.

[0053] The optimized resource allocation scheme draft was obtained by simulating resource flow paths using a graph neural network algorithm model, including: The multi-dimensional resource situation map model is transformed into a graph neural network algorithm model G = (V, E, A) that can be processed by graph neural networks. Here, V is the set of nodes (each node represents a resource entity, such as a substation or hospital), E is the set of edges (representing the interaction relationships between resources), and A is the attribute feature matrix of the nodes (containing information such as the key node states extracted from the S104 pattern and the real-time resource load of each node). The graph neural network algorithm model simulates resource flow through a message passing mechanism. At each layer, each node aggregates messages from neighboring nodes and updates its own state representation. Multiple layers are stacked to ensure that a node can obtain messages from neighboring nodes beyond multiple hops. Resource allocation decisions (such as "scheduling X units of electricity from node A to node C") are modeled as actions on the edges of the graph neural network algorithm model. An action probability is output for each possible allocation edge, and finally, the path with the highest probability is selected to form a draft resource allocation scheme.

[0054] Furthermore, in S106, based on the optimized resource allocation scheme draft, the sequence of control commands generated in the cloud includes: The cloud platform translates the draft resource allocation plan into a series of ordered instructions based on a pre-defined device control protocol library, resulting in a control instruction sequence. For example, the draft resource allocation plan of "dispatch 10% of the surplus power in area A to area B" is transformed into the following instructions: Instruction 1: Send the instruction "reduce output power by X kilowatts" to the energy storage station controller in area A. Instruction 2: Send the instruction "adjust topology, prioritizing area B" to the smart grid routing node. Instruction 3: Send the instruction "prepare to receive additional X kilowatts of power load" to the substation controller in area B. Instruction 4: Send the instruction "fine-tune parameters to ensure power quality" to the voltage regulators along the route. These instructions constitute a control instruction sequence, with strict timing dependencies and logical consistency checks between them, ensuring the safety and reliability of the entire operation process.

[0055] The process of reconfiguring resources and obtaining execution feedback data by sending them to the edge device includes: The cloud sends control command sequences to corresponding end-side devices (such as smart meters, traffic signal controllers, and water pump valve actuators) via secure and reliable communication links (such as MQTT over TLS and dedicated APN IoT channels). Upon receiving the control command sequence, the end-side devices perform authentication and command verification to prevent unauthorized operation. After successful verification, the actuators on the end-side devices (such as door actuators and electronic relays) begin operation, completing the actual reconfiguration of resources. While performing resource configuration, the end-side devices' built-in sensors (such as position sensors, current sensors, and flow meters) collect real-time status data after execution, obtaining execution feedback data. Execution feedback data mainly includes two categories: direct status feedback: such as valve opening degree, power output, and current phase of traffic lights, directly reflecting the execution result of the commands; and indirect effect feedback: collected through IoT sensor networks deployed throughout the city, reflecting the overall status of the system after resource allocation, such as changes in traffic flow and power grid frequency stability in the allocated area.

[0056] If the feedback data shows a deviation, the draft resource allocation plan will be iteratively updated to determine the final resource allocation execution plan, including: The cloud platform compares the execution feedback data with the expected targets of the resource allocation plan draft, calculating the execution deviation. For example, if the expected voltage increase in area B is 5%, but the actual feedback shows only a 2% increase, then there is a deviation. If the execution deviation is below a preset deviation threshold, the plan is considered successfully executed, and the current resource allocation plan draft is determined as the final resource allocation execution plan. If the execution deviation is greater than the preset deviation threshold, a dynamic adjustment mechanism is triggered. The execution feedback data is used as new input to the resource multi-dimensional situation map model and graph neural network algorithm model to iteratively update and re-optimize the resource allocation plan draft, generating an updated resource allocation plan draft. Control command sequences are then generated again, sent to the end-side devices to perform resource reconfiguration, obtain execution feedback data, and calculate the execution deviation until the execution deviation is less than the preset deviation threshold or the maximum number of iterations is reached, resulting in the final resource allocation execution plan.

