A Multidimensional Feature Spatiotemporal Evolution Prediction Method Based on Structure-Aware Partitioning

Abstract

This invention relates to a multidimensional feature-based spatiotemporal evolution prediction method based on structure-aware partitioning, belonging to the field of spatiotemporal data mining and prediction technology. The method includes: acquiring a historical time-series information matrix and a spatial topology matrix based on historical spatiotemporal sequence data; performing topologically-aware segmentation of spatial nodes based on the spatial topology matrix, dividing them into local subgraphs with a constrained upper limit on the number of nodes, and completing them into a unified-dimensional block sequence; constructing an intra-block attention module with logarithmically smoothed adjacency bias to capture spatial dependencies between nodes within the local subgraphs; treating the local subgraphs as supernodes and constructing an inter-block attention module to capture long-distance global spatiotemporal dependencies across subgraphs; fusing local and global features, and outputting the spatiotemporal sequence prediction results for future time steps through a prediction layer. This invention balances local correlation and global dependency, reduces the computational complexity of large-scale spatiotemporal networks, and improves prediction accuracy and reliability, making it suitable for scenarios such as traffic flow and meteorological evolution.

Description

Technical Field

[0001] This invention relates to a multidimensional feature spatiotemporal evolution prediction method based on structure-aware partitioning, which is applicable to scenarios that require capturing spatiotemporal dynamic correlations, such as traffic flow prediction, meteorological evolution simulation, and epidemic transmission trend prediction. It belongs to the field of spatiotemporal data mining and prediction technology. Background Technology

[0002] In fields such as intelligent transportation, weather early warning, and public health, accurate prediction of multidimensional spatiotemporal sequences is a core technology supporting scientific decision-making, optimal resource allocation, and early warning of abnormal conditions. With the widespread deployment of sensing devices, spatiotemporal data exhibits characteristics of massive scale, complex dimensions, and strong spatiotemporal correlations, gradually revealing the significant limitations of traditional prediction methods. First, traditional spatiotemporal forecasting methods (such as ARIMA time series analysis and shallow machine learning models like support vector machines) treat spatiotemporal data as a collection of independent observation points, neglecting the inherent topological relationships between nodes in geographic space. For example, traffic flow on adjacent road segments in urban road networks exhibits strong coupling, and meteorological monitoring station data forms spatial correlations due to atmospheric diffusion. However, traditional methods cannot effectively model these dependencies, resulting in a lack of spatial continuity in the forecast results and an accuracy that fails to meet practical needs.

[0003] Secondly, existing deep learning models suffer from the problem of not being able to simultaneously address local and global dependencies: standard convolutional neural networks (CNNs) are limited by the size of their receptive field and can only capture local spatial features, failing to model long-distance global spatiotemporal dependencies across regions; while graph neural networks (GNNs) can explicitly express topological relationships, when directly processing large-scale spatiotemporal networks, the surge in the number of nodes leads to an exponential increase in computational complexity and excessive memory consumption, failing to meet the timeliness requirements of real-time prediction scenarios.

[0004] Furthermore, spatiotemporal data typically contains multidimensional features such as temporal dynamics (e.g., peak-hour traffic patterns), spatial heterogeneity (e.g., differences in meteorological sensitivity across different regions), and event-driven characteristics (e.g., the impact of sudden accidents on traffic). Existing methods often process single-dimensional features in isolation, lacking a unified feature representation mechanism, leading to insufficient fusion of spatiotemporal features and further limiting the improvement of prediction accuracy.

[0005] Therefore, developing a spatiotemporal sequence prediction method that can take into account topological correlation, local and global dependencies, low computational complexity, and multidimensional feature fusion has become an urgent technical problem to be solved in this field. Summary of the Invention

[0006] This invention provides a multidimensional feature spatiotemporal evolution prediction method based on structure-aware partitioning. It effectively captures spatiotemporal correlations through dynamic block division and attention mechanism to achieve accurate prediction and solves three core problems existing in spatiotemporal sequence prediction technology: 1) Ignoring spatial topological correlations, resulting in poor spatial continuity of prediction results; 2) Difficulty in balancing local feature extraction and long-distance global dependency modeling, and high computational complexity of large-scale networks; 3) Insufficient fusion of multidimensional features and lack of a unified feature representation mechanism, which limits prediction accuracy.

