Land space knowledge evolution model construction method

Abstract

The invention provides a territorial space knowledge evolution model construction method, and belongs to the technical field of territorial space knowledge evolution models.The territorial space knowledge evolution model construction method includes the steps that a distributed graph storage framework is constructed, sub-graph partitions are divided through a community detection algorithm, and external event information is collected and converted into a graph operation instruction sequence; positioning a k-hop neighborhood sub-graph by using an incremental sub-graph matching algorithm and executing local updating to generate an intermediate knowledge graph, calculating node state embedding representation of a fusion time coding vector, capturing a time sequence dependence mode through a time perception graph attention aggregation algorithm, executing semantic consistency constraint detection to identify conflicting nodes, and obtaining a time perception graph attention aggregation algorithm; when the conflict degree exceeds a threshold value, the multi-version concurrency control protocol is triggered to roll back to the snapshot version, otherwise, the intermediate knowledge graph is reserved as final output, and the problem that it is difficult to give consideration to updating efficiency and semantic consistency guarantee when the territorial space knowledge graph responds to an external event for evolution updating is solved.

Description

Technical Field

[0001] This invention belongs to the technical field of territorial spatial knowledge evolution model, and specifically relates to a method for constructing a territorial spatial knowledge evolution model. Background Technology

[0002] In the field of territorial spatial planning, knowledge graphs are relied upon to store geographic entities, policy rules, and their relationships. Traditional knowledge graph construction methods employ static modeling, generating an initial graph structure by batch importing historical data. When external events such as policy updates or land use changes occur, the knowledge graph needs to be adjusted accordingly to reflect the latest state. Existing graph update technologies mainly employ full-map reconstruction or simple node and edge modification operations. While full-map reconstruction ensures structural integrity, it incurs enormous computational costs, while simple modification operations lack precise control over the scope of the update's impact and fail to consider the temporal evolution of events, leading to semantic conflicts and logical inconsistencies in the updated graph. In the current context of digital transformation in territorial spatial planning, due to the frequent triggering of external events and their complex and intertwined impacts, traditional methods struggle to achieve efficient incremental updates while maintaining semantic consistency. In other words, existing technologies present a technical challenge in balancing update efficiency and semantic consistency when updating territorial spatial knowledge graphs in response to external events. Summary of the Invention

[0003] In view of this, the present invention provides a method for constructing a land and space knowledge evolution model, which can solve the technical problem in the prior art that it is difficult to balance update efficiency and semantic consistency when the land and space knowledge graph evolves and updates in response to external events.

[0004] This invention is implemented as follows: A method for constructing a territorial spatial knowledge evolution model is provided. This method constructs a distributed graph storage architecture to divide the initial territorial spatial knowledge graph into multiple subgraph partitions. External event information is collected and structured as event tuples. An event parser transforms these event tuples into graph operation instruction sequences. Based on the graph operation instruction sequences, the affected k-hop neighborhood subgraphs are located. An incremental subgraph matching and update algorithm is used to identify the target node set and target edge set. A local update operation is performed to generate an intermediate knowledge graph. A time-encoded vector is calculated for each node in the target node set, fusing the time interval between the event timestamp and the node's historical update timestamp. A time-aware graph attention aggregation algorithm is used to calculate the node state embedding representation. The territorial spatial knowledge evolution model performs semantic consistency constraint detection based on the node state embedding representation and the intermediate knowledge graph to identify conflicting node sets. When the conflict level exceeds a preset threshold, a multi-version concurrency control protocol is triggered to roll back to the snapshot version. When the conflict level does not exceed the preset threshold, the intermediate knowledge graph is retained, and the node state embedding representation is updated as the final output.

[0005] Among them, the community detection algorithm uses the Louvain algorithm to optimize the modularity of the initial knowledge graph of the national land space. By iteratively merging node communities, the connection density within the community is maximized while the connection between communities is sparsed.

[0006] The memory cache uses an LRU eviction policy to maintain hotspot subgraph partitions, and the disk-mapped index uses a B+ tree structure to store the mapping relationship between node identifiers and disk addresses.

[0007] The external event information comes from the policy release system, land management system and regional planning system. The event parser performs semantic understanding of the event content text based on a pre-trained language model.

[0008] The incremental subgraph matching and update algorithm locates the seed node set based on the event influence range parameter, and performs a k-hop breadth-first traversal from the seed node set to obtain the k-hop neighborhood subgraph.

[0009] The time encoding vector is calculated using a time decay function, which is an exponential kernel function that calculates the time interval between the current moment and the node's historical update timestamp and converts it into a time decay factor.

[0010] The time-aware graph attention aggregation algorithm calculates a spatial attention score and a time decay factor for each neighboring node of a node, and multiplies the spatial attention score and the time decay factor to obtain the joint weight.

[0011] Among them, the time-aware graph attention aggregation algorithm uses a multi-head attention mechanism to compute multiple sets of joint weights in parallel to capture dependency patterns at different time scales, and generates node state embedding representations by weighting and summing the features of neighboring nodes according to the joint weights.

[0012] The land space knowledge evolution model consists of a graph structure layer, a spatiotemporal coding layer, a state update layer, and a consistency detection layer. The graph structure layer stores intermediate knowledge graphs, and the spatiotemporal coding layer calculates the time decay factor.

[0013] The semantic consistency constraint set contains logical rules in the field of territorial spatial planning. These logical rules are represented in first-order predicate logic form and define the constraint relationships between node attributes and the legality conditions of edge types.

[0014] The consistency detection layer checks whether nodes and edges in the intermediate knowledge graph violate the logical rules in the semantic consistency constraint set by traversing them, and marks nodes or edges that violate the rules as members of the conflict node set.

[0015] Among them, the multi-version concurrency control protocol creates an independent snapshot version for each graph update transaction, uses an optimistic locking mechanism to detect write conflicts when a transaction is committed, and performs causal ordering of concurrent transactions based on event timestamps.

[0016] In the k-hop neighborhood subgraph, the value of k is dynamically determined based on the event impact range parameter. For local land parcel change events, k is set to 2 to 3, and for regional policy adjustment events, k is set to 4 to 6.

[0017] Among them, the node state embedding means that the historical trajectory information is maintained through the gating loop unit update mechanism. The gating loop unit controls the fusion ratio between the historical state and the current input through the reset gate and the update gate.

[0018] Among them, the pre-trained language model is fine-tuned on the text corpus of the land and space planning field using the BERT or RoBERTa architecture, and learns domain-specific vocabulary and semantic patterns.

[0019] The final land space knowledge evolution model will be output to the land space planning decision-making system for the purpose of evaluating planning schemes, providing early warning of land use conflicts, and analyzing policy impacts.

