Multi-dimensional space-time prediction method based on dual-scale space-time collaboration

Abstract

The invention discloses a multi-dimensional space-time prediction method based on dual-scale space-time collaboration. The method comprises the following steps: acquiring historical space-time sequence data and constructing a space relation matrix; performing dual-scale feature reconstruction on the spatial relationship, and respectively extracting global spatial structure features and local spatial interaction features; decomposing the space-time sequence in a time dimension, constructing short-term change features and long-term evolution features, and matching time features with corresponding spatial scale features through a spatial scale adaptive decision-making mechanism; further, bidirectional interaction modeling of time features and space features is realized through a layered spatio-temporal feature collaborative fusion mechanism; and outputting prediction results of a plurality of time steps in the future based on the fused spatial-temporal features. The method reduces the calculation complexity while guaranteeing the prediction precision, has good universality and expansibility, and can be widely applied to various space-time prediction scenes of traffic, meteorology, public health, energy and environment monitoring.

Description

Technical Field

[0001] This invention relates to the field of empty sequence prediction technology, and specifically to a multidimensional spatiotemporal prediction method based on dual-scale spatiotemporal coordination. Background Technology

[0002] Spatiotemporal series forecasting technology is playing an increasingly important role in dynamic system state analysis and decision support. It can effectively predict system state changes, helping to optimize resource allocation, improve operational efficiency, and enhance risk response capabilities and system sustainability. Traditional statistical models are only suitable for basic time series forecasting scenarios, while modern time series forecasting models perform well when handling univariate or low-dimensional series. However, when faced with high-dimensional, multivariate spatiotemporal data, their linear assumptions and inherent limitations in independently handling spatiotemporal structure significantly restrict their predictive performance. Spatiotemporal graphical neural networks, by jointly modeling spatiotemporal correlations, provide a new technical path for spatiotemporal series forecasting.

[0003] Existing research mainly constructs joint representations by combining graph convolutional networks with sequence models, but the following problems still exist: First, the multi-scale structural dependency capture efficiency under sparse spatiotemporal data matrices is low, which restricts the model's scalability; second, spatiotemporal features are often decoupled, making it difficult to fully integrate structural information into the temporal evolution coding process, thus limiting the model's representation depth. In terms of structural dependency modeling, traditional graph convolutional methods often face computational efficiency bottlenecks. On the one hand, the sparsity of the data matrix leads to incomplete capture of inter-node relationships, affecting the accuracy of topological representation; on the other hand, the high computational complexity introduced by global attention mechanisms or dynamic graph modeling restricts the model's practicality in large-scale network structures. Existing improvement schemes often struggle to balance modeling accuracy and computational efficiency. At the spatiotemporal feature processing level, existing methods mostly adopt spatiotemporal serial or spatiotemporal parallel architectures. The former extracts spatiotemporal features sequentially in stages, which can independently enhance information in each dimension, but fails to fully stimulate the interaction and complementarity between spatiotemporal features; the latter, while attempting joint modeling, relies heavily on static fusion mechanisms, making it difficult to adapt to the dynamic correlation characteristics between system states, thus limiting the comprehensive capture of complex spatiotemporal dependencies. Summary of the Invention

[0004] The purpose of this invention is to provide a multidimensional spatiotemporal prediction method based on dual-scale spatiotemporal coordination to solve the problems mentioned in the background art.

[0005] The present invention provides the following technical solution: a multidimensional spatiotemporal prediction method based on dual-scale spatiotemporal coordination, comprising the following operational steps:

[0006] Step S1: Obtain historical spatiotemporal sequence data and construct a spatiotemporal information matrix and a spatial relationship matrix.

[0007] Preferably, historical spatiotemporal sequence data is represented as a tensor. Where T is the number of historical time steps, N is the number of spatial nodes, and C is the feature dimension; the spatial relationship matrix is ​​a static adjacency matrix. Its elements Represents a node With nodes The connection relationships or distance weights between them.

[0008] Step S2: Introduce a dual-scale spatial feature reconstruction mechanism to reconstruct the spatial relationship on two scales, and then extract the global spatial structure features and local spatial interaction features respectively.

