A multi-target coordination-based urban resource optimization configuration method and system

Abstract

The present application relates to resource optimization configuration technical field, especially in kind of city resource optimization configuration method and system based on multi-objective coordination, method, including: the pre-processing of the obtained city resource related data;Based on the pre-processed city resource related data constructs dynamic space-time knowledge graph;Construct reinforcement learning-neural network hybrid optimization model, dynamic space-time knowledge graph as the embedded representation of the model;Adopt quantum computing and classical optimization algorithm to carry out multi-objective training optimization to the model;Based on the optimized model generates resource allocation decision. Through the construction of multi-source data acquisition network, the integration of space-time, structured, text and image and other multi-modal heterogeneous data, and the use of space-time graph convolution network, BERT model, improved ResNet model and the like for feature extraction and fusion, the city resource information can be more comprehensively and accurately obtained.

Description

A method and system for optimizing urban resource allocation based on multi-objective collaboration Technical Field

[0001] This invention relates to the field of resource optimization and allocation technology, and in particular to a method and system for urban resource optimization and allocation based on multi-objective collaboration. Background Technology

[0002] In recent years, the rapid development of emerging technologies such as big data, artificial intelligence, and the Internet of Things has provided new technical means for the optimal allocation of urban resources. IoT devices can collect real-time data on urban resource usage and environmental conditions, providing rich real-time data for resource allocation; big data technology can integrate multi-source heterogeneous data and uncover underlying patterns; and artificial intelligence algorithms have demonstrated powerful capabilities in solving complex problems and optimizing decisions.

[0003] However, the application of these technologies in the field of urban resource optimization and allocation still faces many challenges. Existing technologies mostly focus on a single domain or a single objective, lacking systematic integration; data fusion and processing technologies are still imperfect, making it difficult to effectively handle multimodal heterogeneous data; the algorithm models lack adaptability and robustness in the complex and dynamic urban environment, and the interpretability of the models is poor, making it difficult to meet the actual needs of urban resource allocation decision-making.

[0004] There is an urgent need for a new solution to break through the limitations of traditional technologies, fully utilize the advantages of emerging technologies, solve key problems in the optimal allocation of urban resources, achieve efficient, intelligent, and collaborative allocation of urban resources, and promote sustainable urban development. Summary of the Invention

[0005] To address the aforementioned problems, this invention provides a method and system for optimizing urban resource allocation based on multi-objective collaboration.

[0006] Firstly, the present invention provides a method for optimizing the allocation of urban resources based on multi-objective collaboration, which adopts the following technical solution:

[0007] A method for optimizing urban resource allocation based on multi-objective collaboration includes:

[0008] Obtain data related to urban resources;

[0009] Preprocess the acquired urban resource-related data;

[0010] A dynamic spatiotemporal knowledge graph is constructed based on preprocessed urban resource-related data;

[0011] A reinforcement learning-neural network hybrid optimization model is constructed, and a dynamic spatiotemporal knowledge graph is used as the embedded representation of the model;

[0012] Quantum computing and classical optimization algorithms are used to perform multi-objective training and optimization of the model;

[0013] Resource allocation decisions are generated based on the optimized model.

[0014] Furthermore, the acquisition of urban resource-related data includes constructing a multi-source data collection network, adding urban micro-environment sensor and social media data in addition to traditional data sources, and classifying the data into four categories: spatiotemporal data, structured data, text data, and image data.

[0015] Furthermore, the preprocessing of the acquired urban resource-related data includes: using a spatiotemporal graph convolutional network to extract node feature matrices from the spatiotemporal graph of urban resources composed of nodes and edges, and performing convolution operations through convolution kernels; using a Transformer-based BERT model to extract features from text data, converting the text input into a sequence of word vectors, and calculating the context representation of each word through a multi-head attention mechanism; and using an improved ResNet model to extract features from image data, introducing a self-attention module, adding attention weight calculation in the residual block, and enhancing the ability to extract key image features.

[0016] Furthermore, the construction of a dynamic spatiotemporal knowledge graph based on preprocessed urban resource-related data includes using an attention-based BiLSTM-CRF model to perform entity recognition on text data and employing a joint learning framework for complex relation extraction; designing a multi-dimensional graph architecture, using an R-tree index structure in the spatial dimension, storing the graph state corresponding to the time interval of the time slice in the temporal dimension, and associating the graph entities in the image data domain, using a deep learning model to extract image feature vectors and entity feature vectors, and calculating the correlation degree through cosine similarity.

[0017] Furthermore, the construction of a dynamic spatiotemporal knowledge graph based on preprocessed urban resource-related data also includes setting a triggering mechanism driven by multi-source data, triggering graph updates through priority weights and setting trigger thresholds; and after the graph update, using a graph editing distance algorithm to calculate the difference before and after the graph update, using a graph neural network (GNN) to update node features based on the graph differences, and predicting potential entities and relationships through multi-layer iterative learning to update the knowledge graph and achieve evolutionary reasoning.

