Government affair space digital twin governance system based on multi-source data fusion

Abstract

The application discloses a government affair space digital twin governance system based on multi-source data fusion, and the system comprises a data acquisition module, which acquires initial multi-source data and pre-processes the multi-source data; a physical twin module, which inputs a city state real vector and a disturbance vector into a city mechanism model and performs correction processing to obtain a digital twin result; a social simulation module, which inputs an attribute vector, an environment constraint vector and initial social text data into a large language model, generates an initial behavior decision vector and performs iterative processing to obtain a social simulation set; a verification module, which extracts initial space-time statements from the large language model, performs deduction verification by using the city mechanism model, and generates a final space-time statement set; and a governance decision module, which selects a maximum benefit strategy as an optimal governance scheme. The system realizes deep causal coupling between traditional digital twin simulation and social simulation through the verification module.

Description

Technical Field

[0001] This invention relates to the field of digital twin technology, specifically to a digital twin governance system for government spaces based on multi-source data fusion. Background Technology

[0002] The government space digital twin governance system based on multi-source data fusion is a new generation of intelligent digital twin platform designed to meet the needs of modern urban government governance. The system adopts a structure including data acquisition module, physical twin module, social simulation module, verification module and governance decision-making module. The system can be widely used in multiple core areas such as urban traffic management, environmental control and emergency management.

[0003] In practical applications, existing technologies often cause disturbances to physical systems due to social behavior, which prevents deep causal coupling between urban physical and social systems. Furthermore, the lack of real-time observation data for continuous model simulation and adjustment significantly reduces the reliability of decision-making information. Social behavior simulations are unable to respond to dynamic changes in the urban physical environment, resulting in a serious disconnect between the generated governance strategies and the actual operational needs of the city, making it impossible to effectively solve various complex problems and emergencies in urban operations. Summary of the Invention

[0004] In view of the shortcomings of the existing technologies, the purpose of this invention is to provide a digital twin governance system for government space based on multi-source data fusion, so as to solve the problem that the existing urban physical system and social system cannot achieve deep causal coupling.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] The present invention relates to a digital twin governance system for e-government spaces based on multi-source data fusion, the system comprising:

[0007] Module 1: Data Acquisition Module, which collects initial multi-source data and preprocesses the multi-source data to obtain a spatiotemporal data base. The multi-source data includes initial three-dimensional ground-space data, initial social text data, and initial government affairs thematic data.

[0008] Module 2: Physical Twin Module. Semantic nodes are extracted from the spatiotemporal data base to generate real and perturbation vectors of the city state. The real and perturbation vectors of the city state are input into the city mechanism model and corrected to obtain the digital twin result.

[0009] Module 3: Social Simulation Module. Based on the aforementioned spatiotemporal data foundation, this module constructs social behavioral subjects and attribute vectors; introduces an environmental constraint mechanism to the social subjects to generate environmental constraint vectors; inputs the attribute vectors, environmental constraint vectors, and initial social text data into the large language model to generate initial vectors for behavioral decisions and performs iterative processing to obtain a social simulation set.

[0010] Module 4: Verification Module. Based on the digital twin results and the social simulation set, a causal graph node set and a causal structure graph are constructed sequentially. The social simulation set is fed back to the urban mechanism model for updating, resulting in an updated social simulation set. The updated social simulation set, digital twin results, and causal structure graph are summarized to generate a comprehensive situation set. Initial spatiotemporal statements are extracted from the large language model and deduced and verified using the urban mechanism model to generate a final spatiotemporal statement set.

[0011] Module 5: Governance Decision Module. Based on the comprehensive situation set and the final spatiotemporal statement set, the module calculates the benefits of behavioral strategies and selects the strategy with the maximum benefit as the optimal governance solution.

[0012] Furthermore, the preprocessing in module one includes: inputting the multi-source data into a spatiotemporal mapping function for spatiotemporal registration and inputting it into a semantic encoding function to generate semantic vectors corresponding to each of the registered multi-source data; summarizing and combining the different semantic vectors to generate a semantic node set and inputting any two nodes into an association weight function to generate cross-node association weights; establishing an association edge set based on the semantic node set and association weights; constructing a spatiotemporal graph using the semantic node set and association edge set; and building a spatiotemporal index set based on the spatiotemporal graph and combining it with the spatiotemporal graph to establish a spatiotemporal data foundation.

[0013] Furthermore, the correction process in module two includes: discretizing the urban mechanism model through an explicit time-progression method to obtain a complete urban state projection vector; extracting the observation state vector from the spatiotemporal data base and calculating the error between the urban state projection vector and the observation state vector to generate a prediction error; inputting the prediction error and the urban state projection vector together into a correction function to obtain an urban state correction vector; and summarizing and combining all the urban state correction vectors to generate a digital twin result.

