A method for simulating spatial evolution of commercial blocks based on space-time analysis

Abstract

This invention discloses a method for simulating the spatial evolution of commercial districts based on spatiotemporal analysis, relating to the field of smart city technology. The method includes: collecting multi-source spatiotemporal data to construct an initial spatiotemporal knowledge graph, and using the initial spatiotemporal knowledge graph to obtain a baseline simulation model; collecting real-time multi-source data streams and incrementally updating the initial spatiotemporal knowledge graph based on the real-time multi-source data streams to form a dynamic spatiotemporal knowledge graph; detecting state changes in the dynamic spatiotemporal knowledge graph, and when any rate of change of key entities or key relationships exceeds a predetermined change threshold, triggering and executing incremental simulation of a local area to obtain a local predicted state graph, and updating the local predicted state graph to the dynamic spatiotemporal knowledge graph to obtain a fused predicted knowledge graph. This invention achieves accurate capture and rapid response to micro-level disturbances.

Description

Technical Field

[0001] This invention relates to the field of smart city technology, and in particular to a method for simulating the spatial evolution of commercial districts based on spatiotemporal analysis. Background Technology

[0002] With the deepening integration of smart city and spatiotemporal knowledge graph technologies, dynamic modeling and intelligent inference of complex spatial systems in commercial districts have become a cutting-edge direction in geographic information science and urban computing. Existing technologies mainly rely on static knowledge graphs combined with traditional spatiotemporal analysis methods. This involves constructing a network of entity relationships within a district by integrating multi-source data and using graph neural networks and temporal prediction models to analyze evolution trends. In recent years, dynamic knowledge graph technology has made significant progress, enabling temporal updates of entity attributes and relationship strengths, providing a structured representation foundation for the evolution of district states. Simultaneously, novel algorithms such as continuous-time graph networks based on neural differential equations and spatiotemporal hypergraph fusion have further improved the accuracy and continuity of spatiotemporal correlation modeling.

[0003] Existing knowledge graph-based commercial street spatial simulation technologies face two major bottlenecks. First, there is a lack of efficient real-time interaction mechanisms between the dynamic knowledge graph and the simulation engine. The current approach, relying on periodic full graph updates and offline model training, struggles to respond quickly and provide targeted simulations for sudden localized commercial events, resulting in the inability to accurately predict the spatial chain reactions caused by micro-level disturbances. Second, there is a lack of a closed-loop co-evolutionary framework between the knowledge graph and the simulation model. The absence of a systematic comparison, attribution, and feedback mechanism between predicted results and actual observation data prevents the synchronous calibration of graph data confidence and online learning of model parameters based on discrepancy analysis. This hinders the continuous improvement of long-term prediction accuracy and limits the system's adaptive capabilities. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a method for simulating the spatial evolution of commercial districts based on spatiotemporal analysis to address the problems of insufficient real-time interaction between dynamic knowledge graphs and simulation engines, as well as the lack of a closed-loop co-evolution mechanism between knowledge graphs and simulation models.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] This invention provides a method for simulating the spatial evolution of commercial districts based on spatiotemporal analysis. The method includes: collecting multi-source spatiotemporal data to construct an initial spatiotemporal knowledge graph, and using the initial spatiotemporal knowledge graph to obtain a baseline simulation model; collecting real-time multi-source data streams and incrementally updating the initial spatiotemporal knowledge graph based on the real-time multi-source data streams to form a dynamic spatiotemporal knowledge graph; detecting state changes in the dynamic spatiotemporal knowledge graph, and when the rate of change of any key entity or key relationship exceeds a predetermined change threshold, triggering and executing incremental simulation of a local area to obtain a local predicted state graph, and then updating the local predicted state graph. The predicted state graph is updated to the dynamic spatiotemporal knowledge graph to obtain a fused predicted knowledge graph. Based on the fused predicted knowledge graph, the baseline simulation model is invoked to perform full-scale simulation prediction according to the preset prediction cycle, generating a full-scale predicted graph. Actual observation data is collected to construct a real state graph, and the full-scale predicted graph is compared and the real state graph is analyzed to generate a difference analysis report. Based on the difference analysis report, the confidence level of the dynamic spatiotemporal knowledge graph is calibrated, and the baseline simulation model is subjected to online incremental learning to obtain the updated dynamic spatiotemporal knowledge graph and the baseline simulation model.

[0008] As a preferred embodiment of the commercial district spatial evolution simulation method based on spatiotemporal analysis described in this invention, the specific steps for collecting multi-source spatiotemporal data to construct an initial spatiotemporal knowledge graph are as follows:

[0009] Collect multi-source spatiotemporal data and perform real-time analysis and feature extraction to generate structured event streams;

[0010] Based on structured event streams, a learnable parameterized function is constructed through spatiotemporal hashing and multilayer perceptron, and the structured event streams are optimized and trained to generate an implicit neural field representation of the continuous spatiotemporal of the street blocks.

[0011] Spatial voxelization sampling and semantic decoding are performed on the implicit neural field representation. Three-dimensional spatial entities are extracted by the moving cube algorithm, and the relationship decoder is used to generate the relationship between entities.

[0012] By using three-dimensional spatial entities as nodes, the relationships between entities as edges, and combining them with corresponding attribute information, an initial spatiotemporal knowledge graph is constructed.

[0013] As a preferred embodiment of the spatiotemporal analysis-based commercial district spatial evolution simulation method of the present invention, the specific steps for obtaining the baseline simulation model using an initial spatiotemporal knowledge graph are as follows:

[0014] Extract graph snapshots of each time slice within the historical cycle from the initial spatiotemporal knowledge graph and construct a spatiotemporal graph sequence;

[0015] Based on spatiotemporal graph sequences, a neural differential equation graph network based on continuous-time memory integration and periodic attention mechanisms is constructed as the architecture of the baseline simulation model.

[0016] Using spatiotemporal graph sequences as training data, the parameters of the neural differential equation graph network are optimized and trained using the adjoint method to obtain a baseline simulation model with continuous-time extrapolation capabilities.

[0017] As a preferred embodiment of the spatiotemporal analysis-based commercial district spatial evolution simulation method of the present invention, the specific steps for forming a dynamic spatiotemporal knowledge graph are as follows:

[0018] Collect real-time multi-source data streams and abstract them into high-order event streams with business semantics;

[0019] The state update amount of entity nodes in the initial spatiotemporal knowledge graph is calculated by calculating the state update amount of the high-order event flow through the spatiotemporal hypergraph incremental fusion, and an intermediate semantic representation of the update amount and the influence strength of the relationship is generated.