[0018] Furthermore, the construction of the reinforcement learning-neural network hybrid optimization model uses a dynamic spatiotemporal knowledge graph as the embedded representation of the model. This includes introducing a dual-network structure and a competitive network structure on the basis of DQN. The dual-network structure reduces the problem of Q-value overestimation by separating the target network and the evaluation network. The Q-network is divided into an advantage function and a value function through the network structure, thereby calculating the Q-values ​​of the target network and the evaluation network, which is used to efficiently learn the value of the state and the advantage of the action.

[0019] Furthermore, the construction of the reinforcement learning-neural network hybrid optimization model, which uses the dynamic spatiotemporal knowledge graph as the model's embedded representation, also includes optimizing the Transformer encoder for the long sequence dependencies and complex semantic relationships of urban resource allocation data. In the multi-head attention mechanism, a dynamic weight adjustment strategy is introduced to assign dynamic weights to the attention calculations of different heads based on the importance and association strength of entities in the knowledge graph. Finally, the embedded representation of the dynamic spatiotemporal knowledge graph is deeply fused with other data features. Specifically, a graph attention network (GAT) is used to learn the embedding of the knowledge graph to obtain node embedding vectors, and the knowledge graph embedding vectors are fused with data features through a gating mechanism.

[0020] Furthermore, the multi-objective training and optimization of the model using quantum computing and classical optimization algorithms includes introducing a hybrid framework of quantum computing and classical optimization algorithms, leveraging the parallelism of quantum computing to accelerate the search process; employing a quantum genetic algorithm to perform a global search for model parameters, using qubit encoding to represent model parameters, and representing the state of quantum individuals with probability amplitudes; combining a dynamic spatiotemporal knowledge graph to perceive the spatiotemporal changes of urban resources in real time; and introducing a dynamic weight adjustment strategy for the multi-objectives in urban resource allocation, adjusting model parameters to ensure that the allocation scheme always approaches the optimal Pareto front, thereby achieving dynamic balance optimization of multiple objectives.

[0021] Furthermore, the process of generating resource allocation decisions based on the optimized model includes inputting preprocessed real-time data into an optimized reinforcement learning-neural network hybrid model. The model generates multiple candidate resource allocation schemes based on the learned strategies and the current state of urban resources. A multi-dimensional evaluation index system is constructed to comprehensively evaluate the candidate schemes. The Analytic Hierarchy Process (AHP) is used to determine the weights of each evaluation index, calculate the comprehensive score of each candidate scheme, rank the candidate schemes according to the comprehensive scores, and select the scheme with the highest score as the final resource allocation decision scheme.

[0022] Secondly, a multi-objective collaborative urban resource optimization and allocation system includes:

[0023] The data acquisition module is configured to acquire data related to urban resources;

[0024] The preprocessing module is configured to preprocess the acquired urban resource-related data.

[0025] The graph module is configured to construct a dynamic spatiotemporal knowledge graph based on preprocessed urban resource-related data;

[0026] The embedding module is configured to construct a reinforcement learning-neural network hybrid optimization model, using a dynamic spatiotemporal knowledge graph as the embedded representation of the model;

[0027] The optimization module is configured to perform multi-objective training optimization of the model using quantum computing and classical optimization algorithms.

[0028] The decision module is configured to generate resource allocation decisions based on the optimized model.

[0029] Thirdly, the present invention provides a computer-readable storage medium storing a plurality of instructions adapted for loading and execution by a processor of a terminal device of the aforementioned method for optimizing urban resource allocation based on multi-objective collaboration.

[0030] Fourthly, the present invention provides a terminal device, including a processor and a computer-readable storage medium, wherein the processor is used to implement various instructions; the computer-readable storage medium is used to store multiple instructions, the instructions being adapted to be loaded and executed by the processor to provide the aforementioned method for optimizing urban resource allocation based on multi-objective collaboration.

[0031] In summary, the present invention has the following beneficial technical effects:

[0032] This technical solution focuses on the optimal allocation of urban resources, integrating multi-source data to construct a knowledge graph and hybrid model, achieving multi-objective optimization and dynamic decision-making. Its technical effects are significant, as detailed below:

[0033] 1. By constructing a multi-source data acquisition network, integrating multimodal heterogeneous data such as spatiotemporal, structured, text, and images, and employing spatiotemporal graph convolutional networks, BERT models, and improved ResNet models for feature extraction and fusion, urban resource information can be obtained more comprehensively and accurately. Simultaneously, the construction of a dynamic spatiotemporal knowledge graph enables refined extraction of multi-granular entities and relationships. Through intelligent dynamic update and maintenance mechanisms, it reflects the spatiotemporal changes of urban resources in real time, providing rich and accurate knowledge support for subsequent decision-making.

[0034] 2. A reinforcement learning-neural network hybrid optimization model, combining a dual-network structure, a competitive network structure, and optimization of the Transformer encoder, enhances the model's ability to learn urban resource allocation strategies, focuses more on key resource information, and improves the accuracy of strategy generation. During training, methods such as priority experience replay, curriculum learning strategies, and adversarial training are employed to accelerate training convergence and enhance the model's generalization ability. Hierarchical policy generation and multi-agent collaborative decision-making mechanisms make resource allocation decisions more systematic and rational, while a knowledge graph-based decision explanation mechanism improves the interpretability of decisions and enhances decision-makers' trust.