[0014] Furthermore, the environmental constraint mechanism in module three includes: mapping the urban state correction vector to the social environmental constraint space to generate an environmental constraint vector.

[0015] Furthermore, the iterative processing in module three involves: iterating the attribute vector to generate the attribute vector for the next time step; re-inputting the attribute vector for the next time step, the environmental constraint vector, and the initial social text data into the preset large language model to generate a behavior decision correction vector; re-aggregating the behavior decision correction vector to generate an average behavior correction vector; and summarizing and combining all the average behavior correction vectors to generate a social simulation set.

[0016] Furthermore, the update process in module four includes: calculating a social behavior disturbance vector using the feedback matrix and the social simulation set; re-inputting the social behavior disturbance vector and the real disturbance vector into the urban mechanism model to generate an urban state update vector; remapping the urban state update vector into the social environment constraint space to generate a corrected environmental constraint vector; re-inputting the attribute vector, the corrected environmental constraint vector, and the initial social text data into a preset large language model to generate a behavior decision update vector; and re-aggregating and combining the behavior decision update vector to generate an updated social simulation set.

[0017] Furthermore, the deduction and verification in module four includes: calculating the causal influence strength between any two causal nodes in the causal graph node set; statistically analyzing the causal influence strength to obtain a causal consistency index; comparing the causal consistency index with a preset judgment threshold; if the causal consistency index is greater than the judgment threshold, the updated social simulation set and the digital twin result are causally consistent; if the causal consistency index is less than the judgment threshold, the social behavior perturbation vector is recalculated until the causal consistency index is greater than the judgment threshold; and summarizing the updated social simulation set, the digital twin result, and the causal structure graph to generate the final spatiotemporal statement set.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0019] The system of this invention proposes a verification module, which obtains the causal influence strength by constructing a causal graph containing physical state and social behavior nodes. It realizes the perturbation of the physical system by social behavior through a feedback matrix, forming a two-way closed-loop inference, realizing the deep causal coupling between the urban physical system and the social system. This solves the core problem that existing technologies cannot realistically simulate the interaction of complex urban systems. At the same time, it establishes a physical verification mechanism based on natural language statements, which greatly improves the reliability of decision-making information.

[0020] The system of this invention proposes a physical twin module and a social simulation module. By introducing an unscented Kalman filter correction mechanism, the simulation state is continuously adjusted using real-time observation data, which effectively reduces model errors and uncertainties. At the same time, the urban physical state is transformed into an environmental constraint vector of social subjects through a feedback matrix, enabling social behavior simulation to respond to changes in the urban physical environment and avoiding the situation where physical simulation and social simulation are independent of each other. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a schematic diagram of the system's workflow in this invention;

[0023] Figure 2 This is a schematic diagram of the overall modular architecture of the system in this invention;

[0024] Figure 3 This is a schematic diagram of the data acquisition module in the system of the present invention, which collects multi-source data. Detailed Implementation

[0025] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to embodiments and accompanying drawings. The content mentioned in the embodiments is not intended to limit the present invention.

[0026] Reference Figure 1 As shown, the present invention provides a digital twin governance system for e-government spaces based on multi-source data fusion. The system includes:

[0027] Module 1: Data Acquisition Module. This module collects initial multi-source data and preprocesses it to obtain a spatiotemporal data base. The source data includes initial 3D ground-space data, initial social text data, and initial government affairs thematic data.

[0028] First, initial multi-source data is collected, including initial 3D ground-space data, initial social text data, and initial government affairs thematic data. The function is shown below:

[0029]

[0030]

[0031]

[0032] in, For the first Initial three-dimensional ground-air data; This is the initial three-dimensional ground-to-air data set; This serves as the initial three-dimensional ground-to-space data index. This represents the initial total number of three-dimensional ground-to-air data. This is the initial set of social text data; For the first Initial social text data; This represents the initial total number of social text data. Index the initial social text data; This is the initial set of government affairs thematic data; For the first Initial government affairs data; This represents the initial total number of government affairs data topics; This serves as the initial index for government affairs thematic data;

[0033] The initial 3D ground-to-air data, initial social text data, and initial government affairs thematic data are input into the spatiotemporal mapping function for spatiotemporal registration, generating registered 3D ground-to-air data, registered social text data, and registered government affairs thematic data. The spatiotemporal mapping function is shown below:

[0034]

[0035]

[0036]

