CN114519932B

CN114519932B - Regional traffic condition integrated prediction method based on space-time relation extraction

Info

Publication number: CN114519932B
Application number: CN202210021913.6A
Authority: CN
Inventors: 于家傲; 彭磊
Original assignee: Shenzhen Institute of Advanced Technology of CAS
Current assignee: Shenzhen Institute of Advanced Technology of CAS
Priority date: 2022-01-10
Filing date: 2022-01-10
Publication date: 2023-06-20
Anticipated expiration: 2042-01-10
Also published as: CN114519932A

Abstract

The invention discloses an area traffic condition integrated prediction method based on space-time relation extraction. The method comprises the following steps: constructing a road network topological graph of a road and a parking lot topological graph of a parking lot aiming at a target area, wherein for the road network topological graph, the road is used as graph nodes, the connection rule between the nodes follows the natural connection rule of the road, the average speed information of each road is used as the characteristic of the node, for the parking lot topological graph, the parking lot is used as the node, the shortest path is used as the connection rule of the node, and the parking lot parking space occupation quantity is used as the node characteristic; based on the road network topological graph and the parking lot topological graph, the average vehicle speed prediction result of each road and the prediction result of the parking space occupation of each parking lot are obtained by utilizing a constructed multi-task model, wherein the multi-task model framework comprises a space-time extraction network, a heterogeneous graph neural network and a time sequence model. The invention can realize the integrated prediction of traffic flow and parking condition, and improves the prediction precision.

Description

Regional traffic condition integrated prediction method based on space-time relation extraction

Technical Field

The invention relates to the technical field of traffic prediction, in particular to an area traffic condition integrated prediction method based on space-time relation extraction.

Background

At present, an intelligent traffic system plays an important role in urban life, and the travel efficiency is improved. Traffic flow prediction is an important component of intelligent traffic, and can help traffic managers to know traffic information changes in advance, so that corresponding control strategies are formulated. In addition, reasonable travel plans can be formulated for the travelers according to the prediction results. Both driving and stopping belong to the state of the vehicle, and therefore traffic prediction should include road traffic prediction and stopping condition prediction. Both road traffic and parking saturation have an effect on overall traffic, which together determine the traffic conditions of an area. Almost all current traffic predictions take into account road traffic flow predictions and parking situation predictions separately. However, road traffic flows are closely related to parking situations, especially in the vicinity of some hot points of interest, such as hospitals, scenic spots, shopping malls, etc. On the one hand, the huge traffic flow puts pressure on the parking lot. For example, during holidays, parking lots for some attractions are fully saturated. On the other hand, a limited parking space results in a low-speed cruising of the vehicle on the road. According to investigation, the ineffective traffic flow caused by searching for a parking lot or a parking space accounts for 15% of the urban road traffic flow. Furthermore, studies have shown that 30% of urban traffic congestion is due to vehicles seeking empty spaces. These vehicles often travel at low speeds, causing traffic jams. Thus, there is a dense and inseparable link between road traffic and parking conditions, and if predicted separately, traffic characteristics cannot be fully grasped.

In the existing researches, traffic flow prediction and parking space occupation prediction of a parking lot can be roughly divided into two types. One is a predictive model based on a single time series, and the other is a predictive model based on multi-factor effects. The model based on the single time sequence only uses the historical data of the prediction target, for example, the early prediction adopts a model based on statistics, but the method has low anti-interference capability and inaccurate prediction result.

In recent years, because the deep learning model has strong feature extraction capability and sample space fitting capability, many students begin to develop traffic flow and parking space occupation prediction method research based on the deep learning model, and only enough historical data is needed to predict traffic flow and parking condition at the next moment, so that the prediction effect is improved compared with an early model as a whole. The predictive model based on multi-factor influence considers the influence of other factors on the predicted target in addition to utilizing the historical data of the predicted target. For example, the influence of weather and air quality is taken into consideration when predicting traffic flow, and the influence of traffic speed, charge and holidays is taken into consideration when predicting parking occupancy. In general, such solutions currently occupy a relatively small number and the influencing factors studied are limited. Although some predictive models take into account external factors, they still fall into a single-task model, where interactions between certain factors are underutilized. Whereas road traffic and parking conditions are related closely to each other as described above.

In summary, the existing traffic flow prediction and parking situation prediction or independent predictions of each other only form a unidirectional effect, ignoring the interaction relationship between the two traffic activities. Moreover, there is currently no clear way to quantify the spatiotemporal relationship between traffic flow and parking conditions. In addition, most of the existing neural networks adopt a splicing mode in a space feature fusion method, but feature dimensions after splicing are increased, so that a sparse matrix in a high-dimensional space is changed, and gradient is difficult to find when a model is trained. And if two matrices with some kind of correlation are stitched together as a new feature representation there must be information redundancy.