[0057] Furthermore, in S107, based on the final resource allocation execution plan, the implementation effect is continuously tracked through the monitoring module, including: The monitoring module accesses IoT sensors at distributed urban nodes, business systems (such as traffic signal control systems and power grid dispatching systems), and execution feedback data from edge devices. It calculates resource utilization, response speed, and business continuity to obtain the implementation effect. Resource utilization includes metrics such as parking space utilization and power grid load rate. Response speed is measured by the time from issuing a dispatch command to the edge device producing an actual effect. Business continuity is measured by the rate of decrease in critical service interruptions due to resource shortages. A feedback loop mechanism is used to correct deviations in the final resource allocation execution plan, resulting in a real-time optimized urban resource management status, including: The implementation results are compared with the expected goals of the final resource allocation execution plan to calculate the implementation effect deviation. When the implementation effect deviation exceeds a preset threshold, the feedback loop mechanism is activated. Root cause analysis is performed on the implementation effect deviation, abnormal data is correlated, the source of the deviation is located, and the root cause analysis results are obtained.

[0058] Based on the nature of the implementation effect deviations and the root cause analysis results, different levels of correction strategies are initiated: For local, small-scale implementation effect deviations, step S105 is re-executed to generate a local optimization plan and issue it for implementation. For global, large-scale deviations, the nature of the implementation effect deviations and the root cause analysis results are fed back to the resource multidimensional situation map model or graph neural network algorithm model, and steps S104-S106 are re-executed to obtain a global optimization plan and issue it for implementation. Finally, a real-time optimized urban resource management status is obtained.

[0059] In one application scenario, it is necessary to allocate traffic resources in a smart city.

[0060] Establish a comprehensive monitoring system: Monitoring data sources: traffic flow sensors, speed detectors, GPS floating car data, current traffic light timing status, and variable message sign operation status at major intersections.

[0061] Real-time data acquisition: This data is transmitted to the cloud monitoring module in real time via IoT protocols.

[0062] Calculate Key Performance Indicators (KPIs) and compare them with expected results: Expected results: After the implementation of the plan, it is hoped that the average vehicle speed in the city center area will increase from 20km / h to 25km / h, and the congestion index (delay time) will decrease by 15%.

[0063] Real-time calculation: The monitoring module calculates the actual average vehicle speed and congestion index based on real-time data.

[0064] Comparison: Assuming actual data shows an average vehicle speed of 22 km / h, the congestion index decreases by 5%. This deviates from expectations.

[0065] Identify and assess performance deviations: Deviation calculation: The deviation rate of average vehicle speed is |22-25| / 25 = 12%, and the deviation rate of congestion index decrease is |5%-15%| / 15% ≈ 66.7%.

[0066] Threshold judgment: If the threshold is set to 20%, the average vehicle speed deviation is within an acceptable range, but the congestion index deviation exceeds the threshold and needs to be corrected.

[0067] Analyze the root causes of the deviation: Correlation analysis: The monitoring module found that after the green light time for outbound traffic increased, the queue length for inbound traffic unexpectedly increased, leading to new congestion points in adjacent areas.

[0068] Model backtesting: Examining the graph neural network model in S105, it was found that the model underestimated the elasticity of traffic flow in the direction of entering the city during the simulation (i.e., the traffic flow in the direction of entering the city did not decrease due to congestion, but instead continued to increase due to the rigid demand during the morning rush hour).

[0069] Root cause: The current traffic light timing scheme imposes excessive restrictions on inbound traffic, leading to an imbalance in the load on local road networks.

[0070] Activate the graded correction mechanism: Secondary correction (scheme iteration): Because the deviation is large and the root cause is that the model assumptions do not match the actual situation, the scheme needs to be optimized again.

