Smart city management system integrating artificial intelligence and big data

Abstract

This invention discloses a smart city management system integrating artificial intelligence and big data, relating to the field of smart city management technology. It includes: a data acquisition and feature extraction module for acquiring and processing multi-source urban data, extracting road network topology features, functional area distribution features, and pedestrian flow dynamics features through graph neural networks, generating and storing structured feature vectors; a cross-domain knowledge transfer model construction module for building a city simulator, generating virtual data through adversarial generative networks and reinforcement learning, bridging the source and target domains based on the virtual data, and utilizing graph matching and optimal transport theory to achieve knowledge transfer from the source domain to the target domain; and a model optimization and alignment module for aligning the feature distributions of the source and target domains through adversarial training with a domain discriminator and gradient inversion layer, dynamically adjusting data weights to achieve seamless transfer from virtual distributions to the real distributions of the target domain, and realizing seamless knowledge transfer from mature urban areas to newly built urban areas.

Description

Technical Field

[0001] This invention relates to the field of smart city management technology, specifically to a smart city management system that integrates artificial intelligence and big data. Background Technology

[0002] With the rapid development of smart city construction, artificial intelligence and big data technologies are increasingly being applied in urban management. Existing technologies typically train AI models based on historical data from mature urban areas. By analyzing multi-source data such as road network topology, POI distribution, and pedestrian flow dynamics, they enable intelligent management of tasks such as traffic flow prediction and functional area planning. For example, traditional methods use graph neural networks to extract urban spatial features and combine them with time series models to capture dynamic patterns, which performs well in mature urban areas with sufficient data.

[0003] However, existing technologies face significant shortcomings: First, newly developed urban areas lack historical data accumulation, resulting in missing spatiotemporal characteristics. Traditional AI models struggle to accurately capture their dynamic evolution patterns, leading to substantial biases in prediction results. Second, when models trained in mature urban areas are directly transferred to newly developed urban areas, their performance deteriorates significantly due to domain offset issues, creating a small-sample learning dilemma. For example, the road network structure and pedestrian flow distribution in newly developed urban areas differ significantly from those in mature urban areas, and existing models cannot effectively adapt to these distributional differences. Furthermore, existing technologies lack sufficient coverage of long-tail scenarios. Newly developed urban areas may experience data sparsity or uneven distribution in their early stages of development, making it difficult for traditional methods to learn highly generalizable features from limited data, resulting in prediction failures for rare scenarios, such as sudden pedestrian peaks. Although some studies have attempted to alleviate this problem through data augmentation or transfer learning, the lack of systematic integration of virtual data generation and cross-domain knowledge transfer makes it difficult to achieve seamless knowledge transfer from mature urban areas to newly developed urban areas. Summary of the Invention

[0004] (a) Technical problems to be solved

[0005] To address the shortcomings of existing technologies, this invention provides a smart city management system that integrates artificial intelligence and big data. By constructing a cross-domain knowledge transfer model and a virtual data generation mechanism, it solves the problem of model performance degradation caused by data scarcity in newly built urban areas, realizes seamless knowledge transfer from mature urban areas to newly built urban areas, and significantly improves the prediction accuracy and adaptability of the smart city management system in data-scarce scenarios.

[0006] (II) Technical Solution

[0007] To achieve the above objectives, the present invention provides the following technical solution: a smart city management system integrating artificial intelligence and big data, comprising:

[0008] The data acquisition and feature extraction module is used to acquire and process multi-source urban data, including road network topology data, POI data and mobile phone signaling data. It extracts road network topology features, functional area distribution features and dynamic features of pedestrian flow through graph neural networks, generates structured feature vectors and stores them.

[0009] The cross-domain knowledge transfer model construction module is used to build a city simulator. It generates virtual data through adversarial generative networks and reinforcement learning, and bridges the source domain and target domain based on the virtual data. It uses graph matching and optimal transport theory to realize the transfer of knowledge from the source domain to the target domain.