[0009] Preferably, a dual-channel static topology encoder is used to perform dual-scale feature reconstruction; the dual-channel static topology encoder takes the static spatial relation matrix as its only input and processes the static adjacency matrix. Perform initial feature mapping, and then perform the embedding representation after initial feature mapping. Input two independent multilayer perceptron channels: ,in, For global spatial structure features, For local spatial interaction features, For global spatial feature channels, For local spatial feature channels, and They have the same structure but do not share parameters.

[0010] Step S3: Decompose the historical spatiotemporal sequence in the time dimension to construct short-term change features and long-term evolution features, and match the time features with the corresponding spatial scale features through a spatial scale adaptive decision-making mechanism.

[0011] Preferably, the method for decomposing the spatiotemporal sequence in the time dimension is as follows: [The text abruptly ends here, likely due to an incomplete sentence or a formatting error.] Classified as short-term series With long-term sequences ,in, For short-term series time steps, For long-term series time steps, The number of spatial nodes, For feature dimension, and For short-term sequences Encoders based on multi-head self-attention extract short-term variation features And model the time coding structure for long-term sequences. Perform linear projection to align the feature dimensions.

[0012] Preferably, the spatial scale adaptive decision-making mechanism is implemented through the following steps: introducing a spatial scale adaptive decision-making mechanism to address the global spatial structure features respectively. and local spatial interaction features Spatial dimension compression is performed to obtain the global summary vector. and local summary vector ; Long-term evolutionary characteristics Time-dimensional compression is performed to obtain a time-series summary vector. The next step is to concatenate the three summary vectors and input them into a small multilayer perceptron to generate a two-dimensional decision vector. ,in, This represents vector concatenation. and These represent the weights of whether the model should emphasize global or local spatial features.

[0013] Preferably, the method for matching temporal features with spatial scale features is as follows: long-term evolutionary features are input into an optional feature enhancer, which contains at least one self-attention layer to mine complex dependency patterns related to spatial topology in the long-term sequence; during the model training phase, decision weights are used to perform a weighted summation of the features before and after enhancement to achieve soft fusion, resulting in the fused long-term evolutionary features. : During the model inference stage, comparison and The size, if If the feature size is larger, the long-term evolutionary features will be processed by a feature enhancer before being output; otherwise, the original long-term evolutionary features will be output directly. .in, It is an enhancement module consisting of a multi-head self-attention layer and a feedforward network.

[0014] Step S4: Through a hierarchical spatiotemporal feature collaborative fusion mechanism, perform bidirectional interactive modeling of temporal and spatial features.

[0015] Preferably, short-term change characteristics Long-term evolution characteristics after modulation Global spatial features With local spatial features The features are concatenated along the feature dimension to form preliminary fused features, which serve as input to the spatiotemporal feature collaborative fusion module. Represented as: Subsequently, the data is fed into a fusion network consisting of multiple residual-connected feedforward network blocks for deep integration. Simultaneously, the spatiotemporal feature collaborative fusion module has an explicit feedback path from temporal features to spatial features, projecting the integrated temporal feature information back into the spatial feature domain for iterative optimization of the spatial feature representation, thus initially fusing the features. Input a fusion network consisting of multiple residual MLP blocks. Deep interaction is performed, and a jump connection from temporal features to spatial features is introduced as a feedback path, ultimately resulting in an optimized joint spatiotemporal feature representation. ,in, This represents a linear projection layer.

[0016] Step S5: Output the prediction results for multiple future time steps based on the fused spatiotemporal features.

[0017] Preferably, the joint spatiotemporal features are represented Mapped to the future through a linear regression layer. Predicted sequence at each time step ,in, , and For learnable parameters, To predict the target dimension, the prediction result is further output as follows: ,in, This represents the regression mapping function composed of multi-level residual MLPs.

[0018] On the other hand, a multi-dimensional spatiotemporal prediction method based on dual-scale spatiotemporal collaboration is applicable to any spatiotemporal sequence prediction task in traffic flow prediction, meteorological index prediction, epidemic transmission prediction, or population density prediction, with spatial nodes corresponding to sensors, regional grids, traffic intersections, or monitoring stations.