[0020] Based on the influence strength of relationships in the intermediate semantic representation, the topological structure of the initial spatiotemporal knowledge graph is inferred through the confidence propagation algorithm to generate a set of structural changes;

[0021] The state update values ​​and structural change sets of entity nodes are atomically applied to the initial spatiotemporal knowledge graph to generate a dynamic spatiotemporal knowledge graph.

[0022] As a preferred embodiment of the spatiotemporal analysis-based commercial district spatial evolution simulation method of the present invention, the specific steps for obtaining the local predicted state map are as follows:

[0023] Based on dynamic spatiotemporal knowledge graphs, calculate the rate of change of attributes of key entities and key relationships within recent time windows;

[0024] Centered on the anomaly source, local subgraphs associated with the anomaly source are extracted from the dynamic spatiotemporal knowledge graph and used as the target region for incremental simulation.

[0025] Incremental simulations are performed on the target area, and the future state of the target area is deduced to generate a local predicted state map.

[0026] As a preferred embodiment of the spatiotemporal analysis-based commercial district spatial evolution simulation method of the present invention, the specific steps for obtaining the fusion prediction knowledge graph are as follows:

[0027] Identify state differences and logical conflicts between the local predicted state graph and the dynamic spatiotemporal knowledge graph, and generate a set of differences and conflicts.

[0028] Based on the difference conflict set, the semantic space of the local predicted state graph is adapted by the manifold alignment algorithm to obtain the aligned predicted graph representation.

[0029] Based on the aligned prediction graph representation, the fused state of each node in the dynamic spatiotemporal knowledge graph is obtained through probabilistic graph fusion reasoning.

[0030] The merged state is atomically updated to the dynamic spatiotemporal knowledge graph to generate a fusion prediction knowledge graph.

[0031] As a preferred embodiment of the spatiotemporal analysis-based commercial district spatial evolution simulation method of the present invention, the specific steps for generating the full-scale prediction map are as follows:

[0032] Uncertainty quantification and integration of external events are performed on the fusion predictive knowledge graph to generate enhanced input states;

[0033] Based on the enhanced input state, the baseline simulation model is driven to perform multi-scenario parallel inference to generate multiple future state trajectories;

[0034] Cluster analysis and probability fusion are performed on multiple future state trajectories to generate a full-scale prediction map.

[0035] As a preferred embodiment of the spatiotemporal analysis-based commercial district spatial evolution simulation method of the present invention, the specific steps for collecting actual observation data to construct a real-state map are as follows:

[0036] Collect actual observation data and perform reliable fusion of the actual observation data to obtain entity state estimates;

[0037] Conflict resolution and truth discovery are performed on the preliminary entity state estimates to obtain the entity state truth values ​​with high confidence.

[0038] Spatiotemporal consistency completion is performed on the true values ​​of entity states with high confidence to construct a true state map.

[0039] As a preferred embodiment of the spatiotemporal analysis-based commercial district spatial evolution simulation method of the present invention, the specific steps for generating the difference analysis report are as follows:

[0040] The difference between the full-scale predicted map and the real-state map is quantified by graph alignment and multi-dimensional difference measurement to generate a difference set.

[0041] Based on the difference set, the key root causes and diagnostic information leading to the differences are obtained through causal diagnostic analysis.

[0042] Integrate the set of differences, key root causes, and diagnostic information, and generate a difference analysis report.

[0043] As a preferred embodiment of the spatiotemporal analysis-based commercial district spatial evolution simulation method of the present invention, the specific steps for obtaining the updated dynamic spatiotemporal knowledge graph and baseline simulation model are as follows:

[0044] Based on the attribution analysis and calibration recommendations in the difference analysis report, a confidence adjustment strategy for relevant data points in the dynamic spatiotemporal knowledge graph is generated.

[0045] Based on the training sample pairs composed of the fused predictive knowledge graph and the real state graph, and combined with the confidence adjustment strategy, the baseline simulation model is subjected to online incremental learning to generate model parameter update data.

[0046] By applying confidence adjustment strategies and updating model parameters, the dynamic spatiotemporal knowledge graph and the baseline simulation model are updated synchronously, resulting in the updated dynamic spatiotemporal knowledge graph and the baseline simulation model.

[0047] The beneficial effects of this invention are as follows: Through real-time data stream-driven dynamic knowledge graph incremental updates and local trigger simulation, accurate capture and rapid response to micro-perturbations are achieved, making the knowledge graph a truly dynamically evolving "digital twin"; through the co-evolution mechanism of knowledge graph and simulation model with difference diagnosis, decoupled diagnosis and collaborative repair of model bias and data quality problems are achieved, and a sustainable evolutionary "perception-cognition-decision" intelligent agent is established. Attached Figure Description

[0048] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0049] Figure 1 This is a flowchart of a method for simulating the spatial evolution of commercial districts based on spatiotemporal analysis.

[0050] Figure 2 A flowchart for forming a dynamic spatiotemporal knowledge graph.

[0051] Figure 3 The flowchart for generating full-scale prediction maps.

[0052] Figure 4 This is an overview of the error in the target area.

[0053] Figure 5 This is a heatmap of spatial errors. Detailed Implementation

[0054] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0055] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0056] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0057] Reference Figures 1-5 As one embodiment of the present invention, this embodiment provides a method for simulating the spatial evolution of commercial districts based on spatiotemporal analysis, comprising the following steps:

[0058] S1. Collect multi-source spatiotemporal data to construct an initial spatiotemporal knowledge graph, and use the initial spatiotemporal knowledge graph to obtain a baseline simulation model.

[0059] Collect multi-source spatiotemporal data and perform real-time analysis and feature extraction to generate structured event streams.

[0060] The specific process includes simultaneously acquiring multi-source spatiotemporal data with precise timestamps and geographic coordinates from various sources such as urban sensing devices, mobile terminal location information, commercial transaction records, traffic monitoring video streams, and social media check-in logs. The multi-source spatiotemporal data is then processed in real time using streaming processing. Natural language processing methods are used to identify location and time entities in the text. Computer vision algorithms are used to detect pedestrian trajectories and store pedestrian density from the video stream. Combined with signal processing, spatial location change sequences are extracted from sensor data. Based on a unified spatiotemporal ontology standard, the parsing results from different sources are mapped into event elements with standard semantic labels. Finally, the events are organized into a structured event stream according to their temporal order and spatial relationships.

[0061] It should be noted that multi-source spatiotemporal data includes video streams from urban cameras, signaling data from mobile communication base stations, public transportation card swipe records, ride-hailing order trajectories, device signal data collected by Wi-Fi probes, user-posted content with geotags on social media, and real-time location search and navigation paths in map applications.

