Abstract
Existing traffic flow prediction methods generally only consider the spatiotemporal characteristics of traffic flow. However, in addition to the spatiotemporal characteristics, the interference of various external factors needs to be considered in traffic flow prediction, including severe weather, major events, traffic control, and metro failures. The current research still cannot fully use the information contained in these external factors. To address this issue, we propose a novel metro traffic flow prediction method (KGR-STGNN) based on knowledge graph representation learning. We construct a knowledge graph that stores factors related to metro traffic networks. Through the knowledge graph representation learning technology, we can learn the influence representation of external factors from the traffic knowledge graph, which can better incorporate the influence of external factors into the prediction model based on the spatiotemporal graph neural network. Experimental results demonstrate the effectiveness of our proposed model.
1. Introduction
Accurate traffic prediction has become a core component of intelligent transportation system (ITS) construction, providing a decision-making basis for traffic management and helping improve residents’ travel efficiency. Many studies have shown that there are many external factors (such as weather, events, accidents, etc.) that have an impact on traffic flow. Different types of POIs around traffic nodes can also lead to different traffic flow patterns. With the continuous development of intelligent transportation systems, more and more factors affecting traffic flow can be collected. However, existing traffic flow prediction methods generally consider learning spatiotemporal features from traffic data to predict future traffic flow changes. These methods ignore the exploitation of the influence of external factors on traffic flow. Therefore, it is a challenging research problem to consider these factors to obtain accurate traffic prediction results fully.
The development of deep learning methods brings new solutions to traffic forecasting. Generally, deep learning-based methods build deep networks by stacking different basic modules. Many architectures have been developed to handle complex spatiotemporal data. Graph convolutional networks can be used to model data in nonEuclidean space, which is more suitable for road network structure or metro structure in traffic flow prediction. Existing GCN-based methods usually use a purely data-driven approach to learn the spatiotemporal dependencies of traffic flow. Several studies attempt to exploit information from external factors in traffic flow forecasting. However, these methods usually directly encode external factors into representation vectors. The representation of the traffic flow is directly concatenated with the obtained representation vector in the prediction model. In large-scale traffic data, there are complex correlations between external factors and the traffic flow of nodes. Existing representation vectors of external factors are still deficient in expressing their influence on traffic flow.
In order to deal with the limitations of data-driven forecasting models on the utilization of external factors, one solution is to take a traffic knowledge graph composed of various traffic factors into account when constructing a predictive model. Knowledge graphs store real-world concepts and knowledge in the form of graph structures. The intelligent transportation system has a wide range of data sources and rich semantic information. The construction of a traffic knowledge graph can help solve traffic problems by using the traffic information contained in the data. The external factor representation obtained by the existing methods cannot fully express the impact on the traffic flow in the prediction model. Knowledge graph representation learning provides a new research direction for obtaining more useful semantic representations. Knowledge graph representation can obtain low-dimensional dense vectors, which provide better entity representation and prior knowledge for machine learning or deep learning models for downstream tasks.
The metro knowledge graph contains many influencing factors related to the operation of the metro network, such as weather, events, and POI (points of interest). We propose a metro traffic flow prediction method based on knowledge graph representation learning and spatiotemporal graph neural network (KGR-STGNN). The knowledge graph representation learning technique can be used to obtain the low-dimensional vector representation of the knowledge graph, and the representation of the entities and relationships of the entire metro network can be obtained at regular intervals. These semantic representations can fully represent the correlation characteristics between external factors and metro networks. Besides, this method utilizes Graph Neural Network (GNN) to learn the spatial dependencies between traffic flow sequences. Temporal Convolutional Network (TCN) is used to learn the temporal features of traffic flow. The dynamic semantic representation learned in each time period is used as a part of the input, which is fused with the traffic flow representation obtained through the spatiotemporal learning module. Finally, a traffic flow representation including the influence of complex external factors is obtained, to achieve a more accurate metro traffic flow prediction.
The main goal of this research is to combine a knowledge graph representation containing metro operation information with a spatiotemporal traffic flow prediction model. The main contributions of this study are summarized as follows: (i)A metro traffic knowledge graph is constructed, and we propose a new metro traffic flow prediction model based on this knowledge graph, which effectively improves the modeling ability of the impact of traffic events on traffic prediction(ii)The knowledge graph representation model is used to obtain the traffic knowledge representation at different times, which is significantly different from the external factor utilization method in the previous prediction model(iii)A new metro traffic flow prediction model KGR-STGNN based on knowledge graph representation and spatiotemporal graph network is proposed. Experimental results demonstrate the effectiveness of using traffic knowledge in traffic forecasting
2. Related Work
In this section, we review related research on knowledge graph and traffic flow prediction.
