Traffic index prediction method and device

Abstract

The present application discloses a traffic index prediction method, which obtains historical traffic index data of a target traffic area in multiple historical time windows; classifies the historical traffic index data to obtain discrete fixed feature data and continuous time-varying feature data; inputs the discrete fixed feature data into a first embedding model in a traffic index prediction model to obtain a latent vector of discrete features, and inputs the continuous time-varying feature data into a second embedding model in a traffic index prediction model to obtain a high-dimensional latent vector of continuous features; uses a feature fusion layer in the traffic index prediction model to fuse the latent vector of discrete features with the high-dimensional latent vector of continuous features to obtain spatiotemporal features; uses a prediction layer in the traffic index prediction model to predict spatiotemporal features to obtain traffic index prediction data of the target traffic area in a future time window. The present application improves the training speed of the model and effectively improves the prediction effect of traffic indicators.

Description

Technical Field

[0001] The present application relates to the field of intelligent transportation technology, and in particular to a method and device for predicting traffic indicators. Background Art

[0002] Traffic indicators are various quantitative indicators used to describe the status of transportation systems, traffic efficiency and traffic quality. Traffic indicators play an important role in analyzing and evaluating the transportation status and traffic congestion level of a city, region or transportation network. By measuring and analyzing traffic indicators, transportation planners, traffic managers and decision makers can better understand the traffic situation so as to optimize and improve it.

[0003] In related technologies, the prediction of traffic indicators is mainly achieved by pure time series models and space-time graph models. The pure time series model establishes a time series prediction model by mining the relationship between the traffic indicators of the previous time window and the traffic indicators of the current time window in historical data to achieve the prediction of traffic flow in the future. This solution often cannot make good use of various forms of traffic indicator data, and the performance bottleneck of the model is caused by insufficient information; the space-time graph model often requires complete data of all facilities in the area as model input. When data is sparse and missing, the model performance will be affected. At the same time, it is not compatible with changes in traffic networks and road conditions. When the traffic network changes or the road conditions change suddenly, the space-time graph model needs to be expanded and updated, which may bring additional computing and time costs. Summary of the invention

[0004] The purpose of this application is to provide a traffic indicator prediction method that makes full use of historical data information, efficiently extracts the characteristic information implicit in the data, and effectively improves the prediction effect of traffic indicators.

[0005] Based on the above purpose, the present application provides a traffic index prediction method, which includes:

[0006] Obtain historical traffic index data of the target traffic area within multiple historical time windows;

[0007] Preprocessing the historical traffic index data to obtain discrete fixed feature data and continuous time-varying feature data;

[0008] Input the discrete fixed feature data into the first embedding model in the traffic index prediction model to obtain the latent vector of the discrete feature, and input the continuous time-varying feature data into the second embedding model in the traffic index prediction model to obtain the high-dimensional latent vector of the continuous feature;

[0009] The feature fusion layer in the traffic index prediction model is used to fuse the latent vector of discrete features with the high-dimensional latent vector of continuous features to obtain spatiotemporal features.

[0010] The prediction layer in the traffic index prediction model is used to predict the spatiotemporal characteristics to obtain the traffic index prediction data of the target traffic area in the future time window, wherein the historical time window and the future time window are continuous in time.

[0011] Furthermore, the discrete fixed feature data includes the facility data of the target traffic area and the timestamp information of each historical time window. The discrete fixed feature data is input into the first embedding model in the traffic index prediction model to obtain the latent vector of the discrete feature, including:

[0012] The facility data is input into the first embedding model to perform feature dimension reduction processing to obtain a latent vector of the facility data;

[0013] The timestamp information of each historical time window is input into the first embedding model for feature dimensionality reduction processing to obtain the latent vector of the timestamp information.

