CN-115271172-B - High-order principal component decomposition-based space-time diagram convolution traffic speed prediction method

Abstract

The invention discloses a space-time diagram convolution traffic speed prediction method based on high-order principal component decomposition, which is characterized in that a space-associated adjacency matrix of nodes is constructed based on the distance between nodes of stations, the node attribute is from the average running speed of vehicles at the represented stations, in addition, the attribute of the node has dynamic time sequence, the whole traffic network is modeled as a space-time diagram structure, the space-time diagram is decomposed into a core tensor and a factor matrix in each dimension based on the high-order principal component decomposition, the expression form of the space-time diagram convolution network after the high-order principal component decomposition is deduced, and a space-time diagram convolution network model of unified modeling of space-time association is realized. The invention applies the high-order principal component decomposition algorithm to the space-time diagram convolution network for the first time, and provides a space-time diagram convolution model based on the high-order principal component truncation decomposition from the perspective of tensor space high-dimensional association modeling, which has the advantages of reducing data dimension, parallel calculation and noise suppression.

Inventors

• CUI ZHEN
• XU XURAN
• ZHANG TONG
• XU CHUNYAN

Assignees

• 南京理工大学

Dates

Publication Date

20260512

Application Date

20220624

Claims (2)

1. 1. The traffic speed prediction method based on the high-order principal component decomposition space-time diagram convolution is characterized by comprising the following steps of: step 1, acquiring speed data information of vehicles running through each station at a plurality of moments on a map, wherein each moment corresponds to one graph structure data, and according to the time sequence evolution characteristics of a traffic network, splicing to obtain a time-space graph characteristic tensor One time per minute, said Referred to as a time slice; Step 2, acquiring position information of all stations on a traffic map, and constructing a spatial relationship adjacency matrix according to spatial position relations among different positions on the map; step 3, according to the dynamic sequence of the traffic map in the time dimension, each node is connected with the past The time slices are connected by themselves, and a dynamic time sequence adjacency matrix of a time-space diagram is constructed; Step 4, constructing a graph structure from the spatial association in each time slice and the time sequence association of each node at different moments for the space-time graph feature tensor constructed in the step 1 based on the two types of the association matrixes constructed in the step 2 and the step 3, and constructing a multi-scale space-time graph convolution network model; step 5, applying a principal component decomposition algorithm of a high-order tensor to the space-time diagram characteristic tensor constructed in the step 1 to perform decomposition calculation to obtain a core tensor and a factor matrix corresponding to each dimension; Step 6, carrying out convolution operation on the core tensor obtained by decomposition in the step 5 and the factor matrix corresponding to each dimension by utilizing the multi-scale space-time diagram convolution network model established in the step 4; Step 7, performing linear regression on the node embedding learned in the step 6 to obtain the predicted speed value, Said step 4 comprises the steps of, Step 4.1, introducing the space adjacency matrix constructed in the step2 For the space-time diagram characteristic tensor constructed in the step 1 Performing spatial domain correlation information transfer for the first Feature matrix of airspace map of each time slice The expression of the spatial correlation filtering is as follows: , Step 4.2, introducing the dynamic time sequence adjacent matrix constructed in the step3 For the space-time diagram characteristic tensor constructed in the step 1 Performing timing related information transfer for the first Feature matrix of a timing diagram of individual nodes The expression of the time sequence correlation filtering is as follows: step 4.3, introducing two operations of multidimensional tensors, Definition of the first calculation tensor Matrix The tensor matrix n-mode product is defined as, , ; Defining a second calculation: The three-dimensional tensor batch multiplication is defined as, , ; The multiscale space-time diagram convolutional network is defined as, , Wherein, in the graph network, nodes A kind of electronic device Order neighbor representation node Through the process of Nodes reachable by steps, contiguous matrix And The power of the power includes And The fusion of the order-neighbor information, The higher-order principal component decomposition algorithm in the step 5 is as follows, , Wherein, the Representing the decomposed feature kernel tensor, , And Representing a factor matrix in three dimensions of space, features and time sequence after decomposition, wherein 、 And 。
2. 2. The method for predicting traffic speed based on high-order principal component decomposition space-time diagram convolution according to claim 1, wherein the specific operation of step 6 is to approximate the multi-scale space-time diagram convolution network defined in step 4 to be, , Since the tensor matrix n-mode product satisfies the exchange and combining laws, the above equation is equal to: In the above equation, the spatial correlation filtering, the temporal correlation filtering and the feature space mapping are exchanged for filtering operations on the corresponding dimension factor matrix after decomposition, denoted as, , , , It is further deduced that the expression of the space-time diagram convolutional network based on the higher-order principal component decomposition is, , Finally, for a specific node The expression of the space-time information aggregation of the space-time diagram convolution model is as follows: ; Upper visual expression node Is characterized by information aggregation of nodes that are spatially close on the map and information aggregation of past moments of the nodes.