[0035] 3. A hybrid framework combining quantum computing and classical optimization algorithms is introduced, along with meta-learning-guided transfer optimization, to quickly find the optimal solution for model parameters and improve model adaptability. Spatiotemporally dynamic parameter adjustment and multi-objective dynamic balance optimization enable the model to adapt promptly to spatiotemporal changes in urban resources and multi-objective needs. Multi-dimensional evaluation and feedback optimization integrates cross-domain knowledge and combines human-machine collaborative feedback to continuously optimize resource allocation schemes. Through a complete process of resource allocation scheme implementation, monitoring, effect evaluation, and feedback optimization, a closed-loop system is formed to continuously improve the efficiency and effectiveness of urban resource allocation. Attached Figure Description

[0036] Figure 1 is a schematic diagram of a method for optimizing urban resource allocation based on multi-objective collaboration according to Embodiment 1 of the present invention. Detailed Implementation

[0037] The present invention will be further described in detail below with reference to the accompanying drawings.

[0038] Example 1

[0039] Referring to Figure 1, a method for optimizing urban resource allocation based on multi-objective collaboration in this embodiment includes:

[0040] Obtain data related to urban resources;

[0041] Preprocess the acquired urban resource-related data;

[0042] A dynamic spatiotemporal knowledge graph is constructed based on preprocessed urban resource-related data;

[0043] A reinforcement learning-neural network hybrid optimization model is constructed, and a dynamic spatiotemporal knowledge graph is used as the embedded representation of the model;

[0044] Quantum computing and classical optimization algorithms are used to perform multi-objective training and optimization of the model;

[0045] Resource allocation decisions are generated based on the optimized model.

[0046] Specifically:

[0047] Complete technical solutions for optimal allocation of urban resources

[0048] I. Multimodal Heterogeneous Data Fusion and Preprocessing

[0049] (I) Data Collection and Classification

[0050] A multi-source data acquisition network was constructed, adding urban micro-environment sensors (such as air quality and noise monitoring equipment) and social media data (discussions by residents about resource needs) to traditional data sources. The data was divided into four categories: spatiotemporal data (spatial information on land coordinates and infrastructure distribution), structured data (project approval data and real estate registration information), text data (policy documents and public opinion feedback), and image data (satellite remote sensing images and building exterior pictures).

[0051] (II) Feature Extraction and Fusion

[0052] (1) Spatiotemporal data: A spatiotemporal graph convolutional network (ST-GCN) is used. For the spatiotemporal graph of urban resources G=(V,E) composed of nodes V and edges E, the node feature matrix X∈R N×C (N is the number of nodes, C is the feature dimension), through convolution kernel W∈R K×C×F (K is the kernel size, F is the output feature dimension) Perform convolution operation:

[0053]

[0054] Where Yi,f is the output of node i in feature dimension f, and N(i) is the set of neighboring nodes of node i.

[0055] (2) Text data: Feature extraction was performed using the Transformer-based BERT model, transforming the text input into a word vector sequence T = [t1, t2, ..., tn]. The contextual representation of each word was calculated using a multi-head attention mechanism.

[0056]

[0057] Where Q, K, and V are the query, key, and value vectors, respectively, and dk is the dimension of the key vector.

[0058] (3) Image data: Using the improved ResNet model, a self-attention module is introduced to increase the calculation of attention weights in the residual block, thereby enhancing the ability to extract key image features.

[0059] The extracted features are then fused using a gated fusion mechanism:

[0060] F = σ(G1X1 + G2X2 + G3X3 + G4X4) where X1, X2, X3, and X4 are spatiotemporal, structured, text, and image data features, respectively; G1, G2, G3, and G4 are learnable gating matrices; and σ is the activation function.

[0061] II. Construction of Dynamic Spatiotemporal Knowledge Graph

[0062] (I) Fine-grained extraction of multi-granular entities and relations

[0063] Hierarchical entity extraction: Entity recognition is performed using a BiLSTM-CRF model based on an attention mechanism. Let the input text sequence be x = [x1, x2, ..., xn]. The BiLSTM forward propagation calculates the hidden states. Backpropagation to compute hidden state Obtain the bidirectional hidden state Multi-head attention mechanism calculates attention weights αt,i:

[0064]

[0065] e t,i =score(h t ,h i )

[0066] Where 'score' is the similarity calculation function, and the attention output is obtained through weighted summation. Input zt into the CRF layer and calculate the probability of the label sequence y = [y1, y2, ..., yn]:

[0067]

[0068] Where A is the transition matrix, P is the emission matrix, and Y is the transition matrix. x The set of all possible label sequences.

[0069] (2) Complex Relation Extraction: A joint learning framework is adopted, with entity extraction loss function denoted as Lentity, relation classification loss function denoted as Lrelation, and joint loss function denoted as Lrelation.