2.1. Traffic Knowledge Graph
In the field of knowledge graph, early research mainly involves semantic networks that contain a lot of information, such as WordNet [1], Google Knowledge Graph [2], etc. The general knowledge graph is built on the basis of a large amount of real-world data and mainly focuses on the modeling and storage of knowledge in a wide range and multiple domains. Domain knowledge graph [3] is oriented to a single industry, which can realize knowledge modeling and reasoning and provide help for the analysis of industry problems and the formulation of strategies, such as geographic information graph [4], medical knowledge graph [5], and so on. In recent years, domain knowledge graphs for a single industry have emerged continuously. Grainger et al. [6] automatically and dynamically build domain semantic networks based on corpus data. Pan et al. [7] use the linked data of large organizations to build and deploy industry-oriented data. Li et al. [8] constructed an e-commerce knowledge graph based on e-commerce data to help understand user needs and improve recommendation effects.
The traffic field has a wide range of data sources and rich semantic information. Building a traffic knowledge graph can help solve traffic problems by using the traffic information contained in the data. Muppalla et al. [9] used traffic surveillance video data to build a video information graph to help understand traffic patterns. Zhang et al. [10] proposed an urban knowledge graph neural network model (Urban Knowledge Graph Neural Network, UKG-NN) to combine urban knowledge graph and neural network to solve the problem of traffic accident reasoning and optimizing cargo storage. Keller [11] builds an air traffic knowledge graph by fusing various aviation data, which provides help for air traffic management. Zhou and Chen [12] constructed an urban traffic knowledge graph containing various types of urban knowledge, and combined the deep spatiotemporal convolutional neural network (STCNN) to extract features from the urban knowledge graph to achieve urban congestion.
2.2. Precise Prediction of the Area.
Zhu et al. [13] proposed a traffic flow prediction method based on graph convolution on the basis of knowledge representation driven. By constructing a traffic knowledge graph and using an entity-attribute relation-based knowledge graph representation model (KR-EAR) combines knowledge graph and information in the transportation network, the accuracy of traffic flow prediction is effectively improved. Sun et al. [14] constructed a traffic knowledge graph and an affair graph based on open-source data and further used the traffic knowledge graph to identify relevant traffic events in social media data.
In general, the metro knowledge graph is of great significance to the construction of intelligent transportation systems, but there are few studies on the metro knowledge graph, and the application form of the knowledge graph in the transportation field is simple.
2.3. Knowledge Graph Representation
The purpose of knowledge graph representation learning techniques is to express semantic entities and relationships in knowledge graphs as low-dimensional dense vectors [1]. The vector representations obtained through knowledge graph representation learning can be used to compute complex semantic associations between different entities and relationships in the knowledge graph. On the basis of knowledge graph representation learning technology, researches such as knowledge graph completion, knowledge reasoning, and link prediction are developing continuously [15]. Through the obtained distribution representation, the heterogeneous information in the knowledge graph can be fused to obtain a more complex representation, which provides help for downstream tasks.
Inspired by word embedding techniques in the field of natural language processing such as word2vec [16], knowledge graph representation learning methods have been continuously proposed and practically applied. The most widely used knowledge graph representation learning direction is the TransE model proposed by Bordes et al. [17]. The idea of this model comes from word vector representation, and words with high semantic similarity also have a relatively close distance in a low-dimensional vector space. Bordes et al. project entities and relations into a low-dimensional vector space and predict new knowledge entities by summing entity vectors and relation vectors to obtain new vectors. Although the TransE model has the advantages of simplicity, effectiveness, and low computational complexity, it still has shortcomings in the representation of complex relationships such as one-to-many, many-to-one, and many-to-many. In order to improve the representation of complex semantic relations, many researchers have proposed a series of improved methods based on the TransE model. Wang et al. [18] proposed the TransH model to learn semantic representations of complex relations between entities and by projecting knowledge entities onto different planes related to the relations. Lin et al. [19] proposed the TransR model based on the TransH model. The core idea of this model is to learn different types of entity representations and relational representations in different types of vector spaces, and then map the entity vectors to the corresponding relational vector spaces for computation.