[0010] The model optimization and alignment module is used to align the feature distributions of the source and target domains through adversarial training with a domain discriminator and a gradient inversion layer, and dynamically adjust the data weights to achieve seamless transfer from the virtual distribution to the real distribution of the target domain.

[0011] Furthermore, in the data acquisition and feature extraction module, the road network topology data includes road node coordinates, road edge geometry information, road grade and connectivity; POI data includes POI name, geographic coordinates, category and operating status; and mobile signaling data includes unique device identifiers, timestamps and base station locations.

[0012] Furthermore, in the data acquisition and feature extraction module, the graph neural network includes a node embedding layer, a multi-scale aggregation layer, and a feature output layer. The node embedding layer converts road nodes, POIs, and signaling base stations into multi-dimensional vectors. The multi-scale aggregation layer extracts road network topology features, functional area distribution features, and pedestrian flow dynamic features. The feature output layer integrates the extracted features into structured feature vectors, stores them in the database, and establishes an index using spatial grid ID and timestamp as joint primary keys.

[0013] Furthermore, in the cross-domain knowledge transfer model construction module, the city simulator includes a data input layer, a generation module, a reinforcement learning module, and an output layer. The generation module uses a Wasserstein adversarial generative network with gradient penalty to generate virtual data, including virtual road networks, virtual POI distributions, and virtual pedestrian flow dynamics. The reinforcement learning module uses the similarity between the distribution of virtual data and real data as a reward signal and uses a deep deterministic policy gradient algorithm to train the generator parameters.

[0014] Furthermore, in the cross-domain knowledge transfer model construction module, the source domain is a mature urban area and the target domain is a newly built urban area. By constructing a graph structure of the source domain, target domain, and virtual data, spatial grids are used as basic nodes. The node features include the average road network level, the proportion of POI categories, and the peak of pedestrian traffic. The edge weight is the feature similarity between adjacent grids. The Louvain community discovery algorithm is used to aggregate adjacent nodes to form superpoints. The feature of the superpoint is the average of the features of the basic nodes it contains.

[0015] Furthermore, in the cross-domain knowledge transfer model construction module, a virtual graph is used to bridge the source domain and the target domain, calculate the superpoint matching degree and retain the correspondence where the matching degree is not lower than the matching degree threshold; a domain adaptation algorithm based on graph matching is used to calculate the initial correspondence between the superpoints of the source domain and the target domain, and the feature mapping is optimized through the optimal transfer theory to weighted map the superpoint features of the source domain to the target domain, thereby realizing knowledge transfer.

[0016] Furthermore, in the model optimization and alignment module, the domain discriminator is a fully connected neural network that takes source domain superpoint features and target domain enhancement features as input and outputs the probability of the domain to which the features belong. A gradient reversal layer is inserted between the multi-scale aggregation layer and the domain discriminator. During backpropagation, the gradient is multiplied by a fixed coefficient of -1, forcing the multi-scale aggregation layer to learn domain-independent general features.

[0017] Furthermore, in the model optimization and alignment module, the goal of adversarial training is to minimize the feature extraction loss and domain adversarial loss; adjust the weights according to the amount of real data in the target domain; fuse data features from different sources using a weighted average method, and input the model features as a weighted sum of source domain features, virtual features, and real features in the target domain; the training objective is to minimize the prediction error until the model's accuracy on the target domain validation set converges, achieving seamless knowledge transfer.

[0018] An electronic device includes a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to realize a smart city management system integrating artificial intelligence and big data.

[0019] A computer-readable storage medium storing a computer program that, when executed, implements the aforementioned smart city management system integrating artificial intelligence and big data.

[0020] (III) Beneficial Effects

[0021] This invention provides a smart city management system that integrates artificial intelligence and big data, which has the following beneficial effects:

[0022] (1) By integrating multi-source urban data and using graph neural networks to extract structured feature vectors, the standardization and efficient processing of urban data are realized, the data quality is improved, the feature expression capability is enhanced, and a unified and indexable high-dimensional feature foundation is provided for subsequent cross-domain knowledge transfer, thereby significantly improving the analysis accuracy and decision-making efficiency of the smart city management system.