[0037] in, For the spatiotemporal mapping function, GDAL (coordinate transformation) and PostGIS (spatiotemporal indexing and spatial matching) are used for processing the initial 3D ground-space data. For the initial social text data, spaCy (spatiotemporal entity extraction), ERNIE (Chinese spatiotemporal semantic recognition), and Pandas (timestamp alignment) are used for processing. For the initial government affairs thematic data, algorithms based on geohash spatial matching and time window-based event alignment are used. To register 3D ground-to-air data; To register social text data; To ensure accurate matching of government affairs data;

[0038] The registered 3D ground-to-air data, registered social text data, and registered government topic data are input into the semantic encoding function, so that different types of data are mapped to the same semantic space, generating semantic vectors for the registered 3D ground-to-air data, registered social text data, and registered government topic data. The semantic encoding function is shown below:

[0039]

[0040]

[0041]

[0042] in, For semantic encoding functions, GeoBERT (joint encoding of geographic coordinates and semantics) is used for registering 3D geospatial data, BERT-base-Chinese (Chinese text semantic encoding) is used for registering social text data, and TabNet (semantic embedding of tabular data) is used for registering government affairs thematic data. To register semantic vectors for 3D ground-to-air data; To register semantic vectors for social text data; To register semantic vectors for government affairs thematic data;

[0043] Secondly, the different semantic vectors are aggregated and combined to generate a set of semantic nodes, as shown below:

[0044]

[0045] in, A collection of semantic nodes;

[0046] Select any two nodes from the semantic node set and input them into the association weight function to generate cross-node association weights. The association weight function is shown below:

[0047]

[0048]

[0049] in, For association weights; For the first One semantic node; For the first One semantic node; It is an L2 norm; and All are semantic node indexes;

[0050] Based on the semantic node set and association weights, an association edge set is established, as shown below:

[0051]

[0052] in, For the set of associated edges;

[0053] Finally, a spatiotemporal graph is constructed using the set of semantic nodes and the set of associated edges, as shown below:

[0054]

[0055] in, For spatiotemporal mapping;

[0056] Based on the spatiotemporal graph, a spatiotemporal index set is constructed to achieve fast localization of any semantic node. The spatiotemporal index set is shown below:

[0057]

[0058] in, A set of spatiotemporal indexes; For spatial indexing; For time indexing; For relational indexes;

[0059] Based on the spatiotemporal map and spatiotemporal index set, a spatiotemporal data foundation is established, as shown below:

[0060]

[0061] in, As a spatiotemporal data foundation;

[0062] Among these, the system uses a drone-based LiDAR system (DJI Matrice300RTK + Zenmuse L1) to collect 3D point cloud data and orthophotos of the city; a ground-based 3D LiDAR system (Velodyne AlphaPrime) to create high-precision 3D models of road intersections, business districts, and industrial parks; a BeiDou positioning terminal (Hexin Xingtong UB4B0M) to collect spatiotemporal reference calibration data; a networked data acquisition gateway (Huawei AR651-S) to crawl text data from social media and news websites; and a government data access gateway (Qianxin Wangshen SecSIS3000) to securely isolate and exchange data between the government's private network and the system's intranet.

[0063] The system utilizes a semantic encoding GPU server (Dell PowerEdge R760xa) for multimodal data semantic encoding and spatiotemporal map construction, a spatiotemporal database server (Inspur NF5280M6) for storing spatiotemporal data, and a distributed object storage system (Huawei OceanStorPacific9550) for storing unstructured data such as raw point clouds, images, and text.

[0064] Module 2: Physical Twin Module. Semantic nodes are extracted from the spatiotemporal data base to generate real and perturbation vectors of the city status. The real and perturbation vectors of the city status are then input into the city mechanism model and corrected to obtain the digital twin result.

[0065] First, all semantic nodes related to urban physical operation are extracted from the spatiotemporal data base and represented in set form, as shown in the function below:

[0066]

[0067] in, It is the set of all semantic nodes related to the physical operation of the city. For semantic node indexing;

[0068] The set of all semantic nodes related to urban physical operation is input into the state mapping function to generate the true urban state vector. The state mapping function is as follows:

[0069]

[0070]

[0071] in, for The real vector of the city's state at any given time. It is a time variable; The state mapping function typically employs a graph convolutional layer (GCN) to aggregate local semantic features and form a complete real vector of city state. For the first One state variable, This serves as an index for state variables, which typically include traffic flow, average road speed, pollutant concentration, temperature and humidity, river flow, river level, pipeline pressure, and energy load. This represents the total number of state variables.