In conclusion, knowledge graph representation learning provides convenience for the application of knowledge graphs in downstream tasks. The use of metro traffic knowledge graph also requires the help of knowledge graph representation learning.
2.4. Traffic Flow Prediction
Traffic flow prediction is an important research problem in the intelligent transportation system. Traditional flow traffic prediction methods are mainly based on statistical models. HA [20], ARIMA [21], and VAR [22] are classic time series analysis methods based on statistical models, widely used in traffic flow prediction. However, these methods are generally designed for small datasets and are not suitable for processing time-series data with complex dynamic characteristics in large-scale datasets. Machine learning methods such as SVR model [23], Bayesian model [24], and k-nearest neighbor model [25] can deal with modeling on complex data and can better reflect the nonlinear and uncertain characteristics of traffic data. However, traditional machine learning methods usually only consider the temporal information but ignore the spatial dependence in the traffic data. Moreover, these methods rely on the ability to design feature models manually and still have limitations in modeling complex and dynamic traffic flows.
Recently, deep learning methods have shown superiority in traffic flow prediction problems. RNN [26] and its variants are widely adopted as a component of traffic flow prediction models to capture the temporal dependence of traffic data. LSTM [27] and GRU [28] further enhance the RNN’s ability to model long-range time-series traffic data. In addition, CNN [29] has also been proven to be effective in time modeling. STGCN [30] uses convolutional networks to extract temporal features from graph-structured traffic data. GSTNet [31] and Graph Wavenet [32] use dilated causal convolution to capture the temporal dependence of transportation network nodes. Dilated causal convolution [33] is a special CNN that effectively captures long-term dependence by changing the size of the receptive field. In terms of spatial dependency modeling, main studies use CNN to model spatial correlation, but CNN is mainly designed for modeling data in Euclidean space. Many researchers have adopted GCN [34] to effectively handle traffic data in nonEuclidean space to address this limitation. For example, STGCN [30] and TGCN [35] use GCN to learn the topological structure of the traffic road network.
In summary, traditional statistical models and deep learning models for traffic flow prediction are constructed based on the analysis of spatiotemporal features of traffic data. However, the semantic information contained in the traffic data also reflects the specific laws and patterns of traffic flow, which can be utilized to enhance traffic prediction results.
3. Traffic Knowledge Graph
The knowledge graph is a semantic structure that can express entities, concepts, and relationships in the real world. The knowledge graph can conveniently store data in a form that is more recognizable to humans, providing a better format for data management. The knowledge graph is developed on the basis of semantic network. In essence, a knowledge graph is a more structured and abstract semantic network. The knowledge graph is mainly composed of entities and relationships. Entities represent individuals or concepts in the objective world, and relationships represent various semantic associations between entities.
In order to make more effective use of the semantic information contained in metro data, we construct a metro knowledge graph on the basis of processing and analyzing multisource heterogeneous traffic data. The data sources of the knowledge graph constructed in this paper are mainly Beijing metro card data, Beijing metro road network data, map data, weather data, and social media data. The entities of the knowledge graph mainly include track sites, routes, regions, map points of interest (POI), weather, traffic events, etc. The relationship mainly includes the attribution relationship between site and line, the spatial positioning relationship between site and latitude and longitude value, the attribution relationship between site and area, the attribution relationship between line and region, the spatial positioning relationship between POI (points of interest) and latitude and longitude value, and the influence of site and weather. Relationships and affected relationships of site-traffic events, etc.: in addition, the main entity has various attribute information, such as the site’s passenger flow value, the traffic event’s time, etc.