[0020] This invention constructs a distributed graph storage architecture and employs an incremental subgraph matching algorithm, performing update operations only on local areas affected by events. By locating k-hop neighbor subgraphs, it avoids full-graph scanning, significantly reducing computational complexity. It introduces time-encoded vectors and a time-aware graph attention aggregation algorithm, fusing event timestamps with node historical update information. This allows the node state embedding representation to reflect temporal evolution characteristics, solving the semantic inadequacy problem caused by traditional methods neglecting the time dimension. A semantic consistency constraint detection mechanism and a multi-version concurrency control protocol are established. Logical rule verification is performed before the update is completed, and when semantic conflicts are detected, it automatically rolls back to the snapshot version, ensuring that the graph always meets the constraints of the land spatial planning field. In summary, this invention, by combining an incremental update strategy with a time-aware consistency detection mechanism, solves the technical problem mentioned in the background art of the difficulty in balancing update efficiency and semantic consistency guarantees when the land spatial knowledge graph evolves and updates in response to external events. Attached Figure Description

[0021] Figure 1 This is a flowchart of the method of the present invention.

[0022] Figure 2 This is a schematic diagram of the structure of the territorial spatial knowledge evolution model.

[0023] Figure 3 This is a flowchart of event-driven incremental subgraph matching and updating.

[0024] Figure 4 This is a schematic diagram of the multi-head attention mechanism of the time-aware graph attention aggregation algorithm.

[0025] Figure 5A statistical graph showing the distribution of conflicting nodes and rule violations in the semantic consistency detection layer.

[0026] Figure 6 A visualization of the evolution trajectory of key node states under continuous events. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below.

[0028] like Figure 1 The diagram shown is a flowchart of a method for constructing a land spatial knowledge evolution model provided by the present invention. This method includes the following steps:

[0029] S01. Construct a distributed graph storage architecture, divide the initial knowledge graph of national land space into multiple subgraph partitions according to the community detection algorithm, each subgraph partition contains a set of nodes and a set of edges, and establish an independent memory cache and disk mapping index for each subgraph partition;

[0030] S02. Collect external event information and represent it in a structured form as an event tuple. The event tuple includes an event type identifier, event content text, event timestamp, and event impact range parameter. The event tuple is then converted into a sequence of graph operation instructions by an event parser.

[0031] S03. Locate the affected k-hop neighborhood subgraph according to the graph operation instruction sequence, use the incremental subgraph matching and update algorithm to identify the target node set and target edge set that need to be modified, and only perform add, delete or attribute modification operations on the target node set and the target edge set to generate an intermediate knowledge graph.

[0032] S04. Calculate a time encoding vector for each node in the target node set. The time encoding vector integrates the time interval information of the event timestamp and the node's historical update timestamp. Calculate the node state embedding representation through the time-aware graph attention aggregation algorithm and output it to the national spatial knowledge evolution model.

[0033] S05. The territorial spatial knowledge evolution model performs semantic consistency constraint detection based on the node state embedding representation and the intermediate knowledge graph, identifies the set of conflicting nodes according to the set of semantic consistency constraints, and triggers a multi-version concurrency control protocol to roll back to the snapshot version when the degree of conflict of the set of conflicting nodes exceeds a preset threshold.

[0034] S06. When the degree of conflict does not exceed a preset threshold, the territorial spatial knowledge evolution model retains the intermediate knowledge graph and updates the node state embedding representation, and uses the updated knowledge graph and the node state embedding representation as the final territorial spatial knowledge evolution model.

[0035] The community detection algorithm employs the Louvain algorithm to optimize the modularity of the initial land spatial knowledge graph. It iteratively merges node communities to maximize the internal connection density while sparsening the connections between communities, ultimately decomposing the entire graph into several tightly connected subgraph partitions. The memory cache uses an LRU eviction policy to maintain hot subgraph partitions. High-frequency access subgraph partitions remain resident in memory, improving query efficiency, while cold data subgraph partitions are loaded from disk on demand. The disk-mapped index uses a B+ tree structure to store the mapping relationship between node identifiers and disk addresses, supporting logarithmic-level node location operations. The initial land spatial knowledge graph originates from historical data in the field of land spatial planning and includes geographic entity nodes, policy rule nodes, and relation edges.

[0036] The external event information originates from the policy release system, land use management system, and regional planning system. This external event information includes policy update events, land use change events, and regional adjustment events. The event parser performs semantic understanding of the event content text based on a pre-trained language model, extracting information on geographic entities, attribute changes, and relationship adjustments involved in the event. The extracted results are mapped to node addition, edge deletion, or attribute modification instructions in the graph operation instruction sequence. Each instruction in the graph operation instruction sequence includes an operation type, an operation object identifier, and operation parameter values. The event impact range parameter is defined as the geographic area or policy coverage area affected by the event, used to narrow the search space for subsequent subgraph matching.

[0037] The incremental subgraph matching and update algorithm first locates the seed node set based on the event impact range parameter. Starting from the seed node set, it performs a k-hop breadth-first traversal to obtain the affected k-hop neighborhood subgraph. The k-hop neighborhood subgraph contains all nodes and edges whose distance to the seed node set does not exceed k hops. An incremental variant of the VF2 algorithm is used to perform pattern matching in the k-hop neighborhood subgraph to identify the target node set and target edge set that match the graph operation instruction sequence. Only the target node set and target edge set are locally updated to generate the intermediate knowledge graph. Simultaneously, the local reconstruction of the disk mapping index is maintained, and a version linked list records the timestamp and operation content of each modification, achieving incremental updates while avoiding full graph reconstruction.

[0038] The time-encoded vector is calculated using a time decay function, which takes the form of an exponential kernel function. The formula for calculating the time-encoded vector of node i at time t is as follows: First, calculate the current time t and the node's historical update timestamp. time interval Then, the time interval is converted into a dimensionless time parameter through normalization. ,in The maximum time window is set for the system, and the time decay factor is calculated last. ,in The decay rate hyperparameter is obtained by multiplying the time decay factor by the position encoding vector. The node history update timestamp is derived from the most recent modification time of the target node set recorded in the version list.

[0039] The time-aware graph attention aggregation algorithm extends the time dimension based on the graph attention network framework, calculating a spatial attention score and a time decay factor for each neighbor node j of node i. The spatial attention score is obtained by calculating the similarity between node features using a learnable attention weight matrix, and the time decay factor is calculated based on the historical interaction times of neighbor node j. The spatial attention score and the time decay factor are multiplied to obtain a joint weight. A multi-head attention mechanism is used to compute multiple sets of joint weights in parallel to capture dependency patterns at different time scales. Finally, the features of all neighbor nodes are weighted and summed according to the joint weights to generate the node state embedding representation of node i. The node state embedding representation integrates the node's historical state, neighborhood aggregation information, and the time encoding vector, and maintains historical trajectory information through a gated recurrent unit update mechanism.

