A digital twin three-dimensional simulation system of a smart city

Abstract

The application discloses a kind of digital twin three-dimensional simulation systems of wisdom city, it is related to wisdom city digital technology field, the system includes multi-source heterogeneous data space-time semantic fusion module, city-level multilevel three-dimensional scene dynamic construction module, adaptive LOD rendering scheduling module, physical information neural network simulation deduction module, low-delay virtual-real two-way mapping and closed-loop control module and cloud edge end collaborative computing power scheduling module, the precision fusion and semantic consistency alignment of multi-source heterogeneous data are realized by coordinate conversion and city ontology knowledge graph semantic reasoning, the synchronous guarantee of rendering frame rate and visual quality under large-scale urban scene is realized by LOD decision network model, physical mechanism partial differential equation constraint is embedded in neural network loss function and adaptive adjustment weight coefficient is guaranteed high-fidelity simulation deduction under extreme working condition, and real-time index and complexity index quantitative calculation realize the efficient collaborative scheduling of cloud edge end computing power.

Description

Technical Field

[0001] This invention relates to the field of smart city digital technology, specifically a digital twin 3D simulation system for smart cities. Background Technology

[0002] Digital twin technology, by constructing a high-fidelity digital copy of a physical city, enables real-time perception, simulation, and closed-loop control of the city's operational status. It has become a core technological support for the construction of new smart cities. With the explosive growth of the scale of urban Internet of Things terminals and the rapid development of 3D rendering technology, city-level digital twin 3D simulation systems are continuously evolving towards large scale, high precision, and strong real-time performance.

[0003] However, existing urban 3D simulation systems focus on static visualization, with low integration of multi-source heterogeneous data. Data from IoT sensors, video surveillance, geographic information systems, building information models, and urban operation systems are isolated from each other, with inconsistent time and spatial references, making it difficult to converge in real time to form a consistent information view. This results in lag and deviation between the digital twin and the physical city. At the same time, urban models at different scales, such as macro-level traffic flow models, meso-level regional environment models, and micro-level building internal personnel evacuation models, often use independent modeling and simulation tools, which are fragmented and cannot perform cross-scale linkage simulation. When it is necessary to go from global situation analysis to local refined deduction, it is necessary to switch systems and reset boundary conditions, which is cumbersome and prone to introducing inconsistencies.

[0004] Therefore, there is an urgent need for a digital twin 3D simulation system for smart cities that can achieve high-precision fusion of multi-source data and low-latency bidirectional mapping between virtual and real data, in order to solve the above-mentioned technical problems. Summary of the Invention

[0005] To address the technical challenges in existing smart city 3D simulation systems, such as low accuracy in multi-source heterogeneous data fusion, difficulty in balancing rendering performance and visual quality in large-scale scenes leading to inaccurate predictions under extreme conditions, and large latency in virtual-real mapping with a lack of closed-loop control capabilities that fails to meet real-time requirements, this invention designs a digital twin 3D simulation system for smart cities. This system includes a multi-source heterogeneous data spatiotemporal semantic fusion module, a city-level multi-level 3D scene dynamic construction module, an adaptive LOD rendering scheduling module, a physical information neural network simulation and deduction module, a low-latency virtual-real bidirectional mapping and closed-loop control module, and a cloud-edge-device collaborative computing power scheduling module. Through coordinate transformation and semantic reasoning from the city ontology knowledge graph, it achieves accurate fusion and semantic consistency alignment of multi-source heterogeneous data. The LOD decision network model ensures simultaneous guarantee of rendering frame rate and visual quality in large-scale urban scenes. Physical mechanism partial differential equation constraints are embedded into the neural network loss function, and adaptive adjustment of weight coefficients ensures high-fidelity simulation and deduction under extreme conditions. Virtual-real deviation detection and NSGA-II multi-objective optimization reverse control execute low-latency closed-loop regulation. Quantitative calculation of real-time and complexity indicators enables efficient collaborative scheduling of cloud-edge-device computing power.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a digital twin 3D simulation system for smart cities, the system comprising:

[0007] The multi-source heterogeneous data spatiotemporal semantic fusion module is used to connect to multiple types of data sources in the city, perform spatiotemporal benchmark normalization and semantic alignment processing based on the city ontology knowledge graph, and generate a city full-element fusion dataset.

[0008] A city-level multi-level 3D scene dynamic construction module is used to construct a city 3D digital twin scene model containing static base and dynamic elements based on the city's full-element fusion dataset.

[0009] The adaptive LOD rendering scheduling module is used to obtain viewpoint features and scene semantic weights, dynamically allocate the optimal LOD level to 3D objects in the scene through a pre-trained LOD decision network model, and perform multi-threaded parallel rendering.

[0010] The physical information neural network simulation and deduction module is used to embed the constraints of physical mechanism partial differential equations into the loss function of the data-driven prediction model to construct the PINN fusion simulation model, simulate and deduce the urban operation status, and map the simulation results to the urban three-dimensional digital twin scene model.

[0011] The low-latency virtual-real bidirectional mapping and closed-loop control module is used to acquire a snapshot of the physical city status at a preset acquisition cycle, incrementally update the digital twin space, calculate the virtual-real deviation, and generate reverse control commands for the physical city execution devices.

[0012] The cloud-edge-device collaborative computing power scheduling module is used to dynamically allocate computing tasks to edge preprocessing nodes, regional aggregation nodes, or central cloud nodes for execution based on the real-time requirements and computational complexity indicators of the computing tasks.

[0013] Furthermore, the spatiotemporal reference normalization process of the multi-source heterogeneous data spatiotemporal semantic fusion module is as follows:

[0014] The original urban data in the WGS-84 coordinate system, CGCS2000 coordinate system, local independent coordinate system and engineering coordinate system are uniformly converted to the preset city-level spatiotemporal reference coordinate system.

[0015] The original urban data of different precision levels are labeled with precision levels, and the multi-precision data of the same spatial location are weighted and fused based on the reciprocal weighted fusion rule of precision level;

[0016] The NTP protocol is used to synchronize and calibrate timestamps from different data sources, and each standardized data is marked with a unified spatiotemporal stamp containing four dimensions: longitude, latitude, elevation, and UTC time.

[0017] Furthermore, the urban ontology knowledge graph includes an urban spatial entity ontology layer, an urban operational event ontology layer, and an urban business rule ontology layer. The semantic alignment process performed by the multi-source heterogeneous data spatiotemporal semantic fusion module is as follows:

[0018] Based on the urban spatial entity ontology layer, a BERT-BiLSTM-CRF-based named entity recognition model is used to semantically annotate spatial entities in standardized urban data, generating spatial entity data with semantic tags.