[0071] a. Update the resource multidimensional situation map: Update the map with the latest traffic flow data, queue length, etc.

[0072] b. Re-simulation: Based on the latest data, the graph neural network recalculates the traffic light timing scheme, taking into account the rigid demand of inbound traffic flow. This may result in a more balanced scheme (such as appropriately increasing the green light time for inbound traffic, but at the same time suggesting guiding some traffic flow to be dispersed through other routes).

[0073] c. Generate a new draft: Obtain a new draft of the traffic light timing and guidance scheme.

[0074] Perform the correction and confirm the effect: The new scheme generates control commands through step S106 and sends them to traffic lights and variable message signs.

[0075] The monitoring module immediately began tracking traffic indicators under the new scheme.

[0076] The average vehicle speed reached 24 km / h, the congestion index decreased by 12%, and the deviation narrowed to an acceptable range.

[0077] The system records the successful experiences of this correction for future model training.

[0078] The final resource allocation and execution plan is as follows: During the morning rush hour, in order to alleviate congestion in the city center, the timing of traffic lights on major roads leading to the city center will be adjusted to increase the green light time for outbound directions, while variable message signs will be used to guide vehicles to detour.

[0079] This application also provides a smart city resource allocation system based on edge-cloud collaboration, the system comprising: The data acquisition and edge computing module is used to collect multi-dimensional, heterogeneous resource status data from scattered nodes in the city in real time by deploying IoT sensors. The collected data is initially labeled and converted using a standardized protocol to obtain a standardized resource status dataset. Edge computing tasks are executed on the edge devices to filter noise from the standardized resource status dataset, calculate the data dimensionality, and if the data dimensionality exceeds a preset threshold, the K-means clustering algorithm is applied to group the data and extract resource load feature vectors.

[0080] The cloud-based data fusion and analysis module receives resource load feature vectors from edge devices, merges feature vectors from different edge devices through a multi-source data fusion mechanism to obtain a fused vector, and uses a deep learning model to perform correlation analysis on the fused vector to obtain a set of quantitative indicators of the interaction relationship between resources.

[0081] The Resource Situation Modeling and Dynamic Adjustment Module is used to construct a multi-dimensional resource situation map model for a set of quantitative indicators. The model is updated in real time in the cloud and the rate of change of interactive relationship indicators is calculated. If the rate of change is higher than a preset threshold, a dynamic adjustment mechanism is triggered to determine the resource availability distribution pattern.

[0082] The resource allocation optimization and execution module extracts the status of key nodes from the resource availability distribution pattern, simulates resource flow paths using a graph neural network algorithm model, and obtains an optimized draft resource allocation scheme. Based on the draft, a sequence of control commands is generated and sent to the end-side devices to execute resource reconfiguration. The scheme is iteratively updated based on execution feedback data to determine the final resource allocation execution plan.

[0083] The monitoring and feedback module is used to continuously track the implementation effect based on the final resource allocation execution plan, and to correct deviations through a feedback loop mechanism to obtain a real-time optimized status of urban resource management.

[0084] It should be noted that the terms "first," "second," and similar terms used in this application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, "a" or "one," and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. "A plurality" or "several" indicates at least two. Unless otherwise stated, terms such as "front," "back," "left," "right," "lower," and / or "upper" are for illustrative purposes only and are not limited to a location or spatial orientation. Terms such as "comprising" or "including" indicate that the elements or objects preceding "comprising" encompass the elements or objects listed following "comprising" or "including" and their equivalents, and do not exclude other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect.