3.1. Knowledge Graph Construction
The construction of the metro knowledge graph mainly includes the following six steps: (1)Traffic data acquisition. The sources of traffic data related to metro network management and operation are very complex, including metro network data, metro card swiping data, points of interest data, weather data, and social media data. Some of the above data can be obtained through the cooperation channels with the Beijing Municipal Transportation Authority or the Metro Operation Department, and some of the data needs to be obtained through a given application programming interface(2)Data processing. The collected data has different formats and organizational structures, so further processing the acquired data into structured data is necessary(3)Extracting traffic entities and relationships. The key part of the construction of a metro knowledge graph requires artificially using professional knowledge to determine the pattern layer composition of the knowledge graph, and then extracting metro entities and relationships from the processed data(4)Obtaining the attribute value of the traffic entity. Consistent with the operation in the previous step, the required attribute value key-value pairs are extracted from the processed data and filled in as the attribute value of the knowledge graph entity(5)Storage in the knowledge graph storage tool. Install the corresponding knowledge graph storage software, store the constructed entities, relationships, and attributes, and use the graph database as a graph storage tool(6)Query usage of traffic knowledge graph. Based on the storage in the previous step, using the query algorithm, the entities, attributes, and relationships in the graph can be retrieved
The primary task of metro knowledge graph construction is to define the mode layer, which is the knowledge graph’s core part and key content. The first step is clarifying the metro traffic network’s core elements and knowledge. The crucial elements in metro traffic system are stations, routes, and environments. Stations and lines are the metro network’s basic elements and structural elements. Generally speaking, stations and lines do not change over a period of time and belong to the fixed and static part of the knowledge graph. The environment contains two parts, one part is static elements related to traffic, such as POI (points of interest), and the other part is dynamic elements, such as weather and traffic events. Figure 1 shows the overall process of building a knowledge graph. We use the “site-belonging-line” triplet as an example, and show the process of extracting entities, relationships, and attributes from raw data. The knowledge graph is finally stored in the graph database.
3.2. Main Components of Traffic Knowledge Graph
For the static part of the metro knowledge graph, its main entities include stations, routes, regions, and POI (points of interest). Station entities have attributes such as name, number, latitude and longitude, and transfer information. Line entities have attributes such as name and number. POI entities have attributes such as id, name, type, location, latitude and longitude, and the area to which they belong. The relationship between entities mainly includes: the adjacent relationship between stations in space, the attribution relationship between stations and lines, the adjacent relationship between station and POI. The upper part of Figure 1 shows the static entities and relationships of the traffic knowledge graph.
Unlike the static part, the dynamic part of the metro knowledge graph is mainly composed of elements that change over time, such as weather, traffic events, and metro station passenger flow. The entities in the dynamic part mainly include weather, events, and passenger flow status. The weather entity has attributes such as recording time, temperature, humidity, pollutant indicators, weather conditions, and wind conditions; the event entity has attributes such as event type, occurrence time, and event location; and the passenger flow status has attributes such as time and passenger flow. After realizing the design of the dynamic part entity and the static part entity, it is necessary to define the relationship between them further. As shown in the schematic diagram of the dynamic part of the metro knowledge graph in Figure 2, there is an “affected” relationship between stations and weather or events.
The traffic knowledge triples in the graph are defined as , where and represent the head entity and tail entity, respectively, and represents the relationship. The properties of the entity are defined as . The entire metro knowledge graph can be expressed as , represents the set of knowledge entities, represents the set of association relationships, and represents the attribute collection. Table 1 shows the entities and attributes of the metro traffic knowledge graph, and Table 2 shows the edges in the metro traffic knowledge graph. Figure 3 shows the constructed traffic knowledge graph stored in the graph database.
4. Methodology
4.1. Knowledge Graph Representation
TransE is the most representative translation-based knowledge graph representation learning model. Inspired by the translation-invariant feature of word vectors in the word2vec model [16], the TransE model was proposed and used in the representation learning of knowledge graph entities and relations [17]. Specifically, for the triple representing the knowledge ontology, and represent the head entity and the tail entity, respectively, and represents the association between entities. The TransE model has the following settings: the embedding vector of the tail entity should be equal to the sum of the embedding vector of the head entity and the embedding vector of the relationship , that is, . When designing the knowledge graph representation learning model, the scoring function is used to measure the reliability of the triples, and it is an important tool to measure the rationality of the representation results. The scoring function of the TransE model is:
indicates that the calculation method of vector distance is norm. TransE regards entities and relationships in the knowledge graph as two matrices. The structure of the entity matrix is , where represents the number of entities, represents the dimension of each entity vector, and each row in the matrix represents the word vector of an entity; and the relationship matrix structure is , where represents the number of relationships and represents the dimension of each relation vector. For the knowledge graph trained by TransE, if two vectors are extracted from the entity matrix and the relationship matrix, respectively, and operation is performed on these two vectors, the obtained result will be similar to the vector of another entity in the entity matrix. Finally, the existing triples in the knowledge graph are represented in the form of word vectors. Figure 4 shows the basic principle of the TransE model.