[0062] Based on structured event streams, a learnable parameterized function is constructed through spatiotemporal hashing and multilayer perceptron, and the structured event stream is optimized and trained to generate an implicit neural field representation of the continuous spatiotemporal of the street blocks.

[0063] The specific process includes inputting the timestamps and geographic coordinates of each event in the structured event stream into a spatiotemporal hashing encoding method. This method maps high-dimensional sparse spatiotemporal locations into low-dimensional dense encoding vectors. The encoding vectors are then concatenated with the semantic features of the events and input into a multilayer perceptron. The multilayer perceptron is used to construct a learnable parameterized function. This parameterized function takes spatiotemporal location as input and street status attributes as output. During training, real events in the structured event stream are used as supervision signals. The parameters are iteratively optimized by minimizing the error between the predicted and observed values. Ultimately, the learnable parameterized function can continuously interpolate any unobserved spatiotemporal point, generating an implicit neural field representation of the continuous spatiotemporal nature of the street.

[0064] It should be noted that spatiotemporal hashing is a method of jointly converting continuous time and space coordinates into discrete compact codes that preserve local proximity; a multilayer perceptron is a feedforward neural network consisting of an input layer, one or more hidden layers, and an output layer, which fits complex mapping relationships through learnable weights and nonlinear activation functions.

[0065] The implicit neural field representation is spatially voxelized and semantically decoded. Three-dimensional spatial entities are extracted using the moving cube algorithm, and the relationship decoder is used to generate the relationships between entities.

[0066] The specific process includes: performing voxel sampling according to a regular grid within the three-dimensional space covered by the implicit neural field representation of the continuous spatiotemporal street block; obtaining the implicit neural field output value corresponding to each voxel position; mapping the implicit neural field output value to attribute labels with category semantics through semantic decoding; extracting the continuous three-dimensional spatial entity geometric surface based on the field value isosurface of the voxel using the moving cube algorithm; and inputting the extracted three-dimensional spatial entities into the relation decoder, which analyzes the spatial layout, semantic category, and contextual dependency between entities to generate the relationship between entities.

[0067] It should be noted that the moving cube algorithm is a geometric modeling method used to extract isosurface triangular meshes from a three-dimensional scalar field. Its function is to convert voxelized implicit field data into a continuous three-dimensional surface representation. The relation decoder is a method used to infer and generate inter-entity relationships from the semantic and geometric features of entities. Its function is to resolve the semantic and topological dependencies between entities in three-dimensional space.

[0068] By using three-dimensional spatial entities as nodes, the relationships between entities as edges, and combining them with corresponding attribute information, an initial spatiotemporal knowledge graph is constructed.

[0069] The specific process includes treating each three-dimensional spatial entity as a node, representing the relationship between entities as edges connecting the corresponding nodes, and attaching the geometric features, semantic categories, and the type and strength of the relationship between entities to the nodes and edges respectively, thereby forming an initial spatiotemporal knowledge graph containing spatiotemporal semantic structure.

[0070] Extract graph snapshots of each time slice within the historical cycle from the initial spatiotemporal knowledge graph and construct a spatiotemporal graph sequence.

[0071] The specific process includes dividing the initial spatiotemporal knowledge graph according to a preset time granularity, extracting the three-dimensional spatial entities, the relationships between entities and their attribute information that exist in each time point or time period to form a graph snapshot of the corresponding time slice, and then organizing these graph snapshots into an ordered spatiotemporal graph sequence in chronological order.

[0072] It should be noted that the preset time granularity is determined based on the typical cyclical characteristics of the evolution of street space and the frequency of events, and is set by analyzing the temporal distribution patterns of events in the historical structured event flow.

[0073] Based on spatiotemporal graph sequences, a neural differential equation graph network based on continuous-time memory integration and periodic attention mechanisms is constructed as the architecture of the baseline simulation model.

[0074] The specific process includes taking a snapshot of the map of each time slice in the spatiotemporal graph sequence as input, using continuous-time memory integration to model the changes in node states at any time point, capturing repetitive patterns and long-term dependencies in the spatial evolution of the block through a periodic attention mechanism, and embedding these two mechanisms into the dynamic evolution function of the neural differential equation graph network, so that the node representation can evolve smoothly with continuous time and respond to periodic patterns, forming the architecture of a baseline simulation model with temporal continuity and periodic perception capabilities.

[0075] Furthermore, the process of constructing a neural differential equation graph network based on continuous-time memory integration and periodic attention mechanisms involves defining a continuous dynamic process of node state evolution over time, using differential equations to describe the rate of change of each entity node's state, and weighting and accumulating historical states through a memory integration mechanism to characterize long-term dependencies. Combined with a periodic attention mechanism, the interaction weights between nodes at different phases are adaptively obtained based on the temporal context to model the periodic patterns and instantaneous disturbances in event responses. On this basis, all entity nodes and their relationships are organized into a graph structure, and the propagation rules of graph neural networks are used to achieve neighborhood information aggregation and state updates in the continuous-time domain. By optimizing parameters to ensure consistency between the deduction results and observations, a neural differential equation graph network with continuous-time evolution capabilities and periodic perception characteristics is formed.

[0076] It should be noted that the dynamic evolution function refers to a learnable function used in neural differential equation graph networks to describe the continuous change of the state of an entity node over time. The output is the derivative of the node state with respect to time, which drives the node state to undergo differential evolution along the time axis.

[0077] Using spatiotemporal graph sequences as training data, the parameters of the neural differential equation graph network are optimized and trained using the adjoint method to obtain a baseline simulation model with continuous-time extrapolation capabilities.

[0078] The specific process includes using snapshots of the spatiotemporal map of each time slice in the spatiotemporal map sequence as supervision signals, efficiently obtaining the gradient of the loss function relative to the parameters of the neural differential equation graph network during continuous time evolution through the adjoint method, and updating the parameters of the neural differential equation graph network accordingly, so that the neural differential equation graph network can accurately fit the dynamic evolution of the street spatial state in historical time, and obtain a baseline simulation model that can be continuously extrapolated at any time point.

[0079] It should be noted that the adjoint method is a numerical optimization method for efficiently obtaining the gradient of the loss function with respect to the parameters in a neural differential equation. Its purpose is to achieve memory-efficient backpropagation without explicitly storing the intermediate states of forward propagation.

[0080] S2. Collect real-time multi-source data streams and incrementally update the initial spatiotemporal knowledge graph based on the real-time multi-source data streams to form a dynamic spatiotemporal knowledge graph.

[0081] Collect real-time multi-source data streams and abstract them into high-order event streams with business semantics.