[0014] Furthermore, the first embedding model includes a multi-layer perceptron, wherein the facility data is input into the first embedding model for feature dimension reduction transformation to obtain a latent vector of the facility data, including:

[0015] Encoding the facility data to obtain the facility code of the target traffic area;

[0016] Establish an index for the facility code, encode the facility code using one-hot encoding, and obtain the facility code vector;

[0017] Inputting the timestamp information of each of the historical time windows into the first embedding model for feature dimensionality reduction transformation to obtain a latent vector of the timestamp information, including:

[0018] For the timestamp information corresponding to each historical time window, each timestamp is cyclically encoded to determine the day cycle code and week cycle code corresponding to each timestamp;

[0019] Create an index for each day cycle code, and use one-hot encoding to obtain a day cycle code vector for each day cycle code;

[0020] Create an index for each weekly code, and use one-hot encoding to obtain a weekly code vector for each weekly code.

[0021] The facility coding vector, the day cycle coding vector and the period coding vector are input into the multilayer perceptron in parallel to obtain the latent vector of the facility coding, the latent vector of the day cycle coding and the latent vector of the week cycle coding.

[0022] Furthermore, the first embedding model includes a facility encoder, which encodes the facility data to obtain the facility code of the target traffic area, including:

[0023] Based on the facility data, encoding processing is performed through a facility encoder to obtain the facility code of the target traffic area, wherein the facility data is at least one of the following: intersection data, road section data, lane data, direction data, lane type data, and the facility code is at least one of the following: intersection number, road section number, lane number, direction number, lane type number.

[0024] Furthermore, the first embedding model includes a day cycle encoder and a week cycle encoder, and for each timestamp information corresponding to each historical time window, cyclic encoding is performed on each timestamp to determine the day cycle code and week cycle code corresponding to each timestamp, including:

[0025] The day cycle encoder determines the total amount of time windows of one day based on the duration of the historical time window, determines the day number corresponding to each timestamp based on the timestamp corresponding to each historical time window, and cyclically encodes each timestamp according to each day number and the total amount of time windows to determine the day cycle code corresponding to each timestamp;

[0026] The weekly cycle encoder determines the week number corresponding to each timestamp based on the timestamp corresponding to each historical time window, and cyclically encodes each timestamp according to each week number and the total number of days in a week to determine the weekly cycle code corresponding to each timestamp.

[0027] Furthermore, the second embedding model adopts a WaveNet neural network model, which includes a causal convolution layer and a stacked dilated convolution layer, each dilated convolution layer includes a dilated convolution, a gated unit, a skip connection and a residual connection, wherein the continuous time-varying feature data is input into the second embedding model of the traffic index prediction model to obtain a high-dimensional latent vector of the continuous feature, including:

[0028] The continuous time-varying feature data is processed by the causal convolution layer to obtain time series data, which is then input into the stacked dilated convolution layer.

[0029] For each dilated convolution layer, time features of time series data are extracted through dilated convolution. The gating unit inputs the output of the dilated convolution into the hyperbolic tangent function and the S-type function respectively. The output results of the two functions are multiplied element by element to complete the skip-layer connection. The skip-layer connection and the one-dimensional output of the causal convolution layer are added element by element to form a residual. The residual result is input into the next dilated convolution layer. The output results of each dilated convolution layer are connected together and then pass through the activation function and convolution operation to obtain a high-dimensional latent vector of continuous features.

[0030] Furthermore, the feature fusion layer includes L stacked convolutional layers, each of which includes a fully connected layer and a residual connection, wherein the latent vector of the discrete feature is fused with the high-dimensional latent vector of the continuous feature to obtain the spatiotemporal features, including:

[0031] Stack the latent vector of discrete features with the high-dimensional latent vector of continuous features to obtain a fused feature vector;

[0032] For each convolution layer, the fused feature vector is linearly mapped through the fully connected layer to obtain the output vector. The output vector is summed and fused with the vector calculated by the residual connection and mapped to the next convolution layer. The output vector and the vectors output by each convolution layer are connected through a skip layer and then passed through the activation function and convolution operation to obtain the spatiotemporal features.