[0019] Compared with the prior art, the beneficial effects achieved by the present invention are:

[0020] (1) To address the efficiency problem of multi-scale spatial modeling, this invention proposes a bilateral scale spatial topology reconstruction mechanism. By using a global-local dual-channel MLP architecture, the topological associations and dynamic interactions of neighboring nodes at the system-level or global scale are captured respectively, which effectively overcomes the problem of multi-scale dependency modeling under sparse matrices while ensuring computational efficiency.

[0021] (2) To address the problem that spatial information cannot be integrated into time coding in spatiotemporal feature processing, this invention proposes a spatiotemporal feature collaborative fusion mechanism, constructs a bidirectional interactive spatiotemporal coding framework, introduces spatial information marking in the time coding stage, and achieves deep fusion of spatiotemporal features through a hierarchical strategy. Attached Figure Description

[0022] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0023] Figure 1 This is a flowchart of the steps of the multidimensional spatiotemporal prediction method based on dual-scale spatiotemporal coordination provided in the embodiments of the present invention;

[0024] Figure 2 This is a schematic diagram of the end-to-end network structure of the multidimensional spatiotemporal prediction model provided in an embodiment of the present invention;

[0025] Figure 3 A schematic diagram of the end-to-end network structure of the bilateral scale spatial topology reconstruction module provided in an embodiment of the present invention;

[0026] Figure 4 This is a schematic diagram of the end-to-end network structure of the spatiotemporal feature collaborative fusion module provided in an embodiment of the present invention. Detailed Implementation

[0027] 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.

[0028] Combination Figures 1-4 As shown, the present invention provides the following technical solution: a multidimensional spatiotemporal prediction method based on dual-scale spatiotemporal coordination, comprising the following operational steps:

[0029] Step S1: Based on historical spatiotemporal sequence data, obtain the observation information of each spatial node in the transportation system in the continuous time dimension, and construct the spatiotemporal information matrix and spatial relationship matrix.

[0030] In this embodiment, the spatiotemporal system is abstracted as a directed or undirected graph: in, This represents a set of spatial nodes, where each node corresponds to a spatial unit with spatiotemporal observation attributes. Nodes can represent sensors, regional grids, monitoring stations, or functional units. The number of nodes is: , This represents the set of structural relationships between spatial nodes. Historical spatiotemporal sequence data is represented as a tensor. ,in For historical time steps, The number of spatial nodes, This is the feature dimension. Based on this, a spatial relationship matrix is ​​constructed: The spatial relation matrix is ​​a static adjacency matrix, and its elements are... Represents a node With nodes The connection relationships or distance weights between them are determined. A historical spatiotemporal sequence information matrix is ​​also constructed. ,in, Indicates the length of a historical time step. Represents the feature dimension of a node.

[0031] For example, the essence of the spatiotemporal prediction problem is to learn a mapping function that utilizes past data. Historical spatiotemporal information within a time step and the structural relationships between spatial nodes are used to predict the future. The system state at each time step is formally defined as: This step enables a unified modeling input for time-dependent information and spatial topological information, laying the foundation for subsequent multi-scale spatial feature extraction and spatiotemporal co-modeling.

[0032] Step S2: To address the issue that spatial dependencies in complex spatiotemporal systems exhibit significant differences at different scales, a dual-scale spatial feature reconstruction mechanism is introduced to perform multi-scale modeling of the spatial relationship matrix.

[0033] In this embodiment, the spatial relationship matrix is ​​embedded using a nonlinear mapping to map it to a high-dimensional feature space, thereby obtaining an intermediate representation of the spatial features. :

[0034]

[0035] in, It is an activation function. and These are the weight matrix and the bias term, respectively. To discard the mask, the next step, in order to enhance the spatial feature representation and avoid information degradation, is to introduce a feature reconstruction structure with residual connections to enhance intermediate spatial features and obtain the final spatial feature representation. :

[0036]

[0037] in, and These are the weight matrix and bias term of the fully connected layer, respectively. In terms of implementation, two sets of feature reconstruction channels with independent parameters are used to generate global and local spatial features, respectively. Wherein:

[0038] Global spatial features are used to characterize the overall structural dependencies of a system at a macroscopic scale, reflecting the propagation and coupling patterns of state information over a large spatial range;

[0039] Local spatial features are used to characterize the direct interaction relationships between adjacent nodes, reflecting the fine-grained association patterns of the system within local regions.

