CN-121982898-A - Traffic flow prediction method based on space-time multi-hypergraph neural network

CN121982898ACN 121982898 ACN121982898 ACN 121982898ACN-121982898-A

Abstract

The invention discloses a traffic flow prediction method based on a space-time multi-hypergraph neural network, which belongs to the technical field of intelligent traffic systems and specifically comprises the following steps of S1, selecting an area and a time range, acquiring structured data by a traffic monitoring system to construct a data set, dividing the data set into a test set and a training set, S2, modeling spatial dependence in traffic flow data by a spatial characteristic extraction module by adopting a multi-hypergraph neural network MHGNN by using three hypergraphs, fusing output of the three hypergraphs, generating a group of comprehensive spatial characteristics for each road section, S3, capturing the time dependence in the traffic flow data by adopting an LSTM neural network by a time characteristic extraction module, and S4, constructing a space-time multi-hypergraph neural network STMHGNN model to fuse the spatial characteristics. The traffic flow prediction method based on the space-time multi-hypergraph neural network is high in prediction accuracy and efficient in extracting space-time characteristics.

Inventors

LIU PENGJIE
SHAO FENG
SHAO HU
DENG JIADING
SONG ZHENGUO
LI GUOLIANG
LI BOHAN
Yao Hongluan

Assignees

中国矿业大学

Dates

Publication Date: 20260505
Application Date: 20260311

Claims (5)

1. A traffic flow prediction method based on a space-time multi-hypergraph neural network is characterized by comprising the following steps: S1, selecting an area and a time range, acquiring structural data by using a traffic monitoring system to construct a data set, and dividing the data set into a test set and a training set; s2, a spatial feature extraction module adopts a multi-hypergraph neural network MHGNN to model spatial dependence in traffic flow data by using an upstream-downstream adjacent hypergraph, a regional hypergraph and a history mode hypergraph, and the outputs of the three hypergraphs are fused to generate a group of comprehensive spatial features for each road section; S3, capturing time dependence in traffic flow data by a time feature extraction module through an LSTM neural network; And S4, constructing a space-time multi-hypergraph neural network STMHGNN model, splicing the outputs of the MHGNN and the LSTM neural network along the characteristic dimension, and fusing the space-time characteristics.
2. The traffic flow prediction method based on space-time multiple hypergraph neural network according to claim 1, wherein MHGNN uses upstream-downstream adjacent hypergraphs in S2 to construct characteristics based on traffic state propagation along communication paths and forming dynamic bidirectional dependency, each road section is connected with all upstream and downstream adjacent road sections through hyperedges, such adjacent spatial correlation is modeled, and the hyperedge number of the hypergraph is Incidence matrix of upstream-downstream adjacency hypergraph The behavior of the road is superb and listed as a road section; MHGNN capturing spatial dependencies of such regions by grouping segments belonging to the same geographic region into a hyperedge using a region hypergraph, the association matrix of the region hypergraph Is listed as a road segment; MHGNN identifying spatially related road segments showing similar traffic flow patterns by using a history pattern hypergraph, clustering history traffic observation data by adopting a Gaussian Mixture Model (GMM), and representing the traffic flow observation data as a traffic flow observation matrix Wherein Representing the total number of time periods, Representing the total number of road segments, traffic flow observation matrix Each column of (a) A time series of historical traffic flow observations corresponding to a particular road segment.
3. The traffic flow prediction method based on the space-time multiple hypergraph neural network of claim 2, wherein S2 is characterized in that Is one Dimension vector, representing the first A sequence of traffic observations of individual road segments at all points in time, wherein The specific steps of the road section flow clustering based on the Gaussian mixture model are as follows: S2.1, calculating a probability density function of the GMM The characteristics are divided into Class, each feature The probability density function of (2) is calculated as follows: ; Wherein, the Is the first The mixing coefficient of the Gaussian components meets the following conditions , Is the mean vector of the gaussian component, Is the covariance matrix of the gaussian component, Parameters updated for the GMM model; The calculation formula of the gaussian density function is as follows: ; S2.2, an expected step, calculating each sample Belonging to the first The posterior probability of each gaussian component is calculated as follows: ; Wherein, the Is the first The mixing coefficient of the Gaussian components meets the following conditions , Is the mean vector of the gaussian component, Covariance matrix of Gaussian component; s2.3, a maximizing step, namely updating parameters of the Gaussian mixture model according to the posterior probability obtained through calculation, wherein the parameters are specifically as follows: ; ; ; S2.4, each feature is distributed to a cluster corresponding to the Gaussian component with the maximum posterior probability to obtain a clustering result, and the clustering result is specifically shown as follows: ; Wherein, the Is the first The cluster index to which the individual samples are ultimately assigned.
4. The traffic flow prediction method based on the space-time multi-hypergraph neural network according to claim 3, wherein the time dependency in the traffic flow data is modeled by adopting a long-short-term memory network LSTM in the time feature extraction of S3, and the regulation and control of the information flow are realized by introducing a storage unit and three control gate mechanisms of an input gate, a forgetting gate and an output gate, wherein the specific calculation process is as follows: ; ; ; ; ; ; Wherein, the For forgetting the gate, the memory unit at the previous time is controlled Is used for the retention of the (a), For inputting the gate, determine how much new information is Will be added to the state of the cell, To output the gate, determine the current cell state To hidden state Is used for the degree of contribution of (a), Is a time step Is used for the input of the (c) to be processed, In order to be in the hidden state at the previous moment, In order to be the state of the cell at the previous moment, For a leachable weight corresponding to a forget gate, For the input gate corresponding learner weights, Is the leachable weight corresponding to the candidate memory cell, To output the corresponding learnable weights for the gates, Is a leachable weight forgetting the hidden state corresponding to the door, Is a learnable weight of the hidden state corresponding to the input gate, Is the leachable weight of the hidden state corresponding to the candidate memory unit, To output the learnable weight of the hidden state corresponding to the door, For forgetting the bias term corresponding to the gate, For the bias term corresponding to the input gate, Is the bias term corresponding to the candidate memory cell, For outputting the bias term corresponding to the gate Sum function Is that The model function and hyperbolic tangent activation function are calculated as follows: ; ; Wherein, the Is the input value for the activation function.
5. The traffic flow prediction method based on the space-time multiple hypergraph neural network of claim 4, wherein the fused space-time features in S4 are input into a fully connected neural network, the fully connected neural network maps the same prediction feature space to obtain a predicted value, errors between the predicted value and a real observed value are systematically minimized, a mean square error is adopted as a loss function, and a calculation formula is shown as follows: ; Wherein, the In order to be the number of road segments, In order to predict the length of the period, In order to predict the traffic flow rate, Is the real traffic flow.