[0082] The specific process includes continuously receiving camera video streams through urban sensing infrastructure, subscribing to base station signaling data in real time from communication operator interfaces, obtaining card swipe record streams by connecting to public transportation operation platforms, accessing trip trajectory pushes from ride-hailing platforms, and pulling user-posted content with geographic tags through application programming interfaces of social media platforms to form a unified access real-time multi-source data stream. Then, streaming parsing is used to identify spatiotemporal behavioral patterns related to commercial activities, such as sudden increases in store traffic, gatherings caused by promotional activities, and the opening or closing of brand stores. These behavioral patterns are then standardized and described according to a unified commercial semantic ontology to form a high-order event stream that is time-ordered, semantically clear, and expresses complex commercial phenomena.

[0083] It should be noted that real-time multi-source data streams include continuously generated camera video streams, base station signaling data, public transportation card swipe records, ride-hailing vehicle trajectories, and geotagged social media content, etc. Their core characteristics are low latency, continuous arrival, and instant availability, used to reflect dynamic phenomena that are currently occurring. Real-time multi-source data streams are a dynamic subset of multi-source spatiotemporal data in the time dimension, focusing on streaming data that is currently being generated and has time constraints. Multi-source spatiotemporal data has a broader scope, including all spatiotemporal labeled information, including historical archived data, batch-processed data, and real-time streaming data.

[0084] By incrementally fusing the spatiotemporal hypergraph, the state update amount of entity nodes in the initial spatiotemporal knowledge graph by the high-order event flow is calculated, generating an intermediate semantic representation of the update amount and the influence strength of the relationship. The expression is as follows: ;

[0085] in, Indicates the event Caused by entity nodes The state update vector, This represents a higher-order event abstracted from a real-time multi-source data stream. This represents an entity node in a spatiotemporal knowledge graph. Indicates an event For nodes Attention weights This represents the learnable weight matrix. Indicates an event semantic feature vectors, It represents the Hadamah accumulation. This represents the natural exponential function. Represents entity nodes spatial coordinate vector, Indicates an event The spatial coordinate vector of the occurrence, Indicates the spatial attenuation coefficient. Represents pi (π). Indicates the current time point, Indicates an event The time of occurrence This represents the time decay period parameter. This represents the normalized singer function.

[0086] It should be noted that, Indicates an event For nodes Attention weights are calculated by events semantic feature vectors and nodes The similarity between the current state vectors is a scalar value obtained by normalizing the data through a learnable attention mechanism; the spatial decay coefficient is a scalar value obtained by using the natural exponential function based on the Euclidean distance between events and nodes, used to measure the decay effect of spatial proximity on the intensity of influence; the time decay period parameter is a scalar parameter obtained by periodically modulating the difference between the event occurrence time and the current time using a normalized Singer function, based on the analysis of the temporal regularity of business activities in high-order event streams.

[0087] The specific process includes taking high-order events as input, using the spatial distance decay term and time periodicity decay term between the semantic feature vector corresponding to the event and the spatial coordinate vector of the entity node for weighted modulation, combining the attention weight of the event to the node and the learnable weight matrix, associating the event semantics with the node state through the Hadamard product operation, and using the natural exponential function and the normalized Singer function to model the spatial proximity and time periodicity effects respectively, outputting the state update vector of the entity node caused by the event, which is used to subsequently update the semantic representation of the corresponding node in the initial spatiotemporal knowledge graph.

[0088] Based on the influence strength of relationships in the intermediate semantic representation, the topological structure of the initial spatiotemporal knowledge graph is inferred through the confidence propagation algorithm to generate a set of structural changes.

[0089] The specific process includes using the influence strength of the relationships between entities contained in the intermediate semantic representation as the initial confidence input on the edges, using a confidence propagation algorithm to iteratively transmit the semantic support between nodes and edges on the graph structure of the initial spatiotemporal knowledge graph, and judging which relationships between entities are strengthened, weakened, or newly emerged due to the continuous influence of higher-order event flows based on the confidence distribution after propagation convergence. This allows for the identification of nodes and edges that need to be added, deleted, or modified (e.g., shop nodes added due to the event of a new store opening and their connection edges with surrounding high-traffic areas, or commercial association edges marked as to be deleted due to low confidence caused by long-term decline in customer traffic). This forms a structural change set describing the changes in the graph topology.

[0090] It should be noted that the confidence propagation algorithm is a reasoning method that iteratively updates the confidence of nodes and edges on a graph structure through a message passing mechanism. Its function is to infer the reliability or association strength of each element in the global structure based on local observations and graph topological relationships.

[0091] The state update values ​​and structural change sets of entity nodes are atomically applied to the initial spatiotemporal knowledge graph to generate a dynamic spatiotemporal knowledge graph.

[0092] The specific process includes: adding the state update of the entity node to the attributes of the corresponding node in the initial spatiotemporal knowledge graph item by item, and simultaneously executing the added, deleted or modified node and edge operations described in the structural change set on the topology of the initial spatiotemporal knowledge graph in an indivisible atomic transaction manner, so as to ensure that the state update and structural change are logically consistent and take effect at the same time, forming a dynamic spatiotemporal knowledge graph that reflects the latest spatiotemporal state of the commercial district.

[0093] S3. Detect state changes in the dynamic spatiotemporal knowledge graph. When the rate of change of any key entity or key relationship exceeds a predetermined change threshold, trigger and execute incremental simulation of the local region to obtain the local predicted state graph. Update the local predicted state graph to the dynamic spatiotemporal knowledge graph to obtain the fused predicted knowledge graph.

[0094] Based on a dynamic spatiotemporal knowledge graph, the rate of change of attributes of key entities and key relationships within a recent time window is calculated using the following expression: ;

[0095] in, Indicates key element At the point of time The rate of change of the attribute, Indicates key elements (key entities or key relationships). This indicates the preset recent time window length. Indicates key element At a specific moment State attribute vector, This represents the time integral variable.

[0096] It should be noted that the preset recent time window length is set based on the duration and response speed of typical events in the commercial district, by analyzing the time scale of the evolution of key entity attributes in the historical high-order event flow.

[0097] The specific process includes performing an integral operation on the state attribute vector of the key element in the dynamic spatiotemporal knowledge graph as it evolves over time within a preset recent time window. The rate of change of the state attribute vector is accumulated and summed using the time integral variable, and the integral result is normalized with the time window length to obtain the rate of change of the key element's attributes at the current time point.

[0098] It should be noted that key entities and key relationships refer to entity nodes and relationship edges that have a significant representational or driving effect on the operational status of commercial districts in a dynamic spatiotemporal knowledge graph. The definition is determined by comprehensively considering the frequency of occurrence, scope of influence, sensitivity to attribute changes, and centrality index in the graph topology of the entity or relationship in the historical high-order event flow.

