A method, medium, equipment, and product for dynamic prediction of geological hazards based on spatiotemporal knowledge graphs.

Abstract

This invention discloses a method, medium, equipment, and product for dynamic prediction of geological disasters based on spatiotemporal knowledge graphs, belonging to the field of geological disaster prediction technology. The method includes: constructing a knowledge graph based on historical monitoring data of geological disasters; extracting entities and relationships from real-time monitored geological disaster data; and dynamically updating the knowledge graph. The spatiotemporal attributes in the knowledge graph are encoded to obtain temporal and spatial feature vectors. Attention weights are constructed for entities and their neighboring entities, and neighbor information is aggregated based on these attention weights to obtain the entity's neighborhood context vector. The spatial feature vector, the current temporal feature vector, and the current entity's neighborhood context vector are fused with the entity's feature vector from the previous time step to obtain the current entity's feature vector. A geological disaster susceptibility index is then calculated, and geological disasters are predicted using this index. This invention can achieve dynamic and high-precision prediction of geological disaster susceptibility.

Description

Technical Field

[0001] This invention relates to the field of geological disaster prediction technology, and in particular to a method, medium, equipment and product for dynamic prediction of geological disasters based on spatiotemporal knowledge graphs. Background Technology

[0002] Existing geological hazard susceptibility assessment studies, whether knowledge-driven empirical and mechanical models or data-driven statistical and machine learning models, are mostly static assessments. These methods primarily determine the inherent susceptibility level of a region based on unchanging or slowly changing hazard-inducing environmental factors, thus belonging to spatial prediction. The occurrence of geological hazards is a complex nonlinear dynamic system evolution process, controlled not only by static hazard-inducing conditions but also by the coupling effects of spatiotemporally correlated dynamic triggering factors (such as rainfall, earthquakes, and human engineering activities). Existing susceptibility assessment results are insufficient to flexibly address the uncertain, multi-stage deformation and evolution processes of geological hazards, and cannot predict dynamic trends at specific points in time or under specific external conditions; that is, they lack spatiotemporal dynamic prediction capabilities, making it difficult to effectively support the needs of deeper tasks such as precise monitoring and early warning, and risk prediction and prevention.

[0003] At the knowledge representation level, while existing research has begun to utilize ontology and knowledge graphs to structurally and uniformly describe the elements and relationships of geological hazards, these constructions primarily focus on the abstraction and correlation of static geographical and geological environmental elements, observational data, and geological hazard characteristics related to geological hazards. Their limitation lies in the fact that existing knowledge graphs, as a schema layer, while providing an interface for machine learning to understand prior knowledge of geological hazard backgrounds, lack explicit characterization of the temporal dimension and dynamic evolution process. The spatiotemporal relationship between multi-site, heterogeneous observational data of geological hazards and environmental factors does not simply follow statistical regression laws. Therefore, how to construct and apply a "spatiotemporal knowledge graph" to uniformly store and express the dynamic characteristics of geological hazards, inducing factors, and the temporal changes and spatial correlations of multi-site observational data, thereby guiding machine learning in dynamic feature mining, is a pressing problem that needs to be solved.

[0004] Therefore, existing technologies in the same field still have significant defects and shortcomings in the structured expression of spatiotemporal knowledge and the effective integration of dynamic data during the transition from static evaluation to spatiotemporal dynamic prediction. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing geological hazard prediction methods in terms of structured representation of spatiotemporal knowledge and fusion of dynamic data, and to propose a dynamic geological hazard prediction method based on spatiotemporal knowledge graphs, comprising the following steps:

[0006] S1. Define the entity types, relation types, and spatiotemporal attributes of geological hazards. Use entity recognition and relation extraction techniques to extract triples from historical monitoring data of geological hazards and construct a knowledge graph. Extract entities and relations from real-time monitored geological hazard data and dynamically update the knowledge graph. S2. Encode the spatiotemporal attributes in the knowledge graph to obtain temporal feature vectors and spatial feature vectors; S3. Based on the temporal feature vector and spatial feature vector, construct the attention weights of the entity and its neighboring entities, and aggregate the neighbor information of the entity based on the attention weights to obtain the entity's neighborhood context vector. S4. The spatial feature vector, the current time feature vector, and the neighborhood context vector of the current entity are fused with the feature vector of the entity at the previous time step to obtain the feature vector of the current entity. S5. Pass the feature vector of the current entity through a fully connected output layer and map it to the desired value using the Sigmoid function. The geological hazard susceptibility index is obtained from the interval; S6. Predict geological disasters using the geological disaster susceptibility index.