[0070] L=λ1L entity +λ2L relation , where λ1 and λ2 are weighting coefficients.

[0071] For the derivation of implicit relationships, a probabilistic graphical model, the Bayesian network, is used. Let the set of variables be X = X1, X2, ..., Xn. The conditional probability P(Xi|Xi) is calculated using Bayes' theorem:

[0072]

[0073] Where X -i Indicates the difference from X i The set of other variables besides.

[0074] (II) Design of High-Dimensional Heterogeneous Dynamic Spatiotemporal Knowledge Graph Structure

[0075] (1) Multi-dimensional graph architecture: In the spatial dimension, an R-tree index structure is adopted. For a spatial object o, its minimum bounding rectangle (MBR) is MBR(o) = (x min ,y min ,x max ,y max In an R-tree, node n contains a set of child node pointers P and an MBR(n) covering the child nodes. In the time dimension, time slice t corresponds to the time interval [start]. t end t ], stores the spectral state St within this interval.

[0076] (2) Heterogeneous data fusion structure: For the association between image data and map entities, a deep learning model is used to extract the image feature vector fimage and the entity feature vector fentity, and the association degree is calculated by cosine similarity. A connection is established when the sim value exceeds the threshold.

[0077] (III) Intelligent Dynamic Update and Maintenance Mechanism

[0078] (1) Triggering mechanism driven by multi-source data: Let the data type set be D = d1, d2, ..., dm, and the corresponding priority weights be w = w1, w2, ..., wm. The priority of data di is pi = wi * si, where si is the data importance score. When data di arrives and pi exceeds the trigger threshold T, the graph update is triggered.

[0079] (2) Incremental Update and Conflict Resolution Strategy: The graph edit distance algorithm is used to calculate the difference between the graph before and after the update. Let graphs G1 = (V1, E1) and G2 = (V2, E2). The graph edit distance d(G1, G2) is obtained by minimizing the cost of node and edge insertion, deletion, and modification operations.

[0080]

[0081] Where ops is the sequence of operations, and cost(op) is the cost of operation op.

[0082] (3) Automatic Evolution and Reasoning of Knowledge Graphs: Reasoning is performed using Graph Neural Networks (GNNs), assuming the node feature matrix H... l For the features of the l-th layer nodes, update the node features through graph convolution operations:

[0083]

[0084] Where N(u) is the set of neighboring nodes of node u, Cu,v is the normalization constant, and W l Let b be the weight matrix. lThis is the bias vector. Through multi-layer iterative learning, latent entities and relationships are predicted, and the knowledge graph is updated.

[0085] III. Reinforcement Learning-Neural Network Hybrid Optimization Model

[0086] (I) In-depth design of model architecture

[0087] (1) Deep Q-Network (DQN) Enhancement Structure

[0088] Building upon traditional DQN, this paper introduces a double-network structure (DoubleDQN) and a competing-network structure (DuelingDQN). The double-network structure reduces Q-value overestimation by separating the target network and the evaluation network. Let the parameters of the evaluation network be θ, and the parameters of the target network be θ₀. - Every certain number of steps, the parameters of the evaluation network are copied to the target network. When calculating the target Q-value, the evaluation network selects an action, and the target network calculates the Q-value.

[0089]

[0090] The competitive network structure divides the Q-network into two parts: an advantage function A(s,a;θ) and a value function V(s;θ). This is achieved through the formula...

[0091]

[0092] Calculate the Q-value, where |A| is the size of the action space. This structure enables more efficient learning of the value of states and the advantages of actions, improving the model's ability to learn urban resource allocation strategies.

[0093] (2) Transformer encoder optimization

[0094] To address the long-sequence dependencies and complex semantic relationships in urban resource allocation data, the Transformer encoder is optimized. A dynamic weight adjustment strategy is introduced into the multi-head attention mechanism. Dynamic weights are assigned to the attention of different heads based on the importance and association strength of entities in the knowledge graph. Let the importance score of entity ei be Iei, and the weights are calculated using the formula... Calculate the weight βh of the h-th attention head, where αh,i is the weight of the h-th attention head when calculating the attention of entity ei. The final attention output is:

[0095] This allows the model to focus more on key resource information and improves the accuracy of strategy generation.

[0096] (3) Knowledge graph embedding and fusion

[0097] The embedding representation of the dynamic spatiotemporal knowledge graph is deeply integrated with other data features. A Graph Attention Network (GAT) is used to learn the embedding of the knowledge graph, obtaining the node embedding vector hv. For each node v, the embedding vector is calculated using formula e. ij =LeakyReLU(a T [Wh i ||Wh j Calculate the attention coefficients eij between nodes i and j, where a is the attention vector and W is the learnable weight matrix. The attention weights are then obtained after normalization. This leads to node embedding. Knowledge graph embedding vectors and data features are fused using a gating mechanism: F fusion =σ(G kg h kg +G data F data ), where hkg is the knowledge graph embedding vector, Fdata is other data features, and Gkg and Gdata are learnable gating matrices, realizing the organic combination of knowledge graph information and resource data, and providing richer information for model decision-making.