4.2. Traffic Knowledge Graph Pretraining
The traffic knowledge graph contains a large amount of traffic semantic information related to the operation, scheduling, and management of the metro network, such as traffic events that affect subway traffic flow, bad weather, and information about the metro system itself (metro network structure and POI). This information can be collectively referred to as external factors affecting subway traffic flow. In order to utilize these external factors in the traffic flow prediction model, the knowledge graph representation learning method is adopted to pretrain the traffic knowledge graph, so as to obtain the traffic semantic representation expressing the influence of external factors. In the real world, metro traffic flow will be disturbed by different external factors at different times, and the dynamic characteristics of external factors need to be fully considered in pretraining.
As shown in Figure 5, the TransE model is used to pretrain the subgraphs at different times. According to half an hour as a time period, the subknowledge graph under each time period is extracted from the metro knowledge graph. Table 3 shows the main entities and relationships in the pretrained knowledge graph, which represent the external factors that we incorporate in implementing traffic predictions. Each subknowledge graph consists of two parts; one part is the static road network information and POI. The other part is dynamic traffic events and weather information. Through the representation learning of each subknowledge graph, the semantic representations () of 325 metro station entities in Beijing under different time periods are obtained, and the dimension of each semantic representation is 30 dimensions. When using knowledge representation technology to obtain the representation of station entities, the external factors affecting metro operation will also be mapped to the low-dimensional vector space at the same time, and the relationship between station entities and external factor entities also follows the semantic rules of the TransE model. Therefore, the semantic information of external factors is also included in the representation results. The formula for knowledge graph pretraining is shown below:
where, represents the semantic representation of 325 metro stations at time , are the station entities, and represent relationships and tail entities related to station entities at time .
4.3. Spatial-Temporal Modeling of Traffic Flow
Modeling the spatiotemporal characteristics of metro traffic flow is a key task to achieving subway traffic flow prediction. In order to effectively learn the spatiotemporal characteristics of metro traffic flow, this paper uses graph convolutional neural networks and temporal convolutional neural networks to model spatial and temporal features, respectively. The specific usage of graph convolutional neural network is described in detail below.
Metro networks generally have a graph structure. Convolutional neural networks used in the image field are generally used to extract local spatial features of data in Euclidean space. Graph Neural Network (GCN) is more suitable for data in nonEuclidean space, and GCN can capture more realistic spatial features of metro traffic flow. The frequency domain-based graph convolution operation requires the use of the Laplacian matrix of the graph:
where, represents the identity matrix of order , is the adjacency matrix of the orbital network graph, and is the degree matrix of the orbital graph calculated from the adjacency matrix. Therefore, the definition of the graph convolution operation on the graph signal of the orbital network using the convolution kernel is: where, represents the operator of the graph convolution operation, is the Laplacian matrix of the orbital graph, means the matrix composed of the eigenvectors of the matrix , and represents the diagonal matrix composed of the eigenvalues of . Due to a large number of metro stations, the decomposition of the matrix in the above formula requires a large amount of calculation. Therefore, this paper uses the Chebyshev polynomial approximation graph convolution kernel to solve the graph convolution operation:
represents the k-dimensional coefficient vector of the Chebyshev polynomial. The recursive formula of this polynomial is: . In the above formula, , and is the maximum value of the eigenvalues of the matrix . The Chebyshev polynomial is used to complete the approximation of the convolution operation on the graph, which reduces the computational complexity of the graph convolution operation. The spatial modeling model used in this chapter is based on ChebNet [36] simplification to obtain first-order Chebyshev graph convolution, and its formula is:
In the formula, is the input feature of graph convolution, the adjacency matrix and the degree matrix are added to the node’s own feature.
The recurrent neural network is generally used to model the traffic flow time series in the traffic flow prediction problem. However, the prediction of each state in the recurrent neural network is closely dependent on the previous state, and the computational complexity is high. In many studies, time-series data is regarded as one-dimensional structured data, and it is more computationally efficient to use convolutional neural networks to extract time-series features on time-series data. This section uses Gated Temporal Convolutional Network (Gated TCN) to extract and model temporal features in metro traffic flow. A Gated TCN consists of two layers: a 1D convolutional, gated linear unit GLU that learns temporal properties. Its calculation formula is as follows:
The width of the convolution kernel of the one-dimensional convolution is , and the temporal convolution network performs convolution operations on the neighbors of the input traffic flow, and each operation shortens the traffic flow sequence by . The input of each track site is represented as a sequence of length , the number of channels is , and the input is denoted as . The convolution kernel is , the input is convolved to obtain the output , and are the input of the gated linear unit (GLU), represents the Hadamard product operation, and is the activation function.