Disclosure of Invention

The invention aims to overcome the defects of the prior art, and provides an integrated prediction method for regional traffic conditions based on space-time relation extraction, which can realize the integrated prediction of traffic flow and parking conditions and improve the prediction precision.

The technical scheme of the invention is as follows: the regional traffic condition integrated prediction method based on space-time relation extraction comprises the following steps:

constructing a road network topological graph of a road and a parking lot topological graph of a parking lot aiming at a target area, wherein the road network topological graph takes the road as a graph node, the connection rule among nodes follows the natural connection rule of the road, the average speed information of each road is taken as the characteristic of the node, the parking lot is taken as the node in the parking lot topological graph, the shortest path method is taken as the connection rule of the node, and the parking lot parking space occupation quantity is taken as the node characteristic;

based on the road network topological graph and the parking lot topological graph, an average vehicle speed prediction result of each road in the road network and a prediction result of each parking lot parking space occupation in the parking lot network are obtained by utilizing a constructed multi-task model framework, wherein the multi-task model framework comprises a space-time extraction network, a heterogeneous graph neural network and a time sequence model.

Compared with the prior art, the method has the advantages that the average speed of the vehicles on the road and the occupied number of the parking spaces in the parking lot are used as indexes of road traffic flow and parking conditions respectively, a multi-task prediction model is built, the regional traffic conditions are predicted in an integrated mode, and the regional traffic condition integrated prediction based on space-time relation extraction is realized. Compared with the existing single-task prediction model, the traffic flow and parking in the prediction process can not generate interaction relation, the method synchronously predicts traffic conditions in an area, including road traffic conditions and parking conditions, and the provided integrated model has wider visual field. In addition, the method carries out space feature fusion according to the space-time relationship between traffic flow and parking, thereby not only quantifying the correlation between the traffic flow and the parking, but also further optimizing the model and remarkably improving the prediction effect of regional traffic conditions.

Other features of the present invention and its advantages will become apparent from the following detailed description of exemplary embodiments of the invention, which proceeds with reference to the accompanying drawings.

Drawings

The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments of the invention and together with the description, serve to explain the principles of the invention.

FIG. 1 is a flow chart of an integrated prediction method for regional traffic conditions based on spatiotemporal relation extraction according to an embodiment of the present invention;

FIG. 2 is a road and parking lot visualization of a certain patch of Shenzhen Luo lake region in the city according to one embodiment of the present invention;

FIG. 3 is a road network topology and parking lot network topology according to one embodiment of the invention;

FIG. 4 is a schematic diagram of a spatio-temporal relationship extraction network according to one embodiment of the invention;

FIG. 5 is a schematic diagram of a spatial feature extraction and fusion model according to one embodiment of the invention;

FIG. 6 is a schematic diagram of an integrated predictive model architecture in accordance with one embodiment of the invention;

FIG. 7 is a schematic illustration of an experimental scenario in accordance with one embodiment of the present invention;

FIG. 8 is a graph showing average vehicle speed predictions throughout the day for a web street, according to one embodiment of the present invention;

FIG. 9 is a schematic diagram of a prediction result of the number of occupied parking spaces in one day in a Huarun Wanning parking lot according to an embodiment of the invention;

FIG. 10 is a graph of accuracy contrast over the course of a day for a multi-tasking integrated prediction model and a single-tasking prediction model in accordance with one embodiment of the present invention;

FIG. 11 is a comparative schematic of mutual information of real data and predicted data regarding road traffic and parking according to one embodiment of the present invention;

FIG. 12 is a graph illustrating accuracy as a function of mutual information between road traffic and parking according to one embodiment of the present invention.

Detailed Description

Various exemplary embodiments of the present invention will now be described in detail with reference to the accompanying drawings. It should be noted that: the relative arrangement of the components and steps, numerical expressions and numerical values set forth in these embodiments do not limit the scope of the present invention unless it is specifically stated otherwise.

The following description of at least one exemplary embodiment is merely exemplary in nature and is in no way intended to limit the invention, its application, or uses.

Techniques, methods, and apparatus known to one of ordinary skill in the relevant art may not be discussed in detail, but are intended to be part of the specification where appropriate.

In all examples shown and discussed herein, any specific values should be construed as merely illustrative, and not a limitation. Thus, other examples of exemplary embodiments may have different values.

It should be noted that: like reference numerals and letters denote like items in the following figures, and thus once an item is defined in one figure, no further discussion thereof is necessary in subsequent figures.