[0007] Furthermore, sinusoidal positional encoding is applied to periodic time attributes to obtain periodic time feature vectors; continuously increasing time attributes are normalized and then linearly transformed through a fully connected layer to obtain continuous time feature vectors; the two time feature vectors are concatenated and input into a multilayer perceptron to generate the final time feature vector. .

[0008] Furthermore, the spatial attributes of entity coordinates are discretized into strings using Geohash encoding, and then mapped to spatial embedding vectors using a trainable embedding layer. A topological embedding vector is obtained by pre-training the spatial properties of the entity topology using the DeepWalk random walk algorithm. ;Will and The vectors are concatenated and then processed by an MLP to generate the final spatial feature vector. .

[0009] Furthermore, S3 specifically refers to: S31. Define the messages passed between an entity and its neighboring entities through relationships as follows:

[0010] in, This represents the message passed between the i-th entity and its j-th neighbor entity through relation r. Represents time t-1 eigenvectors, It is the j-th neighbor entity of the i-th entity. It is for relationships Defined learnable weight matrix, It is for relationships The bias term.

[0011] S32. Calculate the relationship-based attention coefficient between an entity and its neighboring entities, using the following formula:

[0012]

[0013]

[0014]

[0015] in, Let denot , where is the attention coefficient between the i-th entity and its j-th neighbor entity at time t based on relation r, and ... and The transformation matrix represents the query and the key. This represents the feature vector of the i-th entity at time t-1. This represents the gating coefficient between the i-th entity and its j-th neighbor entity at time t, based on relation r. It is the Sigmoid activation function. It is a learnable weight matrix. It is a time decay function. It is a spatial Gaussian kernel function. It is a learnable decay rate with respect to r. and These are the i-th and j-th time feature vectors. and These are the i-th and j-th spatial feature vectors. It is a learnable spatial influence radius with respect to r.

[0016] S33. Obtain attention weights based on the attention coefficients of an entity and its neighboring entities according to their relationships. The formula is as follows:

[0017] in, This represents the attention weights of the i-th entity and its j-th neighbor entity at time t based on relation r. Let i represent the set of neighboring entities of the i-th entity. This represents the set of relationships between the i-th entity and its k-th neighboring entities. Let represent the attention coefficient between the i-th entity and its k-th neighbor entity at time t based on relation r'; S34. Based on attention weights, aggregate neighbor information to obtain the entity's neighborhood context vector, using the following formula:

[0018] in, This represents the neighborhood context vector of the i-th entity at time t. It is the transformation matrix of values.

[0019] Furthermore, the feature vector of the current entity is expressed by the formula:

[0020] in, It is the feature vector of the i-th entity at time t. Indicates a gated loop unit. This represents the feature vector of the i-th entity at time t-1. This represents the neighborhood context vector of the i-th entity at time t. This represents the time feature vector of the i-th entity at time t. This represents the spatial feature vector of the i-th entity.

[0021] Furthermore, the geological hazard susceptibility index is expressed by the formula:

[0022] in, express Geological disaster susceptibility index at all times It is the Sigmoid activation function. and These are the weight and bias coefficients of the output layer. It is the eigenvector of the entity at time t.

[0023] Furthermore, according to steps S2-S5, a dynamic prediction model for geological disasters based on spatiotemporal knowledge graphs is constructed. The total loss of the model training includes graph structure loss, temporal smoothing constraint loss, and spatial proximity constraint loss. Graph structure loss is used to constrain the gap between the embedded vectors learned by the model and the facts in the graph; temporal smoothing constraint loss is used to constrain the gap between adjacent times before and after entity embedding; and spatial proximity constraint loss is used to constrain the embedded vectors learned by the model to follow geographic spatial autocorrelation.

[0024] The present invention also proposes a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the above-described method for dynamic prediction of geological disasters based on spatiotemporal knowledge graphs.