[0099] The attribute change rate is compared with a predetermined change threshold. When any change rate exceeds the predetermined change threshold, the corresponding key entity and key relationship are marked as an anomaly source.

[0100] The specific process includes comparing the attribute change rate with a predetermined change threshold. When any change rate exceeds the predetermined change threshold, the corresponding key entity and key relationship are marked as an anomaly source. The process involves checking the attribute change rate of each key entity and key relationship within the recent time window, comparing the attribute change rate with a pre-set change threshold, and immediately marking the key entity and key relationship corresponding to the attribute change rate as an anomaly source if any attribute change rate is found to be greater than the predetermined change threshold.

[0101] It should be noted that the predetermined change threshold is determined based on the statistical distribution characteristics of attribute changes of key entities and key relationships in the historical operation of the commercial district. The threshold boundary is set by analyzing the mean and standard deviation of the change rate of the same attribute in the historical period in the dynamic spatiotemporal knowledge graph, combined with the business requirements for anomaly sensitivity. The exemplary value range is usually between 5% and 20%, and the specific value depends on the attribute type. For example, the change threshold for customer flow related attributes can be set to 10%, and the change threshold for shop association strength can be set to 15%.

[0102] Centered on the anomaly source, local subgraphs associated with the anomaly source are extracted from the dynamic spatiotemporal knowledge graph and used as the target region for incremental simulation.

[0103] The specific process includes using the key entities and key relationships marked as anomaly sources as starting nodes and edges, traversing the direct connections and indirect neighbors of the anomaly sources in the dynamic spatiotemporal knowledge graph through the semantic relevance range, collecting all entities and relationships that have topological connections or semantic dependencies with the anomaly sources within a preset neighborhood range, forming a connected local subgraph, and defining the local subgraph as the target region focused on by subsequent incremental simulations.

[0104] It should be noted that the semantic relevance range is predefined based on the categories, attributes, and historical co-occurrence patterns of entities and relationships in the dynamic spatiotemporal knowledge graph, and is used to limit the semantic similarity of nodes and edges that can be included in the local subgraph during traversal; the preset neighborhood range is set based on the spatial scale of commercial districts, typical distances of entity interactions, and historical anomaly propagation patterns, by analyzing the spatial decay characteristics of the impact of abnormal events in the dynamic spatiotemporal knowledge graph.

[0105] Incremental simulations are performed on the target area, and the future state of the target area is deduced to generate a local predicted state map.

[0106] The specific process includes inputting the local subgraph corresponding to the target region into the baseline simulation model, using the neural differential equation graph network based on continuous-time memory integration and periodic attention mechanism in the baseline simulation model to continuously evolve the state of each entity node and relation edge in the local subgraph, updating only the state variables within the target region while maintaining consistency with the global dynamic spatiotemporal knowledge graph, deducing the state distribution of the target region at a future specified time point, and organizing the state distribution into a local predicted state graph containing the updated entity attributes and relation strengths.

[0107] It should be noted that the future state of the target area refers to the set of predicted values ​​of the attributes of each entity node and the strength of relation edges of the local subgraph extracted with the anomaly source as the center at a specified future time point. It reflects the possible evolution of semantic attributes of commercial elements in the target area, such as pedestrian density, shop activity, and spatial correlation strength, over time under current observation and historical evolution patterns.

[0108] like Figure 4 The target region error overview chart is used to show the change of the mean absolute error of the target region over time before and after the occurrence of "micro-disturbance". It uses multiple curves to compare the "complete solution (incremental update + local trigger)" with "no local trigger / no co-evolutionary closed loop / no incremental update (batch processing)" scenarios. The overview chart shows the error trend over the entire time window, and the local magnified chart is marked with peak positions and the maximum difference point Δ, highlighting that when the change rate of key entities / relationships exceeds the predetermined change threshold and triggers local incremental simulation, the target region error can fall back faster and the peak value is lower. This directly reflects the beneficial effect of "real-time data stream-driven dynamic knowledge graph incremental update and local trigger simulation to achieve accurate capture and rapid response to micro-disturbance".

[0109] Identify state differences and logical conflicts between the local predicted state graph and the dynamic spatiotemporal knowledge graph, and generate a set of differences and conflicts.

[0110] The specific process includes comparing the attribute values ​​and relation edge strengths of each entity node in the local predicted state graph with the state of the corresponding element in the dynamic spatiotemporal knowledge graph at the same time point, obtaining the degree of deviation of attribute values, and detecting logical conflicts such as semantic category inconsistency, topological connection contradiction, or temporal causality violation. All state differences that exceed the allowable error range and logical conflict entries that violate the knowledge graph ontology constraints are summarized to form a structured set of differences and conflicts.

[0111] It should be noted that the allowable error range refers to the maximum deviation limit that entity attribute values ​​or relationship strengths are allowed to exist without being considered abnormal when comparing local predicted state maps and dynamic spatiotemporal knowledge graphs. This range is based on the statistical analysis of the natural fluctuation amplitude of similar attributes in historical observation data, and is usually set by adding or subtracting a certain number of standard deviations from the mean. For example, the allowable error range for the passenger flow attribute can be set to plus or minus 15% of the historical average passenger flow for the same period, and the allowable error range for the shop association strength can be set to twice the standard deviation of the historical association strength.

[0112] Based on the difference conflict set, the semantic space of the local predicted state graph is adapted by the manifold alignment algorithm to obtain the aligned predicted graph representation.

[0113] The specific process includes using the state differences and logical conflicts identified in the difference conflict set as alignment constraints, and using the manifold alignment algorithm to gradually adjust the semantic vector space embedded in the local predicted state graph to match the semantic manifold of the dynamic spatiotemporal knowledge graph while maintaining the consistency of the internal structure of the local predicted state graph. This makes the semantic representations of corresponding entities and relations in the two graphs tend to be consistent in terms of geometric distance and topological proximity, and outputs the aligned predicted graph representation after semantic space correction.

[0114] It should be noted that manifold alignment algorithms are geometric learning methods that map semantic structures in different datasets or representation spaces onto a shared low-dimensional manifold. Their purpose is to achieve semantic alignment across spaces while preserving their respective local neighborhood relationships.

[0115] Based on the aligned prediction graph representation, the fused state of each node in the dynamic spatiotemporal knowledge graph is obtained through probabilistic graph fusion reasoning.