The invention provides an area traffic condition integrated prediction method based on space-time relation extraction, which can synchronously predict the average speed of vehicles on a road and the occupied number of parking spaces of a parking lot. Referring to fig. 1, the provided method includes the following steps.

And step S110, constructing topological graph characterization of regional traffic conditions, wherein the topological graph characterization comprises a road network topological graph and a parking lot network topological graph.

The figure is a nonlinear data structure, and the adjacency matrix contains quite rich connection relation information. Traffic prediction problems have a strong spatial dependence, where the road network and parking network involved are well suited to be characterized by a graph structure. For example, the flow rate of vehicles on a road is affected by the spatial topology of the road, and vehicles on congested road segments may spread to other interconnected segments. Similarly, a certain space topological relation can be formed between the parking lots, and after a certain parking lot has no empty parking space, the vehicle can be transferred to the surrounding parking lots. The roads and parking lots involved therein are well suited for characterization by a graph structure.

For visual and easy understanding, a small map is intercepted to introduce the establishment process of the road network topological graph and the parking lot network topological graph.

The left diagram of fig. 2 is a screenshot of a small area of the shenzhen roasters area on the OpenStreetMap, and the right diagram is a display of vector data of the road and parking lot in the frame in the ArcGIS. The road and the parking lot may be constructed as a map structure, respectively.

The left graph in fig. 3 is a road network topology graph, and the road is regarded as a graph node. The connection rules between nodes follow the natural connection rules of the road. If two roads intersect, two nodes are connected. Marking a road network topological graph as G _r ＝(V _r ,E _r )，V _r Representing a set of all nodes. V (V) _r ＝{v ₁ ,v ₂ ,…,v _N N is the number of nodes, E _r Representing allA collection of edges. In addition, an adjacent matrix A for connection relation between nodes _r Representation, A _r ∈R ^N×N . The average speed information on each road is used as the characteristic of the node. Each node is characterized by x _r ＝[x ₁ ,x ₂ ,…,x _t ,…,x _T ]. T is the length of the historical time series, x _t The node characteristics at time t are represented,

all node characteristics representing time t, +.>

The right diagram in fig. 3 is a network topology diagram of the parking lot, and the parking lot is regarded as a node. For example, euclidean distance method may be used as a rule for connecting nodes, but this way road network information is lost. Preferably, a shortest path method is used as a connection rule of the node. The shortest road distance between two parking lot nodes is calculated, and if the distance is smaller than a set threshold (e.g. 600 meters), the two nodes are connected. Similarly, the road network topology is marked G _p ＝(V _p ,E _p )。V _p ＝{v ₁ ,v ₂ ,…,v _M M is the number of parking lots. A for adjacency matrix _p Representation, A _p ∈R ^M×M . Taking the occupied number of parking spaces in the parking lot as a node characteristic. The characteristics of each node during the T time period are marked as x _p ＝[x ₁ ,x ₂ ,…,x _t ,…,x _T ]。

All node characteristics representing time t, +.>

In this step, road network and parking lot network topologies including adjacency matrix A can be obtained _r And A _p Node characteristic representation at time t

And->

Etc.

And step S120, extracting the space-time relationship of traffic flow and parking situation based on the road network topological graph and the parking lot network topological graph.

In some cases, traffic flow on the road is germane to parking conditions in a parking lot. For example, roads in the vicinity of hospitals are often congested due to the influence of parking demands. There are a large number of vehicles rushing into the parking lot on the road near the office building during commute peak hours. While residential parking lots and roads at late night are not correlated. The present invention utilizes this spatiotemporal relationship to fuse the spatial features of the parking lot and traffic flow, rather than a simple geographic positional relationship.

How to extract the spatiotemporal relationship between the road nodes and the parking lot nodes. The relation extraction method is applied to the field of natural language processing to extract the relation between two named entities, so that a knowledge graph is established. In the embodiment of the invention, each parking lot node and each road node can be regarded as words, and form a sentence of regional traffic conditions together, so that a space-time relationship topological graph of the road node and the parking lot node is established by referring to a method for extracting entity relationships in natural language processing.