[0025] The present invention also proposes an electronic device, including a processor and a memory, wherein the processor and the memory are interconnected, wherein the memory is used to store a computer program, the computer program including computer-readable instructions, and the processor is configured to invoke the computer-readable instructions to execute the above-described method for dynamic prediction of geological disasters based on spatiotemporal knowledge graphs.

[0026] The present invention also proposes a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps of the above-described method for dynamic prediction of geological disasters based on spatiotemporal knowledge graphs.

[0027] The beneficial effects of the technical solution provided by this invention are: This invention constructs a dynamically updated knowledge graph based on entities, relationships, and spatiotemporal attributes. It integrates neighborhood relationship perception and dynamic knowledge representation vectors of time and spatial distance into the entity representation vector, addressing the shortcomings of existing geological disaster prediction methods in the structured representation of spatiotemporal knowledge and the fusion of dynamic data. This enables dynamic and high-precision prediction of geological disaster susceptibility. Attached Figure Description

[0028] Figure 1 This is a flowchart of a dynamic prediction method for geological disasters based on spatiotemporal knowledge graphs according to an embodiment of the present invention; Figure 2 This is a map showing the geological conditions and potential hazards in the study area according to an embodiment of the present invention. Figure 3 These are the prediction results of geological disaster-prone areas in the study area of ​​this invention embodiment; Figure 4 This is a schematic diagram of the predicted heat map of the geological disaster-prone area and the highest subsidence risk point in the study area of ​​this invention embodiment; Figure 5 This is a schematic diagram illustrating the causal tracing of the explainability of geological disaster susceptibility according to an embodiment of the present invention; Figure 6 This is a block diagram of an electronic device according to an exemplary embodiment of the present invention. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0030] The flowchart of the dynamic prediction method for geological disasters based on spatiotemporal knowledge graphs in this invention is as follows: Figure 1 Specifically, it includes the following steps: S1. Define the entity types, relation types, and spatiotemporal attributes of geological hazards. Specifically, the entity types of geological hazards include: (1) Disaster-causing entities: geological disaster bodies and potential geological disaster areas; (2) Environmental entities: geological units, water systems, roads, elevation, slope, etc.; (3) Triggering entities: rainfall events, earthquakes, human engineering activities, groundwater levels, etc.; (4) Monitoring entities: monitoring stations and deformation observation data.

[0031] Relationship types include: (1) Static relationships: located at, adjacent to, contained; (2) Dynamic relationship: induce, influence, monitor, occur.

[0032] In the spatiotemporal attributes, the time attribute is the timestamp or time interval of the entity and dynamic relationship; the spatial attribute is the spatial coordinates of the entity or geometric objects (such as points, lines, and surfaces).

[0033] This invention employs entity recognition and relation extraction techniques to extract triples from historical geological disaster monitoring data (such as historical disaster records, geological reports, remote sensing image interpretation results, and meteorological and hydrological monitoring data) to construct a knowledge graph. Entities and relations are extracted from real-time monitored geological disaster data. Specifically, real-time rainfall, groundwater levels, and other data are used to determine early warning thresholds and encapsulate events to generate new triggering entities and dynamic relationships, thereby dynamically updating the knowledge graph.

[0034] The system uses Neo4j, which supports spatiotemporal indexing and graph queries, in conjunction with PostGIS or a dedicated spatiotemporal graph database for storage, ensuring efficient spatiotemporal proximity and path queries.

[0035] S2. Encode the spatiotemporal attributes in the knowledge graph to obtain temporal feature vectors and spatial feature vectors.

[0036] Sine positional encoding is applied to periodic time attributes (such as months and hours) to obtain periodic time feature vectors. This can be expressed by the formula:

[0037]

[0038] in, and Represents the periodic time feature vector. It refers to the time position (such as the day of the month). It is a dimensional index. It is the embedded dimension.

[0039] After normalizing the continuously growing time attribute, a linear transformation is performed through a fully connected layer to obtain the continuous time feature vector.

[0040] The two time feature vectors are concatenated and input into a multilayer perceptron to generate the final time feature vector. .

[0041] The spatial attributes of entity coordinates are discretized into strings using Geohash encoding, and then mapped to spatial embedding vectors using a trainable embedding layer. A topological embedding vector is obtained by pre-training the spatial properties of the entity topology using the DeepWalk random walk algorithm. .Will and The vectors are concatenated and then processed by an MLP to generate the final spatial feature vector. .