[0040] The proposed territorial spatial knowledge evolution model consists of four layers: a graph structure layer, a spatiotemporal encoding layer, a state update layer, and a consistency detection layer. The graph structure layer stores the node and edge sets of the intermediate knowledge graph. The spatiotemporal encoding layer receives the time-encoded vector and calculates the time decay factor. The state update layer calculates the node state embedding representation using the time-aware graph attention aggregation algorithm. The consistency detection layer performs conflict detection based on the semantic consistency constraint set and outputs the set of conflicting nodes. This four-layer structure enables the complete evolutionary process from event input to graph update to conflict detection. The node state embedding representation is stored in the hidden state storage unit of the state update layer for subsequent semantic reasoning and decision support.

[0041] The semantic consistency constraint set contains logical rules in the field of land spatial planning. These logical rules are represented in first-order predicate logic form, defining the constraint relationships between node attributes and the legality conditions of edge types. The consistency detection layer checks whether any logical rule in the semantic consistency constraint set is violated by traversing the nodes and edges in the intermediate knowledge graph. When a combination of node attributes does not meet the constraint conditions or an edge connection violates the type restriction, the node or edge is marked as a member of the conflict node set. The degree of conflict is calculated by statistically analyzing the number and severity of violations of logical rules, with the severity determined according to the priority weight of the logical rules. The semantic consistency constraint set is derived from land spatial planning regulations and industry standards.

[0042] The multi-version concurrency control protocol creates an independent snapshot version for each graph update transaction, which records the graph state at the start of the transaction. An optimistic locking mechanism is used to detect write conflicts during transaction commit, where multiple transactions simultaneously modify the same node or edge. Concurrent transactions are causally ordered based on event timestamps, prioritizing transactions with earlier timestamps. When a write conflict is detected or the conflict severity exceeds a preset threshold, the current transaction is rolled back to the snapshot version, all executed graph operations are undone, and the transaction status is marked as failed. The preset threshold is set according to the fault tolerance requirements of the land spatial planning application scenario, typically set to the maximum allowed number of logical violations. The preset threshold ranges from 3 to 10 violating nodes.

[0043] Optionally, the system may also output the final territorial spatial knowledge evolution model to a territorial spatial planning decision-making system. The territorial spatial planning decision-making system receives the updated knowledge graph and the node state embedding representation output by the territorial spatial knowledge evolution model, and uses them to perform planning scheme evaluation, land use conflict early warning, and policy impact analysis. The updated knowledge graph contains the latest geographic entity relationships and policy constraints, and the node state embedding representation is used to calculate the semantic similarity and association strength between nodes, supporting intelligent query and knowledge reasoning functions.

[0044] The value of k in the k-hop neighborhood subgraph is dynamically determined based on the event impact range parameter. For local parcel change events, k is set to 2 to 3; for regional policy adjustment events, k is set to 4 to 6; and for global policy update events, k is set to 7 to 10. Dynamically adjusting the k value balances update accuracy and computational efficiency, avoiding unnecessary large-scale subgraph traversal. The version linked list uses a doubly linked list structure to store historical snapshots, supporting forward tracing and backward recovery operations. Each linked list node stores a snapshot timestamp, modification operation records, and graph state difference information.

[0045] The gated recurrent unit update mechanism controls the fusion ratio of historical states and current inputs through reset and update gates. The reset gate determines how much historical information is forgotten, and the update gate determines how much historical state is retained. When node states change drastically, the update gate increases its openness; when node states are stable, the reset gate closes to maintain long-term memory. The gated recurrent unit achieves adaptive state updates through a nonlinear activation function and a learnable parameter matrix, solving the gradient vanishing problem in deep graph neural networks.

[0046] The multi-head attention mechanism employs 4 to 8 attention heads, each independently calculating attention weights between nodes to capture interaction patterns across different semantic spaces. The outputs of all attention heads are concatenated and passed through a linear transformation layer to generate the final node state embedding representation. A residual connection structure is used to add the node state embedding representation to the node features of the previous layer, preserving shallow information transmission channels and preventing oversmoothing in deep networks. Oversmoothing refers to the problem where node representations tend to become identical and lose distinguishability after stacking multiple layers of graph convolutions.

[0047] The pre-trained language model employs a BERT or RoBERTa architecture and is fine-tuned on text corpora in the field of land spatial planning, learning domain-specific vocabulary and semantic patterns. The fine-tuning process trains the model to identify entities and relationships in event texts using labeled samples, outputting structured event tuples. The labeled samples include historical policy documents, land use change records, and planning texts; each labeled sample includes the event type, involved entities, and attribute change information. The output of the pre-trained language model is processed by the event parser to generate the graph operation instruction sequence.

[0048] The spatial attention score in the time-aware graph attention aggregation algorithm is calculated as follows: for node i and its neighboring node j, the node feature vector is weighted by a learnable attention weight matrix W. and Mapping to the attention space, calculate the similarity score. Where 'a' is the attention coefficient vector, || denotes the vector concatenation operation, and then the spatial attention score is obtained by softmax normalization. ,in Let i represent the set of neighboring nodes of node i.

[0049] The incremental variant of the VF2 algorithm in the incremental subgraph matching and update algorithm employs a state-space search strategy. It maintains a set of partially matched states and finds the complete isomorphic mapping of the subgraph through backtracking. In each iteration, the state-space search strategy selects unmatched node pairs for expansion, checking if they satisfy node label consistency and edge connectivity consistency constraints. If satisfied, the nodes are added to the partially matched state set; otherwise, the search backtracks to the previous state. This incremental variant performs the search only within the k-hop neighborhood subgraph, reducing computational complexity compared to full graph matching.

[0050] The conflict detection algorithm of the consistency detection layer traverses the node set of the intermediate knowledge graph, extracts the attribute value set and associated edge type set of each node, and matches them with the logical rules in the semantic consistency constraint set. The logical rules represent the implication relationship between premises and conclusions; a node is considered in conflict when it satisfies the premises but not the conclusion. The degree of conflict is calculated by accumulating the severity weights of all rule violations. The severity weights are set according to the mandatory nature and scope of influence of the rules; the weights for mandatory rules are set to 5 to 10, and the weights for advisory rules are set to 1 to 3.