[0085] The singular forms “a,” “the,” and “the” used in this application specification and appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0086] It should be understood that those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A smart city resource allocation method based on end-cloud collaboration, characterized in that, The method comprises: S101, collecting resource state data from city dispersed nodes in real time by deploying Internet of Things sensors, the resource state data including multi-dimensional heterogeneous format data of energy, transportation, medical care and education types, performing preliminary labeling and format conversion on the collected resource state data by using a standardized protocol to obtain a standardized resource state data set; S102, performing an edge computing task on an end-side device, filtering noise of the standardized resource state data set, calculating a data dimension of the standardized resource state data, if the data dimension of the standardized resource state data is greater than a data dimension preset threshold, applying a K-means clustering algorithm to group process the standardized resource state data to determine a preliminary extracted resource load feature vector; S103, uploading the preliminary extracted resource load feature vector to a cloud end, fusing resource load feature vectors from different end-side devices by a multi-source data fusion mechanism to obtain a fusion vector, performing correlation analysis on the fusion vector by using a deep learning model to obtain a quantitative index set of interaction relationships between resources; S104, constructing a resource multi-dimensional situation graph model for the quantitative index set, performing real-time updating operation on the resource multi-dimensional situation graph model in the cloud end, and calculating an interaction relationship index change rate, if the interaction relationship index change rate is higher than a change rate preset threshold, triggering a dynamic adjustment mechanism to judge a resource availability distribution pattern; S105, extracting a key node state from the resource availability distribution pattern, simulating a resource flow path by a graph neural network algorithm model to obtain an optimized resource allocation scheme draft; S106, generating a control instruction sequence in the cloud end according to the optimized resource allocation scheme draft, issuing the control instruction sequence to the end-side device to execute resource reconfiguration and obtain execution feedback data, if the execution feedback data shows deviation, iteratively updating the resource allocation scheme draft to determine a final resource allocation execution plan; S107, continuously tracking implementation effects by a monitoring module according to the final resource allocation execution plan, correcting the deviation part of the final resource allocation execution plan by using a feedback cycle mechanism to obtain a real-time optimized city resource management state. 2.The smart city resource allocation method based on end-cloud cooperation according to claim 1, characterized in that, In S101, the preliminary labeling and format conversion of the collected resource state data by using the standardized protocol comprises: converting unstructured or semi-structured data into unified structured format data; mapping and converting fields with different sources but different meanings, or different names but synonymous meanings; labeling data with metadata tags of source, timestamp and device ID. 3.The smart city resource allocation method based on end-cloud collaboration of claim 1, wherein, In S102, the noise filtering of the standardized resource state data set comprises: for numerical data, using statistical filtering or a simple digital filter to filter noise; for image and video data, using Gaussian filtering or median filtering to perform convolution operation on the image to remove noise points; setting a data quality confidence preset threshold, only when the data quality confidence is lower than the data quality confidence preset threshold, triggering the noise filtering of the standardized resource state data set; The K-means clustering algorithm is applied to grouping processing of the standardized resource state data, and the preliminary extracted resource load feature vector includes: randomly selecting K data points as initial clustering centers, calculating the distance of each data point to the K center points; recalculating the mean of all data points in each cluster, and taking the mean as the new clustering center; repeating the calculation until the center point changes less than the preset threshold of the center point change or reaches the maximum iteration number; after grouping processing, each cluster represents a class of data patterns with similar resource load states, and the key information is extracted from each cluster to form the resource load feature vector, and the resource load feature vector includes: clustering center, number of data points in the cluster, and radius or variance of the cluster.