[0042] S3. Based on the temporal and spatial feature vectors, construct the attention weights of the entity and its neighboring entities, and aggregate the neighbor information of the entity based on the attention weights to obtain the entity's neighborhood context vector.

[0043] Specifically: S31. Define the messages passed between an entity and its neighboring entities through relationships as follows:

[0044] in, This represents the message passed between the i-th entity and its j-th neighbor entity through relation r. Represents time t-1 eigenvectors, It is the j-th neighbor entity of the i-th entity. It is for relationships Defined learnable weight matrix, It is for relationships The bias term.

[0045] S32. Calculate the relationship-based attention coefficient between an entity and its neighboring entities, using the following formula:

[0046]

[0047]

[0048]

[0049] in, This represents the attention coefficient between the i-th entity and its j-th neighbor entity at time t, based on relation r. Spacetime Gating Adjustment, where 'a' represents the attention vector. and The transformation matrix represents the query and the key. This represents the feature vector of the i-th entity at time t-1. This represents the gating coefficient between the i-th entity and its j-th neighbor entity at time t, based on relation r. It is used to quantify the impact of spatiotemporal distance on the intensity of information transmission. It is the Sigmoid activation function. It is a learnable weight matrix. It is a time decay function; the longer the time, the lower the influence of the information. It is a spatial Gaussian kernel function, and its influence decreases with increasing spatial distance. It is a learnable decay rate with respect to r. and These are the i-th and j-th time feature vectors. and These are the i-th and j-th spatial feature vectors. It is a learnable spatial influence radius with respect to r.

[0050] S33. Obtain attention weights based on the attention coefficients of an entity and its neighboring entities according to their relationships. The formula is as follows:

[0051] in, This represents the attention weights of the i-th entity and its j-th neighbor entity at time t based on relation r. Let i represent the set of neighboring entities of the i-th entity. This represents the set of relationships between the i-th entity and its k-th neighboring entities. Let represent the attention coefficient between the i-th entity and its k-th neighbor entity at time t based on relation r'.

[0052] S34. Based on attention weights, aggregate neighbor information to obtain the entity's neighborhood context vector, using the following formula:

[0053] in, This represents the neighborhood context vector of the i-th entity at time t. It is the transformation matrix of values.

[0054] S4. The spatial feature vector, the current time feature vector, and the neighborhood context vector of the current entity are fused with the feature vector of the entity at the previous time step to obtain the feature vector of the current entity.

[0055] The feature vector of the current entity is expressed by the formula:

[0056] in, It is the feature vector of the i-th entity at time t. Indicates a gated loop unit. This represents the feature vector of the i-th entity at time t-1. This represents the neighborhood context vector of the i-th entity at time t. This represents the time feature vector of the i-th entity at time t. This represents the spatial feature vector of the i-th entity.

[0057] The GRU (Gated Recurrent Unit) automatically learns to retain historical information and incorporate new information by updating and resetting gates. This step is responsible for fusing the entity's own historical state with newly aggregated neighborhood information to capture the entity's temporal evolution patterns.

[0058] Reset door It determines how much historical information is forgotten.

[0059]

[0060] Update Gate : Determines how much new information to accept.

[0061]

[0062] Candidate hidden state :

[0063] Final hidden state :

[0064] in , , All are learnable parameters. This represents the Hadamard product (element-by-element multiplication). This represents the reset gate at time t. This represents the input at time t. This represents the hidden state at time t-1. This represents the update gate at time t. Let represent the candidate hidden state at time t. This represents the hidden state at time t.

[0065] S5. Pass the feature vector of the current entity through a fully connected output layer and map it to the desired value using the Sigmoid function. The geological hazard susceptibility index is obtained by dividing the interval into ranges, and is expressed by the formula:

[0066] in, express Geological disaster susceptibility index at all times It is the Sigmoid activation function. and These are the weight and bias coefficients of the output layer. It is the eigenvector of the entity at time t.

[0067] S6. Predict geological disasters using the geological disaster susceptibility index.