The space-time relation extraction can be implemented by using various time sequence models, such as LSTM (long short term memory network) and GRU (round robin gating unit). Fig. 4 is a network structure of spatiotemporal relation extraction, which is composed of biglu (bi-directional gating cyclic unit) and attention mechanism. A GRU is a recurrent neural network that excels in modeling sequence data. However, the common circulation god has the defects of gradient disappearance, gradient explosion, excessive parameters, difficult training and the like of the network model. LSTM and GRU successfully solve the above problems. The two principles are similar, namely, the gating mechanism is used for controlling information such as input, memory and the like, so that prediction is made in the current time step. But the GRU is less parametric than LSTM and thus more easy to train. The GRU can be considered as a variant of LSTM. Preferably, the GRU is utilized to extract a spatiotemporal relationship between the parking lot node and the road node. The GRU model is updated as follows:

r _t ＝σ(W _r ·[h _t-1 ,x _t ]) (1)

z _t ＝σ(W _z ·[h _t-1 ,x _t ]) (2)

wherein x is _t Is the input data at the time t and is formed by the average speed of a road at the time t

And the occupied number of parking spaces in a certain parking lot +.>

Spliced into a->

h _t-1 Is the hidden state at time t-1, which contains the relevant state of the previous node. r is (r) _t To reset the gate for controlling the degree to which state information at a previous time is ignored. z _t To update the gate, the state information for controlling the previous time is brought to the extent in the current state. />

Is a candidate hidden state, i.e. the memory information at the current moment. h is a _t It is the output state at time t that will be passed on to the next node.

To fully extract features using prior and posterior probabilities, in one embodiment, biglu (bi-directional gating loop unit) is employed. It is determined by both counter-directed state units.

The GRU () function is a nonlinear transformation of input data, w _t And v _t Respectively, are forward hidden states

And a backward hidden state->

Weight of t e [1, T)]。

In fig. 4, the role of the Attention layer (Attention) is to distinguish the importance of multiple hidden states and assign different weights, and the output v of biglu is obtained by weighted summation of the hidden states, expressed as:

u _t ＝tanh(w _a1 h _t ) (8)

v＝∑ _t α _t h _t (10)

the final output of the spatio-temporal relationship extraction network is obtained by the following formula, which represents the relationship between nodes.

In one embodiment, four relationships are defined, namely:parking lots affect traffic flow; traffic flow affects parking; both traffic flow and parking interact; traffic flow and parking are not affected by each other. Not all data is used for all labels in the training set, so the training is performed in a semi-supervised manner. The space-time relation extraction network can obtain a heterogeneous topological graph consisting of parking lot nodes and road nodes in a region. This is a weighted directed graph, with weights derived from the softmax function output. The heterogeneous topology map represents the space-time relationship between traffic flow and parking, and can be represented by weighted directed adjacency matrix A _rp Representation, A _rp ∈R ^(N+M)×(N+M) 。

And S130, respectively extracting the spatial characteristics of the road and the parking lot by using the graph convolution neural network, and fusing the spatial characteristics of the road and the parking lot.

How to fuse two types of heterogeneous traffic activity data is the first problem to be solved. Urban geospatial is a natural tie that associates all urban activities, and traffic activities in the same geospatial will exhibit a particular association law during the same time period. Fusion modeling of spatial features of multiple traffic activities is therefore a good solution. Firstly, the spatial features are respectively extracted, then, the spatial feature fusion is carried out by utilizing the heterogeneous graph convolution neural network according to the extracted space-time relationship on the feature level instead of adopting a simple splicing fusion mode, and a displayed attention mechanism is constructed in the heterogeneous graph neural network to improve the model training speed.

Considering that traffic flow and parking conditions have certain spatial dependence, firstly, the spatial features are respectively extracted, and then, spatial feature fusion is carried out according to the space-time relationship of two types of nodes on the feature level.

Specifically, the spatial feature extraction task is completed first. The object of traditional convolutional neural network research is data with European structure. The most notable feature of European structure data is that they have a regular spatial structure, e.g. image data is a regular two-dimensional matrix and speech data is a regular one-dimensional matrix. The graph structure is a complex non-European structure, the number of neighbors and the order of the nodes of each node are uncertain, and each node has own characteristics. In the task of extracting the spatial semantic features, the graph convolution neural network solves the limitation of the traditional convolution neural network, and can simultaneously consider the feature information of the nodes and the connection information between the nodes.

For example, a two-layer GCN (graph convolutional neural network) model is represented as follows:

x is the feature matrix and A is the adjacency matrix. Self-loops are added to each node in order to preserve its own information in aggregating node features. Specifically, this can be achieved by adding the adjacency matrix a and the identity matrix I.

Then pair->

Normalization processing is performed>

Wherein->

For the degree matrix->

W ₀ And W is ₁ Is a weight matrix, σ (·) represents the activation function, typically using Relu () as the activation function.

The feature matrix obtained after the road nodes and the parking area nodes extract the spatial features by the GCN network is as follows:

representing advanced feature representation of all road nodes at time t, then the feature representation of a certain road node is +.>

Similarly, the characteristic of a parking area node is expressed as +.>

Next, the spatial features represented by the two feature matrices are fused. If the two types of spatial features do not have correlation, the two types of spatial features can be spliced after being flattened. But it is considered that if there is a certain correlation between the two, the concatenation will cause information redundancy. Feature fusion is performed according to the spatio-temporal relationship of the two types of nodes. The weighted directed topology of the spatio-temporal relationship and its characterization a have been obtained in the previous step _rp It consists of two types of heterogeneous nodes. Correspondingly, two types of spatial features are fused by using the heterogeneous graph neural network.