[0116] The specific process includes identifying the consistency and differences between the aligned predicted graph representation and the entities and relationships in the dynamic spatiotemporal knowledge graph, and establishing cross-graph links using semantic similarity and spatial proximity as weights to construct a joint probability graph. Using probabilistic graph fusion inference, Bayesian fusion is performed on the observed and predicted states of each node on the joint probability graph, comprehensively considering the confidence and uncertainty of the aligned predicted graph representation and the dynamic spatiotemporal knowledge graph, and obtaining the posterior probability distribution of each node in a unified semantic space. The expected value or maximum a posteriori estimate of the posterior probability distribution is used as the fused state of each node in the dynamic spatiotemporal knowledge graph.

[0117] The merged state is atomically updated to the dynamic spatiotemporal knowledge graph to generate a fusion prediction knowledge graph.

[0118] The specific process involves writing the fused state into the attribute fields of the corresponding nodes in the dynamic spatiotemporal knowledge graph one by one using indivisible atomic operations. This ensures that the state update of each node is logically complete and does not interfere with other updates, thereby forming a fused prediction knowledge graph that integrates the aligned prediction information and the current observation information while maintaining the consistency of the graph.

[0119] S4. Based on the fusion prediction knowledge graph, call the baseline simulation model, perform full-scale simulation prediction according to the preset prediction period, and generate a full-scale prediction graph.

[0120] Uncertainty quantification is performed on the fusion predictive knowledge graph, and external events are integrated to generate enhanced input states.

[0121] The specific process includes obtaining entropy or variance based on the posterior probability distribution of the attributes of each node in the fusion prediction knowledge graph to characterize the degree of uncertainty; mapping newly emerging external events from real-time multi-source data streams that have not yet been included in the higher-order event stream to relevant nodes or regions in the fusion prediction knowledge graph according to spatiotemporal coordinates and semantic types; and concatenating or weighting the semantic features of these external events with the uncertainty quantification results of the corresponding nodes to form an enhanced input state containing prediction confidence and new event perturbation information.

[0122] Based on the enhanced input state, the baseline simulation model is driven to perform parallel extrapolation of multiple scenarios, generating multiple future state trajectories.

[0123] The specific process includes inputting the enhanced input state as the initial condition into the baseline simulation model, and applying different reasonable perturbation combinations to key uncertainty dimensions (such as the intensity of external events, the sensitivity of entity responses, or the direction of relationship evolution) during the extrapolation process to form multiple distinct but semantically consistent scenario settings. The baseline simulation model, based on a neural differential equation graph network structure with continuous-time memory integration and periodic attention mechanisms, independently performs continuous-time evolution for each scenario, simultaneously generating multiple future state trajectories that are different in time, space, and semantics.

[0124] Cluster analysis and probability fusion are performed on multiple future state trajectories to generate a full-scale prediction map.

[0125] The specific process includes: vectorizing the entity node states and relation edge strengths at corresponding time points in multiple future state trajectories; using clustering analysis to identify trajectory clusters with similar evolution patterns; obtaining corresponding probability weights based on the distribution density and frequency of occurrence of trajectories within each cluster; weighting and fusing the central trajectories of each cluster according to probability weights on each time slice to obtain a unified state representation covering different probabilities and their confidence levels; and integrating the fusion results of all time slices into a full-scale prediction map covering the entire spatiotemporal range and containing uncertain semantics.

[0126] It should be noted that the significance of the full-scale prediction map lies in providing a unified prediction view that covers the entire spatiotemporal range and integrates multiple scenario possibilities and uncertain semantics, supporting comprehensive perception and robust decision-making regarding the future state of commercial districts.

[0127] S5. Collect actual observation data to construct a true state map, and compare and analyze the differences between the full-scale predicted map and the true state map to generate a difference analysis report.

[0128] Collect actual observation data and perform reliable fusion of the actual observation data to obtain entity state estimates.

[0129] The specific process includes collecting actual observation data from sources such as camera video streams, base station signaling data, public transportation card swipe records, ride-hailing vehicle trajectories, and geotagged social media content. Based on the confidence indices of each data source in terms of spatiotemporal coverage, semantic consistency, and historical reliability, a reliable fusion method such as weighted averaging or Bayesian fusion is used to integrate multi-source observations of the same entity within the same time window, eliminating redundancy and conflicts, and outputting an entity state estimate with unified semantics that reflects the current real situation.

[0130] It should be noted that actual observation data includes real-time or recent events collected by physical or digital sensors, such as camera video streams, base station signaling data, public transportation card swipe records, ride-hailing vehicle trajectories, and geotagged social media content. Its core characteristic is that it reflects the objective state that has occurred or is occurring in the real world, emphasizing authenticity and verifiability.

[0131] Conflict resolution and truth discovery are performed on the preliminary entity state estimates to obtain the entity state truth values ​​with high confidence.

[0132] The specific process includes addressing numerical biases, semantic contradictions, or logical inconsistencies in the preliminary entity state estimates generated from different observation sources or fusion methods, using methods based on voting mechanisms, confidence weighting, or... Figure 1 The conflict resolution strategy for consistency constraints identifies and eliminates abnormal or low-confidence estimates, and integrates the reliability weights of each source estimate with historical accuracy records through a truth discovery algorithm. Iteratively optimizes the consensus values ​​of each entity attribute and outputs the true values ​​of entity states that are semantically and numerically consistent and have high confidence.

[0133] It should be noted that the truth discovery algorithm is a method of identifying and inferring the most credible facts from multiple potentially conflicting information sources. Its role is to automatically discover a consensus result that is close to the truth by modeling the reliability of each information source and aggregating the provided observations in the presence of noise, bias, or errors.

[0134] Spatiotemporal consistency completion is performed on the true values ​​of entity states with high confidence to construct a true state map.

[0135] The specific process includes using high-confidence entity state truth values ​​as anchor points, filling in the missing state at the time dimension through interpolation or extrapolation methods, inferring the reasonable state of unobserved areas based on the topological proximity and semantic correlation between entities in the spatial dimension, and applying smoothness and continuity checks to the completion results in combination with the physical constraints and periodic laws of street space evolution to ensure that the state changes of all entities in time and space conform to real logic, forming a real state map that covers the complete spatiotemporal range and is internally consistent.

[0136] It should be noted that physical constraints refer to the basic conditions that must be met for changes in the physical state of a street space, such as the inability of pedestrian flow to move instantly and the restriction of shop business status by day and night rhythms; periodic patterns refer to the repetitive patterns of commercial street activities in time, such as the difference in pedestrian flow between weekdays and weekends, and the periodic peaks in customer flow brought about by holiday promotions.

[0137] The difference between the full-scale predicted map and the real-state map is quantified by graph alignment and multi-dimensional difference measurement, generating a difference set.