[0068] Based on the predicted geological hazard susceptibility index, the index is divided into several levels (e.g., extremely low, low, medium, high, and extremely high susceptibility zones) using either the natural discontinuity method or the frequency ratio method based on historical hazard points to characterize the probability of geological hazards occurring in different spatial units. The geological hazard susceptibility levels for different prediction time windows (e.g., the next 24 hours) are overlaid onto an ArcGIS spatial base map to generate a time-series dynamic geological hazard zoning map. When the prediction results show that the geological hazard susceptibility level of a certain area reaches "high" or "extremely high," the system automatically generates an early warning message and pushes it through a visual interface, SMS, or App.

[0069] Based on steps S2-S5, a dynamic prediction model for geological disasters based on a spatiotemporal knowledge graph can be constructed. The model is then trained using steps S2-S5, employing a multi-task loss function. The training objective is to ensure that the embedding vectors simultaneously maintain consistency in graph structure, time, and space. The loss function is:

[0070] in, and It is a hyperparameter used to balance the relative importance of the three tasks. It is the graph structure loss, which forces the model to learn embedding vectors that can accurately represent and reconstruct facts in the knowledge graph. Its core task is to enable the model to distinguish between "real facts" and "fake facts". It is a time-smoothing constraint loss, a physical regularizer. The state evolution of entities such as geological hazards and geological units is usually gradual (except for sudden events such as earthquakes). Therefore, the state evolution of an entity in... The embedding of time should not be related to its position If the embeddings at different times differ too much, the temporal smoothing constraint loss is used to ensure the temporal continuity of the embedding vectors. It is a spatial proximity constraint loss, a geospatial regularizer that forces the model to learn embedding vectors that follow geospatial autocorrelation. Entities of the same type that are close to each other in physical space (e.g., two adjacent raster cells or two adjacent geological units) should also have their embedding vectors close to each other in vector space.

[0071] Using the above method, this invention can learn high-quality dynamic knowledge representation vectors that integrate semantic, topological, temporal, and spatial characteristics.

[0072] This invention can use knowledge graphs to trace the causal relationships of predicted geological disasters. The specific method is as follows: 1. Locate entity e with a high GSI value.

[0073] 2. Find the dynamically induced entity with the highest attention weight for e (calculated according to the method in step 3). (e.g., heavy rainfall event A).

[0074] 3. Query The attribute (rainfall is) ).

[0075] 4. Query the entities with key static relationships for e (e.g., e is located in "mudstone geological unit B").

[0076] Output: "Entity e located in 'Mudstone Geological Unit B' has an 'Extremely High' susceptibility to geological hazards in the next 24 hours, mainly due to: '150mm heavy rainfall event A'." This invention realizes a complete closed loop of real-time map construction, dynamic feature aggregation, temporal state update and interpretable prediction, and achieves high-precision dynamic prediction of geological disaster susceptibility.

[0077] To verify the effectiveness of the method of the present invention, the geological disaster results in the study area were predicted and evaluated. Figure 2 This is a map showing the geological conditions and potential hazards in the study area according to an embodiment of the present invention. Figure 3 This is the prediction result of geological disaster-prone areas in the study area of ​​this invention. From... Figure 2 and Figure 3 It can be seen that: The predicted high-risk areas are highly consistent with the actual distribution of potential hazards: Figure 3 The displayed high-risk and extremely high-risk landslide areas are spatially distributed similarly to... Figure 2 The locations of the landslide hazard points identified in the field investigation showed a very high degree of agreement. This proves that the prediction method based on spatiotemporal knowledge graphs proposed in this invention can effectively learn the complex nonlinear relationship between disaster-prone environmental factors and landslide occurrence, achieving precise spatial location of the disaster.

[0078] The introduction of dynamic triggering factors significantly improved the sensitivity of prediction: compared to Figure 2 Static background conditions Figure 3 The prediction results not only consider basic factors such as topography and geology, but also capture the instantaneous contribution of real-time dynamic inducing factors to the susceptibility of geological disasters. Figure 2 Areas that were originally classified as low to medium risk may experience a change in risk levels if affected by dynamic factors such as real-time rainfall. Figure 3 The region can be accurately identified as a dynamic high-risk area, effectively making up for the shortcomings of traditional static assessment methods in dealing with spatiotemporally related dynamic inducing factors.