[0040] Understandably, different spatial scales exhibit different dependency patterns, making it difficult for single-scale spatial modeling methods to simultaneously consider both global structural information and local interaction information. This embodiment employs a dual-channel static topology encoder to perform dual-scale feature reconstruction; this encoder uses the static spatial relation matrix as its sole input and processes the adjacency matrix... Initial feature mapping is performed to obtain the initial spatial embedding representation. :

[0041]

[0042] in, , For learnable parameters, For the hidden layer dimension; subsequently, Input two independent multilayer perceptron channels:

[0043]

[0044] in, For global spatial structure features, For local spatial interaction features, For global spatial feature channels, For local spatial feature channels, and The structures are identical, but the parameters are not shared. Through a dual-scale spatial feature reconstruction mechanism, different spatial dependencies can be effectively expressed, improving the flexibility and accuracy of spatial feature modeling.

[0045] Step S3: To address the issue that time series exhibit different evolutionary patterns at different time scales, the historical spatiotemporal series is decomposed along the time dimension to construct short-term change characteristics and long-term evolutionary characteristics.

[0046] In this embodiment, the method for decomposing the spatiotemporal sequence in the time dimension is as follows: [The text abruptly ends here, likely due to an incomplete sentence or a formatting error.] Classified as short-term series With long-term sequences ,in, For short-term series time steps, For long-term series time steps, The number of spatial nodes, For feature dimension, and Short-term variation characteristics are used to characterize the rapid fluctuations in system state over short timescales, such as sudden disturbances, local anomalies, or short-period changes. This applies to short-term sequences. Encoders based on multi-head self-attention extract short-term variation features And the time coding structure is modeled, which are represented as follows:

[0047]

[0048] Long-term evolutionary features are used to characterize the stable changing trends of a system's state over a longer timescale, such as periodic evolution, long-term growth, or decay processes. To maintain the stability and interpretability of long-term trend features, we first analyze long-term sequences... Perform linear projection to align the feature dimensions:

[0049]

[0050]

[0051] in, For long-term temporal context representation, The feature dimension projection matrix of the long-term sequence. , These are learnable parameters.

[0052] For example, a spatial scale adaptive decision-making mechanism is introduced to address the global spatial structure features. and local spatial interaction features Spatial dimension compression is performed to obtain the global summary vector. and local summary vector ; Long-term evolutionary characteristics Time-dimensional compression is performed to obtain a time-series summary vector. The next step is to concatenate the three summary vectors and input them into a small multilayer perceptron to generate a two-dimensional decision vector.

[0053]

[0054] in, This represents vector concatenation. and These represent the weights that the model should emphasize for global or local spatial features, respectively. Then, based on the spatial feature distribution corresponding to the current sample, an adaptive determination is made to match long-term temporal features with global or local spatial scale features. Specifically, the long-term evolutionary features are input into an optional feature enhancer, which contains at least one self-attention layer to uncover complex dependency patterns related to spatial topology in the long-term sequence. During model training, decision weights are used to perform a weighted summation of the features before and after enhancement, achieving soft fusion to obtain the fused long-term evolutionary features. :

[0055]

[0056] In the model inference stage, comparison and The size, if If the feature size is larger, the long-term evolutionary features will be processed by a feature enhancer before being output; otherwise, the original long-term evolutionary features will be output directly. .in, It is an enhancement module consisting of a multi-head self-attention layer and a feedforward network. The single-layer calculation process is as follows:

[0057]

[0058] in, To enhance the intermediate hidden states within the module, It is a feedforward neural network. This is a layer normalization operation.