[0023] (2) By constructing a city simulator through virtual data generation and reinforcement learning, and combining graph matching and optimal transport theory, the knowledge of mature urban areas is effectively transferred to newly built urban areas, solving the problem of data scarcity in newly built urban areas, reducing the impact of domain offset, and improving the generalization ability of the model, thereby significantly improving the planning accuracy and adaptability of the smart city management system in data-scarce scenarios.

[0024] (3) Adversarial training is carried out through domain discriminator and gradient inversion layer to effectively align the feature distribution of source domain and target domain, reduce the impact of domain offset, dynamically adjust data weights, ensure that the model smoothly transitions from virtual distribution to real distribution, generate domain-independent general features, enhance model generalization, and significantly improve the prediction accuracy and migration stability of smart city management system in data-scarce scenarios by weighted fusion of multi-source data. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the smart city management system that integrates artificial intelligence and big data according to the present invention.

[0026] Figure 2 This is a schematic diagram of the model optimization and alignment module of the present invention. Detailed Implementation

[0027] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0028] Please see Figures 1 to 2 This invention provides a smart city management system integrating artificial intelligence and big data, comprising: a data acquisition and feature extraction module, a cross-domain knowledge transfer model construction module, and a model optimization and alignment module, wherein:

[0029] The data acquisition and feature extraction module is used to acquire and process multi-source urban data, including road network topology data, POI data and mobile phone signaling data. It extracts road network topology features, functional area distribution features and dynamic features of pedestrian flow through graph neural networks to generate structured feature vectors.

[0030] Specifically, by using the road network data API interface provided by the map open platform, such as the Gaode Map open API interface, the road network topology data of the target city can be obtained, including attributes such as road node coordinates, road edge geometry information, road level, road length and connectivity. Among them, the road level includes expressway, main road, secondary road and branch road. The collection frequency is set to once a day to dynamically update temporary construction or road diversion information.

[0031] Through urban geographic information public service platforms and Points of Interest (POI) databases, such as Meituan and Dianping POI open data, we collect POI data reflecting the distribution of urban functional areas. POI categories include residential, commercial, public service, and transportation hub categories. Among them, residential categories include houses and communities, commercial categories include shopping malls and restaurants, public service categories include schools and hospitals, and transportation hub categories include subway stations and bus stations. Data fields include POI name, geographic coordinates, category, and opening status. Valid POIs are filtered during collection, and records that have been cancelled or closed are removed.

[0032] In cooperation with telecommunications operators, we obtain anonymized mobile signaling data, which has removed user identification and retains only the device's unique identifier, timestamp, and base station location to reflect population dynamics. The collection scope is clearly defined as the administrative area of ​​the target city, with a time granularity of once every 15 minutes, and includes the spatiotemporal location sequence of users.

[0033] For road network data, outliers are identified using the standard deviation method. The mean μ and standard deviation σ of road lengths are calculated, and outlier road records with lengths greater than μ+3σ or less than μ-3σ are removed. Isolated nodes (nodes without connected roads) are removed through topology verification. For POI data, POIs exceeding the administrative boundaries of the target city are removed through geographic coordinate verification. Duplicate records are removed by matching duplicate names with coordinates, such as multiple annotations of the same POI. For mobile signaling data, abnormal locations caused by signal drift are removed, such as unreasonable movement records with a location span exceeding 50 kilometers within 15 minutes, as well as null records caused by equipment failure. Multi-source city data is standardized and uniformly converted to the WGS84 coordinate system to ensure spatial consistency.