[0098] (II) Innovation and Optimization of Training Process

[0099] (1) Priority Experience Review

[0100] The experience replay mechanism is improved by adopting Prioritized Experience Replay. A priority is assigned to each experience sample based on the sample's TD error δ=|yQ(s,a;θ)|. Where β is a parameter controlling the degree of influence of priority (0 < β ≤ 1). During sampling, samples are drawn from the experience pool according to the priority ratio, as shown in the formula: Here, α is a parameter that adjusts the sampling bias (0 < α ≤ 1). This approach enables the model to learn important experience samples more frequently, accelerating training convergence, and is particularly suitable for learning complex and critical decision-making scenarios in urban resource allocation.

[0101] Course learning strategies

[0102] (2) A curriculum learning strategy is introduced, designing a sequence of training tasks from simple to complex based on the difficulty of the urban resource allocation problem. In the early stages of training, simple resource allocation scenarios, such as land use adjustment in a single area, are set up to allow the model to quickly learn basic strategies. As training progresses, the difficulty of the tasks is gradually increased, introducing complex scenarios such as multi-area collaborative allocation and consideration of multiple resource types. This approach helps the model gradually build complex decision-making capabilities, avoiding learning difficulties or getting stuck in local optima due to excessively high task difficulty in the early stages of training.

[0103] (3) Adversarial training enhances generalization

[0104] Adversarial training is employed to improve the model's generalization ability. An adversarial network is introduced, which attempts to generate perturbation data that causes the main model to make incorrect decisions. The main model then strives to identify the perturbation data and maintain correct decisions. During training, the main model's loss function is augmented with an adversarial loss term, Ladv, and the main model and adversarial network are optimized through a minimax game. The adversarial loss function is Ladv = \mathbbEx\simpdata[\log(1-D(x))] + \mathbbEx'\simpadv[\log(D(x'))], where D is the adversarial network (discriminator), pdata is the distribution of the real data, and padv is the distribution of perturbation data generated by the adversarial network. Adversarial training enhances the model's robustness against uncertainties and anomalies in urban resource allocation.

[0105] (III) Strategy Generation and Decision Optimization

[0106] (1) Generation of hierarchical strategies

[0107] A hierarchical strategy generation framework is constructed, dividing urban resource allocation decisions into three levels: macro, meso, and micro. The macro level formulates an overall resource allocation strategy, such as the resource allocation ratio for different functional areas of the city; the meso level, based on the macro strategy, determines the resource allocation plan for specific areas, such as the project portfolio of a commercial district; and the micro level, for specific projects, formulates detailed resource usage plans, such as the material procurement and construction arrangements for individual buildings.

[0108] The generation of strategies at each level is handled by different sub-networks, which collaborate through information transmission and feedback mechanisms. The macro-level strategy layer provides guidance to the meso-level, while the meso-level strategies are further refined and fed back to adjust the macro-level strategies. The micro-level strategies formulate specific implementation plans based on the requirements of the meso-level and feed back the implementation status to the meso and macro levels, forming a multi-level, collaborative strategy generation system that improves the systematicness and rationality of resource allocation decisions.

[0109] (2) Multi-agent cooperative decision-making

[0110] A multi-agent reinforcement learning mechanism is introduced, treating different types of resources (such as land, funds, and manpower) or different management departments as independent agents. Each agent makes decisions under its own goals and constraints, while interacting with other agents through communication and cooperation mechanisms. A joint action space A = A1 × A2 × … × A is adopted. n The joint reward function represents the combination of actions of all agents. Taking into account the rewards of each agent, where Ri is the reward for the i-th agent and wi is the weighting coefficient, the agents are trained to cooperate and jointly optimize the allocation of urban resources. For example, the land resource agent and the financial agent collaborate to allocate funds rationally while ensuring the quality of land development, thereby maximizing resource utilization efficiency.

[0111] (3) Explainable decision support

[0112] To improve the interpretability of decision-making, a knowledge graph-based decision explanation mechanism is introduced. After generating resource allocation strategies, the entity relationships and rules in the knowledge graph provide an explanatory basis for the decision. For example, when the model decides to increase the construction of public facilities in a certain area, relevant information such as the population density of the area, the service coverage of existing facilities, and policy and regulatory requirements for public facility allocation are extracted from the knowledge graph. Natural language generation technology is then used to explain the reasons for the decision to decision-makers in an easy-to-understand way, enhancing their trust and understanding of the model's decisions and facilitating the implementation and adjustment of decisions in practical applications.