[0072] Factors influencing urban operations are extracted from the spatiotemporal data foundation to construct a perturbation vector, as shown below:

[0073]

[0074] in, for The actual perturbation vector at any given moment; For the first One perturbation variable, This serves as an index for disturbance variables, which typically include rainfall, wind speed, holiday activities, large gatherings, accidents, emergency control measures, and fluctuations in energy supply. This represents the total number of disturbance variables;

[0075] Secondly, the actual urban state vector and the actual disturbance vector are input into the preset urban mechanism model for digital twin simulation to obtain the urban state simulation vector. The urban mechanism model consists of multiple sub-models, including a traffic flow sub-model, an environmental diffusion sub-model, and a hydrodynamic sub-model. The unified expression of the urban mechanism model is shown below:

[0076]

[0077] in, The overall instantaneous rate of change of the state vector of multiple urban elements; Mechanism mapping function;

[0078] Traffic flow sub-model:

[0079]

[0080] in, For traffic flow entering the flow; Traffic outflow; The instantaneous rate of change of the number of vehicles over time; These are state variables belonging to traffic flow;

[0081] Environmental diffusion sub-model:

[0082]

[0083] in, The rate of change of pollutant concentration over time; The atmospheric molecular diffusion coefficient; This represents the two-dimensional spatial diffusion term of the pollutants. For wind field advection transport; This represents the wind speed vector of the regional wind field. For pollution source emission items; These are state variables belonging to pollutant concentration;

[0084] Hydrodynamic sub-model:

[0085]

[0086] in, The rate of change of river flow over time; The inflow volume; The outflow volume of water; These are state variables belonging to the river flow rate;

[0087] By discretizing the urban mechanism model using an explicit time-progression method, a complete urban state projection vector is obtained, as shown in the function below:

[0088]

[0089]

[0090]

[0091]

[0092]

[0093] in, The initial real vector of the city state; This represents the actual perturbation vector at the initial moment; The discrete time step represents the unit interval of the continuous time division; for The vector representing the city's state at any given moment; This is the vector representing the city's state at the previous moment; This is the inferred perturbation vector from the previous moment; Total duration;

[0094] A correction mechanism is introduced for the city state projection vector, as shown below:

[0095] The observed state vector is extracted from the spatiotemporal data base, and the error between the city state projection vector and the observed state vector is calculated to generate the prediction error. The function is as follows:

[0096]

[0097] in, This represents the prediction error; This is the observed state vector, which belongs to manually observed values; for The vector representing the city's state at any given moment;

[0098] The prediction error and the city state projection vector are input together into the correction function to obtain the city state correction vector. The correction function is shown below:

[0099]

[0100] in, for The city state correction vector at any given time; The state correction gain matrix is ​​calculated by recursively calculating the prediction error covariance matrix and the observation error covariance matrix using the unscented Kalman filter algorithm. The state correction gain matrix is ​​calculated, and the state correction gain matrix is ​​recalculated once a new set of observation state vectors is obtained.

[0101] Finally, all the city status correction vectors are aggregated and combined to generate the digital twin result, as shown below:

[0102]

[0103] in, For digital twin results;

[0104] The urban mechanism model includes an input layer, a hidden layer, and an output layer. The input of the input layer is the true urban state vector and the true disturbance vector. The output of the output layer is the urban state projection vector. The hidden layer of the traffic flow sub-model adopts the macroscopic LWR framework, the hidden layer of the environmental diffusion sub-model adopts the two-dimensional convection-diffusion equation, and the hidden layer of the hydrodynamic sub-model adopts the Saint-Venant equations.

[0105] The activation function for the traffic flow sub-model is the Greenshields quadratic function, the activation function for the environmental diffusion sub-model is the threshold activation function, and the activation function for the hydrodynamic sub-model is the weir flow formula.

[0106] The loss function of the urban mechanism model adopts the data loss function, which is the mean square error between the inferred value and the observed value;

[0107] The evaluation indicators for the traffic flow sub-model are average vehicle speed, traffic volume, and congestion prediction; the evaluation indicators for the environmental diffusion sub-model are pollution concentration, ozone correlation coefficient, and accuracy of exceedance warning; and the evaluation indicators for the hydrodynamic sub-model are river flow and water level.

[0108] During training, the physical parameters of each sub-model were calibrated using historical data and optimized using a genetic algorithm. The three sub-models were then coupled together, and the data was assimilated using an ensemble Kalman filter algorithm, while updating the parameters and states. The historical data consisted of continuous observations over the past three years, with a temporal resolution greater than one hour and a spatial coverage greater than 90%.