[0059] This invention introduces an enhanced temporal modeling structure when matching global spatial scales to capture long-term temporal dependencies across spatial units; when matching local spatial scales, it directly uses linearly aligned temporal features, thereby reducing computational complexity while ensuring modeling effectiveness. Through the aforementioned temporal decomposition and spatial scale adaptive matching mechanism, consistent modeling of temporal and spatial features at the semantic level is achieved, avoiding modeling biases caused by the simple superposition of features at different scales.

[0060] Step S4: Input the dual-scale spatial features obtained in step S2 and the multi-scale temporal features obtained in step S3 into the spatiotemporal feature fusion module to construct a bidirectional interactive modeling mechanism between the temporal and spatial dimensions.

[0061] In this embodiment, the input to the spatiotemporal feature collaborative fusion module includes short-term variation features. Long-term evolution characteristics after modulation Global spatial features With local spatial features By splicing along the feature dimensions, preliminary fused features are formed. :

[0062]

[0063] Input a fusion network consisting of multiple feedforward network blocks with residual connections. Deep interactive modeling is performed; simultaneously, the spatiotemporal feature collaborative fusion module has an explicit feedback path from temporal features to spatial features, which back-maps the integrated temporal feature information to the spatial feature domain through linear projection for iterative optimization of the spatial feature representation; through the combined action of multi-layer residual MLP blocks and skip connections, the optimized joint spatiotemporal feature representation is finally obtained. :

[0064]

[0065] in, This represents a linear projection layer.

[0066] During the fusion process, spatial features participate in temporal feature modeling as structural constraint information, enabling temporal features to perceive different spatial dependency patterns. Simultaneously, the dynamic change information contained in the temporal features influences the spatial feature modeling process, dynamically adjusting the weight distribution of different spatial features at different times. Through multi-level feature interaction and collaborative updates, deep coupling between temporal and spatial features is achieved, enhancing the model's ability to express complex spatiotemporal dependencies.

[0067] Step S5: Input the spatiotemporal fusion features obtained in step S4 into the prediction module, predict the system state for multiple future time steps through the regression mapping function, and output the prediction results.

[0068] In this embodiment, the joint spatiotemporal feature representation is used. Mapped to the future through a linear regression layer. Predicted sequence at each time step ,in, , and For learnable parameters, To predict the target dimension, the prediction result is further output as follows:

[0069]

[0070] in, This represents the regression mapping function composed of multi-level residual MLPs.

[0071] Based on a comprehensive consideration of the system's temporal evolution characteristics and spatial structural features, this invention enables multi-step prediction of future states. This method is applicable to various application scenarios with complex spatiotemporal dependencies, such as, but not limited to, traffic system prediction, meteorological element prediction, and public health situation evolution prediction.

[0072] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0073] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A multidimensional spatiotemporal prediction method based on dual-scale spatiotemporal coordination, characterized in that: The following steps are included: Step S1: Obtain historical spatiotemporal sequence data and construct a spatiotemporal information matrix and a spatial relationship matrix; Step S2: Introduce a dual-scale spatial feature reconstruction mechanism to reconstruct the spatial relationship using dual-scale features, and then extract the global spatial structure features and local spatial interaction features respectively; Step S3: Decompose the historical spatiotemporal sequence in the time dimension, construct short-term change features and long-term evolution features, and match the time features with the corresponding spatial scale features through a spatial scale adaptive decision-making mechanism; Step S4: Through a hierarchical spatiotemporal feature collaborative fusion mechanism, perform bidirectional interactive modeling of temporal and spatial features; Step S5: Output the prediction results for multiple future time steps based on the fused spatiotemporal features.

2. The multidimensional spatiotemporal prediction method based on dual-scale spatiotemporal coordination according to claim 1, characterized in that: In step S1, the historical spatiotemporal sequence data is represented as a tensor, and the spatial relation matrix is ​​a static adjacency matrix.