[0034] A graph neural network is constructed, comprising a node embedding layer, a multi-scale aggregation layer, and a feature output layer. The node embedding layer converts entities such as road nodes, points of interest (POIs), and signaling base stations from multi-source urban data into multi-dimensional vectors. The multi-scale aggregation layer extracts road network topology features, functional zone distribution features, and pedestrian flow dynamics features. The road network topology features include road grade features and connectivity features. The graph neural network embeds the road grade enumeration values ​​from the road network data, outputting a road grade vector as the road grade feature. Based on the road network adjacency matrix, the graph neural network calculates the connectivity probability between nodes, generating a connectivity heatmap.

[0035] The standardized road grade enumeration values ​​are converted into one-hot encoded vectors. For example, road grade enumeration values ​​are 1=expressway, 2=arterial road, 3=secondary arterial road, and 4=local road. The one-hot encoded vectors are [1, 0, 0, 0] for expressways, [0, 1, 0, 0] for arterial roads, and so on, forming the initial grade feature vector. In the node embedding layer of the multi-scale graph neural network, the embedding dimension is set, and the embedding layer weight matrix is ​​a trainable parameter matrix. The initial grade feature vector is mapped to a multi-dimensional embedding vector through matrix multiplication. The formula is: embedding vector = initial grade feature vector × weight matrix + bias term.

[0036] Using road intersections as nodes, a road network adjacency matrix is ​​constructed. The element values ​​of the road network adjacency matrix follow these rules: if nodes are directly connected (i.e., there is a physical road), the value is the road's level weight: expressway = 4, main road = 3, secondary road = 2, and local road = 1; if they are not directly connected, the value is 0. A graph convolutional network (GCN) is used to calculate the connectivity probability. The GCN contains two hidden layers: the first layer has an output dimension of 64, and the second layer has an output dimension of 32. The activation function for both layers is ReLU. The input is the adjacency matrix. For any two nodes, the connectivity probability is the dot product of the second layer's output vectors, normalized by the softmax function, ranging from 0 to 1.

[0037] The functional area distribution characteristics include POI heatmaps and business type ratios. At a 1km×1km grid scale, the number of POIs within the grid is aggregated through a graph neural network to generate a spatial heatmap matrix for each POI category. By calculating the proportion of each type of POI within each grid, such as the number of commercial POIs / the total number of POIs in the grid, a business type ratio vector is output as the business type ratio.

[0038] The dynamic characteristics of pedestrian flow include the OD matrix and spatiotemporal trajectory. Based on mobile phone signaling data, the number of people flowing between each grid within 15 minutes is counted through graph neural network to generate the OD matrix, where rows = starting grid, columns = ending grid, and values ​​= number of people flowing. The spatiotemporal trajectory is extracted by performing time-series modeling on the signaling location sequence of a single device to extract the stop points and movement paths of the trajectory. The stop points are positions with a duration of ≥30 minutes.

[0039] The feature output layer integrates the extracted road network topology features, functional area distribution features, and pedestrian flow dynamic features into a structured feature vector, stores it in the database, and establishes an index with spatial grid ID and timestamp as the joint primary key to form basic urban structural data.

[0040] By integrating multi-source urban data and using graph neural networks to extract structured feature vectors, the standardization and efficient processing of urban data are achieved, improving data quality, enhancing feature expression capabilities, and providing a unified, indexable, high-dimensional feature foundation for subsequent cross-domain knowledge transfer. This significantly improves the analytical accuracy and decision-making efficiency of the smart city management system.

[0041] The cross-domain knowledge transfer model construction module is used to build a city simulator. It generates virtual data through adversarial generative networks and reinforcement learning, and bridges the source domain and target domain based on the virtual data. It uses graph matching and optimal transport theory to realize the transfer of knowledge from the source domain to the target domain.

[0042] Specifically, a city simulator is constructed that includes a data input layer, a generation module, a reinforcement learning module, and an output layer. The input layer receives city structure data; the generation module and the reinforcement learning module work together to output virtual data.