[0051] like Figure 2 As shown, the training process of the territorial spatial knowledge evolution model adopts a supervised learning approach. The training dataset includes historical event sequences and corresponding knowledge graph snapshot sequences. The learnable parameter matrix of the state update layer is optimized using the backpropagation algorithm. The loss function consists of three parts: node classification loss, edge prediction loss, and consistency detection loss. The node classification loss measures the cross-entropy between the predicted node state embedding representation and the real node category; the edge prediction loss measures the binary cross-entropy between the predicted edge connection probability and the real graph structure; and the consistency detection loss measures the F1 score between the predicted set of conflicting nodes and the labeled set of conflicting nodes. After training, the parameters of the territorial spatial knowledge evolution model are fixed and used to process new external event information.

[0052] The graph structure layer uses an adjacency matrix and a node feature matrix to represent the intermediate knowledge graph, where the adjacency matrix has a dimension of [missing information]. Where N is the total number of nodes, and matrix elements of 0 or 1 indicate whether there are edges connecting the nodes. The dimension of the node feature matrix is... Where D is the node feature dimension, and each row represents the attribute feature vector of a node. The spatiotemporal coding layer concatenates the temporal coding vector with each row of the node feature matrix to generate a dimension of... The enhanced feature matrix is ​​denoted by T, where T is the temporal encoding dimension. The state update layer receives the enhanced feature matrix and the adjacency matrix, and outputs a dimension of T through the temporal-aware graph attention aggregation algorithm. The node state embedding representation matrix, where H is the hidden state dimension.

[0053] The member nodes of the conflict node set are marked as pending correction in the updated knowledge graph. The territorial spatial planning decision-making system generates correction suggestions based on the nodes in the pending correction state. The correction suggestions include operation schemes such as deleting conflict nodes, modifying node attribute values, or adjusting edge connection relationships. The correction suggestions are automatically generated by analyzing the causes of the conflict, prioritizing the correction scheme with the smallest impact to avoid triggering cascading modifications.

[0054] Optionally, the present invention also provides a method for forming a territorial spatial knowledge evolution system by means of a computer, wherein the computer is provided with a readable storage medium, the readable storage medium stores program instructions, and the program instructions are used to execute the above-described method when the computer is run.

[0055] The specific implementation methods of the above steps are described in detail below.

[0056] The specific implementation of step S01 involves first reading all node and edge data of the initial knowledge graph of the national spatial planning database. The node data includes the spatial coordinates, attribute labels, and creation timestamps of geographic entities, while the edge data includes the relationship type and relationship strength weights between nodes. The Louvain algorithm is used to optimize the modularity of the initial knowledge graph. This algorithm iteratively optimizes the network modularity index based on a greedy strategy. The modularity index measures the difference between the internal connection density of a community and the expected value of a random network. It iterates by repeatedly merging nodes into neighboring communities that maximize the modularity gain until the modularity no longer increases, at which point the iteration terminates. The Louvain algorithm outputs a node affiliation mapping table for multiple subgraph partitions. Based on this mapping table, the original node set and edge set are allocated to the corresponding subgraph partition storage units. A fixed-size cache is allocated in memory for each subgraph partition. The cache size is dynamically set according to the number of nodes in the subgraph partition, typically configured to accommodate 100 to 500 nodes. The cache is managed using an LRU (Least Recently Used) eviction policy, recording the most recent access timestamp of each subgraph partition. When the cache is full, the subgraph partition with the earliest access timestamp is evicted, and its data is written back to disk. A B+ tree index structure is established to map node identifiers to disk storage addresses. The leaf nodes of the B+ tree index store the complete node identifier and disk sector address, while non-leaf nodes store the index key value range for quick location of the target leaf node. The purpose of these steps is to reduce the memory footprint of large-scale knowledge graphs and improve data access efficiency through graph partitioning and hierarchical storage.

[0057] The specific implementation of step S02 involves subscribing to event notification messages from the message queues of external policy release systems, land use management systems, and regional planning systems. These event notification messages contain the original text description and metadata information of the event. The received event notification messages are structured and parsed. An event type identifier field is extracted to determine whether it belongs to a policy update event, land use change event, or regional adjustment event. An event content text field is extracted to obtain a detailed description of the event. An event timestamp field is extracted to record the exact time the event occurred. An event impact range parameter field is extracted to obtain the geographic area code or policy coverage identifier involved in the event. The extracted structured fields are assembled into an event tuple data structure, which includes four required fields: event type identifier, event content text, event timestamp, and event impact range parameter. The event tuple is then input into an event parser module, which is implemented based on the pre-trained language model BERT architecture. This parser learns the semantic representation of domain vocabulary by fine-tuning it on text corpora in the field of land spatial planning. The event parser module segments the event content text into words and inputs them into a pre-trained language model to obtain the context-related embedding vector for each word. It then identifies geographical entity names, policy clauses, and attribute values ​​in the text through a named entity recognition layer, and identifies the relationship types and directions between entities through a relation extraction layer. Based on the identified entities and relationships, it generates a graph operation instruction sequence, which includes five operation types: adding nodes, deleting nodes, adding edges, deleting edges, and modifying attributes. Each instruction carries an operation object identifier and operation parameter values. The purpose of these steps is to transform external events described in natural language into structured graph operation instructions, thereby achieving event-driven knowledge graph updates.

[0058] The specific implementation of step S03 involves retrieving a matching set of seed nodes from the initial knowledge graph of the national land space based on the event impact range parameter. This set of seed nodes contains all nodes whose geographic region codes or policy identifiers match the event impact range parameter. A breadth-first search algorithm is executed starting from each node in the seed node set. This algorithm maintains a queue of nodes to be visited and a set of visited nodes. Initially, seed nodes are added to the queue. Each time, a node is removed from the queue, marked as visited, and its unvisited neighbor nodes are added to the queue. Expansion stops when the traversal depth reaches k hops. The value of k is dynamically set according to the event type identifier: k is set to 2 for local land parcel change events, 5 for regional policy adjustment events, and 8 for global policy update events. The breadth-first search algorithm outputs a k-hop neighborhood subgraph data structure containing all nodes and edges within k hops. An incremental subgraph matching and update algorithm is used to locate the target node set and target edge set in the k-hop neighborhood subgraph. This algorithm is based on the VF2 algorithm's state space search strategy, maintaining a partial matching state stack to record the currently matched node pair set. Each time, a candidate node is selected from the unmatched nodes to pair with a pattern graph node. Node label consistency and edge connection consistency constraints are checked. If the constraints are satisfied, the node pair is pushed onto the state stack, and the algorithm recursively matches the next node. If the constraints are not satisfied, the top element of the state stack is popped to try other candidate nodes. A complete match is completed when the number of node pairs in the state stack equals the number of pattern graph nodes. The matched nodes and edges are marked as the target node set and target edge set, respectively. According to the graph operation instruction sequence, corresponding add, delete, or attribute modification operations are performed on the target node set and target edge set to generate an intermediate knowledge graph and update the affected node mapping items of the B+ tree index structure. A new version node is appended to the version list, recording the event timestamp of this update, the list of operation instructions, and the differences in node states before and after the modification. The purpose of these steps is to avoid full graph scanning by matching local subgraphs, reduce the computational complexity of incremental updates, and retain historical version information.