3. The multidimensional spatiotemporal prediction method based on dual-scale spatiotemporal coordination according to claim 2, characterized in that: Step S2, which involves dual-scale feature reconstruction of spatial relationships, includes: A dual-channel static topology encoder is used to perform dual-scale feature reconstruction. The dual-channel static topology encoder takes the static spatial relation matrix as the only input, performs initial feature mapping on the static adjacency matrix, and embeds the initial space into the multilayer perceptron channel that represents the two independent input parameters.

4. The multidimensional spatiotemporal prediction method based on dual-scale spatiotemporal coordination according to claim 3, characterized in that: Step S2 involves extracting global spatial structure features and local spatial interaction features, including: By embedding features into the spatial relationship matrix through nonlinear mapping, the matrix is ​​mapped to a high-dimensional feature space to obtain an intermediate representation of the spatial features. Furthermore, a feature reconstruction structure based on residual connections is introduced to enhance the intermediate spatial features, resulting in the final spatial feature representation. Simultaneously, two sets of feature reconstruction channels with independent parameters are used to generate global spatial features and local spatial features respectively.

5. A multidimensional spatiotemporal prediction method based on dual-scale spatiotemporal coordination according to claim 4, characterized in that: Step S3, which involves decomposing the historical spatiotemporal sequence along the time dimension, includes: Historical sequences are divided into short-term and long-term sequences; For short-term sequences, a multi-head self-attention encoder is used to extract short-term variation features and model the temporal coding structure. For long-term sequences, a linear projection is performed to align the feature dimensions.

6. The multidimensional spatiotemporal prediction method based on dual-scale spatiotemporal coordination according to claim 5, characterized in that: The operation of the spatial scale adaptive decision-making mechanism in step S3 includes: A spatial scale adaptive decision-making mechanism is introduced to compress the spatial dimension of global spatial structure features and local spatial interaction features respectively, so as to obtain global summary vector and local summary vector. The long-term evolutionary features are compressed in time dimension to obtain a time-series summary vector; The global summary vector, local summary vector, and temporal summary vector are concatenated and input into a small multilayer perceptron to generate a two-dimensional decision vector.

7. A multidimensional spatiotemporal prediction method based on dual-scale spatiotemporal coordination according to claim 6, characterized in that: Step S3, matching the temporal features with the corresponding spatial scale features, includes: Long-term evolutionary features are input into an optional feature enhancer, which contains at least one self-attention layer to mine complex dependency patterns related to spatial topology in long-term sequences. During the model training phase, decision weights are used to perform a weighted summation of the features before and after enhancement to achieve soft fusion and obtain the long-term evolutionary features after fusion. During the model inference phase, the weights of the model's emphasis on global spatial features and the model's emphasis on local spatial features are compared. When the model's emphasis on global spatial features is greater, the long-term evolutionary features are processed by the feature enhancer and then output; otherwise, the original long-term evolutionary features are output directly.

8. A multidimensional spatiotemporal prediction method based on dual-scale spatiotemporal coordination according to claim 7, characterized in that: The model training phase also includes an enhancement module consisting of a multi-head self-attention layer and a feedforward network.

9. A multidimensional spatiotemporal prediction method based on dual-scale spatiotemporal coordination according to claim 8, characterized in that: The implementation of the hierarchical spatiotemporal feature collaborative fusion mechanism in step S4 includes: Short-term change features, modulated long-term evolution features, global spatial features, and local spatial features are spliced ​​along the feature dimension to form preliminary fusion features, which serve as inputs to the spatiotemporal feature collaborative fusion module. A fusion network consisting of multiple residual-connected feedforward network blocks is used to perform deep interactive modeling of the preliminary fusion features; The spatiotemporal feature collaborative fusion module has an explicit feedback path from temporal features to spatial features, which back-projects the integrated temporal feature information to the spatial feature domain for iterative optimization of the spatial feature representation. An optimized joint spatiotemporal feature representation is obtained through the combined action of multi-layer residual MLP blocks and skip connections.

10. A multidimensional spatiotemporal prediction method based on dual-scale spatiotemporal coordination according to claim 9, characterized in that: The output prediction results in step S5 include: mapping the joint spatiotemporal feature representation through a linear regression layer to a prediction sequence for multiple future time steps, thus obtaining the final prediction results.

Images