[0109] Specifically, a high-performance computing node (Lenovo ThinkSystem SR670V) is used to solve traffic flow, environmental diffusion, and hydrodynamic sub-models in parallel. A real-time data access server (Dell PowerEdge R750xs) is used to access real-time observation data from traffic checkpoints, environmental monitoring stations, and hydrological stations. A Kalman filter accelerator card (Xilinx Alveo U280FPGA) is used to perform real-time calculation of the state correction gain matrix and assimilation of observation data.

[0110] Module 3: Social Simulation Module. Based on the spatiotemporal data foundation, it constructs social behavioral subjects and attribute vectors; introduces an environmental constraint mechanism to the social subjects to generate environmental constraint vectors; inputs the attribute vectors, environmental constraint vectors, and initial social text data into the large language model to generate initial vectors for behavioral decisions and performs iterative processing to obtain the social simulation set.

[0111] First, based on the spatiotemporal data foundation, a set of social behavioral agents is constructed, as shown below:

[0112]

[0113] in, A collection of social actors; For the first These are social entities, including residents, businesses, tourists, netizens, and administrators. An index for social entities; The total number of social entities;

[0114] Construct attribute vectors corresponding to social entities, as shown below:

[0115]

[0116] in, In order to be in Time of the first The attribute vector corresponding to each social subject; Risk perception level; As a measure of resource demand; As a measure of social trust; The degree of willingness to act; It is the transpose symbol;

[0117] An environmental constraint mechanism is introduced for social entities, as shown below:

[0118] The urban state correction vector is mapped to the social environment constraint space to generate the environmental constraint vector, as shown in the function below:

[0119]

[0120] in, In order to be in Time of the first An environmental constraint vector for each social entity; The environmental mapping matrix is ​​constructed by extracting the most correlated typical variable pairs based on the linear correlation between urban state variables (traffic, environment, hydrology, etc.) and social environmental constraint variables (travel accessibility, resource supply capacity, risk exposure, etc.).

[0121] Secondly, the attribute vector, environmental constraint vector, and initial social text data are input into a pre-defined large language model to generate initial vectors for behavioral decisions. The large language model is shown below:

[0122]

[0123] in, This is the initial vector for behavioral decisions, and the initial vector for behavioral decisions is described in natural language. Operators for large language models of social behavior; Initial social text data associated with social entities; Intensity of travel behavior; For the purpose of information dissemination; Intensity of resource acquisition behavior;

[0124] The initial vectors for all behavioral decisions are aggregated to generate an average initial vector for behavioral decisions, as shown in the function below:

[0125]

[0126] in, The initial vector for average behavior; Overall travel intensity; The overall intensity of information dissemination; The intensity of overall resource demand;

[0127] Finally, based on the continuous evolutionary nature of social behavior, an iterative mechanism is introduced into the attribute vector until the preset maximum number of iterations is reached to generate the attribute vector for the next time step. The iterative mechanism is as follows:

[0128]

[0129] in, This is the attribute vector for the next time step; In order to be in Time of the first The attribute vector corresponding to each social subject; This is the behavioral feedback coefficient; The environmental impact coefficient is determined by: first, constructing a multiple linear regression model using the least squares method to solve for the behavioral feedback coefficient and the environmental impact coefficient; then, using the accuracy of social behavior as the reward function, iteratively adjusting the behavioral feedback coefficient and the environmental impact coefficient through deep reinforcement learning (DRL); finally, domain experts fine-tune the coefficients for specific scenarios (such as emergencies or large-scale events) by combining their experience in sociology and psychology.

[0130] The attribute vector, environmental constraint vector, and initial social text data for the next time step are re-input into the pre-defined large language model to generate behavioral decision correction vectors, as shown in the function below:

[0131]

[0132] in, This is the behavioral decision correction vector;

[0133] The behavior decision correction vector is re-aggregated to generate an average behavior correction vector, as shown in the function below:

[0134]

[0135] in, This is the average behavior correction vector;

[0136] All average behavior correction vectors are aggregated and combined to generate a social simulation set, as shown below:

[0137]

[0138] in, A collection of social simulations;

[0139] The large language model consists of an input layer, a hidden layer, and an output layer. The input layer takes attribute vectors, environmental constraint vectors, and initial social text data as inputs. The output layer outputs behavioral decision initial vectors and spatiotemporal statement text. The hidden layer uses a 12-layer Transformer encoder with 12 attention heads per layer and uses the Swiglu activation function. The output layer that outputs the decision initial vectors uses the Sigmoid activation function, and the output layer that outputs the spatiotemporal statement text uses the Softmax activation function.

[0140] The loss functions of the large language model are the behavior prediction loss function and the text generation loss function. The behavior prediction loss function is the mean squared error between the observed value and the inferred value, and the text generation loss function is the cross-entropy between the true statement and the inferred statement.