The heterogeneous graph neural network can process topological graphs containing nodes of different categories, and feature fusion among connected nodes is completed. The heterograph neural network model converts different types of nodes into the same semantic space, and feature aggregation of neighbor nodes is completed through an attention mechanism. The attention coefficient reflects the importance of different neighbor nodes relative to the current node, and the importance is obtained by repeated training. To increase training speed, in one embodiment, a priori knowledge is introduced as the attention factor. The weight between nodes in the time-space relation topological graph is used as the attention coefficient. Only the neighbor features are fused when the neighbor nodes are fused. Therefore, the characteristics of the node i after being fused with the neighbor node are as follows:

is characteristic of the current node i. j is the neighbor node pointing to i, +.>

Is characteristic of node j. e is the set of edges consisting of j and i. Alpha _e Is the weight of the edge, alpha _e ＝A _rp [j][i]。/>

And merging the features of the neighbor nodes with the degree of penetration for the node i. After each node fuses the features of the neighboring nodes according to the above operation, the features of all nodes are expressed as +.>

The process of spatial feature extraction and fusion is shown in fig. 5, wherein the spatial feature extraction and fusion is input into an adjacent matrix and a feature matrix of a traffic flow network and a parking network, node features become advanced feature representation after GCN (graph convolution neural network) is passed, and feature fusion between two types of nodes is completed by a heterogeneous graph neural network and priori knowledge. The output is a fused integral traffic characteristic representation

And step S140, extracting time features and predicting traffic conditions based on the fused spatial features.

Traffic predictions also have time dependence. Both the road vehicle flow rate and the parking space occupation number of the parking lot change with time. And a certain change rule is presented in a certain period, for example, a peak period occurs in a certain period in a day, and weekly data change trends are approximately the same. A recurrent neural network is a sharp tool that processes time series, for example, with a GRU (gated loop unit) for time feature extraction and for the completion of the final prediction.

The above step S130 completes the extraction and fusion of the spatial features. In this step S140, the temporal feature is first extracted. The GRU is suitable for processing sequence data as described above, and is less prone to training in terms of parameters than LSTM. The temporal semantic extraction is preferably performed using a GRU.

Specifically, the overall model structure of the traffic flow and parking condition integrated prediction proposed by the present invention is shown in fig. 6. This is a multiple-input multiple-output model, also a multitasking model. The input of the model is road historical average speed information

And parking area historical parking space occupation data->

The GCN (graph convolution neural network) firstly completes the extraction of spatial features, and meanwhile, the BiGRU extracts the space-time relationship between the road nodes and the parking lot nodes to obtain a space-time relationship topological graph of all the nodes. And then combining the heterographic neural network (HGNN) with the space-time relationship to complete the spatial feature fusion of the two types of nodes. Finally, the GRU (gate control loop unit) extracts the time characteristics and completes the prediction.

In FIG. 6, the output of the model

Predicting the result of average speed for each road in the road network,/->

The method is a prediction result of the occupation of the parking spaces of each parking lot in the parking lot network. The final predictor expression is:

wherein h is _t-1 Is the hidden state at the moment t-1, h _t It is the output state at time t that will be passed on to the next node.

In order to further verify the effect of the invention, the part of parking lot in Shenzhen Rohu area and the surrounding roads are selected as experimental scenes for experiments. In the experiment, average speed information of each road and parking space occupation data of each parking lot within 30 days are collected, and a Huarun Wansheng parking lot and adjacent Wansheng street thereof are selected as prediction objects in a plurality of parking lots and roads. The experimental scene is shown in fig. 7, each sign is a Huarun Wansheng parking lot, and the black lines are Wansheng streets.

In the experiment, the prediction Task of the traffic condition is completed in two ways, the first is to respectively predict the average speed of the road and the occupied number of parking spaces in a parking lot through a Single Task model (Single Task) proposed by other scholars, and the second is to complete the prediction work by utilizing the multi-Task integrated prediction model proposed by the invention.

Fig. 8 and 9 show 24-hour real data and predicted results for the vanity street and the vanity city parking lot, respectively. In fig. 8 and 9, real data, prediction data of the multitasking integrated prediction model proposed by the present invention, and prediction data based on a single-tasking model (comparative experiment) are shown, respectively. From the real data, the average speed fluctuation of one day on the street is large, and the change of the occupied number of the parking lot is relatively gentle. In addition, when the number of vehicles in the parking lot varies significantly in a short time, the average speed of the street is low, which indicates that the influence of parking in the burst mode on road traffic is great regardless of whether the vehicles enter or leave the parking lot.