[0033] Furthermore, the historical traffic index data is preprocessed to obtain discrete fixed feature data and continuous time-varying feature data, including:

[0034] Based on the data attributes of historical traffic index data, it is determined that the data that can be exhaustively enumerated in the historical traffic index data is discrete fixed feature data, and the data that changes in real time in the historical traffic index data is continuous time-varying feature data, wherein the continuous time-varying feature data is at least one of the following: traffic data, speed data, density data, delay data, green-to-letter ratio data, saturation data, and travel time data.

[0035] Based on the above purpose, the present application provides a traffic index prediction device, which includes:

[0036] An acquisition module, used to acquire historical traffic index data of a target traffic area within multiple historical time windows;

[0037] A preprocessing module is used to preprocess the historical traffic index data to obtain discrete fixed feature data and continuous time-varying feature data;

[0038] A feature extraction module is used to input discrete fixed feature data into a first embedding model in a traffic index prediction model to obtain a latent vector of discrete features, and input continuous time-varying feature data into a second embedding model in a traffic index prediction model to obtain a high-dimensional latent vector of continuous features;

[0039] A feature fusion module is used to use the feature fusion layer in the traffic index prediction model to fuse the latent vector of discrete features with the high-dimensional latent vector of continuous features to obtain spatiotemporal features;

[0040] The result prediction module is used to predict the spatiotemporal characteristics using the prediction layer in the traffic index prediction model to obtain traffic index prediction data for the target traffic area in the future time window, wherein the historical time window and the future time window are continuous in time.

[0041] Based on the above objectives, the present application provides a computer device, including:

[0042] A processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other via the communication bus;

[0043] Memory, used to store computer programs;

[0044] The processor is used to implement the steps of the traffic index prediction method as described above when executing the computer program stored in the memory.

[0045] The present application uses the first embedding model and the second embedding model of the traffic index prediction model to perform parallel encoding processing on discrete fixed feature data and continuous time-varying feature data respectively, and uses the feature fusion layer to perform feature fusion processing on the feature vectors output by the two different embedding models. This multi-feature and multi-mode embedding modeling method not only makes full use of historical data information and efficiently extracts the implicit feature information in various data, but also improves the training speed of the model and reduces the complexity of the model, thereby effectively improving the prediction effect of traffic indicators. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 It is a schematic diagram of the system architecture provided according to an embodiment of the present application;

[0047] Figure 2 It is a first flow chart of the traffic index prediction method provided in an embodiment of the present application;

[0048] Figure 3 is a first structural schematic diagram of a traffic index prediction model provided according to an embodiment of the present application;

[0049] Figure 4 is a second structural schematic diagram of a traffic index prediction model provided according to an embodiment of the present application;

[0050] Figure 5 is a second flow chart of the traffic index prediction method provided in an embodiment of the present application;

[0051] Figure 6 is a third flow chart of the traffic index prediction method provided in an embodiment of the present application;

[0052] Figure 7is a schematic diagram of a network structure of a second embedding model provided according to an embodiment of the present application;

[0053] Figure 8 is a schematic diagram of the network structure of the feature fusion layer provided according to an embodiment of the present application;

[0054] Fig. 9 is a schematic diagram of the structure of a traffic index prediction device provided according to an embodiment of the present application;

[0055] Fig.10 It is a schematic diagram of the structure of a computer device provided according to an embodiment of the present application. DETAILED DESCRIPTION

[0056] The present application will be described in detail below in conjunction with the specific implementation modes shown in the accompanying drawings, but these implementation modes do not limit the present application. Structural, methodological, or functional changes made by ordinary technicians in the field based on these implementation modes are included in the protection scope of the present application.

[0057] Please refer to Figure 1 , the system architecture of traffic index prediction provided by the embodiment of the present application. Figure 1 As shown, the traffic index prediction method provided in the embodiment of the present application can be applied in the system, and the data acquisition device 101 is used to collect training data. For the traffic index prediction method in the embodiment of the present application, the training data can be historical traffic index data. After collecting the training data, the data acquisition device 101 can store the training data in the database 102, and the training device 103 can obtain the target model based on the training data maintained in the database 102. For example, the data acquisition device 101 can be a traffic detection device installed on the side of the road to obtain the traffic index data of the road section.