[0007] A method for predicting the spatiotemporal evolution of multidimensional features based on structure-aware partitioning includes the following steps: Step 1: Obtain the historical time series information matrix based on historical spatiotemporal information data. and spatial topological matrix Where T represents the time step and N represents the number of spatial nodes. Representing the intersection set of spatial relationships Let E be the set of edges. It is an adjacency matrix; Step 2: Based on the spatial topology matrix The spatial nodes are topologically segmented into several local subgraphs with a constrained upper limit on the number of nodes, and each subgraph is completed to form a block sequence with a unified dimension. Step 3: Construct graph bias The intra-block attention module extracts features from nodes within each local subgraph. As a smoothing constant, it captures the local spatial dependencies between nodes. ,in , These are the query matrix and the key matrix, respectively. For the attention head dimension, For learnable parameters, It is a mask matrix; Step 4: Construct an inter-block attention module, treating each local subgraph as a whole node, calculating the correlation between subgraphs, and capturing long-distance global spatiotemporal dependencies in space. , where Z is a learnable bias term; Step 5: Fuse local features With global features The prediction layer outputs time series prediction results for future time steps.

[0008] The topology-aware segmentation in step two specifically includes: Step 2.1: Calculate the shortest path matrix for all node pairs using the multi-source Dijkstra algorithm. ,in Represents a node arrive Shortest path length:

[0009] In the formula For all from arrive The set of paths For the edge Length; Step 2.2, Initialization seed nodes Perform the following operations through iterative optimization: Node allocation: For each node Distance according to the map Sort the candidate centers in ascending order and assign them to the first one that satisfies the condition. subgraph , The maximum number of nodes is preset. Center Update: Subgraph The geometric center is projected onto the nearest actual sensor node: In the formula For nodes The coordinates; Connectivity repair: If subgraph If disconnected components are present, retain the included components. The principal component is used to redistribute other components to the nearest neighbor feasible subgraph; Step 2.3: For each subgraph ,like By adding virtual nodes and padding with zero values, the length is increased to a specified value. The sequence forms a block sequence. ,in Complete the segmentation.

[0010] The in-block attention module for constructing the graph bias in step three specifically includes: for each block sequence Design a self-attention mechanism with topological bias: calculate the base attention score. Its expression is To incorporate topological information, a log-smoothed adjacency bias matrix is ​​introduced. Matrix elements According to node pairs The connection status is determined as follows: If The actual set of edges in the graph ,but ,in For the edge The weight, Let be the smoothing constant; if If the corresponding virtual node is not connected, then The basic attention score, adjacency bias matrix, and mask matrix are used to calculate the basic attention score, adjacency bias matrix, and mask matrix. Combined to form the final attention score ,Right now ,in These are learnable parameters.

[0011] The construction of the inter-block attention module in step four specifically includes: Will A block sequence Treating each supernode as a supernode to compute inter-block attention, first, for each supernode... Through mean pooling operation Obtain its global representation; then calculate the inter-block attention score, let Then attention score Then, normalization is performed using the softmax function, i.e. ,in This aggregates information from all supernodes, and finally outputs the future through linear projection of the output layer. The prediction results for each step of the time step.

[0012] The prediction layer in step five includes at least one fully connected layer and a nonlinear activation function, which is used to learn the mapping relationship between fused features and future spatiotemporal sequences.

[0013] Compared with existing technologies, the present invention, through the innovative combination of topology-aware block division and dual attention mechanism, has the following advantages: 1. Topology-aware partitioning achieves a balance between relevance and efficiency: A dynamic partitioning method based on spatial topology ensures strong topological relevance between nodes within each local subgraph, while simultaneously reducing computational complexity by breaking down large-scale networks into smaller subgraphs. O(N 2 ) Down to This significantly reduces computing power consumption and meets the needs of real-time prediction.

[0014] 2. Improving the accuracy of local dependency capture through attention within graph bias blocks: By introducing a log-smooth adjacency bias based on topology, the attention mechanism fully incorporates the actual connection relationships of nodes when extracting local features. At the same time, it uses a mask matrix to shield virtual node interference, which solves the problem of traditional attention mechanisms ignoring spatial topology and improves the accuracy of local feature extraction.