[0043] The generation module uses a Wasserstein adversarial generative network with gradient penalty (WGAN-GP) as the core of the generation network to avoid pattern collapse. The generator takes random noise vectors and structured feature vectors, such as business type ratio vectors, as inputs and outputs virtual data, including virtual road networks, virtual POI distributions, and virtual pedestrian flow dynamics. The discriminator takes real data and virtual data as inputs. The real data is standardized multi-source urban data. The continuity of the discriminator is constrained by a gradient penalty term. Its loss function consists of three parts: the expected discrimination result of real data, the expected discrimination result of virtual data, and the gradient penalty term, with the penalty coefficient set to 10.

[0044] The reinforcement learning module uses the similarity between the distribution of virtual data and real data as the reward signal. The state space is the feature distribution of the currently generated virtual data, and the action space is the parameter adjustment of the generator. The reward value is obtained by measuring the difference between the distribution of real data and virtual data through JS divergence and taking the negative exponent of the difference. The deep deterministic policy gradient algorithm is used for training. The Actor network outputs the parameter adjustment and the Critic network evaluates the value of the action. The number of iterations is set to 5000, and the sample size for each iteration is 1024.

[0045] When the discriminator loss of the adversarial generative network converges and the reward value of reinforcement learning is not less than 0.8, the loss convergence means that the loss fluctuation does not exceed 0.01 for 50 consecutive iterations. The output is virtual data, which contains structured features from the same source as the real data, such as virtual road network topology, virtual POI heat map, virtual OD matrix, etc., forming a virtual city copy that can match the source domain and the target domain.

[0046] The source domain is defined as mature urban areas, i.e., those built no less than 10 years ago, and the target domain is newly built urban areas, i.e. those built no more than 3 years ago. The source domain data and the initial target domain data are separated from the basic urban structure data and associated with the generated virtual data, based on the spatial grid ID.

[0047] Graph structures are constructed for both the source and target domains, using a 500m × 500m spatial grid as the basic nodes. Node features include the average road network level, POI category ratio, and peak pedestrian flow at the 15-minute level within the grid, extracted from the source and target domain data. The edge weights are calculated based on the feature similarity between two adjacent grids using cosine similarity. If the physical distance between two grids exceeds 1 kilometer, the edge weight is set to 0, indicating they are not adjacent. Similarly, a virtual graph structure is constructed using virtual data to generate a virtual graph.

[0048] Adjacent nodes are aggregated in the source and target domains to form superpoints, representing urban functional units such as commercial areas and residential areas. The Louvain community discovery algorithm is used for aggregation, with the goal of maximizing modularity. The modularity threshold is set to 0.5. Each superpoint contains no less than 5 basic nodes. The feature of the superpoint is the average feature of the basic nodes it contains. Finally, both the source and target domains form a two-layer structure of basic nodes and superpoints. The number of superpoints is 1 / 10 to 1 / 5 of the number of basic nodes.

[0049] By using a virtual graph as an intermediate bridge, the relationship between the source domain and the target domain is established. The superpoint matching degree between the source domain and the virtual graph, and the superpoint matching degree between the virtual graph and the target domain are calculated. The superpoint matching degree is calculated based on the cosine similarity of the superpoint feature vectors. The superpoint correspondences with matching degrees not lower than the matching degree threshold are retained, forming a bimodal graph structure of source domain → virtual graph → target domain. The matching degree threshold is set based on the actual situation, and the threshold that maximizes the prediction accuracy, minimizes the domain adversarial loss, and is stable is selected, such as 0.7.

[0050] A domain adaptation algorithm based on graph matching is adopted to calculate the node-level correspondence of the superpoint layer of the source domain and the target domain, extract the superpoint adjacency matrix of the source domain and the target domain, and then calculate their Laplacian matrix respectively. The Laplacian matrix is ​​the difference between the degree matrix and the adjacency matrix. The eigenvectors corresponding to the first 20 largest eigenvalues ​​are obtained through eigenvalue decomposition, which constitute the feature matrix of the source domain and the target domain.