[0019] Based on the urban operation event ontology layer, a BERT-based event extraction model is used to semantically annotate dynamic events in the standardized urban data, identify event type, event subject, event time, and event location, and generate dynamic event data with semantic tags.

[0020] Based on the city business rule ontology layer, the SWRL rule reasoning engine is used to perform association reasoning on spatial entity data with the semantic tags and dynamic event data with the semantic tags, establish semantic association relationships between spatial entities and dynamic events, and generate the city full-element fusion dataset.

[0021] Furthermore, the construction process of the city's three-dimensional digital twin scene model is as follows:

[0022] Based on the GIS data and oblique photogrammetry data in the urban full-element fusion dataset, a three-dimensional mesh model of the urban basic terrain at LOD0 level is generated.

[0023] For buildings with existing BIM data, extract the building's exterior geometric information and building attribute information from the BIM data to generate a detailed 3D solid model of the building at LOD3 level.

[0024] For buildings with only LiDAR point cloud data, the building outline is extracted using a point cloud semantic segmentation model based on PointNet++ network, and then a parametric modeling method is used to generate a LOD2 level 3D solid model of the building.

[0025] For buildings without BIM data and LiDAR point cloud data, oblique photogrammetry data and GIS data are used to generate LOD1 level 3D solid models of the buildings through procedural modeling methods.

[0026] The city's basic terrain 3D mesh model, refined building 3D entity model, and building 3D entity model are spatially registered in the city-level spatiotemporal reference coordinate system. Semantic links between the models are established through the semantic relationships in the city ontology knowledge graph, and a complete city 3D digital twin scene model is assembled.

[0027] Furthermore, the LOD decision network model is a decision network trained based on deep reinforcement learning, and the LOD levels include LOD0 macro-city level, LOD1 street level, LOD2 building level, and LOD3 component level.

[0028] The LOD decision network model is trained using a weighted composite score of rendering frame rate, visual quality, and video memory usage as the reward function.

[0029] Furthermore, the semantic importance weight value in the adaptive LOD rendering scheduling module The acquisition process is as follows:

[0030] Query 3D objects from the preset scene semantic rule base Basic semantic weights under the current task scenario type The basic semantic weights take values ​​in the range of [0, 1].

[0031] Calculate context-related weights based on the current user's task context information. ,in, For three-dimensional objects The Euclidean distance between the user's current focus area and the center point of the area they are currently interested in. For the Gaussian decay radius parameter of the region of interest, As an indicator function, when a 3D object Type The set of object types that users pay attention to The value is 1 if the condition is met, and 0 otherwise.

[0032] Calculate user interaction-related weights based on user interaction behavior. ,in, For three-dimensional objects The set of user interaction behaviors that occur on the platform For the first The basic weights of various interactive behaviors For the first The time interval between each interactive action and the current moment. The time decay constant of the interaction weights;

[0033] The semantic importance weight is obtained by weighting and fusing the basic semantic weight, context-related weight, and user interaction-related weight. ,in, β and γ are preset weighting coefficients and .

[0034] Furthermore, the steps for the physical information neural network simulation and deduction module to construct the PINN fusion simulation model are as follows:

[0035] Constructing a basic deep neural network, using spatiotemporal coordinates The city's operational status parameters serve as the input and the output.

[0036] Select a system of partial differential equations corresponding to the current simulation task, calculate the partial derivatives of the output of the basic deep neural network with respect to the input, and construct the residual terms of the system of partial differential equations as the physical constraint loss function. The fitting error term of the observed data is used as the data loss function. Construct the joint loss function: ,in, and These are adaptive weighting coefficients, all with an initial value of 1.0.

[0037] During the training process, every The data loss function is monitored in each training round. gradient magnitude and the physical constraint loss function gradient magnitude The gradient magnitude is calculated as follows: ,in, The number of layers in the neural network. For loss function For the first Layer parameters gradient vector, It is an L2 norm;

[0038] when and The ratio deviates from the preset range At that time, according to Adjust the adaptive weight coefficient and ,in, The PINN fusion simulation model is obtained by training the model by minimizing the joint loss function, where is the learning rate factor.

[0039] Furthermore, the low-latency virtual-real bidirectional mapping and closed-loop control module calculates the physical entity position. Location of digital twin model Spatial position deviation between ,when When the spatial position deviation exceeds a preset threshold, it is marked as an abnormal position deviation.

[0040] Calculate the operating status parameters of physical entities State parameters of digital twin model Deviation of state parameters between ,in, To prevent division by zero constant, when When the deviation exceeds a preset state parameter deviation threshold, it is marked as an abnormal state parameter deviation.

[0041] Constructing the physical topology adjacency matrix and digital twin topological adjacency matrix Calculate the topological relationship deviation ,in, For XOR operation, when The time marker is marked as an anomaly in topological relationship deviation;

[0042] Constructing a multi-objective optimization objective function Among them, the objective function of urban operation status deviation , Let be the target state vector. In order to control strategy The state vector predicted by the PINN fusion simulation model described below is used to control the energy consumption objective function. , For the first Unit control energy consumption coefficient of each actuator For the first Control quantity of an actuator Total number of devices to be executed;

[0043] Control robustness objective function , For the first Predicted state vector under various perturbation scenarios This represents the predicted state vector under a undisturbed scenario. This represents the total number of disturbance scenarios;

[0044] The Pareto optimal solution set is obtained by solving the multi-objective optimization objective function. The comprehensive optimal solution is selected from the Pareto optimal solution set as the final control strategy, and the reverse control command is generated and sent to the physical city execution device.

[0045] Furthermore, the real-time performance requirements in the cloud-edge-device collaborative computing power scheduling module... The calculation is based on the task deadline constraint and the data timeliness decay function. ,in, To calculate the deadline requirement for the task, This is the preset data timeliness attenuation coefficient. To calculate the waiting time of a task in the queue;

[0046] The computational complexity index The calculation is based on the floating-point operation volume and memory access volume of the computation task. Where FLOPs represents the number of floating-point operations required to perform the computation task. The total amount of memory access required for the computation task. The current available memory bandwidth, and These are the floating-point operation weighting coefficient and the memory access weighting coefficient, respectively. ;

[0047] Based on the aforementioned real-time requirement indicators and the computational complexity index Computing power is allocated according to the following rules:

[0048] when Higher than the first real-time threshold and When the complexity is below the first complexity threshold, the computation task is assigned to the edge preprocessing node;

[0049] when Higher than the second real-time threshold and When the complexity is between the first and second complexity thresholds, the computational task is assigned to the region aggregation node;

[0050] when Below the second real-time threshold or If the complexity exceeds the second complexity threshold, the computation task will be assigned to the central cloud node.