[0113] IV. Model Optimization and Dynamic Adjustment

[0114] (I) Algorithm Integration, Innovation and Optimization

[0115] (1) A hybrid framework combining quantum computing and classical optimization algorithms is introduced, leveraging the parallelism of quantum computing to accelerate the search process. A quantum genetic algorithm (QGA) is used for global search of the model parameters, with qubits encoded to represent these parameters, and the states of individual quantum units represented by probability amplitudes. In quantum gate operations, the probability amplitudes of the qubits are adjusted via a rotation gate U(θ):

[0116]

[0117] Where α and β are the probability amplitudes of the qubits, and θ is the rotation angle. In each generation of evolution, the quantum state is collapsed into a classical solution through quantum measurement, and the solution is locally optimized by combining classical local search algorithms (such as simulated annealing), forming a quantum-classical co-optimization mechanism. This mechanism quickly finds the optimal solution for the model parameters, improving the model's adaptability in complex urban resource configuration scenarios.

[0118] (2) Meta-learning-guided transfer optimization

[0119] A meta-learning framework is constructed to learn optimization patterns in resource allocation tasks across different city regions or time periods. The gradient-based meta-learning algorithm MAML is employed. During the meta-training phase, for multiple source tasks Tsource, the adapted parameters θ' are obtained through a small number of gradient updates.

[0120]

[0121] Where α is the learning rate. This is the loss function for the source task. When faced with a new target task... At that time, the parameters θ' obtained by meta-learning are used as initial parameters, and the target task data are combined for rapid optimization. Transfer learning is used to accelerate the convergence speed of the model in new scenarios, reducing training time and data requirements.

[0122] (II) Dynamic Adaptive Adjustment Mechanism

[0123] (1) Parameter adjustment of spatiotemporal dynamic sensing

[0124] By combining a dynamic spatiotemporal knowledge graph, the spatiotemporal changes of urban resources are perceived in real time. A spatiotemporal attention mechanism is designed to dynamically assign weights to model parameters based on the spatiotemporal association information of entities in the knowledge graph. For the time dimension, the attention weight ωt for different time slices is calculated:

[0125]

[0126] Where ε is the set of entities in the knowledge graph, and TimeScore(e,t) represents the time-related score of entity e at time t. For the spatial dimension, the influence weights of spatial neighborhood entities are calculated based on the R-tree index structure, and finally, a comprehensive spatiotemporal weight is obtained, which is used to adjust the step size and direction of model parameter updates, so that the model can adapt to the new demands arising from spatiotemporal changes in urban resource allocation in a timely manner.

[0127] (2) Multi-objective dynamic equilibrium optimization

[0128] To address the multiple objectives (economic, social, and ecological benefits) in urban resource allocation, a dynamic weight adjustment strategy is introduced. An objective importance assessment model is established, using a combination of the Analytic Hierarchy Process (AHP) and fuzzy comprehensive evaluation method based on real-time data and policy guidance to dynamically calculate the weights of each objective. When external environmental changes occur (such as policy adjustments or unforeseen events), the objective weights are reassessed, and the model's optimization direction is adjusted. Simultaneously, a Pareto front dynamic tracking algorithm is employed to monitor changes in the Pareto front in the objective space in real time. By adjusting model parameters, the allocation scheme is kept close to the optimal Pareto front, achieving dynamic balance optimization across multiple objectives.

[0129] (III) Multi-dimensional evaluation and feedback optimization

[0130] (1) Cross-domain knowledge integration assessment

[0131] A cross-domain knowledge fusion evaluation system is constructed, integrating knowledge from economics, sociology, ecology, and other fields in addition to traditional resource allocation indicators. Complex network analysis methods are introduced, treating the urban resource allocation system as a complex network, and evaluating the importance and influence of resource nodes in the system by calculating indicators such as degree centrality and betweenness centrality. Simultaneously, ecological models (such as the ecological footprint model) are used to assess the impact of resource allocation on the ecological environment. These multi-domain evaluation results are fed back into model optimization, enabling the model to not only focus on short-term benefits but also comprehensively consider long-term system stability and sustainability.

[0132] (2) Human-machine collaborative feedback optimization

[0133] A human-machine collaborative feedback mechanism is established, allowing decision-makers to intuitively view the model's resource allocation schemes and evaluation results through a visual interface, and provide feedback based on practical experience and professional knowledge. Natural language processing technology is used to transform the decision-makers' feedback into parameter adjustment signals that the model can understand. Combined with the model's own evaluation results, this forms a two-way feedback optimization loop. For example, if a decision-maker believes that a public facility allocation scheme for a certain area does not conform to the local residents' living habits, the feedback information is processed, the model adjusts the relevant parameters, and regenerates the allocation scheme, achieving complementary advantages between humans and machines and improving the rationality and feasibility of the resource allocation scheme.

[0134] V. Resource Allocation Decision Generation Based on Optimization Model

[0135] (I) Configuration Scheme Generation

[0136] Preprocessed real-time data is input into an optimized reinforcement learning-neural network hybrid model. Based on the learned strategy and the current state of urban resources, the model generates multiple candidate resource allocation schemes. Each scheme includes detailed information such as land use adjustment planning, project construction priority arrangements, and resource allocation quantities. For example, for land resource allocation, the scheme specifies the development type (residential, commercial, public facilities, etc.) and development intensity of different plots; for project resource allocation, it determines the start time and funding allocation amount for each project.