[0015] 3. Inter-block attention breaks through the limitations of global dependency modeling: The design of treating local subgraphs as super nodes enables the model to model long-distance global spatiotemporal dependencies across subgraph boundaries, making up for the shortcomings of CNN's limited receptive field and the complexity of global computation in traditional GNN. The ability to capture long-distance dependencies is significantly better than existing models.

[0016] 4. Multidimensional feature fusion improves prediction accuracy: By deeply fusing local and global features, the temporal dynamics, spatial heterogeneity, and event-driven nature of spatiotemporal data are fully utilized to construct a unified feature representation mechanism. Compared with existing methods, the prediction error is reduced and the reliability and stability of the prediction results are significantly improved.

[0017] 5. The method of this invention does not require complex hardware deployment and can be directly applied to practical scenarios such as traffic flow prediction, meteorological evolution simulation, and epidemic transmission trend prediction. It has wide applicability and extremely high engineering application value. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is an end-to-end network structure diagram of the multidimensional prediction model of the multidimensional feature spatiotemporal evolution prediction method based on structure-aware partitioning of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Reference Figure 1 A method for predicting the spatiotemporal evolution of multidimensional features based on structure-aware partitioning includes the following steps: Step 1: Obtain the historical time series information matrix based on historical spatiotemporal information data. and spatial topological matrix Where T represents the number of time steps and N represents the number of spatial nodes. Representing the intersection set of spatial relationships Let E be the set of edges. It is an adjacency matrix; The space matrix is ​​represented as . Represents a directed graph, where Represents a set of spatial relationships that intersect; furthermore, This represents the number of nodes. It is the set of edges of all spatial nodes. The time series information matrix is ​​represented as... ,in Indicates the number of time steps. Assume... The number of dimensions representing the features is then integrated into the spatiotemporal information matrix. .

[0022] Understandably, the spatiotemporal sequence prediction problem is defined as learning a mapping function. , using the past Based on the historical spatiotemporal sequence and spatial structure of steps, predict the future. The spacetime step of a step is represented by the formula:

[0023] Step 2: Based on the spatial topology matrix The spatial nodes are topologically segmented into several local subgraphs with a constrained upper limit on the number of nodes, and each subgraph is completed to form a block sequence with a unified dimension. Specifically, segmenting subgraphs is used in large-scale spatiotemporal data prediction tasks, especially for urban road network traffic flow prediction scenarios and large-scale weather forecasting scenarios, in order to effectively balance computational efficiency and prediction accuracy, and accurately capture the spatiotemporal correlations in the data.

[0024] For subsequent segmentation calculations, a shortest path distance matrix between nodes needs to be constructed. The multi-source Dijkstra algorithm is used to calculate the shortest path distance to each node, starting from all nodes, resulting in a matrix. ,in Represents a node To the node The shortest path length. Then, from the set of nodes... Random selection _ distinct nodes are used as the initial partition centers, denoted as _ ... The superscript (0) indicates the initial iteration number. Next, capacity-constrained node allocation is performed, aiming to assign all nodes to... Within each partitioned region, the number of nodes in each partitioned region must be ensured to not exceed a preset capacity limit. The specific process is as follows: First, create... An empty set Used to store the nodes within each segmented region and initialize the node label array. ,in Represents a node The label of the segmented region. Simultaneously, initialize the current node count array for each segmented region. Then for each node compute nodes To all split centers shortest path distance ,in This represents the current iteration number. The segmentation centers are sorted in ascending order of distance to obtain the index sequence. For each split center after sorting. ,if Then the node Assigned to segmented regions ,Right now Then, exit the current loop. Finally, when processing unassigned nodes, if a node exists... Make If these unassigned nodes are not assigned to any partitioned region, then these unassigned nodes will be counted according to the number of nodes in each partitioned region. Nodes are allocated sequentially from smallest to largest capacity to the partitions with the largest remaining capacity. After one round of node allocation, the position of the partition center needs to be updated to ensure a more rational allocation of partitions. For each partition... Calculate its geometric center (in continuous space) or a more suitable representative node (in a discrete graph structure) as the new partition center. In a discrete graph structure, the segmentation regions are calculated. The new split center can be determined by a centrality metric (such as degree centrality or proximity centrality) of all nodes within the region, or simply by selecting the node whose sum of distances to all nodes within the region is minimized. Specifically, for the split region... New split center It can be calculated using the following formula:

[0025] in, Represents a node and The shortest path distance between them.