[0051] The values ​​of the elements in the super-point adjacency matrix are calculated as follows: Extract all the basic nodes contained in the i-th super-point, denoted as set P, and all the basic nodes contained in the j-th super-point, denoted as set Q. Calculate the edge weights of all basic node pairs in sets P and Q, i.e. feature similarity, with a value range of 0 to 1. Take the average of these edge weights as the value of the element. If two super-points are not connected, i.e., the edge weights of all basic node pairs are 0, then the value of the element is 0.

[0052] A matching matrix is ​​constructed using feature vector similarity as a metric. Each element in the matrix is ​​calculated using the Euclidean distance between the feature vectors of the source and target domain hyperpoints, specifically the negative exponent of that Euclidean distance. The larger the value, the more similar the features of the two hyperpoints are. The Hungarian algorithm is used to solve the maximum weight matching of the matching matrix to obtain the initial correspondence between the source and target domain hyperpoints.

[0053] Based on the initial correspondence, the optimal transmission theory is used to optimize the feature mapping from the source domain to the target domain. The source domain feature distribution (superpoint feature distribution) and the target domain feature distribution are defined. Each element in the transmission cost matrix is ​​the Euclidean distance between the source domain superpoint feature and the target domain superpoint feature, which is the square root of the sum of the squares of the differences between corresponding elements of the two feature vectors. The optimal transmission matrix is ​​solved by the Sinkhorn algorithm with 100 iterations and a regularization parameter of 0.01. The optimal transmission matrix satisfies that the row sum equals the source domain distribution, the column sum equals the target domain distribution, and the total transmission cost is minimized.

[0054] Based on the optimal transfer matrix, source domain superpoint features are mapped to the target domain to generate enhanced features of the target domain. That is, each target domain superpoint feature is a weighted sum of source domain superpoint features, with the weights determined by the transfer matrix, thus realizing the initial transfer of knowledge from the source domain to the target domain.

[0055] By constructing a city simulator through virtual data generation and reinforcement learning, and combining graph matching and optimal transport theory, we have achieved effective transfer of knowledge from mature urban areas to newly built urban areas. This solves the problem of data scarcity in newly built urban areas, reduces the impact of domain offset, and improves the model's generalization ability, thereby significantly improving the planning accuracy and adaptability of the smart city management system in data-scarce scenarios.

[0056] The model optimization and alignment module is used to align the feature distributions of the source domain and the target domain through adversarial training with a domain discriminator and a gradient reversal layer, and dynamically adjust the data weights to achieve seamless transfer from the virtual distribution to the real distribution of the target domain.

[0057] A domain discriminator is constructed to distinguish the source of features. The inputs are source domain superpoint features and target domain enhancement features, both of which are structured features at the superpoint level. The output is the judgment result of the domain to which the feature belongs, i.e., the source domain or the target domain. The domain discriminator adopts a 3-layer fully connected neural network structure, with 256, 128 and 1 nodes in each layer, respectively. The activation functions are ReLU, ReLU and Sigmoid, respectively, which are used to output the probability that the feature belongs to the source domain, with the probability between 0 and 1.

[0058] A gradient inversion layer is inserted between the multi-scale aggregation layer and the domain discriminator. The gradient inversion layer does not change the feature data during forward propagation, but only multiplies the gradient by a fixed coefficient -1 during backward propagation, so that the multi-scale aggregation layer receives a gradient signal opposite to that of the domain discriminator.

[0059] Define an adversarial loss function, which includes feature extraction loss and domain adversarial loss. The feature extraction loss adopts cross-entropy loss. A multi-scale aggregation layer is trained based on the labels of the source domain data. The labels are such as known traffic flow and functional area attributes to ensure that it can accurately extract task-related features.

[0060] The domain adversarial loss adopts the binary cross-entropy loss, with the difference between the output of the domain discriminator and the real domain label as the optimization objective. The source domain label is 1 and the target domain label is 0. The gradient reversal layer forces the multi-scale aggregation layer to learn the domain-independent general features, that is, to make the domain discriminator unable to distinguish the source of the features.