[0137] (II) Scheme Evaluation and Selection

[0138] A multi-dimensional evaluation index system will be constructed to comprehensively evaluate candidate solutions. Evaluation indicators include, but are not limited to: economic benefit indicators (such as expected land appreciation gains and project investment return rate), social benefit indicators (such as the number of new jobs created and the degree of improvement in residents' living convenience), ecological benefit indicators (such as the reduction in carbon emissions and the increase in green coverage), and implementation feasibility indicators (such as policy compliance, technological feasibility, and difficulty in raising funds).

[0139] The weights of each evaluation indicator were determined using the Analytic Hierarchy Process (AHP), and the overall score for each candidate solution was calculated.

[0140]

[0141] Where Scorej is the overall score of the j-th candidate solution, wi is the weight of the ith evaluation metric, and Scoreij is the score of the j-th candidate solution on the ith evaluation metric. Candidate solutions are ranked according to their overall scores, and the solution with the highest score is selected as the final resource allocation decision.

[0142] VI. Implementation and Monitoring of Resource Allocation Plan

[0143] (I) Implementation of the Plan

[0144] A detailed implementation plan for resource allocation should be developed, clearly defining the responsibilities, tasks, and timelines for each department. For example, the land management department is responsible for approving land use changes and land transfers; the construction department is responsible for project construction supervision; and the finance department is responsible for fund allocation and usage supervision. During implementation, a cross-departmental collaborative working mechanism should be established to ensure close cooperation among all stages and guarantee the smooth progress of the resource allocation plan.

[0145] (II) Real-time monitoring

[0146] By leveraging IoT devices, big data platforms, and dynamic spatiotemporal knowledge graphs, the implementation process of resource allocation plans is monitored in real time. Information such as land development progress, project construction status, and resource usage data is collected in real time and compared with the planned schedule. For example, IoT sensors monitor construction progress at building sites, comparing actual progress with planned progress; big data platforms analyze fund usage to ensure funds are allocated and used reasonably according to the budget. If deviations from the plan are detected, early warning signals are issued promptly, and emergency adjustment mechanisms are activated.

[0147] VII. Effectiveness Evaluation and Feedback Optimization

[0148] (I) Effectiveness Evaluation

[0149] After the resource allocation plan is implemented, a comprehensive effectiveness evaluation will be conducted. A combination of quantitative and qualitative methods will be used to assess the achievement of economic, social, and ecological benefits. For example, economic benefits will be evaluated by statistically analyzing actual land appreciation and project revenue; social benefits will be assessed by understanding residents' satisfaction with the improved living environment through questionnaires and on-site interviews; and ecological benefits will be assessed by monitoring changes in environmental indicators (such as air quality and water quality).

[0150] (II) Feedback Optimization

[0151] Based on the effectiveness evaluation results, the problems and shortcomings of the resource allocation plan are analyzed. The evaluation results are fed back to various stages such as data collection, model training, and optimization adjustments to iteratively optimize the technical solution. For example, if it is found that the resource allocation in a certain area fails to achieve the expected economic benefits, the data characteristics of that area are re-analyzed, model parameters or optimization algorithms are adjusted, and a new resource allocation plan is generated, forming a closed-loop urban resource optimization and allocation system to continuously improve the efficiency and effectiveness of urban resource allocation.

[0152] This solution fully presents the entire process from data processing to resource optimization and configuration. If you have any ideas for adjusting the steps or algorithms in the solution, or would like to add specific details, please feel free to raise them.

[0153] Example 2

[0154] This embodiment provides a multi-objective collaborative urban resource optimization and allocation system, including:

[0155] The data acquisition module is configured as follows:

[0156] A computer-readable storage medium storing a plurality of instructions adapted for loading and execution by a processor of a terminal device of the aforementioned method for optimizing urban resource allocation based on multi-objective collaboration.

[0157] A terminal device includes a processor and a computer-readable storage medium, the processor being used to implement various instructions; the computer-readable storage medium being used to store multiple instructions, the instructions being adapted to be loaded and executed by the processor to provide a method for optimizing urban resource allocation based on multi-objective collaboration.