[0141] The evaluation metric for large language models is the causal consistency metric.

[0142] Specifically, the large language model inference server (Huawei Atlas800TA2) performs parallel inference on the large language model of social behavior, the inference accelerator card (NVIDIA H100NVL94GB) performs high-concurrency behavior decision request processing, and the memory expansion node (Dell PowerEdge R760) performs large-scale social subject attribute vector caching.

[0143] Module 4: Verification Module. Based on the digital twin results and the social simulation set, a causal graph node set and a causal structure graph are constructed sequentially. The social simulation set is fed back to the urban mechanism model for updating, resulting in an updated social simulation set. The updated social simulation set, digital twin results, and causal structure graph are summarized to generate a comprehensive situation set. Initial spatiotemporal statements are extracted from the large language model and deduced and verified using the urban mechanism model to generate a final spatiotemporal statement set.

[0144] First, a causal graph node set based on digital twin results and social simulation sets is constructed, as shown below:

[0145]

[0146] in, It is a set of nodes in a cause-effect graph;

[0147] A causal structure graph is constructed based on the set of nodes in the causal graph. The causal structure graph is shown below:

[0148]

[0149] in, This is a cause-and-effect diagram; The edges of the node set in the causal graph;

[0150] The function for calculating the causal influence strength between any two causal nodes in a causal graph node set is shown below:

[0151]

[0152]

[0153] in, The strength of causal influence; To represent the causal nodes after the intervention was implemented The probability of occurrence For causal budgeting; Let be the probability of a causal node occurring naturally, representing the probability of a causal node occurring without any intervention. The baseline probability of occurrence;

[0154] Secondly, using the feedback matrix and combining it with the social simulation dataset, the social behavior perturbation vector is calculated, as shown in the following function:

[0155]

[0156] in, This represents the social behavior disturbance vector. The feedback matrix is ​​the inverse of the environment mapping matrix. The feedback matrix is ​​obtained by inverting the matrix.

[0157] The social behavior disturbance vector is combined with the real disturbance vector and re-inputted into the urban mechanism model to perform the update process of the real urban state vector, thereby generating the updated urban state vector. The function is as follows:

[0158]

[0159] in, Update the vector for city status;

[0160] The city state update vector is remapped into the social environment constraint space to generate a corrected environment constraint vector, as shown in the function below:

[0161]

[0162] in, To correct the environmental constraint vector; This is the environment mapping matrix;

[0163] The attribute vector, the modified environmental constraint vector, and the initial social text data are re-input into the pre-defined large language model to generate behavioral decision update vectors, as shown in the function below:

[0164]

[0165] in, Update the vector for behavioral decisions;

[0166] The behavioral decision update vectors are then re-aggregated and combined to generate an updated social simulation set, as shown below:

[0167]

[0168] in, For the updated social simulation set; The updated average behavior vector is obtained by re-aggregating the data.

[0169] By statistically analyzing the strength of causal influences, a causal consistency index is obtained, the function of which is shown below:

[0170]

[0171] in, As an indicator of causal consistency; This represents the total strength of causal influence. For the first A valid causal influence strength, where valid means the causal influence strength is non-zero;

[0172] The causal consistency index is compared with the preset judgment threshold:

[0173] When causal consistency index Greater than the judgment threshold Then the updated social simulation set With digital twin results Consistency of cause and effect;

[0174] When causal consistency index Less than the judgment threshold If so, the social behavior disturbance vector is recalculated until the causal consistency index is greater than the judgment threshold.

[0175] Judgment threshold The causal consistency index of all samples is calculated based on historical data. The 95th percentile is selected as the initial threshold, and then relevant experts adjust the judgment threshold appropriately according to different scenarios (such as public safety and emergency management).

[0176] The updated social simulation dataset, digital twin results, and causal structure diagram are summarized to generate a comprehensive situation set, as shown below:

[0177]

[0178] in, For comprehensive situational awareness;

[0179] Finally, an initial set of spatiotemporal statements is extracted from the large language model. The initial spatiotemporal statements are as follows:

[0180]

[0181] in, This is the initial set of spacetime statements; For the first An initial spacetime statement, This serves as an index for initial spatiotemporal statements, such as "congestion is expected in a certain area within the next hour"; This represents the total number of initial spacetime statements;

[0182] The initial spatiotemporal statement is transformed into a corresponding semantic vector using a semantic encoding function. At the same time, based on the spatial location and time range in the initial spatiotemporal statement, the real vector of the current urban state and the real disturbance vector of the corresponding region are retrieved from the spatiotemporal data base through the spatiotemporal index.