[0026] Then, connectivity repair and load balancing are performed. During node allocation and partition center updates, situations may arise where partitioned regions are not connected or the number of nodes in each partitioned region is unbalanced. To solve this problem, for each partitioned region... Check if its induced subgraph is connected. If not, retain the subgraph containing the segmentation center. The connected components are identified, and the other connected components are reassigned to their nearest feasible partition center (the partition center has a current node count of less than 1). If all feasible partition centers have no remaining capacity, move the boundary node from the partition region with the highest node count among adjacent partition regions to the current partition region, while ensuring that the number of nodes in each partition region still meets the capacity constraint after the move. .

[0027] Repeat the above cycle of node allocation, split center update, connectivity repair, and load balancing until a stopping condition is met. The stopping condition is that the maximum split center displacement is less than a preset threshold. Or reach the preset maximum number of iterations During the loop, the distance between each segmentation center is cached to reduce redundant calculations.

[0028] The Voronoi graph segmentation method with topology-aware capacity constraints described above divides the spatial structure into multiple connected, size-constrained regions. Nodes within each region exhibit high topological correlation, providing a more reasonable data preprocessing foundation for subsequent spatiotemporal prediction models. This method fully considers the topological characteristics of spatial information, avoiding computational redundancy and information loss issues that may arise from traditional block-based methods, effectively improving the accuracy and computational efficiency of spatiotemporal prediction.

[0029] Step 3: Construct graph bias The intra-block attention module extracts features from nodes within each local subgraph. As a smoothing constant, it captures the local spatial dependencies between nodes. ,in , These are the query matrix and the key matrix, respectively. For the attention head dimension, For learnable parameters, It is a mask matrix; Specifically, after completing the Voronoi graph segmentation with topologically aware capacity constraints, the spatial structure is divided into: A limited-size and connected local subgraph Each subgraph contains no more than a preset capacity limit number of nodes. To further capture the local spatial dependencies between nodes within each local subgraph, this step constructs a graph bias-enhanced intra-block attention module, as detailed below: First, construct the similarity matrix and graph bias. For each local subgraph... First, construct their similarity matrix. This similarity matrix plays a crucial role in describing the strength or similarity of connections between nodes within a subgraph. For each element in the similarity matrix... There are specific definition rules. When a node or node When the virtual nodes are not filled (the main function of these virtual nodes is to ensure the consistency of matrix dimensions), then... If node or node All are valid nodes. Then based on the node or node The actual connectivity relationships in the spatial network are determined by comprehensively considering factors such as possible spatiotemporal sequences and distances, and must meet certain requirements. This is to maintain the self-connecting property.

[0030] After constructing the similarity matrix, a graph bias is introduced to further enhance the model's ability to perceive the real topology of spatial networks. The calculation formula is as follows: ,here It is a very small positive number, and the main purpose of setting this positive number is to avoid numerical problems during logarithmic operations. This graph bias has a unique function: it can amplify actual edges while suppressing spurious or irrelevant edges. In this way, the attention mechanism can more accurately focus on truly relevant pairs of nodes in the topological relationship.

[0031] In addition, a padding mask matrix needs to be defined to properly handle the padded virtual nodes. The elements in the filling mask matrix It has clear rules for value selection: Node or node For virtual nodes that are filled, then If node and nodes If all nodes are valid, then This mask matrix plays a crucial role in the subsequent attention calculation process. It ensures that the filled virtual nodes do not interfere with the attention allocation of valid nodes, thereby guaranteeing the accuracy and effectiveness of the attention calculation.

[0032] In the completed graph offset With the filling mask matrix After construction, it is introduced into the attention computation process to capture the local spatial dependencies between nodes within a local subgraph. For each local subgraph... First, obtain the query matrix through linear transformation. Bond matrix Next, calculate the in-block attention score. Its formula is .