Description

High-order principal component decomposition-based space-time diagram convolution traffic speed prediction method Technical Field The invention relates to the technical field of artificial intelligence and the technical field of traffic planning and management intersection, in particular to a convolution traffic speed prediction method based on a high-order principal component decomposition space-time diagram. Background Traffic speed prediction is an important item of data information in intelligent traffic systems, and the main purpose of the traffic speed prediction is to predict the average speed of vehicles running at various traffic stations (routes). The future speed information predicted by the technology reflects the traffic condition of the traffic station, can be used for travel time prediction and path planning of a navigation system, can realize intelligent control based on the information, and is important data of an intelligent traffic system. Early, researchers predicted speed by statistical prediction models (differential integrated moving average autoregressive model ARIMA) and traditional machine learning methods (support vector autoregressive model SVR), but the methods only considered timing correlations and only extracted linear shallow data correlation features. Later, the deep learning model shows strong deep association characteristic mining capability in the artificial intelligence field, and a series of deep learning models are applied to a traffic speed prediction model and mainly comprise a cyclic neural network (RNN) and an extended model long and short memory network (LSTM) and a gating circulation unit (GRU). However, the recurrent neural network can only be used for modeling time-series correlation characteristics, and in a traffic network, congestion of surrounding sites is gradually caused after one site is blocked, and the time-series correlation characteristics are called space correlation characteristics. Because the distribution of stations within a traffic network is irregular, convolutional networks do not model such spatially-correlated features well. Later, the proposal of graph roll-up neural networks (GCNs) provided new solutions for modeling such irregular spatial correlation features contained within traffic networks. Then STGCN has been proposed to capture the spatial correlation features and one-dimensional convolution through the GCN to capture the timing correlation feature correlations, and later the T-GCN model has been proposed to combine the GRU and GCN to capture the timing correlation features and the spatial correlation features, respectively, within the traffic data. However, the two-stage scheme separately considers the space-time correlation, because modeling of high-dimensional data faces the problems of huge memory occupation and high computational complexity. In addition, the traffic data collected by the sensor is easy to be interfered by external factors to generate noise, and serious noise interference is generated in the training of the model to influence the stability of the model. Therefore, we make improvements to this and propose a convolution traffic speed prediction method based on a high-order principal component decomposition space-time diagram. Disclosure of Invention In order to solve the technical problems, the invention provides the following technical scheme: the invention relates to a method for predicting the convolution traffic speed based on a high-order principal component decomposition space-time diagram, which comprises the following steps, Step 1, acquiring speed data information of vehicles running through the position at a plurality of moments on a map, wherein each moment corresponds to one graph structure data, and according to the time sequence evolution characteristics of a traffic network, the time-space graph characteristic tensors are obtained by splicing; Step 2, acquiring position information of all stations on a traffic map, and constructing a spatial relationship adjacency matrix according to spatial position relations among different positions on the map; Step 3, connecting each node with the past T time slices according to the dynamic sequence of the traffic map in the time dimension, and constructing a dynamic time sequence adjacency matrix of the space-time diagram; step 4, based on the two types of incidence matrixes constructed in the step 2 and the step 3, establishing a multi-scale space-time diagram convolution network model from the angle that the space incidence in each time slice is a diagram structure and the time sequence incidence of each node at different moments is a diagram structure for the space-time diagram feature tensor constructed in the step 1; step 5, applying a principal component decomposition algorithm of a high-order tensor to the space-time diagram characteristic tensor constructed in the step 1 to perform decomposition calculation to obtain a core tensor and a factor matrix c