[0183] The current real urban state vector and real disturbance vector of the corresponding region, along with the initial projection conditions, are input into the preset urban mechanism model. The urban state at each future time is calculated step by step using an explicit time-progression method, and finally the projection state vector at the projection endpoint is obtained. The function is as follows:

[0184]

[0185] in, To deduce the state vector;

[0186] The inferred state vector is input into a pre-defined large language model to obtain the inferred spatiotemporal statement, as shown in the function below:

[0187]

[0188] in, To extrapolate the spatiotemporal statement, the extrapolation spatiotemporal statement is in natural language text;

[0189] The degree of discrepancy between the deduced spacetime statement and the initial spacetime statement is calculated to generate the degree of contradiction, as shown in the function below:

[0190]

[0191] in, For the degree of contradiction, For the initial spacetime statement, For the purpose of deriving spatiotemporal statements, all statements are in vector form;

[0192] The degree of inconsistency is compared with a preset deviation threshold. If the degree of inconsistency... Greater than the deviation threshold At that time, determine the initial spacetime statement. The spatiotemporal deviation statement indicates that this initial spatiotemporal statement violates the laws of physics. The deviation threshold is determined by calculating the probability distribution of the degree of contradiction of the causal consistency index of all samples based on historical data, selecting the 95th percentile as the initial threshold, and then having relevant experts adjust the deviation threshold appropriately according to different scenarios (such as public safety and emergency management).

[0193] All spacetime deviation statements are replaced with deduced spacetime statements and fused with normal spacetime statements to generate corrected spacetime statements, as shown in the function below:

[0194]

[0195] in, To correct the spacetime statement; A normal spacetime statement is a spacetime statement after removing all spacetime deviation statements from the initial set of spacetime statements;

[0196] The revised spacetime statements are summarized to generate the final set of spacetime statements, as shown below:

[0197]

[0198] in, For the final set of spacetime statements;

[0199] Specifically, a graph computing server (Dell PowerEdge R760xd2) is used for large-scale parallel computation of causal graphs and solving for the intensity of causal influences. A large language model verification server (Tencent Cloud Xingxinghai SA3) is used for spatiotemporal statement generation and contradiction calculation. A high-speed cache cluster (Redis cluster deployed on a 3-node Super Fusion 2288HV7) is used for caching causal nodes and spatiotemporal statements.

[0200] Module 5: Governance Decision Module. Based on the comprehensive situation set and the final spatiotemporal statement set, the module calculates the benefits of behavioral strategies and selects the strategy with the maximum benefit as the optimal governance solution.

[0201] First, establish a set of governance objectives, as shown below:

[0202]

[0203] in, A set of governance objectives; These are separate governance objectives; governance objectives include traffic operation efficiency objectives, public safety objectives, environmental quality objectives, social stability objectives, and government service response objectives. The total number of governance targets;

[0204] Generate corresponding governance strategies for each governance objective, as shown below:

[0205]

[0206] in, A set of governance strategies; This is a separate governance strategy; the governance strategy includes traffic management plans, environmental control plans, emergency resource allocation plans, and online public opinion guidance plans. Total number of governance strategies

[0207] Secondly, the comprehensive situation set, the final spatiotemporal statement set, and the governance strategy are input into the comprehensive benefit function to generate the comprehensive benefit, as shown below:

[0208]

[0209]

[0210] in, For comprehensive benefits; The target weight; For the evaluation function, any governance objective corresponds to an evaluation function. For the transportation field, SUMO (traffic simulation evaluation) and TransModeler (road network operation efficiency calculation) are used for processing. For the environmental field, CALPUFF (atmospheric diffusion simulation evaluation) and SWAT (water environment quality evaluation) are used for processing. For the social field, public opinion sentiment analysis based on a large language model (social stability evaluation) is used for processing. For the first The benefit value of each governance strategy;

[0211] The governance strategy with the highest overall benefit among all governance strategies is selected as the globally optimal governance strategy, as shown in the function below:

[0212]

[0213] in, The optimal governance strategy globally; The operator for the maximum value of the independent variable;

[0214] By integrating the globally optimal governance strategy and overall benefits, a final governance solution is generated, as shown below:

[0215]

[0216] in, For the final governance solution;

[0217] Specifically, a multi-objective optimization server (Dell PowerEdge R760) was used to calculate the comprehensive benefit function and search for the global optimal strategy, while a decision simulation server (Lenovo ThinkSystem SR650V3) was used to pre-simulate the effects of the governance strategy.

[0218] This invention has many specific applications, and the above are only preferred embodiments of this invention. It should be noted that for those skilled in the art, several improvements can be made without departing from the principle of this invention, and these improvements should also be considered within the scope of protection of this invention.