Although the prediction results of the model of the present invention and the existing model are acceptable, they are substantially close to real data. However, experiments have found that when parking lot saturation is high, the predictive power of the single-task predictive model is significantly reduced. To find the cause, the model prediction accuracy is calculated every 15 time slices on the test set and the accuracy over time is obtained. The accuracy is calculated by the following formula:

FIG. 10 is a comparison of accuracy over the course of a day for a multitasking integrated prediction model and a single tasking prediction model. It can be seen that the accuracy of the two models is relatively close between 9 pm and 7 am, while at other times the accuracy of the multi-tasked integrated prediction model is significantly higher than that of the single-tasked prediction model, so that the variation in model accuracy is related to traffic conditions. Between 9 pm and 7 am, there are relatively few cars on the road and sufficient parking space, and the accuracy of the two models is very close. However, as the number of stops increases dramatically from 7:00 a.m., the prediction accuracy of the single-tasked prediction model decreases, and in this case, the accuracy of the multi-tasked integrated prediction model increases slightly. Thus, it is believed that the difference in accuracy of the two models may be caused by the correlation between road traffic and parking.

To confirm this, mutual information is used to measure the correlation between average speed on the road and the number of occupied parking spaces. According to the definition of mutual information, the larger the mutual information value is, the stronger the correlation of two variables is. The change of the mutual information of the actual traffic condition and the predicted traffic condition with time is shown in fig. 11, which illustrates the mutual information of the actual traffic condition, the mutual information of the predicted result of the single task prediction model, and the mutual information of the predicted result of the multi-task integrated prediction model.

As can be seen from fig. 11, the mutual information of the real data and the two predicted data is very small when the traffic flow is small, very close to each other, which means that in this case the road traffic and the parking are independent of each other, while the accuracy of the two models is also very close. When traffic becomes busy, the mutual information of the real traffic condition increases, and the single-task prediction model obviously cannot keep pace with the increase of the real mutual information, but the multi-task integrated prediction model can always keep close to the real mutual information. It is worth noting that at this time the prediction accuracy gap between the two models is just enlarged, so it is reasonable to infer that the mutual information between road traffic and parking plays a key role in traffic condition prediction.

Fig. 12 is a graph of model accuracy as a function of mutual information between road traffic and parking, further confirming this inference. It can be seen from the figure that the accuracy of the two models is not very different when the mutual information is small. When the mutual information is increased, the precision of the multi-task integrated prediction model is obviously better than that of a single-task prediction model. This means that separate predictions are only suitable for handling situations where traffic activity correlation is low, whereas in the face of complex real traffic environments, integrated predictions are more practical because they better understand the potential, subtle correlations between different traffic activities in the same spatio-temporal environment.

Compared with the prior art, the invention has at least the following advantages:

1) The invention considers that regional traffic is not only influenced by traffic flow but also influenced by parking saturation of the same region, and utilizes interaction between the two, provides a multi-task integrated prediction model for integrally predicting road traffic condition and parking condition of one region, wherein the model is a multiple-input multiple-output model, inputs historical data of traffic flow and parking space occupation, and outputs traffic flow and parking condition of the next time step to be predicted.

2) In some cases, there is a strong correlation between road traffic and parking. Especially in the vicinity of some hot points of interest, such as scenic spots, hospitals, traffic around large malls is very complex. How to quantify the association between road traffic and parking is currently not an explicit method. The invention utilizes a bidirectional gating circulation unit to extract the space-time relationship between traffic flow and parking and characterizes a weighted directed topological graph. This approach successfully quantifies the correlation between the spatiotemporal relationship of road traffic and parking, which is also an important reason for improving model accuracy.

3) The existing neural network generally adopts a splicing mode in a space feature fusion method, but the feature dimension after splicing is increased, so that a sparse matrix in a high-dimensional space is changed, and gradient is difficult to find when a model is trained. In addition, if two matrixes with certain association are spliced together to be used as new feature representation, information redundancy exists, and the invention provides a new feature fusion mode, namely spatial feature fusion is carried out according to a space-time relationship.

In summary, the regional traffic condition integrated prediction method based on space-time relation extraction provided by the invention synchronously predicts traffic conditions in a region, including traffic flow prediction and parking condition prediction. Compared with an independent prediction model, the integrated model provided by the invention can sense more information, and the model visual field is wider. In addition, the space-time relationship between traffic flow and parking condition is quantitatively characterized, and the space characteristic fusion is carried out according to the space-time relationship instead of adopting simple splicing. In addition, the spatial characteristics are captured by using the graph convolution neural network, and the spatial topological relation between roads and the topological relation between parking lots are fully considered. Further, the invention also verifies the influence of the correlation between traffic flow and parking on model accuracy.