[0058] The training device 103 trains the preset neural network based on the historical traffic index data until the preset neural network meets the preset conditions to obtain the trained target model. The preset conditions may be: the total loss value of the target loss function is less than the preset value, the total loss value of the target loss function does not change, or the number of training times reaches the preset number of times, etc.

[0059] The above-mentioned target model can be used to implement the traffic index prediction method of the embodiment of the present application. The target model in the embodiment of the present application can specifically be a deep neural network model, for example, a WaveNet network (waveform network). It should be noted that in actual applications, the training data maintained in the database 102 does not necessarily all come from the collection of the data acquisition device 101, but may also be received from other devices. It should also be noted that the training device 103 does not necessarily train the target model entirely based on the training data maintained by the database 102, and it is also possible to obtain training data from the cloud or other places for model training. The above description should not be used as a limitation on the embodiments of the present application.

[0060] The target model obtained by training the training device 103 can be applied to different systems or devices, such as Figure 1 The execution device 104 shown may be a terminal, such as a mobile phone terminal, a tablet computer, a laptop computer, etc., or a server or a cloud, etc. For example, the computing module of the execution device 310 executes the target model, and the target model generates the predicted traffic index in the future time window based on the historical traffic index data corresponding to the historical time window. It is worth noting that the training device 103 can generate corresponding target models based on different training data for different goals or different tasks, and the corresponding target models can be used to achieve the above goals or complete the above tasks, thereby providing the user with the desired results.

[0061] It is worth noting that Figure 1 It is only a schematic diagram of a system architecture provided by an embodiment of the present application. The system architecture and application scenarios described in the embodiment of the present invention are intended to more clearly illustrate the technical solution of the embodiment of the present invention, and do not constitute a limitation on the technical solution provided by the embodiment of the present invention. A person of ordinary skill in the art will know that with the evolution of the system architecture and the emergence of new application scenarios, the technical solution provided by the embodiment of the present invention is also applicable to similar technical problems.

[0062] See also Figure 2 , Figure 2 FIG. 1 is a flow chart of a traffic index prediction method provided by an embodiment of the present application. In a specific embodiment, the traffic index prediction method is applied to Fig. 9 The traffic state prediction device shown in Fig.10 Computer equipment shown.

[0063] Taking a computer device as an example to illustrate the specific process of this embodiment, it can be understood that the computer device used in this embodiment can be a server or a terminal, etc. The server can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN, blockchain, and big data and artificial intelligence platforms. The terminal can be a smart phone, tablet computer, laptop computer, desktop computer, smart speaker, smart watch, smart voice interaction device, smart home appliance, vehicle terminal, aircraft, etc., but is not limited to this.

[0064] The traffic index prediction method includes steps S201 to S205.

[0065] Step S201, obtaining historical traffic index data of a target traffic area in multiple historical time windows.

[0066] Among them, the above-mentioned target traffic area refers to: an intersection on the road to be predicted, or several intersections on the road to be predicted, or multiple intersections on certain roads to be predicted. This can be determined according to actual needs, and this application does not limit this.

[0067] In this embodiment, the historical time window may be a period of historical time including the current time, wherein the duration of the historical time window may be any duration. The duration of the historical time window may be one day or one minute. For example, when the duration is divided by day, the days before the current day are used as the historical time windows.

[0068] The traffic index data of a road intersection represents the traffic conditions of the road intersection within a certain period of time. In an embodiment of the present application, the historical traffic index data may be at least one of the following: information of the target traffic area, traffic volume, driving speed, vehicle density, delay, green-to-letter ratio, traffic saturation of vehicle traffic, and travel time. The green-to-letter ratio is the proportion of time that can be used for vehicle traffic within a signal cycle, that is, the ratio of the effective green light time to the signal cycle. The effective green light time refers to the time that can be used for vehicles to pass at a saturated flow rate within a signal cycle, that is, the sum of the actual green light time, the yellow light time, and the full red time. The green-to-letter ratio can reflect the degree of change of the actual green light time. The traffic volume of a road intersection to be tested can represent the total number of vehicles passing through each historical time window before the current moment. The vehicle density of a road intersection to be tested can represent the ratio of the total number of vehicles passing through each historical time window before the current moment to the length of the road.