[0138] The specific process includes using graph alignment methods to perform node matching and structural alignment of corresponding entities and relationships in the full-scale predicted graph and the real state graph in a unified semantic space. The degree of difference between the aligned elements in the two graphs is obtained from four dimensions: attribute value, topological connection, temporal evolution trend and semantic category. All inconsistencies exceeding the preset tolerance threshold are recorded as difference entries and summarized to form a structured difference set.

[0139] It should be noted that the preset tolerance thresholds are set based on the statistical analysis of the natural deviations between historical predictions and the actual state in various dimensions; the exemplary value ranges include tolerance for attribute value differences of 5% to 15%, tolerance for topological connection differences of less than 2%, tolerance for time offset of several hours, and generally no tolerance for semantic category differences.

[0140] Based on the difference set, causal diagnostic analysis is used to obtain the key root causes and diagnostic information that lead to the differences.

[0141] The specific process includes using the inconsistencies recorded in the set of differences as observational evidence, using counterfactual reasoning to trace potential perturbation factors in the full-scale prediction map generation process, identifying the variables or operations that contribute the most to the differences in the stages of high-order event abstraction, baseline simulation model deduction, external event integration or uncertainty quantification, and outputting structured diagnostic information describing the action path and influence mechanism, thereby locating the key root causes that lead to the prediction deviating from the true state.

[0142] It should be noted that the key root causes and diagnostic information refer to the fundamental reasons and mechanisms that lead to the differences between the full-scale prediction map and the true state map, including the location of the source of the deviation, the path of influence, the data involved, and semantic explanations.

[0143] Integrate the set of differences, key root causes, and diagnostic information, and generate a difference analysis report.

[0144] The specific process includes: structurally associating the inconsistencies between the full-scale predicted map and the actual state map recorded in the difference set, the key root causes of the differences identified by the causal diagnostic analysis, and the corresponding diagnostic information; organizing them into a hierarchical attribution analysis content according to attribution logic; clarifying the causal chain behind each difference phenomenon; and proposing targeted calibration suggestions based on the mechanism of action of the key root causes, thus forming a difference analysis report that includes attribution analysis and calibration suggestions.

[0145] like Figure 5 Spatial error heatmaps are used to visually present the error distribution (e.g., absolute error intensity by grid / region) of the "full-scale prediction map" relative to the "real state map" in a spatial dimension. This allows for a quick look at whether errors are concentrated near the disturbance source and its propagation path, and whether non-disturbed areas maintain low errors. This type of spatial distribution evidence directly supports the necessity and interpretability of the present invention to conduct "comparison and difference analysis and generate difference analysis reports" after the prediction period ends. It also provides a visual basis for subsequent "attribution analysis and calibration recommendations / co-repair," demonstrating the beneficial effect of the "knowledge graph of difference diagnosis and the co-evolution mechanism of simulation model."

[0146] S6. Based on the difference analysis report, perform confidence calibration on the dynamic spatiotemporal knowledge graph and perform online incremental learning on the baseline simulation model to obtain the updated dynamic spatiotemporal knowledge graph and baseline simulation model.

[0147] Based on the attribution analysis and calibration recommendations in the difference analysis report, a confidence adjustment strategy for relevant data points in the dynamic spatiotemporal knowledge graph is generated.

[0148] The specific process includes identifying the key root causes and action paths leading to prediction bias in the difference analysis report, identifying entity nodes or relation edges in the dynamic spatiotemporal knowledge graph affected by these root causes, and adjusting the historical confidence of these relevant data points by combining the explanations on data reliability and event abstraction bias in the calibration recommendations. For example, reducing the current confidence of nodes whose state is distorted due to missed external events, or increasing the confidence weight of observational correlation edges that have been verified as highly consistent through truth discovery, thus forming a confidence adjustment strategy for subsequent reasoning and prediction.

[0149] Based on training sample pairs composed of a fused predictive knowledge graph and a real-state graph, and combined with a confidence adjustment strategy, the baseline simulation model is subjected to online incremental learning to generate model parameter update data.

[0150] The specific process includes using the fused predictive knowledge graph as input and the real state graph as the target output to construct one-to-one training sample pairs. Based on the confidence adjustment strategy, corresponding weights are assigned to the entity nodes and relation edges in each sample to reflect the reliability of the entity nodes and relation edges. As the baseline simulation model continues to receive new data, online learning methods such as gradient descent or variants are used to locally fine-tune the model parameters using only the weighted training samples of the current batch, generating incremental parameter data for updating the baseline simulation model.

[0151] By applying confidence adjustment strategies and updating model parameters, the dynamic spatiotemporal knowledge graph and the baseline simulation model are updated synchronously, resulting in the updated dynamic spatiotemporal knowledge graph and the baseline simulation model.

[0152] The specific process includes writing the new confidence values ​​of entity nodes and relation edges determined in the confidence adjustment strategy into the dynamic spatiotemporal knowledge graph, replacing the original confidence attributes, and simultaneously loading the updated model parameter data into the baseline simulation model to overwrite the corresponding parameters to complete the incremental correction of the model weights. This ensures that the current state cognition reflected by the dynamic spatiotemporal knowledge graph and the evolution prediction capability possessed by the baseline simulation model are consistent at the semantic and numerical levels, forming a synchronously evolving updated dynamic spatiotemporal knowledge graph and baseline simulation model.

[0153] In summary, this invention achieves precise capture and rapid response to micro-perturbations through real-time data stream-driven dynamic knowledge graph incremental updates and local triggering simulations, making the knowledge graph a truly dynamically evolving "digital twin"; and through the co-evolution mechanism of knowledge graph and simulation model with differential diagnosis, it achieves decoupled diagnosis and collaborative repair of model bias and data quality issues, establishing a sustainably evolving "perception-cognition-decision" intelligent agent.

[0154] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for simulating the spatial evolution of commercial districts based on spatiotemporal analysis, characterized in that: include, Collect multi-source spatiotemporal data to construct an initial spatiotemporal knowledge graph, and use the initial spatiotemporal knowledge graph to obtain a baseline simulation model; Collect real-time multi-source data streams and incrementally update the initial spatiotemporal knowledge graph based on the real-time multi-source data streams to form a dynamic spatiotemporal knowledge graph; The dynamic spatiotemporal knowledge graph is subjected to state change detection. When the rate of change of any key entity or key relationship exceeds the predetermined change threshold, the incremental simulation of the local region is triggered and executed to obtain the local predicted state graph. The local predicted state graph is then updated to the dynamic spatiotemporal knowledge graph to obtain the fused predicted knowledge graph. Based on the fusion prediction knowledge graph, the baseline simulation model is invoked to perform full-scale simulation prediction according to the preset prediction period, and a full-scale prediction graph is generated. Collect actual observation data to construct a true state map, and compare and analyze the differences between the full-scale predicted map and the true state map to generate a difference analysis report; Based on the discrepancy analysis report, the confidence level of the dynamic spatiotemporal knowledge graph was calibrated, and the baseline simulation model was subjected to online incremental learning to obtain the updated dynamic spatiotemporal knowledge graph and the baseline simulation model.