[0059] The specific implementation of step S04 involves traversing the target node set to obtain the node's historical update timestamp for each node. The historical update timestamp is obtained by querying the most recent modification record of the node from the version list. The time interval between the event timestamp and the node's historical update timestamp is calculated, with the unit being seconds, by subtracting the two timestamps. A maximum time window parameter is set to 86400 seconds, corresponding to 24 hours. The time interval is divided by the maximum time window parameter to obtain a normalized dimensionless time parameter, with a value range of 0 to 1. An exponential kernel function is used to calculate the time decay factor. The input of the exponential kernel function is the dimensionless time parameter, and the output is a decay weight between 0 and 1. The decay rate hyperparameter in the exponential kernel function is set to 2.0 so that nodes with larger time intervals receive smaller decay factors. A position encoding vector is generated to represent the structural position information of the node in the graph. The position encoding vector uses a sine / cosine function encoding method to encode the number of hops from the node to the seed node set. The time decay factor is multiplied element-wise with the position encoding vector to obtain a time encoding vector, which has a dimension of 128. The time encoding vector is then concatenated with the node feature vector of the target node set to form an enhanced feature vector, which contains one-hot encodings of node attribute features and category labels. A computational graph structure for the time-aware graph attention aggregation algorithm is constructed. First, for each node i in the target node set, its neighbor node set is extracted. The neighbor nodes connected by the outgoing and incoming edges of node i are queried using the adjacency matrix of the intermediate knowledge graph. A spatial attention score is calculated for node i and each neighbor node j. This spatial attention score is used to map node features to the attention space using a learnable attention weight matrix, and then the inner product similarity is calculated. After LeakyReLU activation and softmax normalization, a weight value between 0 and 1 is obtained. The node's historical update timestamp is queried to calculate its corresponding time decay factor. The spatial attention score is multiplied by the time decay factor to obtain the joint weight. Four attention heads are used to compute four sets of joint weights in parallel. Each attention head uses an independent attention weight matrix to learn different semantic patterns. The enhanced feature vectors of the neighbor node set are weighted and summed according to the joint weights. The outputs of the four attention heads are concatenated and mapped through a linear transformation layer to a node state embedding representation with a hidden state dimension of 256. A gated recurrent unit update mechanism is used to fuse the node's historical state. The gated recurrent unit includes two gated units: a reset gate and an update gate. The reset gate controls the proportion of forgotten historical information, and the update gate controls the proportion of retained historical state. The gate value is calculated using a sigmoid activation function and weighted and fused with the historical state and the current input. The updated node state embedding representation is output to the state update layer of the territorial spatial knowledge evolution model for storage.The purpose of these steps is to capture the spatiotemporal evolution characteristics of nodes through time encoding and graph attention mechanisms, and to generate dynamic node representations that incorporate historical trajectories.

[0060] The specific implementation of step S05 involves loading a semantic consistency constraint set from the land and space planning regulations database and the industry standard knowledge base. This set contains several first-order predicate logic rules, each defining the implication relationship between preconditions and conclusions, as well as the rule's priority weight. The consistency detection layer of the land and space knowledge evolution model traverses the node set of the intermediate knowledge graph, extracting the attribute value set and associated edge type set for each node, and performing matching detection with the logic rules in the semantic consistency constraint set. It determines whether a node satisfies the preconditions of a certain logic rule. If the preconditions are met, it further checks whether the conclusion constraints are satisfied. If the conclusion constraints are not satisfied, the node is added to the conflict node set, and the violated logic rule identifier is recorded. The priority weights of the violated logic rules for each node in the conflict node set are accumulated, and the total conflict level value is calculated. The priority weight is set to 8 for mandatory rules and 2 for advisory rules. It is then determined whether the total conflict level value exceeds a preset threshold, which is set to a value between 20 and 30 based on the application scenario. If the total conflict level exceeds the preset threshold, the rollback mechanism of the multi-version concurrency control protocol is triggered. The snapshot version data before the update is read from the version list, the intermediate knowledge graph is restored to its state before the update, all modifications to the target node set and the target edge set are undone, the update transaction is marked as failed, and the reason for failure is recorded. The purpose of these steps is to ensure the logical consistency of the knowledge graph through semantic constraint detection and to prevent erroneous updates from introducing contradictory knowledge.

[0061] The specific implementation of step S06 is as follows: when the total conflict level value does not exceed the preset threshold, the territorial spatial knowledge evolution model retains the intermediate knowledge graph as a valid update result and persistently stores the node set and edge set of the intermediate knowledge graph in the graph database. The node state embedding representation matrix is ​​updated by writing the node state embedding representation of the target node set calculated by the state update layer into the corresponding row of the node state embedding representation matrix, while keeping the embedding representations of other unmodified nodes unchanged. The updated knowledge graph data structure and the node state embedding representation matrix are encapsulated into an output data packet of the territorial spatial knowledge evolution model. The output data packet includes the adjacency matrix and node feature matrix of the graph structure layer, the time encoding vector set of the spatiotemporal encoding layer, the node state embedding representation matrix of the state update layer, and the conflict node set markers of the consistency detection layer. The output data packet is transmitted to the territorial spatial planning decision system through a network interface. After receiving the output data packet, the territorial spatial planning decision system parses the updated knowledge graph structure, performs semantic similarity query and association analysis using the node state embedding representation matrix, and generates minor conflict correction suggestions based on the conflict node set markers for planners' reference. The execution log for this update transaction includes the event identifier, update time, number of modified nodes, and conflict level, which is used for subsequent system performance monitoring and optimization analysis. The purpose of these steps is to output a complete land spatial knowledge evolution model to support downstream applications and form an event-driven knowledge graph closed-loop evolution mechanism.