4. The smart city resource allocation method based on end-cloud collaboration according to claim 1 or 3, characterized in that, In S103, uploading the preliminary extracted resource load feature vector to the cloud includes: the end-side device uploads the preprocessed resource load feature vector to the cloud through a secure communication protocol, and the cloud platform sets up a unified data receiving interface with identity authentication, flow control and data verification capabilities; the data receiving interface performs preliminary format verification and timestamp alignment on the resource load feature vector to ensure the consistency of the data in subsequent processing; the resource load feature vector enters the message queue for buffering to decouple the data receiving and processing processes and cope with traffic peaks; The resource load feature vectors from different end-side devices are fused through a multi-source data fusion mechanism to obtain a fusion vector, which includes: based on the attention mechanism method, the weight of the resource load feature vector uploaded by different end-side devices for the current analysis task is calculated, and a weighted average or weighted splicing is performed to obtain the fusion vector; The correlation analysis of the fusion vector is performed by using a deep learning model to obtain a set of quantitative indicators of the interaction relationship between resources, which includes: a deep learning model based on a graph neural network is used to abstract each type of resource load feature vector as a node in a graph and the potential dependency relationship between resources as an edge, so that a deep learning model is directly established for the resource network; the fusion vector is converted into a tensor format recognizable by the deep learning model; in the hidden layer of the deep learning model, the features of different resource types are combined and transformed nonlinearly; the deep learning model outputs a resource correlation matrix, and each element value in the resource correlation matrix represents the mutual influence degree of two resource nodes in the current state, and by analyzing the internal state of the deep learning model, the indirect coupling degree or risk transmission coefficient between resources is derived, and a set of quantitative indicators of the interaction relationship between resources is obtained.

5. The method of claim 1, wherein, In S104, for the set of quantitative indicators, a resource multi-dimensional situation spectrum model is constructed, which includes: abstracting each type of resource entity in the city as a node in a graph, and abstracting the interaction relationship between the resource entities as an edge, and pre-defining the types, attributes and possible relationships between the entities of each type of resource entity through the ontology model to obtain the resource multi-dimensional situation spectrum model; The real-time updating operation of the resource multi-dimensional situation spectrum model in the cloud includes: the cloud establishes a stable connection with the end-side device, continuously receives the latest resource load feature vector and quantitative indicator set; when a new data stream arrives, the cloud automatically identifies the changed part and only updates these parts; The interaction relationship index change rate is calculated, and if the interaction relationship index change rate is higher than the change rate preset threshold, a dynamic adjustment mechanism is triggered. The resource availability distribution pattern is determined by starting from the resource node with a dramatic change, traversing and analyzing the relationship edges in the multi-dimensional situation graph model, simulating the influence range, and determining which regions in the current city have resources in a state of tension, abundance or bottleneck to obtain the resource availability distribution pattern.

6. The smart city resource allocation method based on end-cloud collaboration according to claim 1 or 5, characterized in that, In S105, the key node state is extracted from the resource availability distribution pattern, including one or more features of resource bottlenecks, connection hubs and vulnerable points; the key degree of the node is quantified by a plurality of centrality indexes of the key node, including degree centrality, betweenness centrality and closeness centrality, to obtain the key node state. The resource flow path is simulated by the graph neural network algorithm model to obtain the optimized resource allocation scheme draft, including: converting the resource multi-dimensional situation graph model into a graph neural network algorithm model that can be processed by the graph neural network; the graph neural network algorithm model simulates resource flow through a message passing mechanism, in each layer, each node aggregates neighbor node messages and updates its own state representation, and through multi-layer stacking, a node obtains neighbor node messages outside the multi-hop; the resource allocation decision is modeled as an action on the edge of the graph neural network algorithm model, and an action probability is output for each possible allocation edge, and finally the path with the highest probability is selected to form the resource allocation scheme draft.