[0033] In the formula It is the traditional scaled dot product attention part. For the attention head dimension, the scaling factor can prevent gradient vanishing. As a learnable nonnegative scalar, it can adjust graph bias. Contribution in the calculation. The subgraph similarity matrix reflects the true topology of the spatial network, allowing the attention mechanism to comprehensively consider node feature similarity and actual connectivity when allocating weights. For example, if two nodes are actually connected in a road and have high traffic volume, This will make them receive more attention in attention score calculations. Used to handle virtual nodes; when a node is a virtual node, its corresponding value is... When calculating attention weights using the softmax function, the weights are made close to 0 to prevent virtual nodes from interfering with the attention allocation of valid nodes.

[0034] Obtain attention score Then, the softmax function is used to convert them into attention weights. The softmax operation is performed along the node dimension of the matrix. Because... The virtual nodes have an attention weight of 0, while the attention weights between valid nodes are allocated reasonably based on their scores. Finally, the value matrix is ​​weighted and summed according to the calculated attention weights to extract node features, thus successfully capturing the local spatial dependencies between nodes within the local subgraph.

[0035] Step 4: Construct an inter-block attention module, treating each local subgraph as a whole node, calculating the correlation between subgraphs, and capturing long-distance global spatiotemporal dependencies in space. , where Z is a learnable bias term; In this module's design, all the previously processed local subgraphs are treated as a single, unified node. The purpose of this is to analyze the relationships between different subgraphs from a more macroscopic perspective, thereby capturing long-distance global spatiotemporal dependencies in space.

[0036] Furthermore, to quantify the correlation between these subgraphs that are nodes in a whole, a specific calculation method is employed. Let's assume that for each attention head... There is a query matrix Bond matrix It is obtained by performing a linear transformation on the input features. The formula for calculating the correlation between subgraphs is: .

[0037] in, This is the classic scaling dot product attention calculation part. and These represent the query and key information of different subgraphs in a specific feature space, and their dot product operation... It measures the degree of feature similarity between different subgraphs. It is the dimension of the attention head, the scaling factor. The purpose is to prevent the dot product result from becoming too large, thus avoiding gradient vanishing or exploding problems in subsequent gradient calculations. The formula contains... It is a learnable bias term or an integrator of additional feature information. It contains some prior knowledge or some adjustment parameters learned by the model, which are used to further adjust the calculation results of the correlation between subgraphs, so that the model can more flexibly capture the complex global spatiotemporal dependencies between different subgraphs.

[0038] Step 5: Fuse local features With global features The prediction layer outputs time series prediction results for future time steps.

[0039] Specifically, after completing the intra-block attention module's capture of local spatial dependencies of nodes within a subgraph, and the inter-block attention module's capture of long-distance global spatiotemporal dependencies in space, we obtain local features. With global features These two features characterize the data from different perspectives: local features focus on the subtle relationships between nodes within a subgraph, while global features grasp the macroscopic relationships between subgraphs.

[0040] The fused features are then input into the prediction layer. The prediction layer typically consists of one or more fully connected layers containing non-linear activation functions. Based on the fused input features, the prediction layer learns the mapping relationship from current features to time series values ​​at future time steps. After training and optimization, the prediction layer can output time series predictions for future time steps, providing valuable references for practical applications such as traffic flow forecasting and weather forecasting. At this point, the entire model completes the complete process from data input to prediction output.

[0041] By integrating these components, the proposed model not only addresses the problem of exponential growth in computational complexity due to the increase in the number of nodes, but also enhances the model's ability to integrate and analyze multidimensional feature sets, thus setting a new standard for spatiotemporal prediction models in complex environments.

[0042] This invention employs a topology-aware segmentation method, which dynamically segments data based on the inherent topological structure of the spatiotemporal data. This fully considers the spatial and temporal correlation of the data, enabling each data block to more reasonably reflect the local spatiotemporal characteristics.

[0043] This invention employs both intra-block attention and inter-block attention mechanisms. Intra-block attention focuses on the internal structure of each data block, deeply exploring local spatiotemporal dependencies; inter-block attention, on the other hand, focuses on the relationships between different data blocks, effectively capturing global spatiotemporal correlations, and comprehensively grasping data characteristics from local to global perspectives.