2. The method for simulating the spatial evolution of commercial districts based on spatiotemporal analysis as described in claim 1, characterized in that: The specific steps for collecting multi-source spatiotemporal data to construct an initial spatiotemporal knowledge graph are as follows. Collect multi-source spatiotemporal data and perform real-time analysis and feature extraction to generate structured event streams; Based on structured event streams, a learnable parameterized function is constructed through spatiotemporal hashing and multilayer perceptron, and the structured event streams are optimized and trained to generate an implicit neural field representation of the continuous spatiotemporal of the street blocks. Spatial voxelization sampling and semantic decoding are performed on the implicit neural field representation. Three-dimensional spatial entities are extracted by the moving cube algorithm, and the relationship decoder is used to generate the relationship between entities. By using three-dimensional spatial entities as nodes, the relationships between entities as edges, and combining them with corresponding attribute information, an initial spatiotemporal knowledge graph is constructed.

3. The method for simulating the spatial evolution of commercial districts based on spatiotemporal analysis as described in claim 2, characterized in that: The specific steps for obtaining the baseline simulation model using the initial spatiotemporal knowledge graph are as follows: Extract graph snapshots of each time slice within the historical cycle from the initial spatiotemporal knowledge graph and construct a spatiotemporal graph sequence; Based on spatiotemporal graph sequences, a neural differential equation graph network based on continuous-time memory integration and periodic attention mechanisms is constructed as the architecture of the baseline simulation model. Using spatiotemporal graph sequences as training data, the parameters of the neural differential equation graph network are optimized and trained using the adjoint method to obtain a baseline simulation model with continuous-time extrapolation capabilities.

4. The method for simulating the spatial evolution of commercial districts based on spatiotemporal analysis as described in claim 3, characterized in that: The specific steps for forming a dynamic spatiotemporal knowledge graph are as follows. Collect real-time multi-source data streams and abstract them into high-order event streams with business semantics; The state update amount of entity nodes in the initial spatiotemporal knowledge graph is calculated by calculating the state update amount of the high-order event flow through the spatiotemporal hypergraph incremental fusion, and an intermediate semantic representation of the update amount and the influence strength of the relationship is generated. Based on the influence strength of relationships in the intermediate semantic representation, the topological structure of the initial spatiotemporal knowledge graph is inferred through the confidence propagation algorithm to generate a set of structural changes; The state update values ​​and structural change sets of entity nodes are atomically applied to the initial spatiotemporal knowledge graph to generate a dynamic spatiotemporal knowledge graph.

5. The method for simulating the spatial evolution of commercial districts based on spatiotemporal analysis as described in claim 4, characterized in that: The specific steps for obtaining the local predicted state map are as follows: Based on dynamic spatiotemporal knowledge graphs, calculate the rate of change of attributes of key entities and key relationships within recent time windows; Centered on the anomaly source, local subgraphs associated with the anomaly source are extracted from the dynamic spatiotemporal knowledge graph and used as the target region for incremental simulation. Incremental simulations are performed on the target area, and the future state of the target area is deduced to generate a local predicted state map.

6. The method for simulating the spatial evolution of commercial districts based on spatiotemporal analysis as described in claim 5, characterized in that: The specific steps for obtaining the fusion prediction knowledge graph are as follows: Identify state differences and logical conflicts between the local predicted state graph and the dynamic spatiotemporal knowledge graph, and generate a set of differences and conflicts. Based on the difference conflict set, the semantic space of the local predicted state graph is adapted by the manifold alignment algorithm to obtain the aligned predicted graph representation. Based on the aligned prediction graph representation, the fused state of each node in the dynamic spatiotemporal knowledge graph is obtained through probabilistic graph fusion reasoning. The merged state is atomically updated to the dynamic spatiotemporal knowledge graph to generate a fusion prediction knowledge graph.

7. The method for simulating the spatial evolution of commercial districts based on spatiotemporal analysis as described in claim 6, characterized in that: The specific steps for generating the full-scale prediction map are as follows. Uncertainty quantification and integration of external events are performed on the fusion predictive knowledge graph to generate enhanced input states; Based on the enhanced input state, the baseline simulation model is driven to perform multi-scenario parallel inference to generate multiple future state trajectories; Cluster analysis and probability fusion are performed on multiple future state trajectories to generate a full-scale prediction map.

8. The method for simulating the spatial evolution of commercial districts based on spatiotemporal analysis as described in claim 7, characterized in that: The specific steps for constructing a true state map by collecting actual observation data are as follows. Collect actual observation data and perform reliable fusion of the actual observation data to obtain entity state estimates; Conflict resolution and truth discovery are performed on the preliminary entity state estimates to obtain the entity state truth values ​​with high confidence. Spatiotemporal consistency completion is performed on the true values ​​of entity states with high confidence to construct a true state map.

9. The method for simulating the spatial evolution of commercial districts based on spatiotemporal analysis as described in claim 8, characterized in that: The specific steps for generating the difference analysis report are as follows: The difference between the full-scale predicted map and the real-state map is quantified by graph alignment and multi-dimensional difference measurement to generate a difference set. Based on the difference set, the key root causes and diagnostic information leading to the differences are obtained through causal diagnostic analysis. Integrate the set of differences, key root causes, and diagnostic information, and generate a difference analysis report.

10. The method for simulating the spatial evolution of commercial districts based on spatiotemporal analysis as described in claim 9, characterized in that: The specific steps for obtaining the updated dynamic spatiotemporal knowledge graph and baseline simulation model are as follows. Based on the attribution analysis and calibration recommendations in the difference analysis report, a confidence adjustment strategy for relevant data points in the dynamic spatiotemporal knowledge graph is generated. Based on the training sample pairs composed of the fused predictive knowledge graph and the real state graph, and combined with the confidence adjustment strategy, the baseline simulation model is subjected to online incremental learning to generate model parameter update data. By applying confidence adjustment strategies and updating model parameters, the dynamic spatiotemporal knowledge graph and the baseline simulation model are updated synchronously, resulting in the updated dynamic spatiotemporal knowledge graph and the baseline simulation model.

Images