[0079] Figure 3 The geological hazard susceptibility prediction results exhibit distinct spatiotemporal evolution characteristics, rather than being simple static patches. Based on the causal chain tracing mechanism of this invention, the formation reasons of specific high-risk areas in the figure can be clearly explained, for example, due to the spatiotemporal coupling of specific lithological entities and sudden rainfall events. This prediction result, which deeply integrates spatiotemporal semantic information, provides more scientific decision support for accurate monitoring, early warning, and risk prevention and control.

[0080] Figure 4 This is a schematic diagram of the predicted heat map and the highest subsidence risk point of the geological disaster-prone area in the study area of ​​this invention embodiment. Figure 4 The highest risk points identified were traced back to their causal origins. A causal traceability diagram illustrating the explainability of geological disaster susceptibility is provided for reference. Figure 5 . Figure 5 The inference analysis report generated by the prediction target is presented through a spatiotemporal knowledge graph, demonstrating the excellent interpretability and verification capabilities of the method of this invention. This report, by combining macroscopic thermal location with microscopic causal reasoning, clearly elucidates the causal logic chain of the model's conclusions: First, the Geological Hazard Susceptibility Index (GSI) of the unit is determined to be 0.96 within the next 30 days, classifying it as belonging to an extremely high risk level. Then, deep mining using the knowledge graph reveals that the unit is located in the core area of ​​a goaf formed by underground coal seam mining, which is the primary cause of the disaster, and automatically links it to the loose Quaternary sedimentary rock in that area. Key static prior knowledge, such as the soil geological background, the structural location of the fault (approximately 245 meters away), and historical collapse records from 2016, combined with the soil saturation aggravation effect caused by recent cumulative rainfall of 320 mm, and deformation evidence of a deformation rate exceeding 5 mm per day captured by InSAR monitoring point JC_03, was used to perform weight aggregation and logical mapping of the above multi-source heterogeneous information through the attention mechanism within the model. This effectively verified the scientific nature of the deep learning prediction results and provided intuitive evidence with causal logic support for disaster prevention and mitigation decisions.

[0081] In one exemplary embodiment, a computer-readable storage medium is included, which stores a computer program that, when executed by a processor, implements the above-described method for dynamic prediction of geological disasters based on spatiotemporal knowledge graphs.

[0082] Please see Figure 6 In one exemplary embodiment, the device further includes an electronic device including at least one processor, at least one memory, and at least one communication bus.

[0083] The memory stores a computer program, which includes computer-readable instructions. The processor calls the computer-readable instructions stored in the memory through the communication bus to execute the above-mentioned dynamic prediction method for geological disasters based on spatiotemporal knowledge graphs.

[0084] In one exemplary embodiment, a computer program product is proposed, including a computer program / instruction that, when executed by a processor, implements the steps of the above-described method for dynamic prediction of geological disasters based on spatiotemporal knowledge graphs.

[0085] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A dynamic prediction method for geological disasters based on spatiotemporal knowledge graphs, characterized in that, Includes the following steps: S1. Define the entity types, relation types, and spatiotemporal attributes of geological hazards. Use entity recognition and relation extraction techniques to extract triples from historical monitoring data of geological hazards and construct a knowledge graph. Extract entities and relations from real-time monitored geological hazard data and dynamically update the knowledge graph. S2. Encode the spatiotemporal attributes in the knowledge graph to obtain temporal feature vectors and spatial feature vectors; S3. Based on the temporal feature vector and spatial feature vector, construct the attention weights of the entity and its neighboring entities, and aggregate the neighbor information of the entity based on the attention weights to obtain the entity's neighborhood context vector. S4. The spatial feature vector, the current time feature vector, and the neighborhood context vector of the current entity are fused with the feature vector of the entity at the previous time step to obtain the feature vector of the current entity. S5. Pass the feature vector of the current entity through a fully connected output layer and map it to the desired value using the Sigmoid function. The geological hazard susceptibility index is obtained from the interval; S6. Predict geological disasters using the geological disaster susceptibility index.

2. The method for dynamic prediction of geological disasters based on spatiotemporal knowledge graphs according to claim 1, characterized in that, For periodic time attributes, sinusoidal positional encoding is used to obtain periodic time feature vectors; for continuously increasing time attributes, normalization is performed followed by a linear transformation through a fully connected layer to obtain continuous time feature vectors; the two time feature vectors are concatenated and input into a multilayer perceptron to generate the final time feature vector. .