[0051] Compared with existing technologies, this digital twin 3D simulation system for smart cities has the following advantages:

[0052] I. This invention achieves multi-coordinate system-to-system transformation through a multi-source heterogeneous data spatiotemporal semantic fusion module. It realizes adaptive fusion of multi-precision data based on the reciprocal weighted fusion rule of precision level, and constructs a city ontology knowledge graph including spatial entity ontology layer, runtime event ontology layer, and business rule ontology layer. It performs semantic annotation on multi-source data through named entity recognition model and event extraction model, and then establishes deep semantic association between spatial entities and dynamic events through SWRL rule inference engine. This solves the problems of low data fusion accuracy and semantic conflict caused by non-uniform coordinate system, inconsistent precision level, and non-standard semantic description of multi-type data. It realizes centimeter-level high-precision fusion and semantic consistency guarantee of urban full-element data under a unified spatiotemporal benchmark.

[0053] Second, this invention uses a physical information neural network simulation and deduction module to embed the residual terms of the partial differential equations of the physical mechanism model into the loss function of the deep neural network, constructing a joint loss function. During training, the gradient magnitudes of data loss and physical constraint loss are monitored in real time. When the ratio of gradient magnitudes deviates from the preset range, the adaptive weight coefficients are automatically adjusted, so that the model training process maintains a dynamic balance between data fitting and physical constraints. This solves the problem that the prediction accuracy of pure data-driven models drops significantly in sparse training data regions and extreme conditions. The simulation results not only conform to historical observation data but also strictly meet physical constraints, improving the simulation fidelity and generalization ability of urban traffic flow prediction, air quality prediction, and urban flooding early warning scenarios.

[0054] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description

[0055] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0056] Figure 1 This is a data flow diagram of a digital twin 3D simulation system for a smart city, as described in an embodiment of the present invention.

[0057] Figure 2 This is a flowchart of the low-latency virtual-real bidirectional mapping and closed-loop control module in an embodiment of the present invention;

[0058] Figure 3This is an overall flowchart of a digital twin 3D simulation system for a smart city, as described in an embodiment of the present invention. Detailed Implementation

[0059] To better understand the above technical solutions, a detailed description of the solutions will be provided below in conjunction with the accompanying drawings and specific embodiments. Obviously, the described embodiments are merely some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0060] This invention provides a digital twin 3D simulation system for smart cities, such as... Figure 1 As shown, the system includes: a multi-source heterogeneous data spatiotemporal semantic fusion module, a city-level multi-level 3D scene dynamic construction module, an adaptive LOD rendering scheduling module, a physical information neural network simulation and deduction module, a low-latency virtual-real bidirectional mapping and closed-loop control module, and a cloud-edge-device collaborative computing power scheduling module.

[0061] In this embodiment, the multi-source heterogeneous data spatiotemporal semantic fusion module is used to connect to multiple types of urban data sources, perform spatiotemporal benchmark normalization and semantic alignment processing based on the urban ontology knowledge graph, and generate a fusion dataset of all urban elements. The input data includes multiple types of urban data from different acquisition devices, specifically including: vehicle trajectory data in the WGS-84 coordinate system acquired by the positioning system receiver, infrastructure positioning data in the CGCS2000 coordinate system, land boundary data in the local independent coordinate system, building construction layout data in the engineering coordinate system, real-time environmental monitoring data acquired by IoT sensors, traffic flow image sequences acquired by video surveillance equipment, three-dimensional point cloud data acquired by LiDAR scanning equipment, multi-view image data acquired by oblique photography cameras, and BIM data provided by the building information model storage server.

[0062] The collected data undergoes spatiotemporal benchmark normalization and semantic alignment based on the city ontology knowledge graph, wherein:

[0063] For spatiotemporal reference normalization, the input data undergoes format parsing and validity verification, eliminating data records with missing fields or values ​​exceeding reasonable ranges. The original urban data from WGS-84, CGCS2000, local independent coordinate systems, and engineering coordinate systems are uniformly converted to a preset city-level spatiotemporal reference coordinate system. After spatial reference conversion, the original urban data of different precision levels are labeled with precision levels. Precision level labeling is determined based on the nominal precision of the data source equipment and the data acquisition method. Data acquired using differential GPS is labeled with centimeter-level precision; data acquired using single-point GPS is labeled with meter-level precision; data acquired using total stations or 3D laser scanners is labeled with millimeter-level precision; and data extracted from remote sensing imagery or topographic maps is labeled with decimeter- to meter-level precision. After precision level labeling, multi-precision data from the same spatial location are weighted and fused based on a reciprocal weighted fusion rule for precision levels. The network time protocol is used to synchronize and calibrate the timestamps of different data sources. Coordinated Universal Time (UTC) is obtained through a GPS receiver as the reference time source, and each data acquisition terminal synchronizes its clock with the time server through the NTP protocol.

[0064] For semantic alignment processing, it is performed based on the city ontology knowledge graph, which includes a city spatial entity ontology layer, a city operation event ontology layer, and a city business rule ontology layer. The city spatial entity ontology layer defines the type system of city spatial entities, including entity types such as roads, buildings, bridges, tunnels, rail transit facilities, underground pipelines, greening facilities, traffic signal facilities, environmental monitoring stations, and camera surveillance equipment. Each entity type defines a corresponding attribute set. The city operation event ontology layer defines the type system of various dynamic events in the city operation process, including traffic congestion events, traffic accident events, environmental anomaly events, weather warning events, municipal facility failure events, and public safety events. Each event type defines slots such as event subject, event time, event location, and event attributes. The city business rule ontology layer defines the business logic constraints and association rules between spatial entities and dynamic events.

[0065] Based on the urban spatial entity ontology layer, a BERT-BiLSTM-CRF-based named entity recognition model is used to semantically annotate spatial entities in standardized urban data. The text description field in the standardized urban data is input into a pre-trained BERT Chinese model to obtain the contextual semantic vector representation of each character. The output sequence of the BERT model is input into a bidirectional long short-term memory (BiLSTM) layer. The forward LSTM and backward LSTM encode the input sequence respectively. The hidden states in the two directions are concatenated to obtain the bidirectional semantic feature vector of each character. The output of the BiLSTM layer is input into a conditional random field (CRF) layer. The CRF layer learns the transition probability between adjacent labels and decodes the globally optimal label sequence through the Viterbi algorithm. After processing by the named entity recognition model, spatial entity data with semantic labels is generated. Each piece of spatial entity data contains a unique entity identifier, entity type label, entity name, spatial coordinates, and a set of attribute key-value pairs.