[0069] For example, the traffic flow at a road intersection to be tested can be obtained based on vehicle images taken by a camera at the road intersection to be tested within a historical time window. For example, target recognition is performed based on the captured vehicle images to obtain the traffic flow at each moment, thereby obtaining the traffic flow within each historical time window.

[0070] In the embodiments of the present application, full use is made of various types of historical traffic index data to make the input data of the traffic index prediction model more sufficient, so that the information implicit in various features can be efficiently extracted through the model, effectively improving the prediction effect of the prediction model.

[0071] Step S202: pre-process the historical traffic index data to obtain discrete fixed feature data and continuous time-varying feature data.

[0072] In this embodiment, the historical traffic index data of each historical time window is preprocessed, and the historical traffic index data is classified according to data attributes of the historical traffic index data to obtain discrete fixed feature data and continuous time-varying feature data.

[0073] Exemplarily, based on the data attributes of historical traffic index data, the data that can be exhaustively enumerated in the historical traffic index data is determined to be discrete fixed feature data, and the data that changes in real time in the historical traffic index data is determined to be continuous time-varying feature data. Continuous time-varying feature data is at least one of the following: traffic data, speed data, density data, delay data, green-to-letter ratio data, saturation data, and travel time data. The characteristics of discrete fixed feature data are features with a limited range of values ​​and exhaustive labels, such as intersection data, road section data, lane data, direction data, and lane type data. These data generally do not change. Continuous time-varying feature data refers to attribute data that changes in real time, and its characteristics refer to features whose values ​​are continuous real numbers. This type of data is often impossible to enumerate. For example, vehicle flow, driving speed, vehicle density, delays, green-to-letter ratio, vehicle flow saturation, and travel time. These data change with time.

[0074] Step S203, input the discrete fixed feature data into the first embedding model of the traffic index prediction model to obtain the latent vector of the discrete feature, and input the continuous time-varying feature data into the second embedding model of the traffic index prediction model to obtain the high-dimensional latent vector of the continuous feature.

[0075] In this embodiment, the traffic index prediction model is a trained model. Figure 3The structural diagram of the traffic index prediction model shown in the figure includes a first embedding model, a second embedding model, a feature fusion layer and a prediction layer. Discrete fixed feature data is input into the first embedding model of the traffic index prediction model, and the latent vector of the corresponding discrete feature is obtained through the encoding process of the first embedding model. Continuous time-varying feature data is input into the second embedding model of the traffic index prediction model, and the high-dimensional latent vector of the corresponding continuous feature is obtained through the encoding process of the second embedding model. The first embedding model is different from the second embedding model.

[0076] In this embodiment, the traffic index prediction model adopts different embedding structures for data with discrete features and data with continuous features. The channels between them are independent, allowing each channel to perform convolution operations independently, thereby reducing the amount of model calculation and the number of model parameters. Because the convolution operation of each channel is relatively simple, the model can converge to a better solution more quickly, making the model easier to optimize and deploy.

[0077] Step S204, based on the feature fusion layer of the traffic index prediction model, the latent vector of the discrete feature is fused with the high-dimensional latent vector of the continuous feature to obtain the spatiotemporal feature.

[0078] In the feature fusion layer of the traffic index prediction model, the latent vector of the discrete feature is fused with the high-dimensional latent vector of the continuous feature to output the spatiotemporal feature. The spatiotemporal feature represents the characteristic information of the traffic state. For example, the latent vector of the discrete feature is stacked with the high-dimensional latent vector of the continuous feature to obtain the spatiotemporal feature.