3. The method for dynamic prediction of geological disasters based on spatiotemporal knowledge graphs according to claim 1, characterized in that, The spatial attributes of entity coordinates are discretized into strings using Geohash encoding, and then mapped to spatial embedding vectors using a trainable embedding layer. A topological embedding vector is obtained by pre-training the spatial properties of the entity topology using the DeepWalk random walk algorithm. ;Will and The vectors are concatenated and then processed by an MLP to generate the final spatial feature vector. .

4. The method for dynamic prediction of geological disasters based on spatiotemporal knowledge graphs according to claim 1, characterized in that, S3 specifically refers to: S31. Define the messages passed between an entity and its neighboring entities through relationships as follows: in, This represents the message passed between the i-th entity and its j-th neighbor entity through relation r. Represents time t-1 eigenvectors, It is the j-th neighbor entity of the i-th entity. It is for relationships Defined learnable weight matrix, It is for relationships The bias term; S32. Calculate the relationship-based attention coefficient between an entity and its neighboring entities, using the following formula: in, Let denot , where is the attention coefficient between the i-th entity and its j-th neighbor entity at time t based on relation r, and ... and This represents the transformation matrix of the query and the key. This represents the feature vector of the i-th entity at time t-1. This represents the gating coefficient between the i-th entity and its j-th neighbor entity at time t, based on relation r. It is the Sigmoid activation function. It is a learnable weight matrix. It is a time decay function. It is a spatial Gaussian kernel function. It is a learnable decay rate with respect to r. and These are the i-th and j-th time feature vectors. and These are the i-th and j-th spatial feature vectors. It is a learnable spatial influence radius with respect to r; S33. Obtain attention weights based on the attention coefficients of an entity and its neighboring entities according to their relationships. The formula is as follows: in, This represents the attention weights of the i-th entity and its j-th neighbor entity at time t based on relation r. Let i represent the set of neighboring entities of the i-th entity. This represents the set of relationships between the i-th entity and its k-th neighboring entities. Let represent the attention coefficient between the i-th entity and its k-th neighbor entity at time t based on relation r'; S34. Based on attention weights, aggregate neighbor information to obtain the entity's neighborhood context vector, using the following formula: in, This represents the neighborhood context vector of the i-th entity at time t. It is the transformation matrix of values.

5. The method for dynamic prediction of geological disasters based on spatiotemporal knowledge graphs according to claim 1, characterized in that, The feature vector of the current entity is expressed by the formula: in, It is the feature vector of the i-th entity at time t. Indicates a gated loop unit. This represents the feature vector of the i-th entity at time t-1. This represents the neighborhood context vector of the i-th entity at time t. This represents the time feature vector of the i-th entity at time t. This represents the spatial feature vector of the i-th entity.

6. The method for dynamic prediction of geological disasters based on spatiotemporal knowledge graphs according to claim 1, characterized in that, The geological hazard susceptibility index is expressed by the formula: in, express The geological disaster susceptibility index at any time It is the Sigmoid activation function. and These are the weight and bias coefficients of the output layer. It is the eigenvector of the entity at time t.

7. The method for dynamic prediction of geological disasters based on spatiotemporal knowledge graphs according to claim 1, characterized in that, According to steps S2-S5, a dynamic prediction model for geological disasters based on spatiotemporal knowledge graphs is constructed. The total loss of the model training includes graph structure loss, temporal smoothing constraint loss, and spatial proximity constraint loss. Graph structure loss is used to constrain the gap between the embedded vectors learned by the model and the graph facts; temporal smoothing constraint loss is used to constrain the gap between adjacent times before and after entity embedding; spatial proximity constraint loss is used to constrain the embedded vectors learned by the model to follow geographic spatial autocorrelation.

8. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, it implements the method as described in any one of claims 1-7.

9. An electronic device, characterized in that, The device includes a processor and a memory interconnected thereto, wherein the memory is used to store a computer program, the computer program including computer-readable instructions, and the processor is configured to invoke the computer-readable instructions to perform the method as described in any one of claims 1-7.

10. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method according to any one of claims 1-7.

Images