[0061] The multi-scale aggregation layer and the domain discriminator were trained together. The training batch size was set to 64, the learning rate was set to 0.001, and the Adam optimizer was used. When the accuracy of the domain discriminator approached 50%, the feature distributions of the source domain and the target domain were aligned. The accuracy fluctuated between 48% and 52% for 30 consecutive iterations.

[0062] The virtual data is defined as an intermediate domain, which is divided into a training set and a validation set in a 7:3 ratio. The data format is consistent with that of the source domain and the target domain, and both contain road network, POI, and dynamic features of pedestrian flow. The multi-scale aggregation layer and the domain adaptation algorithm are trained using the source domain data (data from mature urban areas) and the intermediate domain training set as input. The training objective is to minimize the mapping error between the source domain features and the intermediate domain features, which is measured by mean squared error. The number of iterations is set to 300. The training stops when the validation set error drops below 30% of the initial value, so that the cross-domain knowledge transfer model can initially adapt to the virtual distribution.

[0063] The model was further trained using the intermediate domain validation set and the initial target domain data (new urban area data). The mean squared error was used as a metric to optimize the mapping accuracy of the target domain features. During the training process, the proportion of intermediate domain data was gradually reduced from 70% to 30%, while the proportion of target domain data was increased from 30% to 70%. The number of iterations was set to 200 to achieve a smooth transition from the virtual distribution to the real target domain distribution.

[0064] By using KL divergence to calculate the characteristic distribution differences between the source domain and the intermediate domain, and between the intermediate domain and the target domain, the sum of the differences between the source domain and the intermediate domain, and between the intermediate domain and the target domain is calculated. If the ratio of the direct difference between the source domain and the target domain to the sum of the differences is ≥1.5, the path is considered valid, that is, the gradual migration reduces the domain difference by more than 50% compared with the direct migration.

[0065] In the initial stage of model training, the weight of the source domain data is set to 0.5, the weight of the virtual data (intermediate domain) is set to 0.4, and the weight of the target domain real data is set to 0.1, with a total weight of 1. When the amount of new real data in the target domain reaches 50% of the initial data amount, the first weight adjustment is initiated. Subsequently, every time the amount of new data increases by 30%, i.e. relative to the total amount at the time of the last adjustment, an adjustment is triggered until the amount of real data reaches 80% of the amount of source domain data.

[0066] Each adjustment involves the following weighting: target domain real data weight = current weight + 0.2, with an upper limit of 0.8; source domain data weight = current weight - 0.1, with a lower limit of 0.1; virtual data weight = 1 - real data weight - source domain data weight, with a lower limit of 0.1. A weighted average method is used to fuse data features from different sources, resulting in the final input model features = (source domain features × source domain weight) + (virtual features × virtual data weight) + (target domain real features × real data weight). The fused data is used to update the feature extraction module and the cross-domain mapping model. The training objective is to minimize prediction errors, such as traffic flow prediction errors and planning evaluation biases, until the model's accuracy on the target domain validation set converges, for example, with fluctuations ≤2% over 20 consecutive iterations, achieving seamless knowledge transfer from the source domain to the target domain.

[0067] Adversarial training using a domain discriminator and gradient inversion layer effectively aligns the feature distributions of the source and target domains, reduces the impact of domain offset, dynamically adjusts data weights to ensure a smooth transition of the model from the virtual distribution to the real distribution, generates domain-independent general features, enhances model generalization, and significantly improves the prediction accuracy and transfer stability of the smart city management system in data-scarce scenarios by weighted fusion of multi-source data.

[0068] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored on the memory and running on the processor, wherein the processor executes the computer program to realize a smart city management system integrating artificial intelligence and big data.

[0069] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed, implements a smart city management system integrating artificial intelligence and big data.

[0070] In the application, the various formulas mentioned are all calculated by removing dimensions and taking their numerical values. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The coefficients in the formulas are set by those skilled in the art according to the actual situation.