[0079] Step S205, using the prediction layer in the traffic index prediction model to predict the spatiotemporal characteristics, to obtain traffic index prediction data for the target traffic area in the future time window, wherein the historical time window and the future time window are continuous in time.

[0080] In this embodiment, the future time window may include the next moment of the current moment and a period of time thereafter, that is, the historical time window and the future time window may be continuous in time. The duration of the future time window may be a preset duration. The traffic index prediction data may be at least one of the following: vehicle flow, driving speed, vehicle density, delay, green-to-signal ratio, vehicle flow saturation, and travel time.

[0081] Exemplarily, the prediction layer can be a fully connected layer or a multi-layer perceptron. A multi-layer perceptron is a feed-forward artificial neural network model, mainly composed of an input layer, a hidden layer, and an output layer, each layer is composed of a number of neurons, and each neuron has a set of weights and biases.

[0082] Exemplarily, the spatiotemporal features are processed by activation functions and multi-layer perceptrons to output the final prediction results, that is, the traffic index prediction data of the target traffic area in the future time window is obtained.

[0083] In the embodiment of the present application, the first embedding model and the second embedding model of the traffic index prediction model respectively perform parallel encoding processing on discrete fixed feature data and continuous time-varying feature data, and the feature vectors output by the two different embedding models are subjected to feature fusion processing through the feature fusion layer. The traffic index prediction model adopts a model structure of two embedding models in parallel and feature fusion in series, which not only makes full use of historical data information and efficiently extracts the implicit feature information in various data, but also improves the training speed of the model, reduces the complexity of the model, and effectively improves the prediction effect of the model. This multi-feature and multi-mode embedding modeling method, on the one hand, utilizes the label information of discrete fixed feature data, and on the other hand, improves the extraction capability of continuous time series data, makes full use of the two types of data information, and improves the adaptability of the prediction model to traffic scenarios.

[0084] As an optional implementation method, the discrete fixed feature data includes the facility data of the target traffic area and the timestamp information of each historical time window. The discrete fixed feature data is input into the first embedding model in the traffic index prediction model to obtain the latent vector of the discrete feature, including:

[0085] The facility data is input into the first embedding model to perform feature dimension reduction processing to obtain a latent vector of the facility data;

[0086] The timestamp information of each historical time window is input into the first embedding model for feature dimensionality reduction processing to obtain the latent vector of the timestamp information.

[0087] In this embodiment, the facility data and timestamp information in the discrete fixed feature data are input into the first embedding model to obtain the latent vector of the corresponding facility data and the latent vector of the timestamp information, and to extract the latent vector of the discrete features in the historical traffic index data.

[0088] like Figure 4 The first embedding model shown in FIG. 1 includes a multi-layer perceptron. As an optional implementation, Figure 5 As shown, the facility information is input into the first embedding model for feature dimensionality reduction transformation to obtain a latent vector of the facility data, and the timestamp information of each historical time window is input into the first embedding model for feature dimensionality reduction transformation to obtain a latent vector of the timestamp information, including:

[0089] Step S501, encoding the facility data to obtain the facility code of the target traffic area;

[0090] Step S502, indexing the facility code, encoding the facility code using one-hot encoding, and obtaining a facility code vector;

[0091] Step S503, for each timestamp information corresponding to each historical time window, cyclically encode each timestamp to determine the day cycle code and week cycle code corresponding to each timestamp;

[0092] Step S504, creating an index for each day cycle code, and using one-hot encoding to obtain a day cycle code vector for each day cycle code;

[0093] Step S505, creating an index for each weekly code, and using one-hot encoding to obtain a weekly code vector for each weekly code;

[0094] Step S506, input the facility coding vector, the day cycle coding vector and the period coding vector into the multilayer perceptron in parallel to obtain the latent vector of the facility coding, the latent vector of the day cycle coding and the latent vector of the week cycle coding.