[0066] Based on the aforementioned city operation event ontology layer, a BERT-based event extraction model is used to semantically label dynamic events in standardized city data. This event extraction model adopts a pipeline architecture, including an event trigger word recognition module and an event argument extraction module. The event trigger word recognition module uses a BERT-BiLSTM architecture to identify trigger words in the text that indicate the occurrence of events and determine the event type. The event argument extraction module uses a reading comprehension architecture based on a pre-trained language model to transform the event argument extraction task into a machine reading comprehension task. For each predefined argument role of each event type, a corresponding natural language query question is constructed, and the answer is extracted from the text as the argument filling value. The event instance output by the event extraction model includes four core elements: event type, event subject, event time, and event location. After processing by the event extraction model, dynamic event data with semantic labels is generated.

[0067] Based on the aforementioned city business rule ontology layer, the SWRL rule reasoning engine is used to perform association reasoning on spatial entity data and dynamic event data with semantic tags. The SWRL rule language supports the construction of implication rules on the basis of the OWL ontology. Concepts, attributes, and relationships in the city spatial entity ontology layer, city operation event ontology layer, and city business rule ontology layer are imported into the working memory of the rule reasoning engine. The spatial entity data with semantic tags generated in the first step and the dynamic event data with semantic tags generated in the second step are transformed into OWL individual instances. The predefined SWRL rules include: for any spatial entity and dynamic event, if the event location is within the spatial coverage area of ​​the entity and the event type belongs to the entity's associated event type set, then an "occurrence" relationship is established between the entity and the event. In the context of "relationships," for any two spatial entities, if their spatial coverage overlaps, a "spatial adjacency" relationship is established between them. For any dynamic events, Event 1 and Event 2, if Event 1 occurs earlier than Event 2, their locations are the same, and the propagation distance of Event 1 and the influence range of Event 2 meet the preset causal propagation conditions, a "causal association" relationship is established between them. The SWRL rule reasoning engine establishes semantic associations between spatial entities and dynamic events by constructing a rule network and incrementally performing rule matching when working memory changes. Ultimately, it generates a city-wide fusion dataset. Each data record in this dataset contains standardized spatial coordinate data, unified timestamp data, semantic category labels, and semantic associations with other data records.

[0068] In this embodiment, the city-level multi-level 3D scene dynamic construction module constructs a city 3D digital twin scene model containing a static base and dynamic elements based on the city's full-element fusion dataset. The construction of the city 3D digital twin scene model is performed in a hierarchical order from macro to micro, specifically including the construction of the basic urban terrain at LOD0 level, the construction of refined buildings at LOD3 level, the construction of buildings at LOD2 level, the construction of buildings at LOD1 level, and the spatial registration and semantic linking of models at each level.

[0069] The construction of the LOD0-level urban basic terrain 3D mesh model is based on GIS data and oblique photogrammetry data from the urban full-element fusion dataset. The GIS data includes digital elevation model data and digital orthophoto map data, while the oblique photogrammetry data includes a sequence of urban aerial images taken from multiple angles. Aerial triangulation is performed on the oblique photogrammetry data to restore the exterior orientation elements and camera intrinsic parameters of each image. Based on the exterior orientation elements of the images and the digital elevation model data, dense point clouds of the urban surface are generated through multi-view dense matching, and then converted into an LOD0-level urban basic terrain 3D mesh model through rasterization.

[0070] For buildings with existing BIM data, the building's exterior geometric information and building attribute information are extracted from the BIM data to generate a detailed 3D solid model of the building at LOD3 level. The BIM file is read by the IFC parser to extract the geometric representation and attribute information of the components. The local coordinates of each component are transformed to the city-level spatiotemporal reference coordinate system through a local placement transformation matrix. The attribute information includes the component's material type, structural type, construction year, and functional use. The extracted building exterior geometric information is geometrically simplified and repaired, and the detailed structure of the internal components is removed while retaining the main geometric features of the building facade to generate a detailed 3D solid model of the building at LOD3 level.

[0071] For buildings with only LiDAR point cloud data, the building outline is extracted using a point cloud semantic segmentation model based on the PointNet++ network, and then a parametric modeling method is used to generate a LOD2 level 3D solid model of the building.

[0072] For buildings without BIM data and LiDAR point cloud data, the system uses oblique photogrammetry data and GIS data to generate a LOD1 level 3D solid model of the building through a procedural modeling method.

[0073] After the models at each level are generated, the 3D mesh model of the basic urban terrain, the refined 3D solid model of buildings at LOD3 level, the 3D solid model of buildings at LOD2 level, and the 3D solid model of buildings at LOD1 level are spatially registered in the city-level spatiotemporal reference coordinate system. For each model, the coordinates of its reference point in the city-level spatiotemporal reference coordinate system are extracted. If the model is already in the city-level spatiotemporal reference coordinate system, it is directly retained. If the coordinate reference of the model is a local coordinate system, it is transformed using the previously recorded coordinate transformation parameters. After spatial registration is completed, semantic links between the models are established through the semantic association relationships in the urban ontology knowledge graph. The semantic links are established by querying the "spatial inclusion" relationship between each building entity and the entity of its block in the city's full-element fusion dataset, as well as the "spatial adjacency" relationship between each building entity and its adjacent building entities. These relationships are stored as edge information between the 3D models. The edge information includes the relationship type and the relationship confidence. For two building models with a "spatial adjacency" relationship, a semantic link is generated between them. After spatial registration and semantic linking are completed, a complete 3D digital twin scene model of the city is assembled.

[0074] In this embodiment, the adaptive LOD rendering scheduling module queries the preset scene semantic rule library for each 3D object in the 3D scene to determine the basic semantic weight of the 3D object under the current task scene type. The scene semantic rule library stores the basic semantic weight values ​​of various 3D objects under different task scene types. The basic semantic weights range from [0, 1], and the task scenario types include urban planning browsing scenario, traffic operation monitoring scenario, emergency command and dispatch scenario, and environmental assessment and analysis scenario.

[0075] Calculate context-related weights based on the current user's task context information. ,in, For three-dimensional objects The Euclidean distance between the user's current focus area and the center point of the area they are currently interested in. For the Gaussian decay radius parameter of the region of interest, As an indicator function, when a 3D object Type The set of object types that users pay attention to The value is 1 when the condition is met and 0 otherwise. This indicator function ensures that objects that are not of the user's interest type do not receive high context weight even if they are very close in space, thus avoiding over-rendering of non-interested objects.