[0062] It should be noted that the key technical ideas of this invention include an event-driven incremental graph update mechanism and a time-aware dynamic node representation learning mechanism. The event-driven incremental graph update mechanism parses external events into local graph operation instructions and uses a subgraph matching algorithm to locate the affected neighborhood. It only performs update operations on the affected nodes, avoiding full graph reconstruction. Compared to the full reconstruction method of traditional static knowledge graphs, this mechanism significantly reduces update latency and computational resource consumption, achieving near real-time response capability of the knowledge graph to dynamic events and solving the response lag problem caused by global recalculation in traditional methods. The time-aware dynamic node representation learning mechanism incorporates a time decay factor into the graph attention aggregation process and maintains the historical evolution trajectory of nodes through gated recurrent units. This ensures that node representations not only reflect current structural features but also contain temporal dependency information. Compared to the static representation method of ordinary graph neural networks that ignores the time dimension, this mechanism enhances the model's ability to capture knowledge evolution patterns, improves the accuracy of conflict detection and semantic reasoning, and overcomes the limitation of traditional methods in modeling long-range temporal relationships. The synergistic effect of the two technical approaches lies in the incremental update mechanism continuously maintaining the structural integrity of the knowledge graph, and the dynamic representation mechanism synchronously learning the spatiotemporal evolution patterns of nodes. Together, they construct a closed-loop feedback knowledge evolution system. Compared with existing technologies that rely on manual intervention and periodic batch updates, this invention achieves automated and continuous knowledge graph evolution, adapting to the highly dynamic needs of territorial spatial planning while ensuring semantic consistency.

[0063] It should be noted that this invention also addresses the following technical problem: Traditional knowledge graph update methods lack the ability to model the temporal dimension of events, resulting in an inability to accurately reflect the dynamic characteristics of knowledge evolution over time. This invention introduces time-encoded vectors and a time-aware graph attention aggregation algorithm to encode the time interval information between event timestamps and node historical update timestamps as a time decay factor. Combined with a multi-head attention mechanism, it captures the temporal dependencies between nodes at different time scales, enabling the node state embedding representation to incorporate historical trajectory information, thus solving the problem that static graph representation methods cannot characterize the dynamic evolution process of knowledge. Furthermore, existing large-scale graph update technologies face computational efficiency bottlenecks. This invention uses event influence range parameters to dynamically determine the search range of the k-hop neighborhood subgraph. Through incremental subgraph matching, update operations are performed only on the necessary target node set and target edge set, avoiding the huge overhead of full graph traversal and reconstruction. Simultaneously, it utilizes the memory cache and disk-mapped indexes of a distributed graph storage architecture to achieve fast access to hot data, significantly improving the evolution and update efficiency of large-scale territorial spatial knowledge graphs.

[0064] Specifically, the principle of this invention is as follows: This invention divides a large-scale knowledge graph into tightly connected subgraph partitions using a distributed graph storage architecture. By combining the event impact range parameter, it accurately locates the affected k-hop neighborhood subgraphs, ensuring that update operations only affect necessary local areas rather than the entire graph, effectively reducing computational complexity and improving update efficiency. The time-encoded vector quantifies the time interval between the event timestamp and the node's historical state using a time decay function. The time-aware graph attention aggregation algorithm incorporates the time decay factor into the neighborhood information aggregation process, enabling the node state embedding representation to capture the temporal dependency patterns of knowledge evolution, overcoming the limitation of static representation methods in failing to characterize dynamic evolutionary characteristics. The semantic consistency constraint detection layer immediately executes logical rule verification after the graph is updated. By traversing the intermediate knowledge graph, it checks whether node attributes and edge connections violate domain constraints. Combined with a multi-version concurrency control protocol, it rolls back transactions when the conflict level exceeds a threshold, ensuring that the final output knowledge graph meets semantic consistency requirements, thus achieving reliable assurance of semantic correctness while guaranteeing update efficiency.

[0065] The following provides a specific embodiment 1 of the present invention. The specific implementation methods of steps S01-S03 and S06 in this embodiment 1 are the same as those described above, and will not be repeated in detail here. The specific implementation methods of other steps are described in detail below.

[0066] The specific implementation of step S04 involves calculating a time-encoded vector for each node in the target node set. This time-encoded vector integrates the time interval information between the event timestamp and the node's historical update timestamps. The node state embedding representation is then calculated using a time-aware graph attention aggregation algorithm and output to the national spatial knowledge evolution model. The time-encoded vector is calculated using a time decay function, which employs an exponential kernel function. The formula for calculating the time-encoded vector of node i at time t is as follows:

[0067] ;

[0068] In the formula, The time interval is in seconds. The current time of the event is in seconds; Let be the historical update timestamp of node i, in seconds. The historical update timestamp of a node is derived from the most recent modification time of the target node set recorded in the version list. The formula for calculating the dimensionless time parameter after normalization is as follows:

[0069] ;

[0070] In the formula, It is a dimensionless time parameter; The maximum time window set for the system, in seconds, is typically between 86400 and 604800. The formula for calculating the time decay factor is as follows:

[0071] ;

[0072] In the formula, This is the time decay factor; The attenuation rate hyperparameter has an empirical value of 0.5 to 2.0. It is an exponential function. Time-coded vector. It is obtained by multiplying the time decay factor by the location encoding vector, and the calculation formula is expressed as follows:

[0073] ;

[0074] In the formula, The time-encoded vector for node i; The location encoding vector for node i is used. The time-aware graph attention aggregation algorithm extends the time dimension based on the graph attention network framework, calculating a spatial attention score and a temporal decay factor for each neighbor node j of node i. The spatial attention score is calculated as follows: for node i and its neighbor node j, a learnable attention weight matrix is ​​used... node feature vectors and Mapping to the attention space, calculate the similarity score:

[0075] ;

[0076] In the formula, The similarity score between node i and its neighbor node j; This is the attention coefficient vector; This represents a vector concatenation operation; This is a learnable attention weight matrix; Let i be the feature vector of node i; Let be the feature vector of neighbor node j; The activation function is a linear unit with leakage correction, expressed as follows: ,in The input variable is used. The spatial attention score is obtained through normalization.

[0077] ;

[0078] In the formula, Let be the spatial attention score between node i and its neighbor node j; Represents the set of neighboring nodes of node i; For the set of neighboring nodes The index variable in the table. The time decay factor of neighbor node j. Calculated based on its historical interaction times, the calculation method is the same as the aforementioned time decay factor. The same, in which Replace with , Let be the historical update timestamp of neighbor node j, in seconds. The formula for calculating the joint weight is as follows:

[0079] ;

[0080] In the formula, Let be the joint weight between node i and its neighbor node j. A multi-head attention mechanism is used to compute multiple sets of joint weights in parallel to capture dependency patterns at different time scales. The multi-head attention mechanism uses 4 to 8 attention heads, each independently computing the attention weights between nodes to capture interaction patterns in different semantic spaces. The node state embedding representation of node i is shown below. The calculation formula is as follows: It is generated by weighted summation of the features of all neighboring nodes according to their joint weights.

[0081] ;

[0082] In the formula, Let be the node state embedding representation of node i. The node state embedding representation integrates the node's historical state, neighborhood aggregation information, and time-encoded vector, and maintains historical trajectory information through a gated recurrent unit update mechanism.