The present invention may be a system, method, and/or computer program product. The computer program product may include a computer readable storage medium having computer readable program instructions embodied thereon for causing a processor to implement aspects of the present invention.

The computer readable storage medium may be a tangible device that can hold and store instructions for use by an instruction execution device. The computer readable storage medium may be, for example, but not limited to, an electronic storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of the computer-readable storage medium would include the following: portable computer disks, hard disks, random Access Memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static Random Access Memory (SRAM), portable compact disk read-only memory (CD-ROM), digital Versatile Disks (DVD), memory sticks, floppy disks, mechanical coding devices, punch cards or in-groove structures such as punch cards or grooves having instructions stored thereon, and any suitable combination of the foregoing. Computer-readable storage media, as used herein, are not to be construed as transitory signals per se, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., optical pulses through fiber optic cables), or electrical signals transmitted through wires.

The computer readable program instructions described herein may be downloaded from a computer readable storage medium to a respective computing/processing device or to an external computer or external storage device over a network, such as the internet, a local area network, a wide area network, and/or a wireless network. The network may include copper transmission cables, fiber optic transmissions, wireless transmissions, routers, firewalls, switches, gateway computers and/or edge servers. The network interface card or network interface in each computing/processing device receives computer readable program instructions from the network and forwards the computer readable program instructions for storage in a computer readable storage medium in the respective computing/processing device.

Computer program instructions for carrying out operations of the present invention may be assembly instructions, instruction Set Architecture (ISA) instructions, machine-related instructions, microcode, firmware instructions, state setting data, or source or object code written in any combination of one or more programming languages, including an object oriented programming language such as Smalltalk, c++, python, and the like, and conventional procedural programming languages, such as the "C" programming language or similar programming languages. The computer readable program instructions may be executed entirely on the user's computer, partly on the user's computer, as a stand-alone software package, partly on the user's computer and partly on a remote computer or entirely on the remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer through any kind of network, including a Local Area Network (LAN) or a Wide Area Network (WAN), or may be connected to an external computer (for example, through the Internet using an Internet service provider). In some embodiments, aspects of the present invention are implemented by personalizing electronic circuitry, such as programmable logic circuitry, field Programmable Gate Arrays (FPGAs), or Programmable Logic Arrays (PLAs), with state information for computer readable program instructions, which can execute the computer readable program instructions.

Various aspects of the present invention are described herein with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems) and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and/or block diagrams, and combinations of blocks in the flowchart illustrations and/or block diagrams, can be implemented by computer-readable program instructions.

These computer readable program instructions may be provided to a processor of a general purpose computer, special purpose computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions/acts specified in the flowchart and/or block diagram block or blocks. These computer readable program instructions may also be stored in a computer readable storage medium that can direct a computer, programmable data processing apparatus, and/or other devices to function in a particular manner, such that the computer readable medium having the instructions stored therein includes an article of manufacture including instructions which implement the function/act specified in the flowchart and/or block diagram block or blocks.

The computer readable program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other devices to cause a series of operational steps to be performed on the computer, other programmable apparatus or other devices to produce a computer implemented process such that the instructions which execute on the computer, other programmable apparatus or other devices implement the functions/acts specified in the flowchart and/or block diagram block or blocks.

The flowcharts and block diagrams in the figures illustrate the architecture, functionality, and operation of possible implementations of systems, methods and computer program products according to various embodiments of the present invention. In this regard, each block in the flowchart or block diagrams may represent a module, segment, or portion of instructions, which comprises one or more executable instructions for implementing the specified logical function(s). In some alternative implementations, the functions noted in the block may occur out of the order noted in the figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently, or the blocks may sometimes be executed in the reverse order, depending upon the functionality involved. It will also be noted that each block of the block diagrams and/or flowchart illustration, and combinations of blocks in the block diagrams and/or flowchart illustration, can be implemented by special purpose hardware-based systems which perform the specified functions or acts, or combinations of special purpose hardware and computer instructions. It is well known to those skilled in the art that implementation by hardware, implementation by software, and implementation by a combination of software and hardware are all equivalent.

The foregoing description of embodiments of the invention has been presented for purposes of illustration and description, and is not intended to be exhaustive or limited to the embodiments disclosed. Many modifications and variations will be apparent to those of ordinary skill in the art without departing from the scope and spirit of the various embodiments described. The terminology used herein was chosen in order to best explain the principles of the embodiments, the practical application, or the technical improvements in the marketplace, or to enable others of ordinary skill in the art to understand the embodiments disclosed herein. The scope of the invention is defined by the appended claims.