4.4. Traffic Flow Prediction with Knowledge Graph
After pretraining the traffic knowledge graph to obtain the semantic representation of the metro station, the semantic vector reflecting the influence of external factors needs to be utilized in the prediction model. As shown in Figure 6, the traffic flow sequence data input passes through the spatiotemporal feature modeling module in the upper part, which captures the temporal dynamics and spatial correlation of metro traffic flow. After the knowledge graph representation learning pretraining, the dynamic semantic representation reflecting the influence of external factors is obtained within a time period of half an hour, and the gated fusion mechanism is used to fuse the output of the spatiotemporal feature modeling module with the dynamic semantic representation. Finally, the prediction results of metro traffic flow under the influence of external factors are obtained.
The spatiotemporal feature modeling part consists of two Gated TCN layers and one spatial GCN layer, and the spatial GCN layer in the middle undertakes two Gated TCN layers. The calculation formula of this part is as follows:
and are the temporal convolution kernels of the two Gated TCN layers, represents the spatial graph convolution at the middle layer Convolution kernel, and is the activation function.
The semantic representation vector obtained every half an hour is defined as , The gated fusion mechanism is used in the prediction model to fuse spatiotemporal features and semantic features. During fusion, the time of the semantic representation vector is guaranteed to be consistent with the time of the spatiotemporal feature modeling output, and its calculation formula is as follows:
5. Experiment
5.1. Experiment Settings
The experiments in this chapter need to use the constructed metro knowledge map, so this chapter conducts experiments on the Beijing metro passenger flow dataset (BJMF15), and the Beijing metro knowledge graph is also used in the experiment. The experiments evaluate the prediction performance of the knowledge graph representation learning-based traffic flow prediction model (KGR-STGNN) on the dataset BJMF15.
BJMF15: Passenger flow data collected by Beijing Metro System in 2015. The unprocessed raw data includes daily passenger travel record data at 325 stations on 13 lines of the Beijing Metro in August. The data consists of passenger travel records, including entry routes, entry stations, entry times, exit routes, exit stations, and exit times. Taking every five minutes as a time step, the metro passenger flow data is obtained by processing the original passenger swiping card data.
In practical experiments, the complete BJMF15 dataset is divided into training, validation, and test sets. The split ratio of the dataset is 80%, 10%, and 10%, respectively. Different from the prediction of short-term traffic flow in Chapter 2, the impact of external factors on traffic flow needs to be better reflected in a longer time. The experiments in this chapter use historical traffic data of 12 consecutive time steps (1 hour) to predict 12-time steps (60 minutes), 15-time steps (75 minutes), 18-time steps (90 minutes), and 21-time steps (105 minutes) data.
In the KGR-STGNN model, the learning rate is set to 0.001, the size of each batch of training data, that is, the batch size is set to 32, and the number of training epochs is 100. Experiments on graph convolutional layers set a dropout of . All experiments are performed on an Nvidia RTX 2080ti server with 11GB of video memory, and the server operating system is Linux.
5.2. Evaluating Indicator
In order to compare the prediction accuracy of the proposed model and the basic model, accurate numerical indicators are needed as the evaluation basis. This chapter evaluates the forecasting accuracy of all methods using two commonly used traffic forecasting metrics, including:
(1) Mean Absolute Error (MAE), which is a basic measure that reflects the actual situation of prediction accuracy
(2) Root Mean Square Error (RMSE), which is used to measure the deviation of the predicted value from the true value
where, represents the number of stations in rail transit, represents the actual traffic flow, and represents the predicted traffic flow.