[0083] The specific implementation of step S05 involves the territorial spatial knowledge evolution model performing semantic consistency constraint detection based on node state embedding representation and intermediate knowledge graph. It identifies conflicting node sets according to the semantic consistency constraint set. When the conflict severity of a conflicting node set exceeds a preset threshold, a multi-version concurrency control protocol is triggered to roll back to a snapshot version. The semantic consistency constraint set contains logical rules in the territorial spatial planning domain. These logical rules are represented in first-order predicate logic form, defining the constraint relationships between node attributes and the legality conditions of edge types. The consistency detection layer checks whether any logical rule in the semantic consistency constraint set is violated by traversing the nodes and edges in the intermediate knowledge graph. When a node attribute combination does not meet the constraint conditions or an edge connection violates type restrictions, the node or edge is marked as a member of the conflicting node set. The conflict severity is calculated by statistically analyzing the number and severity of violations of logical rules. The formula for calculating the conflict severity is as follows:

[0084] ;

[0085] In the formula, The degree of conflict; The set of logical rules that are violated; The number of nodes that violated rule q; The severity weight of rule q; For a set of logical rules The index variables are defined in the graph. Severity weights are determined based on the priority weights of logical rules, with mandatory rules weighted at 5 to 10 and advisory rules weighted at 1 to 3. The semantic consistency constraint set is derived from land spatial planning regulations and industry standards. A multi-version concurrency control protocol creates an independent snapshot version for each graph update transaction, recording the graph state at the start of the transaction. An optimistic locking mechanism is used to detect write conflicts during transaction commits; write conflicts refer to multiple transactions simultaneously modifying the same node or edge. Concurrent transactions are causally ordered based on event timestamps, prioritizing transactions with earlier timestamps. When a write conflict is detected or the conflict severity exceeds a preset threshold, the current transaction is rolled back to the snapshot version, all executed graph operations are undone, and the transaction status is marked as failed. (Preset threshold) Based on the fault tolerance requirements of the application scenario of land and space planning, it is usually set as the maximum number of allowed logical violations, with the preset threshold value ranging from 3 to 10 violation nodes.

[0086] It should be noted that the variables involved in this invention are explained in detail in Table 1.

[0087] Table 1. Variable Explanation Table

[0088]

[0089] To better understand and implement this invention, a specific application scenario is provided below as Example 2: A technical team is responsible for building a regional territorial spatial planning knowledge management system. This system needs to manage a knowledge graph containing 8763 geographic entity nodes, 12458 policy rule nodes, and 34576 relationship edges, facing the need for graph evolution due to frequent policy updates and dynamic land use changes. The technical team uses the method of this invention to construct a territorial spatial knowledge evolution model, achieving efficient response to external events and ensuring semantic consistency.

[0090] The technical team first constructed a distributed graph storage architecture and used the Louvain algorithm to perform community detection on the initial spatial knowledge graph containing 21,221 nodes. After 15 iterations of optimization, the graph was divided into 32 subgraph partitions, with an average of 663 nodes per subgraph partition. The internal connection density within a community reached 0.78, and the inter-community connection sparsity was 0.12, meeting the requirements for tight connections. The team established an independent 512MB memory cache for each subgraph partition, used an LRU eviction policy to maintain the top 8 most frequently accessed partitions in persistent memory, and built a B+ tree structure disk mapping index with a depth of 4 levels, supporting an average of 3.2 disk accesses to complete node location operations.

[0091] The technical team deployed an event collection module, connecting three data sources: the policy release system, the land use management system, and the regional planning system. During a continuous 30-day operation period, a total of 487 external event information items were collected, and the event type distribution is shown in Table 2.

[0092] Table 2. Statistics on External Event Types

[0093]

[0094] The technical team configured a pre-trained language model based on the RoBERTa architecture as the event parser. This model was fine-tuned on a corpus containing 580,000 texts in the field of land spatial planning, with 125M parameters, achieving an entity recognition F1 score of 0.894 on the labeled test set. The event parser transforms the event content text into structured event tuples. Each event tuple contains an event type identifier, event content text, event timestamp, and event impact range parameters, thereby generating a sequence of graph operation instructions. The instruction types include adding nodes (38%), deleting edges (27%), and modifying attributes (35%).

[0095] like Figure 3 As shown, the technical team executed an incremental subgraph matching and update algorithm for a policy update event. The impact range parameter of this event covers three administrative divisions. Based on the impact range, the system located a seed node set containing 47 nodes. Since this event is a regional policy adjustment event, the system dynamically sets the k value to 5, performing a 5-hop breadth-first traversal starting from the seed node set to obtain a k-hop neighborhood subgraph containing 512 nodes and 1834 edges. The technical team used an incremental variant of the VF2 algorithm to perform pattern matching within this subgraph, identifying a target node set containing 89 nodes and a target edge set containing 143 edges that need modification. Only these target objects were subjected to local update operations to generate an intermediate knowledge graph. The update operation took 1.7 seconds, which is significantly reduced compared to the average time of 23.5 seconds for full graph reconstruction.

[0096] The technical team calculated time encoding vectors for 89 nodes in the target node set, and the system set a maximum time window. For 90 days, the decay rate exceeds the parameter. Set to 0.05. For a node last updated 42 days ago, the system calculates the time interval. The time is 42 days, and after normalization, a dimensionless time parameter is obtained. The time decay factor is 0.467. The value is 0.977. The technical team configured a time-aware graph attention aggregation algorithm, setting up 6 attention heads to compute joint weights in parallel. Each attention head has a hidden dimension of 64, a node feature dimension D of 256, and a time encoding dimension T of 32. For example... Figure 4 As shown, the algorithm calculates the spatial attention score through a learnable attention weight matrix, multiplies it with the time decay factor to obtain the joint weight, and performs weighted aggregation of the neighbor node features to generate a node state embedding representation with a dimension of 384.

[0097] The land spatial knowledge evolution model constructed by the technical team consists of a graph structure layer, a spatiotemporal coding layer, a state update layer, and a consistency detection layer, with a total of 8.7M parameters. The adjacency matrix of the intermediate knowledge graph stored in the graph structure layer has a dimension of [missing information]. The node feature matrix has a dimension of The spatiotemporal coding layer concatenates the temporal coding vector with the node feature matrix to generate a dimensionless array. The enhanced feature matrix. The state update layer calculates the node state embedding representation through a gated recurrent unit update mechanism. The parameter matrices of the reset gate and update gate of the gated unit are both of dimension 1. The memory capacity for historical trajectory information is maintained at 10 time steps.