[0071] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, and combinations thereof. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0072] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0073] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A smart city management system integrating artificial intelligence and big data, characterized by: include: The data acquisition and feature extraction module is used to acquire and process multi-source urban data, including road network topology data, POI data and mobile phone signaling data. It extracts road network topology features, functional area distribution features and dynamic features of pedestrian flow through graph neural networks, generates structured feature vectors and stores them. The cross-domain knowledge transfer model building module is used to build a city simulator, which includes a data input layer, a generation module, a reinforcement learning module, and an output layer. The generation module uses a Wasserstein adversarial generative network with gradient penalty. It takes random noise vectors and structured feature vectors as inputs and outputs virtual data, including virtual road networks, virtual POI distributions, and virtual pedestrian flow dynamics. The reinforcement learning module uses the similarity between the distributions of virtual and real data as the reward signal and employs a deep deterministic policy gradient algorithm to train the generator parameters. The source domain is a mature urban area, and the target domain is a newly built urban area. By constructing a graph structure of the source domain, target domain, and virtual data, spatial grids are used as basic nodes. The node features include the average road network level, the proportion of POI categories, and the peak traffic flow. The edge weight is the feature similarity between adjacent grids. The Louvain community detection algorithm is used to aggregate adjacent nodes to form superpoints. The feature of the superpoint is the average of the features of the basic nodes it contains. By bridging the source and target domains with a virtual graph, the matching degree of superpoints is calculated and the correspondence with a matching degree not lower than the matching degree threshold is retained. The domain adaptation algorithm based on graph matching is used to calculate the initial correspondence between the superpoints of the source and target domains, and the feature mapping is optimized by the optimal transfer theory. The superpoint features of the source domain are weighted and mapped to the target domain to realize knowledge transfer. The model optimization and alignment module is used to align the feature distributions of the source and target domains through adversarial training with a domain discriminator and a gradient inversion layer, and dynamically adjust the data weights to achieve seamless transfer from the virtual distribution to the real distribution of the target domain.

2. The smart city management system integrating artificial intelligence and big data according to claim 1, characterized in that: In the data acquisition and feature extraction module, the road network topology data includes road node coordinates, road edge geometry information, road grade and connectivity; POI data includes POI name, geographic coordinates, category and operating status; mobile signaling data includes device unique identifier, timestamp and base station location.

3. The smart city management system integrating artificial intelligence and big data according to claim 2, characterized in that: In the data acquisition and feature extraction module, the graph neural network includes a node embedding layer, a multi-scale aggregation layer, and a feature output layer; the node embedding layer converts road nodes, POIs, and signaling base stations into multi-dimensional vectors; the multi-scale aggregation layer extracts road network topology features, functional area distribution features, and pedestrian flow dynamic features; The feature output layer integrates the extracted features into a structured feature vector, stores it in the database, and establishes an index using the spatial grid ID and timestamp as a joint primary key.

4. The smart city management system integrating artificial intelligence and big data according to claim 1, characterized in that: In the model optimization and alignment module, the domain discriminator is a fully connected neural network that takes source domain superpoint features and target domain enhancement features as input and outputs the probability of the domain to which the features belong. A gradient reversal layer is inserted between the multi-scale aggregation layer and the domain discriminator. During backpropagation, the gradient is multiplied by a fixed coefficient of -1 so that the multi-scale aggregation layer receives a gradient signal opposite to that of the domain discriminator.

5. The smart city management system integrating artificial intelligence and big data according to claim 4, characterized in that: In the model optimization and alignment module, the goal of adversarial training is to minimize the feature extraction loss and domain adversarial loss; adjust the weights according to the amount of real data in the target domain; and fuse data features from different sources using a weighted average method, with the input features to the model being a weighted sum of source domain features, virtual features, and real features in the target domain. The training objective is to minimize the prediction error until the model's accuracy on the target domain validation set converges, thereby achieving seamless knowledge transfer.

6. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it performs the steps of the smart city management system integrating artificial intelligence and big data as described in any one of claims 1-5.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed, performs the steps of the smart city management system integrating artificial intelligence and big data as described in any one of claims 1-5.

Images