[0076] Calculate user interaction-related weights based on user interaction behavior. User interaction behaviors are recorded in the system log. Each interaction record includes an interaction timestamp, an interaction object identifier, an interaction behavior type, and interaction behavior parameters. The interaction behavior types include five types: click to select, box selection for highlighting, attribute viewing, view focusing, and marking for favorites. For three-dimensional objects The set of user interaction behaviors that occur on the platform For the first The basic weights of various interactive behaviors For the first The time interval between each interactive action and the current moment. The time decay constant of the interaction weights;

[0077] The semantic importance weight is obtained by weighting and fusing the basic semantic weight, context-related weight, and user interaction-related weight. ,in, β and γ are preset weighting coefficients and In this embodiment, α is 0.4, β is 0.35, and γ is 0.25. The selection principle for this set of weight coefficients is as follows: the basic semantic weights provide a stable semantic importance benchmark, the context-related weights reflect the dynamic requirements of the current task scenario, and the user interaction-related weights reflect the user's personalized attention preferences, so as to avoid personalized factors from excessively affecting the overall rendering strategy of the system.

[0078] The LOD levels include LOD0 (Macro City Level), LOD1 (Street Level), LOD2 (Architecture Level), and LOD3 (Component Level). The LOD0 macro city level model expresses the overall spatial pattern of the city using terrain meshes and major landmark building blocks. The LOD1 street level model expresses the overall form of building groups and streets. The LOD2 architecture level model expresses the detailed geometric form and facade texture of buildings. The LOD3 component level model expresses the detailed components and decorative elements of buildings. The LOD decision network model is a decision network trained based on deep reinforcement learning. This network is trained using a near-end policy optimization algorithm. The input state vector of the LOD decision network includes: the current frame rendering frame rate, memory usage, the current virtual camera viewpoint position, viewpoint orientation vector, the basic semantic weight vector of each 3D object in the scene, the context-related weight vector, and the user interaction-related weight vector. The output of the LOD decision network is the optimal LOD level number assigned to each 3D object in the scene. This number takes values ​​of 0, 1, 2, and 3, corresponding to the four levels from LOD0 to LOD3, respectively. The LOD decision network employs a multilayer perceptron architecture with three hidden layers containing 512, 256, and 128 neurons respectively. The ReLU activation function is used, and the output layer uses the Softmax function to output the probability distribution of each LOD level. The final allocation result is obtained through sampling. The LOD decision network is trained using a weighted comprehensive score of rendering frame rate, visual quality, and memory usage as the reward function. During training, the system randomly generates viewpoint positions and orientations in the virtual city scene. Each training round executes 1000 decision steps, using the PPO algorithm to replace the objective function and update the network parameters. The learning rate is set... The batch size is set to 64. After training convergence, the LOD decision network can dynamically assign the optimal LOD level to each 3D object in the scene in real time based on the current viewpoint features and scene semantic weights.

[0079] In this embodiment, the steps for the physical information neural network simulation and deduction module to construct the PINN fusion simulation model are as follows:

[0080] Constructing a basic deep neural network, using spatiotemporal coordinates The city's operational status parameters serve as the input and the output.

[0081] Select a system of partial differential equations corresponding to the current simulation task, calculate the partial derivatives of the output of the basic deep neural network with respect to the input, and construct the residual terms of the system of partial differential equations as the physical constraint loss function. The fitting error term of the observed data is used as the data loss function. Construct the joint loss function: ,in, and These are adaptive weighting coefficients, all with an initial value of 1.0.

[0082] During the training process, every The data loss function is monitored in each training round. gradient magnitude and the physical constraint loss function gradient magnitude The gradient magnitude is calculated as follows: ,in, The number of layers in the neural network. For loss function For the first Layer parameters gradient vector, It is an L2 norm;

[0083] when and The ratio deviates from the preset range At that time, according to Adjust the adaptive weight coefficient and ,in, The PINN fusion simulation model is obtained by training the model by minimizing the joint loss function, where is the learning rate factor.

[0084] After the PINN fusion simulation model is trained, the simulation results are mapped to a 3D digital twin model of the city. This mapping process is achieved through data binding: the state parameter values ​​output by the simulation are associated with corresponding spatial entities in the 3D digital twin model using unified spatiotemporal stamps. For example, the traffic flow density and speed values ​​for each road segment output by the traffic flow simulation are associated with the road model in the 3D scene using road segment identifiers, and visualized using color mapping or a dynamic particle system. Similarly, the pollutant concentration values ​​at each spatial point output by the air pollutant diffusion simulation are associated with an air cube mesh in the 3D scene using spatial coordinate matching, and visualized as a semi-transparent volumetric cloud. Once mapped, the 3D digital twin model of the city can present the simulated future situation, providing predictive decision support for city managers.

[0085] In this embodiment, as Figure 2As shown, the low-latency virtual-real bidirectional mapping and closed-loop control module is used to acquire physical city state snapshots at a preset acquisition cycle, incrementally update the digital twin space, and calculate the physical entity locations. Location of digital twin model Spatial position deviation between ,when When the spatial location deviation exceeds a preset threshold, it is marked as an abnormal location deviation, triggering a location correction process. The location correction process includes: querying the location measurement confidence in the physical city status snapshot; if the confidence is higher than 95%, the physical location is used to replace the digital twin location; if the confidence is lower than 95%, the physical location and the digital twin location are fused through Kalman filtering to obtain the corrected location estimate.

[0086] Calculate the operating status parameters of physical entities State parameters of digital twin model Deviation of state parameters between , These are the values ​​of the status parameters recorded in the physical city status snapshot. The state parameter values ​​and state parameter deviations of the corresponding virtual entity in the digital twin space. ,in, To prevent division by zero constant, the value is taken as zero in this embodiment. Preset state parameter deviation threshold, when When the deviation exceeds the preset state parameter threshold, the entity is marked as having an abnormal state parameter deviation, triggering a state correction process. The state correction process includes: using physical state parameters... As observed values, digital twin state parameters As predicted values, the Bayesian update method is used to calculate the fused state estimate.