[0158] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for optimizing urban resource allocation based on multi-objective collaboration, characterized in that, include: Obtain data related to urban resources; Preprocess the acquired urban resource-related data; A dynamic spatiotemporal knowledge graph is constructed based on preprocessed urban resource-related data; A reinforcement learning-neural network hybrid optimization model is constructed, and a dynamic spatiotemporal knowledge graph is used as the embedded representation of the model; A multi-objective training and optimization model is performed using quantum computing and classical optimization algorithms. Resource allocation decisions are generated based on the optimized model. The construction of a reinforcement learning-neural network hybrid optimization model uses a dynamic spatiotemporal knowledge graph as the model's embedded representation. This includes introducing a dual-network structure and a competitive network structure on top of DQN. The dual-network structure reduces Q-value overestimation by separating the target network and the evaluation network. The competitive network structure divides the Q-network into an advantage function and a value function, thereby calculating the Q-values ​​of the target network and the evaluation network for efficiently learning the value of states and the advantages of actions. The construction of the reinforcement learning-neural network hybrid optimization model, using a dynamic spatiotemporal knowledge graph as the model's embedded representation, also includes optimizing the Transformer encoder for the long-sequence dependencies and complex semantic relationships in urban resource allocation data. In the multi-head attention mechanism, a dynamic weight adjustment strategy is introduced based on the importance of entities in the knowledge graph. The system assigns dynamic weights to the attention calculations of different heads based on the correlation strength. Finally, it deeply integrates the embedded representation of the dynamic spatiotemporal knowledge graph with other data features. Specifically, it uses a graph attention network (GAT) to learn the knowledge graph's embeddings, obtaining node embedding vectors, and then fuses these embedding vectors with data features through a gating mechanism. The system employs quantum computing and classical optimization algorithms for multi-objective training and optimization. This includes introducing a hybrid framework of quantum computing and classical optimization algorithms, leveraging the parallelism of quantum computing to accelerate the search process; using a quantum genetic algorithm to perform a global search for model parameters, encoding qubits to represent model parameters, and representing the state of quantum individuals with probability amplitudes; combining the dynamic spatiotemporal knowledge graph to perceive real-time spatiotemporal changes in urban resources; and introducing a dynamic weight adjustment strategy for multi-objective urban resource allocation, adjusting model parameters to ensure the allocation scheme always approaches the optimal Pareto front, achieving dynamic balance optimization across multiple objectives.

2. The urban resource optimization allocation method based on multi-objective collaboration according to claim 1, characterized in that, The acquisition of urban resource-related data includes constructing a multi-source data collection network, adding urban micro-environment sensor data and social media data in addition to traditional data sources, and classifying the data into four categories: spatiotemporal data, structured data, text data, and image data.

3. The urban resource optimization allocation method based on multi-objective collaboration according to claim 2, characterized in that, The preprocessing of the acquired urban resource-related data includes: using a spatiotemporal graph convolutional network to extract node feature matrices from the spatiotemporal graph of urban resources composed of nodes and edges, and performing convolution operations through convolution kernels; using a Transformer-based BERT model to extract features from text data, converting the text input into a sequence of word vectors, and calculating the context representation of each word through a multi-head attention mechanism; and using an improved ResNet model to extract features from image data, introducing a self-attention module, adding attention weight calculation in the residual block, and enhancing the ability to extract key image features.

4. The urban resource optimization allocation method based on multi-objective collaboration according to claim 3, characterized in that, The construction of a dynamic spatiotemporal knowledge graph based on preprocessed urban resource-related data includes using an attention-based BiLSTM-CRF model to perform entity recognition on text data and employing a joint learning framework for complex relation extraction; designing a multi-dimensional graph architecture, using an R-tree index structure in the spatial dimension, storing the graph state corresponding to the time interval of the time slice in the temporal dimension, associating image data with graph entities, extracting image feature vectors and entity feature vectors using a deep learning model, and calculating the correlation degree through cosine similarity.

5. The urban resource optimization allocation method based on multi-objective collaboration according to claim 4, characterized in that, The construction of a dynamic spatiotemporal knowledge graph based on preprocessed urban resource-related data also includes setting a triggering mechanism driven by multi-source data, which triggers graph updates by setting priority weights and trigger thresholds. After the graph is updated, the graph editing distance algorithm is used to calculate the difference before and after the graph update. The graph neural network (GNN) is used to update the node features based on the graph difference. Through multi-layer iterative learning, potential entities and relationships are predicted, and the knowledge graph is updated to achieve evolutionary reasoning.

6. The urban resource optimization allocation method based on multi-objective collaboration according to claim 5, characterized in that, The process of generating resource allocation decisions based on the optimized model includes inputting preprocessed real-time data into an optimized reinforcement learning-neural network hybrid model, which generates multiple candidate resource allocation schemes based on the learned strategies and the current state of urban resources; and constructing a multi-dimensional evaluation index system to comprehensively evaluate the candidate schemes. The Analytic Hierarchy Process (AHP) was used to determine the weights of each evaluation indicator, calculate the comprehensive score of each candidate solution, rank the candidate solutions according to the comprehensive scores, and select the solution with the highest score as the final resource allocation decision solution.

7. A multi-objective collaborative urban resource optimization allocation system, executing the multi-objective collaborative urban resource optimization allocation method as described in any one of claims 1-6, characterized in that, include: The data acquisition module is configured to acquire data related to urban resources; The preprocessing module is configured to preprocess the acquired urban resource-related data; the graph module is configured to construct a dynamic spatiotemporal knowledge graph based on the preprocessed urban resource-related data. The embedding module is configured to construct a reinforcement learning-neural network hybrid optimization model, using a dynamic spatiotemporal knowledge graph as the model's embedding representation; the optimization module is configured to perform multi-objective training and optimization of the model using quantum computing and classical optimization algorithms; and the decision module is configured to generate resource allocation decisions based on the optimized model.

Images