[0095] In this embodiment, the facility data and sampling timestamp information in the historical traffic index data are encoded respectively. After the facility code, daily cycle code and weekly cycle code are indexed and one-hot encoded, they are input into a multi-layer perceptron for feature dimensionality reduction to obtain the corresponding latent vectors. This can efficiently extract the feature information implied in the discrete feature data in the historical traffic index data, and effectively improve the prediction effect of the model.

[0096] As an optional implementation, Figure 4As shown, the first embedding model includes a facility encoder, and the discrete fixed feature data includes the facility data of the target traffic area, and the facility data is at least one of the following: intersection data, road section data, lane data, direction data, and lane type data. Based on the facility data of the target traffic area, the facility code of the target traffic area is obtained by encoding through the encoder, wherein the facility code of the target traffic area is at least one of the following: intersection number, road section number, lane number, direction number, and lane type number. Corresponding codes are constructed for the facility data to distinguish different prediction objects for the prediction model. For example, if the traffic index prediction is performed in units of intersections, a unique code needs to be established for the intersection. If the traffic index prediction is performed in units of lanes, a unique code needs to be established for each lane to facilitate the prediction model to distinguish different prediction objects. It is understandable that the facility code of the target traffic area can be multiple. By embedding the facility code in the time series feature to make the model compatible with spatial information, the model can recognize that the prediction samples come from different facilities, and has strong robustness to data conditions, avoiding the disadvantage that once the data quality of a single node in the road network is poor, it will affect the prediction effect of all facilities.

[0097] As an optional implementation, Figure 6 and Figure 4 As shown, the first embedding model includes a day cycle encoder and a week cycle encoder. For the timestamp information corresponding to each historical time window, each timestamp is cyclically encoded to determine the day cycle code and week cycle code corresponding to each timestamp, including:

[0098] Step S601, the day cycle encoder determines the total amount of time windows of one day based on the duration of the historical time window, determines the day number corresponding to each time stamp based on the timestamp corresponding to each historical time window, performs cyclic encoding on each time stamp according to each day number and the total amount of time windows, and determines the day cycle code corresponding to each time stamp;

[0099] Step S602, the weekly cycle encoder determines the week number corresponding to each timestamp based on the timestamp corresponding to each historical time window, cyclically encodes each timestamp according to each week number and the total number of days in a week, and determines the weekly cycle code corresponding to each timestamp.

[0100] Since road traffic is cyclical, traffic index data is a type of data with a periodic change pattern. Therefore, the length of the historical time window can be set according to the actual traffic conditions and the data collection situation. For example, the length of the historical time window is set to 5 minutes. Since traffic flow often operates in cycles of one day and one week, this embodiment constructs two types of cyclic encodings, daily cycle and weekly cycle, based on the time window of data collection. Discrete features can be converted into dense vector form by cyclic encoding, so that the model can better capture the intrinsic characteristics of the data. The cyclic encoding method enables the model to establish the time dependence of sequence data, capture the dynamic characteristics in the time series, fully consider the periodic characteristics of traffic flow, and improve the embedding effect.

[0101] In a specific embodiment of the present application, the calculation formula of the day cycle code is:

[0102] ;

[0103] Among them, T Tid is the day cycle code, T is the total time window, t is the day number corresponding to the historical time window, and the day number is used to indicate the number of the time window corresponding to the historical time window. For example, taking the duration of the historical time window as 5 minutes as an example, a day is 24 hours, then the total time window T is 288, and the timestamp corresponding to a certain historical time window is 6:00-6:05, which corresponds to the 73rd time window, that is, the corresponding day number is 73, and the day cycle code corresponding to the historical time window is 73 / 288=0.253. Based on this implementation, the recognition of the data collection timestamp information can be realized, so that the model can recognize the day cycle period corresponding to the historical traffic index data.