Claims

1. An area traffic condition integrated prediction method based on space-time relation extraction comprises the following steps:

based on the road network topological graph and the parking lot topological graph, obtaining an average vehicle speed prediction result of each road in a road network and a prediction result of each parking lot parking space occupation in a parking lot network by utilizing a constructed multi-task model framework, wherein the multi-task model framework comprises a space-time extraction network, a heterogeneous graph neural network and a time sequence model;

wherein the space-time extraction network is used for extracting average vehicle speed information according to a plurality of historical moments

And parking space occupation data of parking lot->

Extracting space-time relations between road nodes and parking area nodes to obtain a space-time relation topological graph of all nodes; for each moment, the spatial characteristics of road nodes and the characteristics of parking area nodes are respectively extracted by adopting two graph convolution neural networks, and then the spatial characteristic matrix of the road nodes and the characteristic matrix of the parking area nodes are fused through the heterogeneous graph neural networks, so that fusion characteristics of corresponding moments are obtained; based on the fusion characteristics of each moment, extracting time characteristics by adopting the time sequence model and outputting an average speed prediction result of each road in a road network at the subsequent moment and a prediction result of the occupation of each parking place of a parking place in the parking place network;

wherein the road network topology graph is marked as G _r ＝(V _r ,E _r ) Regarding the road as a graph node, if two roads intersect, connecting the two nodes, V _r ＝{v ₁ ,v ₂ ,…,v _N The number of nodes is represented by the number of nodes, E _r Representing a set of all edges, the connection relationship between the nodes adopts an adjacency matrix A _r Representation, A _r ∈R ^N×N Taking the average speed information on each road as the characteristics of the nodes, and marking the characteristics of each node as x _r ＝[x ₁ ,x ₂ ,…,x _t ,…,x _T ]T is the length of the historical time series, x _t Node characteristics representing the time t; marking the parking lot network topology map as G _p ＝(V _p ,E _p ) Regarding parking lots as nodes, connecting two parking lots if the shortest road distance between the two parking lots is smaller than a set threshold value, V _p ＝{v ₁ ,v ₂ ,…,v _M M is the number of parking lots, adjacent matrix A _p ∈R ^M×M Taking the occupied number of parking spaces in a parking lot as node characteristics;

the two graph convolution neural networks respectively extract the spatial characteristics of the road nodes and the characteristics of the parking area nodes according to the following steps:

for a two-layer graph roll-up neural network, the model is uniformly expressed as:

x is the feature matrix, A is the adjacency matrix,

and then (2) is in charge of>

Normalized to obtain->

Degree matrix, W ₀ And W is ₁ Is a weight matrix, σ () represents an activation function;

the spatial feature matrix of the road node and the feature matrix of the parking lot node extracted based on the two-layer graph convolution neural network are expressed as follows:

advanced feature representation representing all road nodes at time t, < >>

Is a high-level characteristic representation of all parking area nodes at the moment t, A _r Is an adjacency matrix of a road network topological graph, A _p Is an adjacency matrix of a parking lot network topology;

the spatial characteristics of the node i fused with the neighbor nodes are expressed as follows:

for the feature of the current node i, j is the neighbor node pointing to i, +.>

For the feature of node j, e is the set of edges consisting of j and i, α _e Is the weight of the edge, alpha _e ＝A _rp [j][i]，/>

A is the feature of the node i after being fused to the feature of the neighbor node _rp Is a weighted directed adjacency matrix in a heterogeneous graph neural network.

2. The method of claim 1, wherein the spatio-temporal extraction network comprises a bi-directional gating loop unit and an attention layer, wherein the attention layer weights different ones of the plurality of hidden states of the bi-directional gating loop unit.

3. The method according to claim 2, characterized in that the output v of the bi-directional gating loop unit is derived from a weighted summation of the respective hidden states, expressed as:

u _t ＝tanh(w _a1 h _t )

v＝∑ _t α _t h _t

wherein h is _t Is the output state at time t, w _a1 、w _a2 And alpha _t The weights of the corresponding terms, respectively, the final output of the spatio-temporal relationship extraction network is expressed as:

wherein W is _v Representing the weight, b _v Representing the bias.

4. The method of claim 1, wherein the time series model is a gated loop unit.

5. A computer readable storage medium having stored thereon a computer program, wherein the program when executed by a processor realizes the steps of the method according to any of claims 1 to 4.

6. A computer device comprising a memory and a processor, on which memory a computer program is stored which can be run on the processor, characterized in that the processor implements the steps of the method according to any of claims 1 to 4 when the program is executed.