Claims

1. A digital twin governance system for e-government spaces based on multi-source data fusion, characterized in that: The system includes: Module 1: Data Acquisition Module, which collects initial multi-source data and preprocesses the multi-source data to obtain a spatiotemporal data base. The multi-source data includes initial three-dimensional ground-space data, initial social text data, and initial government affairs thematic data. Module 2: Physical Twin Module. Semantic nodes are extracted from the spatiotemporal data base to generate real and perturbation vectors of the city state. The real and perturbation vectors of the city state are input into the city mechanism model and corrected to obtain the digital twin result. Module 3: Social Simulation Module. Based on the aforementioned spatiotemporal data foundation, this module constructs social behavioral subjects and attribute vectors; introduces an environmental constraint mechanism to the social subjects to generate environmental constraint vectors; inputs the attribute vectors, environmental constraint vectors, and initial social text data into the large language model to generate initial vectors for behavioral decisions and performs iterative processing to obtain a social simulation set. Module 4: Verification Module. Based on the digital twin results and the social simulation set, a causal graph node set and a causal structure graph are constructed sequentially. The social simulation set is fed back to the urban mechanism model for updating, resulting in an updated social simulation set. The updated social simulation set, digital twin results, and causal structure graph are summarized to generate a comprehensive situation set. Initial spatiotemporal statements are extracted from the large language model and deduced and verified using the urban mechanism model to generate a final spatiotemporal statement set. Module 5: Governance Decision Module. Based on the comprehensive situation set and the final spatiotemporal statement set, the module calculates the benefits of behavioral strategies and selects the strategy with the maximum benefit as the optimal governance solution.

2. The system according to claim 1, characterized in that, The preprocessing in Module 1 includes: inputting the multi-source data into a spatiotemporal mapping function for spatiotemporal registration and inputting it into a semantic encoding function to generate semantic vectors corresponding to each of the registered multi-source data; summarizing and combining the different semantic vectors to generate a semantic node set and inputting any two nodes into an association weight function to generate cross-node association weights; establishing an association edge set based on the semantic node set and association weights; constructing a spatiotemporal graph using the semantic node set and association edge set; and building a spatiotemporal index set based on the spatiotemporal graph and combining it with the spatiotemporal graph to establish a spatiotemporal data foundation.

3. The system according to claim 1, characterized in that, The correction process in Module 2 includes: discretizing the urban mechanism model using an explicit time-progression method to obtain a complete urban state projection vector; extracting the observation state vector from the spatiotemporal data base and calculating the error between the urban state projection vector and the observation state vector to generate a prediction error; inputting the prediction error and the urban state projection vector together into a correction function to obtain an urban state correction vector; and summarizing and combining all the urban state correction vectors to generate a digital twin result.

4. The system according to claim 3, characterized in that, The environmental constraint mechanism in Module 3 includes mapping the urban state correction vector to the social environmental constraint space to generate an environmental constraint vector.

5. The system according to claim 1, characterized in that, The iterative processing in module three involves: iterating the attribute vector to generate the attribute vector for the next time step; re-inputting the attribute vector for the next time step, the environmental constraint vector, and the initial social text data into the preset large language model to generate a behavior decision correction vector; re-aggregating the behavior decision correction vector to generate an average behavior correction vector; and summarizing and combining all the average behavior correction vectors to generate a social simulation set.

6. The system according to claim 1, characterized in that, The update process in module four includes: calculating a social behavior disturbance vector using the feedback matrix and the social simulation set; re-inputting the social behavior disturbance vector and the real disturbance vector into the urban mechanism model to generate an urban state update vector; remapping the urban state update vector into the social environment constraint space to generate a corrected environmental constraint vector; re-inputting the attribute vector, the corrected environmental constraint vector, and the initial social text data into a preset large language model to generate a behavior decision update vector; and re-aggregating and combining the behavior decision update vector to generate an updated social simulation set.

7. The system according to claim 1, characterized in that, The deduction and verification in Module 4 includes: calculating the causal influence strength between any two causal nodes in the causal graph node set; statistically analyzing the causal influence strength to obtain a causal consistency index; comparing the causal consistency index with a preset judgment threshold; if the causal consistency index is greater than the judgment threshold, the updated social simulation set and the digital twin result are causally consistent; if the causal consistency index is less than the judgment threshold, the social behavior perturbation vector is recalculated until the causal consistency index is greater than the judgment threshold; and summarizing the updated social simulation set, the digital twin result, and the causal structure graph to generate the final spatiotemporal statement set.

Images