7. The method of claim 6, wherein the method further comprises: In S106, the control instruction sequence is generated in the cloud according to the optimized resource allocation scheme draft, including: the cloud translates the resource allocation scheme draft into a series of ordered instructions according to the preset device control protocol library to obtain the control instruction sequence; The control instruction sequence is sent to the end-side device to perform resource reconfiguration and obtain execution feedback data, including: the cloud sends the control instruction sequence to the corresponding end-side device; after receiving the control instruction sequence, the end-side device performs identity authentication and instruction verification to prevent illegal manipulation; after verification, the actuator of the end-side device starts to act to complete the actual reconfiguration of the resource; the end-side device built-in sensor collects real-time state data after execution to obtain execution feedback data; If the execution feedback data shows a deviation, the resource allocation scheme draft is iteratively updated to determine the final resource allocation execution plan, including: the cloud compares the execution feedback data with the expected target of the resource allocation scheme draft to calculate the execution deviation; if the execution deviation is greater than the deviation preset threshold, the dynamic adjustment mechanism is triggered; the execution feedback data is sent back to the resource multi-dimensional situation graph model and the graph neural network algorithm model as a new input, the resource allocation scheme draft is iteratively updated and re-optimized, an updated resource allocation scheme draft is generated, and the control instruction sequence is generated again, the resource reconfiguration is performed by the end-side device, the execution feedback data is obtained, and the execution deviation is calculated, until the execution deviation is less than the deviation preset threshold or the maximum iteration number is reached, to obtain the final resource allocation execution plan. 8.The smart city resource allocation method based on end-cloud cooperation of claim 1, wherein, In S107, the implementation effect is continuously tracked according to the final resource allocation execution plan through the monitoring module, including: the monitoring module accesses the execution feedback data of the Internet of Things sensors of the city scattered nodes, the business systems and the end-side devices; the resource utilization rate, the response speed and the business continuity are calculated to obtain the implementation effect; The deviation part of the final resource allocation execution plan is modified by using a feedback cycle mechanism to obtain a real-time optimized city resource management state, including: the implementation effect is compared with the expected target of the final resource allocation execution plan to calculate the implementation effect deviation; when the implementation effect deviation exceeds the preset threshold of the implementation effect deviation, the feedback cycle mechanism is activated; the root cause analysis is performed on the implementation effect deviation, the abnormal data is associated, the source of the deviation is located, and the root cause analysis result is obtained; according to the nature of the implementation effect deviation and the root cause analysis result, different levels of correction strategies are started: for local and small-range implementation effect deviation, step S105 is re-executed to generate a local optimization scheme and issue an execution; for global and large-range deviation, the nature of the implementation effect deviation and the root cause analysis result are fed back to the resource multi-dimensional situation graph model or the graph neural network algorithm model, S104-S106 are re-executed to obtain a global optimization scheme and issue an execution; finally, a real-time optimized city resource management state is obtained.

9. A smart city resource allocation system based on end-cloud collaboration, characterized in that, The system comprises: The data acquisition and edge computing module is used for collecting multi-dimensional heterogeneous format resource state data from city scattered nodes in real time by deploying Internet of Things sensors, and performing preliminary labeling and format conversion on the collected data by using a standardized protocol to obtain a standardized resource state data set; an edge computing task is executed on the end-side device to filter noise of the standardized resource state data set, calculate the data dimension, and if the data dimension is greater than a preset threshold, apply a K-means clustering algorithm to group and process the data to extract a resource load feature vector; The cloud data fusion and analysis module is used for receiving the resource load feature vector from the end-side device, fusing the feature vectors of different end-side devices to obtain a fusion vector by using a multi-source data fusion mechanism, and performing correlation analysis on the fusion vector by using a deep learning model to obtain a quantitative index set of the interaction relationship between resources; The resource situation modeling and dynamic adjustment module is used for constructing a resource multi-dimensional situation graph model according to the quantitative index set, updating the model in real time in the cloud and calculating the change rate of the interaction relationship index, and if the change rate is higher than a preset threshold, triggering a dynamic adjustment mechanism to judge the resource availability distribution pattern; The resource allocation optimization and execution module is used for extracting the key node state from the resource availability distribution pattern, simulating the resource flow path by using a graph neural network algorithm model to obtain an optimized resource allocation scheme draft; generating a control instruction sequence according to the draft and issuing it to the end-side device to execute resource reconfiguration, and iteratively updating the scheme based on the execution feedback data to determine a final resource allocation execution plan; A monitoring and feedback module is used to continuously track the implementation effect through the monitoring module according to the final resource allocation execution plan, and correct the deviation part by using a feedback loop mechanism, so as to obtain a real-time optimized urban resource management state.

Images