[0087] Constructing the physical topology adjacency matrix and digital twin topological adjacency matrix Calculate the topological relationship deviation Physical topological adjacency matrix for The square array, in which The total number of urban spatial entities. Represents entities in physical space With entity There are direct connections or adjacent relationships between them. This indicates that the relationship does not exist; digital twin topological adjacency matrix. Defined in the same way, it represents the topological connection relationship between entities in the digital twin space, and the topological relationship deviation. ,in, When the values ​​are different (i.e., one is 0 and the other is 1), the XOR operation results in 1; otherwise, the result is 0. At time 0, it is marked as an anomaly in topological relationship deviation, triggering the topology correction process. The topology correction process includes: comparison and The algorithm identifies all inconsistent edges. For each inconsistent edge, it checks whether the physical connection actually exists. If it exists, the corresponding edge in the digital twin topology matrix is ​​set to 1; otherwise, it is set to 0. A multi-objective optimization function is then constructed. This is used to generate reverse control commands. The multi-objective optimization problem includes: the objective function of urban operation state deviation. Energy consumption control objective function and control robustness objective function Control variable vector ,in The total number of devices to be executed. For the first Control quantities of each executing device. Objective function for deviation of urban operating status. Defined as target state vector With control strategy The state vector predicted by the PINN fusion simulation model The sum of squared errors between: ,in, Let be the target state vector. The PINN fusion simulation model uses the current state as the initial condition and the control strategy... The future state vector is obtained from simulation prediction when given external input; the control energy consumption objective function is... Defined as the sum of the control energy consumption of all executing devices: ,in, For the first The unit control energy consumption coefficient of each actuator is obtained through parameters on the device nameplate or statistical analysis of historical operating data. For the first Control quantity of each execution device; control robustness objective function Defined as the worst-case performance of control under all considered uncertainty disturbance scenarios, assuming a total of Such a disturbance scenario, in the first Control strategy in various disturbance scenarios After execution of control, the state vector predicted by the PINN fusion simulation model is: The predicted state vector in the undisturbed scenario is Control robustness objective function: The negative sign in this formula indicates that, when solving the minimization problem, a more robust control strategy (i.e., a smaller deviation between the disturbed and undisturbed scenarios) corresponds to... The smaller the value, the more likely the disturbance scenarios are to fall into three categories: sensor measurement noise, actuator response delay, and sudden changes in the external environment. The system employs a multi-objective evolutionary algorithm based on NSGA-II to solve this multi-objective optimization problem, obtaining a Pareto optimal solution set. The parameters of the NSGA-II algorithm are set as follows: population size 100, maximum number of generations 200, crossover probability 0.9, and mutation probability 0.1. After execution, a set of Pareto optimal solutions is obtained, each solution corresponding to a set of control policy vectors. The comprehensive optimal solution is selected from the Pareto optimal solution set as the final control policy. The selection of the comprehensive optimal solution uses the TOPSIS decision method: calculating the Euclidean distance from each solution to the positive ideal solution (the vector composed of the minimum values ​​of all objectives) and the Euclidean distance to the negative ideal solution (the vector composed of the maximum values ​​of all objectives), and selecting the solution with the highest relative proximity as the comprehensive optimal solution. The system generates a reverse control command based on the comprehensive optimal solution. This reverse control command includes the execution device identifier, control quantity value, and execution timestamp. Control commands are sent to physical city execution devices via the MQTT protocol. The execution devices adjust their operating status according to the control commands to achieve closed-loop control of the physical city.

[0088] In this embodiment, the cloud-edge-device collaborative computing power scheduling module adopts a three-layer topology. The edge layer includes smart terminal devices deployed throughout the city, the edge layer includes edge preprocessing nodes and regional aggregation nodes, and the cloud layer includes a central cloud node. When a computing task arrives, it is parsed to extract information such as task type identifier, input data size, expected output format, and task dependencies, as well as the real-time performance requirements of the computing task. ,in, To calculate the deadline requirement for the task, The preset data timeliness attenuation coefficient is set to a value of [value] in this embodiment. This coefficient reflects the rate at which the value of data decays over time. To calculate the waiting time of a task in the queue, this value is calculated from the difference between the time the task was enqueued and the current time. This is a computational complexity metric for the task. FLOPs represents the number of floating-point operations required to compute the task, obtained through algorithm complexity analysis and input data size estimation. For neural network inference tasks, FLOPs are obtained by counting the number of multiplication and addition operations in each layer of the network; for matrix operation tasks, FLOPs are calculated based on the matrix dimension and operation type. This refers to the total memory access required for the computation task, including the amount of input data read, intermediate variable storage, and output data write. The current available memory bandwidth is obtained in real time through the system monitoring module. and These are the floating-point operation weighting coefficient and the memory access weighting coefficient, respectively, and they satisfy... In this embodiment The value is 0.6. The value is 0.4.

[0089] Based on real-time requirements indicators and computational complexity metrics The system allocates computing power according to preset allocation rules, which are as follows:

[0090] when Higher than the first real-time threshold and When the complexity is below a first complexity threshold, the computation task is assigned to an edge preprocessing node. In this embodiment, the first real-time threshold is set to a value of [value missing]. In this embodiment, the first complexity threshold is set to 3.0;

[0091] when Higher than the second real-time threshold and When the computational task falls between the first and second complexity thresholds, it is assigned to the regional aggregation node. In this embodiment, the second real-time threshold is set to [value missing]. In this embodiment, the second complexity threshold is set to 6.0.

[0092] when Below the second real-time threshold or When the complexity exceeds the second complexity threshold, the computation task will be assigned to the central cloud node.

[0093] In this embodiment, as Figure 3 As shown, the working principle of this digital twin 3D simulation system for smart cities is as follows:

[0094] The multi-source heterogeneous data spatiotemporal semantic fusion module continuously receives urban data streams from various data sources, performs spatiotemporal benchmark normalization and semantic alignment processing, and generates a fusion dataset of all urban elements.

[0095] The city-level multi-level 3D scene dynamic construction module builds an initial city 3D digital twin scene model based on the dataset. During system operation, the low-latency virtual-real bidirectional mapping and closed-loop control module acquires physical city state snapshots at a preset acquisition cycle, triggering an incremental update process to keep the digital twin scene model synchronized with the physical city.

[0096] The adaptive LOD rendering scheduling module dynamically adjusts the LOD level of each 3D object based on the user's real-time viewpoint changes and interactive behavior to ensure that the rendering frame rate meets the smoothness requirements.

[0097] The physical information neural network simulation and inference module performs simulation and inference based on the current city status, and maps the prediction results to a three-dimensional scene to show the future situation;

[0098] When the virtual-real deviation exceeds the threshold, the low-latency virtual-real bidirectional mapping and closed-loop control module solves the multi-objective optimization problem, generates reverse control commands and sends them to the execution device to form closed-loop regulation.

[0099] The cloud-edge-device collaborative computing power scheduling module continuously monitors the real-time requirements and computational complexity of each computing task throughout the entire operation process, dynamically allocates computing resources to the optimal node, and ensures low latency and high throughput of the system as a whole.