[0044] The present invention’s reasonable block-segmentation strategy and feature fusion architecture optimize the data processing flow, avoid unnecessary computational redundancy, improve overall computational efficiency, and make the model more practical in real-world applications.

[0045] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings, but the present invention is not limited to the described embodiments. For those skilled in the art, various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and these variations still fall within the protection scope of the present invention.

Claims

1. A method for predicting the spatiotemporal evolution of multidimensional features based on structure-aware partitioning, characterized in that: Includes the following steps: Step 1: Obtain the historical time series information matrix based on historical spatiotemporal information data. and spatial topological matrix Where T represents the number of time steps and N represents the number of spatial nodes. Representing the intersection set of spatial relationships Let E be the set of edges. It is an adjacency matrix; Step 2: Based on the spatial topology matrix The spatial nodes are topologically segmented, dividing them into several local subgraphs with a constrained upper limit on the number of nodes, and each subgraph is completed to form a block sequence with a unified dimension. Step 3: Construct graph bias The intra-block attention module extracts features from nodes within each local subgraph. As a smoothing constant, it captures the local spatial dependencies between nodes. ,in , These are the query matrix and the key matrix, respectively. For the attention head dimension, For learnable parameters, It is a mask matrix; Step 4: Construct an inter-block attention module, treating each local subgraph as a whole node, calculating the correlation between subgraphs, and capturing long-distance global spatiotemporal dependencies in space. , where Z is a learnable bias term; Step 5: Fuse local features With global features The prediction layer outputs the spatiotemporal sequence prediction results for future time steps.

2. The method for predicting the spatiotemporal evolution of multidimensional features based on structure-aware partitioning according to claim 1, characterized in that: The topology-aware segmentation in step two specifically includes: Step 2.1: Calculate the shortest path matrix for all node pairs using the multi-source Dijkstra algorithm. ,in Represents a node arrive Shortest path length: ， In the formula For all from arrive The set of paths For the edge Length; Step 2.2, Initialization seed nodes Perform the following operations through iterative optimization: Node allocation: For each node Distance according to the map Sort the candidate centers in ascending order and assign them to the first one that satisfies the condition. subgraph , The maximum number of nodes is preset. Center Update: Subgraph The geometric center is projected onto the nearest actual sensor node: In the formula For nodes The coordinates; Connectivity repair: If subgraph If disconnected components are present, retain the included components. The principal component is used to redistribute other components to the nearest neighbor feasible subgraph; Step 2.3: For each subgraph ,like By adding virtual nodes and padding with zero values, the length is increased to a specified value. The sequence forms a block sequence. ,in Complete the segmentation.

3. The method for predicting the spatiotemporal evolution of multidimensional features based on structure-aware partitioning according to claim 1, characterized in that: The in-block attention module for constructing the graph bias in step three specifically includes: for each block sequence Design a self-attention mechanism with topological bias: calculate the base attention score. Its expression is To incorporate topological information, a log-smoothed adjacency bias matrix is ​​introduced. Matrix elements According to node pairs The connection status is determined as follows: If The actual set of edges in the graph ,but ,in For the edge The weight, Let be the smoothing constant; if If the corresponding virtual node is not connected, then The basic attention score, adjacency bias matrix, and mask matrix are used to calculate the basic attention score, adjacency bias matrix, and mask matrix. Combined, the final attention score is formed. ,Right now ,in These are learnable parameters.

4. The method for predicting the spatiotemporal evolution of multidimensional features based on structure-aware partitioning according to claim 1, characterized in that: The construction of the inter-block attention module in step four specifically includes: Will A block sequence Treating each supernode as a supernode to compute inter-block attention, first, for each supernode... Through mean pooling operation Obtain its global representation; then calculate the inter-block attention score, let Then attention score Then, normalization is performed using the softmax function, i.e. ,in This aggregates information from all supernodes, and finally outputs the future through linear projection of the output layer. The prediction results for each step of the time step.

5. The method for predicting the spatiotemporal evolution of multidimensional features based on structure-aware partitioning according to claim 1, characterized in that: The prediction layer in step five includes at least one fully connected layer and a nonlinear activation function, which is used to learn the mapping relationship between fused features and future spatiotemporal sequences.

Images