Description

Traffic flow prediction method based on space-time multi-hypergraph neural network Technical Field The invention relates to the technical field of intelligent traffic systems, in particular to a traffic flow prediction method based on a space-time multi-hypergraph neural network. Background Traffic flow is a core index in modern traffic planning and management, which not only quantifies the road use intensity, but also reveals the distribution characteristics and dynamic evolution rules of traffic flow in time-space dimension. Accurate traffic flow prediction relies on big data analysis and machine learning technology, can provide scientific decision basis for traffic control, for example optimize road traffic efficiency through self-adaptive signal control, and then effectively alleviate traffic jam, reduce accident risk. In general, traffic flow prediction is a key technology for improving the efficiency of an urban traffic system, and has important value for promoting intelligent travel development. The traffic flow prediction method mainly comprises a statistical method, a machine learning technology and a deep learning method. Early researches have widely adopted statistical methods such as autoregressive integral moving averages and Markov chains, as they can briefly and effectively capture periodic patterns and time-varying features in traffic data. The method can effectively identify trending, seasonal and other time domain features, and is particularly suitable for modeling regular traffic flow modes. However, statistical methods have limitations in processing large-scale data sets in that their computational efficiency decreases as the data size increases and it is difficult to accurately characterize the complex nonlinear relationships inherent in large traffic networks. With the development of technology, the support vector machine, random forest, decision tree and other machine learning methods show stronger large-scale data adaptability and are more flexible in modeling nonlinear modes. Nevertheless, such methods still suffer from a significant level of space-time dependencies in capturing traffic data. In recent years, a deep learning method has made remarkable progress in the field of traffic flow prediction, and exhibits outstanding capabilities in capturing complex traffic patterns, extracting high-order features, and effectively modeling space-time dependencies. The method is good at processing large-scale traffic data, adapts to dynamically-changed traffic conditions, and continuously improves prediction accuracy by means of a deep network structure. Among the many deep learning approaches, graph neural networks are widely used for traffic prediction tasks, which typically build spatial correlations based on adjacency between sensors. However, while such approaches can take into account direct adjacencies between road segments, more complex higher-order interactions between traffic network nodes, such as regional-level functional relevance, and similarities among multiple road segments that appear in historical traffic patterns, tend to be ignored. Currently, research that explicitly characterizes and exploits such complex spatial relationships in models remains relatively lacking. Disclosure of Invention The invention aims to provide a traffic flow prediction method based on a space-time multi-hypergraph neural network, which is capable of improving prediction accuracy, simultaneously considering time and space characteristics, reducing prediction error (MAPE), being superior to a traditional model, having strong robustness, keeping stable prediction accuracy under different time conditions and traffic scenes, verifying that the method has strong simulation capability and robustness on real world complex traffic behaviors, and efficiently extracting space-time characteristics, wherein the combination of HGNN and LSTM effectively captures complex space-time dependence, and is suitable for a large-scale traffic network. In order to achieve the above purpose, the invention provides a traffic flow prediction method based on a space-time multi-hypergraph neural network, which specifically comprises the following steps: S1, selecting an area and a time range, acquiring structural data by using a traffic monitoring system to construct a data set, and dividing the data set into a test set and a training set; s2, a spatial feature extraction module adopts a multi-hypergraph neural network MHGNN to model spatial dependence in traffic flow data by using an upstream-downstream adjacent hypergraph, a regional hypergraph and a history mode hypergraph, and the outputs of the three hypergraphs are fused to generate a group of comprehensive spatial features for each road section; S3, capturing time dependence in traffic flow data by a time feature extraction module through an LSTM neural network; And S4, constructing a space-time multi-hypergraph neural network STMHGNN model, splicing the outputs of the