[0100] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A digital twin 3D simulation system for smart cities, characterized in that, The system includes: The multi-source heterogeneous data spatiotemporal semantic fusion module is used to connect to multiple types of data sources in the city, perform spatiotemporal benchmark normalization and semantic alignment processing based on the city ontology knowledge graph, and generate a city full-element fusion dataset. A city-level multi-level 3D scene dynamic construction module is used to construct a city 3D digital twin scene model containing static base and dynamic elements based on the city's full-element fusion dataset. The adaptive LOD rendering scheduling module is used to obtain viewpoint features and scene semantic weights, dynamically allocate the optimal LOD level to 3D objects in the scene through a pre-trained LOD decision network model, and perform multi-threaded parallel rendering. The physical information neural network simulation and deduction module is used to embed the constraints of physical mechanism partial differential equations into the loss function of the data-driven prediction model to construct the PINN fusion simulation model, simulate and deduce the urban operation status, and map the simulation results to the urban three-dimensional digital twin scene model. The low-latency virtual-real bidirectional mapping and closed-loop control module is used to acquire a snapshot of the physical city status at a preset acquisition cycle, incrementally update the digital twin space, calculate the virtual-real deviation, and generate reverse control commands for the physical city execution devices. The low-latency virtual-real bidirectional mapping and closed-loop control module calculates the physical entity position. Location of digital twin model Spatial position deviation between ,when When the spatial position deviation exceeds a preset threshold, it is marked as an abnormal position deviation. Calculate the operating status parameters of physical entities State parameters of digital twin model Deviation of state parameters between ,in, To prevent division by zero constant, when When the deviation exceeds a preset state parameter deviation threshold, it is marked as an abnormal state parameter deviation. Constructing the physical topological adjacency matrix and digital twin topological adjacency matrix Calculate the topological relationship deviation ,in, For XOR operation, when The time marker is marked as an anomaly in topological relationship deviation; Constructing a multi-objective optimization objective function Among them, the objective function of urban operation status deviation , Let be the target state vector. In order to control strategy The state vector predicted by the PINN fusion simulation model described below is used to control the energy consumption objective function. , For the first Unit control energy consumption coefficient of each actuator For the first Control quantity of an actuator Total number of devices to be executed; Control robustness objective function , For the first Predicted state vector under various perturbation scenarios This represents the predicted state vector under a undisturbed scenario. This represents the total number of disturbance scenarios; The Pareto optimal solution set is obtained by solving the multi-objective optimization objective function. The comprehensive optimal solution is selected from the Pareto optimal solution set as the final control strategy. The reverse control command is generated and sent to the physical city execution device. The cloud-edge-device collaborative computing power scheduling module is used to dynamically allocate computing tasks to edge preprocessing nodes, regional aggregation nodes, or central cloud nodes for execution based on the real-time requirements and computational complexity indicators of the computing tasks; The modules interact asynchronously through message queues and use a distributed consistency protocol to ensure state synchronization between modules, thereby achieving decoupling and low-latency communication.

2. The digital twin 3D simulation system for smart cities according to claim 1, characterized in that, The spatiotemporal reference normalization process of the multi-source heterogeneous data spatiotemporal semantic fusion module is as follows: The original urban data in the WGS-84 coordinate system, CGCS2000 coordinate system, local independent coordinate system and engineering coordinate system are uniformly transformed to the preset city-level spatiotemporal reference coordinate system. That is, the Bursa seven-parameter model is adopted, and the coordinate transformation equation is constructed through three translation parameters (ΔX, ΔY, ΔZ), three rotation parameters (εx, εy, εz) and one scale parameter m. The optimal values ​​of the parameters are solved by the least squares method to achieve high-precision transformation between different coordinate systems. The original city data at different precision levels are labeled with precision classification, and multi-precision data at the same spatial location are weighted and fused based on a weighted fusion rule of the inverse of the precision level. The weighted fusion is as follows: ,in, For the observation of the k-th data source, Let K be the standard deviation corresponding to the accuracy level of the data source, K be the total number of data sources at the same spatial location, and the variance of the fused location estimate be... ; The NTP protocol is used to synchronize and calibrate the timestamps of different data sources. After calibration, each standardized data is marked with a unified spatiotemporal stamp containing four dimensions: longitude, latitude, elevation and UTC time.

3. The digital twin 3D simulation system for smart cities according to claim 1, characterized in that, The city ontology knowledge graph includes a city spatial entity ontology layer, a city operational event ontology layer, and a city business rule ontology layer. The semantic alignment process performed by the multi-source heterogeneous data spatiotemporal semantic fusion module is as follows: Based on the urban spatial entity ontology layer, a BERT-BiLSTM-CRF-based named entity recognition model is used to semantically label spatial entities in standardized urban data. The named entity recognition model extracts character-level contextual features through BERT, captures bidirectional sequence dependencies through BiLSTM, calculates label transition probabilities through the CRF layer, and decodes the optimal label sequence using the Viterbi algorithm. During training, the sum of the cross-entropy loss function and the negative log-likelihood loss of CRF is minimized to generate spatial entity data with semantic labels. Based on the city operation event ontology layer, a BERT-based event extraction model is used to semantically label dynamic events in the standardized city data, identifying event type, event subject, event time, and event location. The event extraction model adopts a pipeline architecture, where BERT-BiLSTM is used for sequence labeling of event trigger words, and a pointer network based on reading comprehension is used for event argument extraction. Each argument role is constructed as a natural language question, and the answer is extracted from the text as the argument filling value to generate dynamic event data with semantic labels. Based on the city business rule ontology layer, the SWRL rule inference engine is used to perform association inference on spatial entity data with the semantic tags and dynamic event data with the semantic tags. The SWRL rules include: when the location of an event is within the coverage area of ​​an entity and the event type belongs to the entity-associated event type set, then it is an entity-event occurrence relationship; when the spatial coverage areas of two entities overlap, it is a spatial adjacency relationship. The inference engine uses the Rete algorithm to optimize rule matching efficiency and supports incremental inference. When new data is added, only the affected part is recalculated. By establishing semantic relationships between spatial entities and dynamic events, the city's full-element fusion dataset is generated.