5.3. Baselines
In order to verify the effectiveness of the Metro Traffic Flow Prediction Model (KGR-STGNN) based on knowledge graph representation learning, this paper uses the following models as benchmarks: (1)SVR (Support Vector Regression Machine). A method for predicting future traffic flow using a support vector machine(2)LSTM (Long Short-Term Memory Network). A time-series prediction model that models the long-term temporal characteristics of traffic-flow sequences(3)GCN (Graph Neural Network). A prediction method to mine the spatial correlation of traffic flow(4)GRU (Gated Recurrent Unit). Similar to the LSTM structure, part of the structure is optimized, and the amount of model parameters is reduced(5)TGCN. The graph convolutional neural network is used to learn the spatial characteristics of the node traffic flow. The gated recurrent unit is used to model the temporal characteristics of the traffic flow sequence, and then the two are combined to obtain a spatiotemporal traffic flow prediction model(6)STGCN. A traffic flow prediction method that uses convolutional neural networks to learn spatiotemporal characteristics. In terms of spatial characteristics, a graph convolutional neural network is used to obtain spatial dependencies between nodes, and one-dimensional convolution is used to model traffic flow sequences in terms of temporal characteristics
5.4. Prediction Performance
The traffic flow prediction model based on knowledge graph representation learning proposed in this paper is tested on the BJMF15 dataset and compared with five benchmark methods. In the experiments where the prediction step size is set to 12 steps, 15 steps, 18 steps, and 21 steps, the experimental results of each prediction method are shown in Table 4. It can be seen from the results that the traffic flow prediction model (KGR-STGNN) based on knowledge graph representation learning proposed in this paper has better prediction performance. Among them, the prediction results of TGCN, STGCN, and KGR-STGNN are quite different from those of SVR, LSTM, and GCN. The first three models all model the temporal and spatial characteristics of metro traffic flow, proving the spatial and temporal modeling of traffic flow data effectiveness. Compared with STCGN, the traffic flow prediction model (KGR-STGNN) based on knowledge graph representation learning proposed in this paper has similar spatiotemporal characteristic learning methods. Compared with STGCN, KGR-STGNN uses the representation of the influence of external factors obtained through knowledge graph representation learning. With the increase in the prediction step size, the MAE evaluation indicators of the KGR-STGNN model are 8.211, 8.996, 9.624, and 11.403, respectively, and the prediction performance is improved by 1.84%, 6.60%, 13.57%, and 15.96%, respectively, compared with the results of the STGCN model. The experimental results show that adding external factor representation can improve the prediction model more and more with the increase in prediction time. As the prediction time becomes longer, the influence of external factors and traffic knowledge on the traffic flow becomes more obvious, proving the KGR-STGNN model’s effectiveness in predicting medium and long-term metro traffic flow.
Figures 7–9 are the changes of the loss function, MAE, and RMSE of the metro traffic flow prediction model based on knowledge graph representation learning during the training and validation of the model. The red curve in the figure represents the change curve in the training case, and the blue curve represents the change curve in the validation case. It can be seen from the above three figures that the loss function and prediction result of the subway traffic flow prediction model based on knowledge graph representation learning can decrease rapidly and stably and finally tend to be stable and reach the model convergence state. This shows that the design of the prediction model in this paper is reasonable, which can effectively reduce the prediction error and achieve good prediction results.
In order to visually show the prediction effect of the KGR-STGNN model, this paper selects the traffic flow prediction results of Pingguoyuan Station of Metro Line 1 on August 29, 2015, for visualization to compare the gap between the actual value and the predicted value. As shown in Figure 10, the predicted results are very close to the real values, especially the curves in the peak passenger flow period have basically overlapped. The prediction results also prove that the model proposed in this paper can predict the traffic flow.
6. Conclusions
There are a large number of external factors in traffic data that affect traffic flow. Using the information contained in external factors in traffic flow forecasting can effectively improve the forecasting effect. To this end, this paper proposes a metro traffic flow prediction model based on knowledge graph representation learning to build the model. In order to model the influence of these external factors, the model learns the representation of the metro knowledge graph and obtains a semantic representation that contains information about the external factors that affect the metro operation. In addition, the model adopts time-gated convolutional network and graph convolutional network to model traffic flow’s spatiotemporal features. A gated fusion mechanism is used to fuse the semantic representation of the subway network with the output of the spatiotemporal feature learning module. Finally, the spatiotemporal feature representation of subway traffic flow data considering the influence of external factors is obtained. Experiments on Beijing metro traffic flow data show that the prediction model proposed in this chapter outperforms other benchmark models in metro traffic flow prediction, especially long-term traffic flow prediction. The experimental results verify that the knowledge representation obtained through representation learning can effectively improve the performance of metro traffic flow prediction.
Data Availability
The data used to support the findings of this study are available from the corresponding author upon request.
Conflicts of Interest
The authors declare that they have no conflicts of interest.
Acknowledgments
This work was supported in part by the National Natural Science Foundation of China under Grants 62072015, U19B2039, U1811463, 61632006, 61876012, and 61902053; the Natural Science Foundation of Beijing under Grant 4172003; and the China Scholarship Council under Grant 201806540008. The authors recognise those individuals who provided help during the research and preparation of the manuscript.