[0098] The consistency detection layer is configured with a semantic consistency constraint set, which contains 67 logical rules derived from the land and space planning regulations. The distribution of rule types is shown in Table 3.

[0099] Table 3 Distribution of Semantic Consistency Constraint Rule Types

[0100]

[0101] like Figure 5As shown, the consistency detection layer traversed 21,221 nodes in the intermediate knowledge graph, detecting a set of 13 conflicting nodes with 17 rule violations. Of these, 8 were violations of land use constraints, 6 were violations of spatial location constraints, and 3 were violations of policy coverage constraints. The technical team calculated the conflict level to be 121, with land use constraint violations contributing 64, spatial location constraint violations contributing 42, and policy coverage constraint violations contributing 27. The system's preset threshold was set to 150. Since the conflict level of 121 did not exceed the threshold, the multi-version concurrency control protocol determined that this update transaction could be committed, preserving the intermediate knowledge graph and updating the node state embedding representation as the final output.

[0102] The technical team output the final land spatial knowledge evolution model to the land spatial planning decision-making system. The decision-making system receives the updated knowledge graph and node state embedding representations and performs planning scheme evaluation tasks. The system uses the node state embedding representations to calculate the semantic similarity between nodes, identifying 127 geographic entity nodes highly relevant to the new policy, and generating a land use conflict early warning report marking 8 potential conflict areas. During a 90-day operation period, the technical team processed a total of 1342 external events, with an average system response time of 2.3 seconds, a knowledge graph consistency maintenance rate of 97.8%, and 34 conflict rollbacks, representing a rollback rate of 2.5%.

[0103] like Figure 6 As shown, the technical team analyzed the evolution trajectory of the state embedding representation of a key node under 30 consecutive update events. The embedding vector of this node has a trajectory length of 8.67 in 384-dimensional space, with a cumulative directional change angle of 245 degrees, reflecting the dynamic evolution characteristics of the node's state with external events. The technical team visualized the 384-dimensional embedding vector by projecting it into 2-dimensional space using principal component analysis, observing that the embedding representations of nodes of the same type form clusters in the semantic space, while clear boundaries are maintained between nodes of different types.

[0104] Compared to traditional full-graph reconstruction methods, this invention uses an incremental subgraph matching algorithm to update only the local k-hop neighborhood affected by the event, avoiding a full scan and reconstruction of the entire knowledge graph and significantly reducing computational complexity. Traditional methods require reloading all node and edge data and rebuilding the index structure when processing external events, while this invention utilizes the memory cache and disk-mapped index of a distributed graph storage architecture to quickly locate hotspot data, performing only local updates on the target node and edge sets. Traditional methods lack the ability to model the temporal dimension of events, using static node feature representations that fail to reflect the dynamic characteristics of knowledge evolution over time. This invention introduces time-encoded vectors and a time-aware graph attention aggregation algorithm, quantifying the time interval between the event timestamp and the node's historical state through a time decay function, enabling the node state embedding representation to integrate historical trajectory information and temporal dependency patterns. Traditional methods lack a systematic semantic consistency verification mechanism after graph updates, which easily leads to logical conflicts and constraint violations. In contrast, this invention establishes a consistency detection layer to perform conflict detection based on domain logic rules. Combined with a multi-version concurrency control protocol, it automatically rolls back to the snapshot version when the conflict level exceeds a threshold, ensuring that the final output knowledge graph always meets the semantic constraints of the land and space planning domain. In principle, this achieves an organic unity between update efficiency and semantic consistency assurance.

[0105] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for constructing a land spatial knowledge evolution model, characterized in that, A distributed graph storage architecture is constructed to divide the initial knowledge graph of national land space into multiple subgraph partitions; external event information is collected and structured as event tuples, and the event tuples are transformed into graph operation instruction sequences through an event parser; the affected k-hop neighborhood subgraphs are located according to the graph operation instruction sequences, and the target node set and target edge set are identified by an incremental subgraph matching and update algorithm, and local update operations are performed to generate an intermediate knowledge graph; For each node in the target node set, calculate a time-encoded vector that fuses the time interval information between the event timestamp and the node's historical update timestamp, and calculate the node state embedding representation using a time-aware graph attention aggregation algorithm; The territorial spatial knowledge evolution model performs semantic consistency constraint detection and identifies conflicting node sets based on node state embedding representations and intermediate knowledge graphs. When the conflict level exceeds a preset threshold, a multi-version concurrency control protocol is triggered to roll back to the snapshot version. When the conflict level does not exceed the preset threshold, the intermediate knowledge graph is retained and the node state embedding representation is updated as the final output.

2. The method according to claim 1, characterized in that, The community detection algorithm uses the Louvain algorithm to optimize the modularity of the initial knowledge graph of the national land space. By iteratively merging node communities, the connection density within the community is maximized while the connection between communities is sparsed.

3. The method according to claim 2, characterized in that, The memory cache uses an LRU eviction policy to maintain hotspot subgraph partitions, and the disk-mapped index uses a B+ tree structure to store the mapping relationship between node identifiers and disk addresses.

4. The method according to claim 3, characterized in that, External event information comes from policy release systems, land management systems, and regional planning systems. The event parser performs semantic understanding of the event content text based on a pre-trained language model.

5. The method according to claim 4, characterized in that, The incremental subgraph matching and update algorithm locates the seed node set based on the event influence range parameter, and performs a k-hop breadth-first traversal from the seed node set to obtain the k-hop neighborhood subgraph.

6. The method according to claim 5, characterized in that, The time-encoded vector is calculated using a time decay function, which takes the form of an exponential kernel function. The time interval between the current moment and the node's historical update timestamp is calculated and converted into a time decay factor.

7. The method according to claim 6, characterized in that, The time-aware graph attention aggregation algorithm calculates a spatial attention score and a time decay factor for each neighboring node of a node, and multiplies the spatial attention score by the time decay factor to obtain the joint weight.

8. The method according to claim 7, characterized in that, The time-aware graph attention aggregation algorithm uses a multi-head attention mechanism to compute multiple sets of joint weights in parallel to capture dependency patterns at different time scales. It generates node state embedding representations by weighting and summing the features of neighboring nodes according to the joint weights.

9. The method according to claim 8, characterized in that, The land space knowledge evolution model consists of a graph structure layer, a spatiotemporal coding layer, a state update layer, and a consistency detection layer. The graph structure layer stores intermediate knowledge graphs, and the spatiotemporal coding layer calculates the time decay factor.

10. The method according to claim 9, characterized in that, The semantic consistency constraint set contains logical rules in the field of territorial spatial planning. The logical rules are represented in the form of first-order predicate logic and define the constraint relationships between node attributes and the legality conditions of edge types.

Images