4. The digital twin 3D simulation system for smart cities according to claim 1, characterized in that, The construction process of the city's three-dimensional digital twin scene model is as follows: Based on the GIS data and oblique photogrammetry data in the urban full-element fusion dataset, a three-dimensional mesh model of the urban basic terrain at LOD0 level is generated. For buildings with existing BIM data, extract the building's exterior geometric information and building attribute information from the BIM data to generate a detailed 3D solid model of the building at LOD3 level. For buildings with only LiDAR point cloud data, the building outline is extracted using a point cloud semantic segmentation model based on PointNet++ network, and then a parametric modeling method is used to generate a LOD2 level 3D solid model of the building. For buildings without BIM data and LiDAR point cloud data, oblique photogrammetry data and GIS data are used to generate LOD1 level 3D solid models of the buildings through procedural modeling methods. The three-dimensional mesh model of the urban basic terrain, the refined three-dimensional solid model of the building, and the three-dimensional solid model of the building are spatially registered in the city-level spatiotemporal reference coordinate system. The spatial registration adopts the iterative nearest point algorithm, which calculates the rigid body transformation matrix iteratively by minimizing the sum of squared Euclidean distances between the source model point set and the target model point set until the convergence error is less than a preset threshold. After registration, semantic links between models are established through semantic relationships in the city ontology knowledge graph. These semantic links are stored in a graph database. Each model node contains attributes such as model identifier, spatial bounding box, and semantic type. Edges represent spatial inclusion, spatial adjacency, and logical association relationships. A confidence score is assigned to each edge, and a complete three-dimensional digital twin scene model of the city is assembled.

5. The digital twin 3D simulation system for smart cities according to claim 1, characterized in that, The LOD decision network model is a decision network trained based on deep reinforcement learning. The LOD levels include LOD0 macro-city level, LOD1 street level, LOD2 building level, and LOD3 component level. The LOD decision network model is trained using a weighted composite score of rendering frame rate, visual quality, and memory usage as the reward function. The reward function is... Where FPS is the current frame rate. Target frame rate, Q is the visual quality score, and MEM is the video memory usage. This is the maximum amount of video memory. The weights are the nearest-nearest policy optimization algorithm used in training, and the objective function is... ,in, The probability ratio of the strategies. For the estimation of the advantage function, Let be the mathematical expectation of time step t. To prune parameters, network parameters are updated via gradient ascent. During training, viewpoint positions and orientations are generated in a virtual city scene. Each iteration executes 1000 decision steps with a batch size of 64 and a learning rate of 3e-4 until the reward function converges.

6. The digital twin 3D simulation system for smart cities according to claim 1, characterized in that, The semantic importance weight value in the adaptive LOD rendering scheduling module The acquisition process is as follows: Query 3D objects from the preset scene semantic rule base Basic semantic weights under the current task scenario type The basic semantic weights take values ​​in the range of [0, 1]. Calculate context-related weights based on the current user's task context information. ,in, For three-dimensional objects The Euclidean distance between the user's current focus area and the center point of the area they are currently interested in. For the Gaussian decay radius parameter of the region of interest, As an indicator function, when a 3D object Type The set of object types that users pay attention to The value is 1 if the condition is met, and 0 otherwise. Calculate user interaction-related weights based on user interaction behavior. ,in, For three-dimensional objects The set of user interaction behaviors that occur on the platform For the first The basic weights of various interactive behaviors For the first The time interval between each interactive action and the current moment. The time decay constant of the interaction weights; The semantic importance weight is obtained by weighting and fusing the basic semantic weight, context-related weight, and user interaction-related weight. ,in, β and γ are preset weighting coefficients and .

7. The digital twin 3D simulation system for smart cities according to claim 1, characterized in that, The steps for constructing the PINN fusion simulation model using the physical information neural network simulation and deduction module are as follows: Constructing a basic deep neural network, using spatiotemporal coordinates As input, and with city operation status parameters as output, the basic deep neural network adopts a residual network architecture, which includes an input layer, four residual blocks and an output layer. Each residual block contains two fully connected layers, and each layer has 128 neurons. Select a system of partial differential equations corresponding to the current simulation task, calculate the partial derivatives of the output of the basic deep neural network with respect to the input, and construct the residual terms of the system of partial differential equations as the physical constraint loss function. The fitting error term of the observed data is used as the data loss function. Construct the joint loss function: ,in, and These are adaptive weighting coefficients, all with an initial value of 1.

0. During the training process, every The data loss function is monitored in each training round. gradient magnitude and the physical constraint loss function gradient magnitude The gradient magnitude is calculated as follows: ,in, The number of layers in the neural network. loss function For the first Layer parameters gradient vector, It is an L2 norm; when and The ratio deviates from the preset range At that time, according to Adjust the adaptive weight coefficient and ,in, The PINN fusion simulation model is obtained by training the model by minimizing the joint loss function, where is the learning rate factor.

8. The digital twin 3D simulation system for smart cities according to claim 1, characterized in that, The real-time performance requirements in the cloud-edge-device collaborative computing power scheduling module The calculation is based on the task deadline constraint and the data timeliness decay function. ,in, To calculate the deadline requirement for the task, This is the preset data timeliness attenuation coefficient. To calculate the waiting time of a task in the queue; The computational complexity index The calculation is based on the floating-point operation volume and memory access volume of the computation task. Where FLOPs represents the number of floating-point operations required to perform the computation task. The total amount of memory access required for the computation task. The current available memory bandwidth, and These are the floating-point operation weighting coefficient and the memory access weighting coefficient, respectively. ; Based on the aforementioned real-time requirement indicators and the computational complexity index Computing power is allocated according to the following rules: when Higher than the first real-time threshold and When the complexity is below the first complexity threshold, the computation task is assigned to the edge preprocessing node; when Higher than the second real-time threshold and When the complexity is between the first and second complexity thresholds, the computational task is assigned to the region aggregation node; when Below the second real-time threshold or If the complexity exceeds the second complexity threshold, the computation task will be assigned to the central cloud node.

9. The digital twin 3D simulation system for smart cities according to claim 1, characterized in that, The system performs an initialization process upon startup: The spatiotemporal semantic fusion module of multi-source heterogeneous data loads the city ontology knowledge graph and historical data, completes the spatiotemporal benchmark and semantic alignment, and generates the initial city full-element fusion dataset. The city-level multi-level 3D scene dynamic construction module constructs an initial 3D digital twin scene model based on the initial city full-element fusion dataset; The adaptive LOD rendering scheduling module loads the parameters of the pre-trained LOD decision network model and initializes the scene semantic rule library; The physical information neural network simulation and deduction module loads the pre-trained PINN fusion simulation model weights, and the low-latency virtual-real bidirectional mapping and closed-loop control module establishes a communication link with the physical city data acquisition equipment and execution equipment, and initializes the virtual-real deviation threshold parameter. The cloud-edge-device collaborative computing power scheduling module starts health monitoring and task queues for each computing node, and enters normal operation after completing the handshake protocol.

Images