[0104] In a specific embodiment of the present application, the calculation formula of the cycle code is:

[0105] ;

[0106] Among them, T Diw is the weekly cycle code, W is the total number of days in a week, d is the week number corresponding to each historical time window, and the week number is used to indicate the day of the week corresponding to the historical time window. For example, the historical time window corresponds to Monday, and the total number of days in a week is seven days. The weekly cycle code corresponding to the historical time window is 1 / 7=0.143. Based on this implementation, the recognition of the data collection timestamp information can be realized, so that the model can recognize the weekly cycle period corresponding to the historical traffic index data.

[0107] As an optional implementation method, the second embedding model uses a WaveNet neural network model, inputs continuous time-varying feature data into the WaveNet neural network model, extracts time features, and obtains a high-dimensional latent vector of continuous features. There is a temporal correlation between continuous time-varying feature data. The present application uses a WaveNet neural network model to increase the receptive field of the model, that is, to increase the extraction of long-term temporal correlations, and capture the temporal dependence of traffic indicator data through the WaveNet neural network model. The basic structure of the WaveNet neural network model is causal convolution and dilated convolution. The significance of causal convolution is to ensure that the model does not violate the order of data when modeling data, so that the model's prediction at time t does not depend on any future data. Dilated convolution is a convolution method that skips the input value when the convolution kernel is applied to data with an area larger than its own length for convolution. WaveNet can model a longer receptive field without information leakage. On the basis of causal convolution and dilated convolution, residual and parameterized skip connections are used in the network structure of the WaveNet model to accelerate convergence so that gradient descent can propagate in deeper models.

[0108] like Figure 7 The schematic diagram of the WaveNet neural network model structure shown in FIG. 1 shows a WaveNet neural network model including a causal convolution layer and a stacked dilated convolution layer. Each dilated convolution layer includes a dilated convolution, a gating unit, a skip connection, and a residual connection. The continuous time-varying feature data is processed by the causal convolution layer to obtain time series data. The time series data is input into the stacked dilated convolution layer. For each dilated convolution layer, the time feature is extracted through the dilated convolution layer. The gate unit inputs the output of the dilated convolution layer into the hyperbolic tangent function tanh and the S-type function sigmoid (σ) respectively. The output results of the two functions are multiplied element by element to complete the skip connection. The skip connection and the one-dimensional output of the causal convolution layer are added element by element to form a residual. The residual result is input into the next dilated convolution layer. The output results of each dilated convolution layer are all connected through the skip connection and then through the activation function and convolution operation to obtain a high-dimensional latent vector of the continuous feature. The WaveNet neural network model uses a stacked dilated convolution layer so that the model meets causality and has a flexible receptive field size. The continuous time-varying feature data is input into the WaveNet neural network model. After passing through the causal convolution layer, the continuous time-varying feature data enters the stacked convolution layer, and the time information is extracted by expanding the convolution layer inside each stacked convolution layer. The residual and skip layer connection methods can enable information to flow freely, retain the details and semantic information in the original input data, avoid the problem of gradient vanishing or gradient diffusion, and avoid information loss, so as to improve the feature capture performance of the model.

[0109] As an optional implementation method, as shown in the schematic diagram of the feature fusion layer in Figure 8, the feature fusion layer includes L stacked convolutional layers, each convolutional layer includes a fully connected layer and a residual connection, the latent vector of the discrete feature is stacked with the high-dimensional latent vector of the continuous feature to obtain a fused feature vector, and the fused feature vector is obtained by vector stacking, that is, the latent vector of the facility code, the latent vector of the day cycle code, the latent vector of the week cycle code, and the high-dimensional latent vector of the continuous feature are stacked, and the fused feature vector is linearly mapped through the fully connected layer to obtain an output vector, and the output vector is summed and fused with the vector calculated by the residual connection and mapped to the next convolutional layer, and the output vector and the vector output by each convolutional layer are skipped and connected together, and then successively pass through the activation function and convolution operation to obtain the spatiotemporal feature. Based on this embodiment, the spatiotemporal information implicit in the high-dimensional space can be further extracted through stacked convolution operations and residual connections.

[0110] The traffic index prediction method provided in the embodiment of the present application includes training of a preset traffic index prediction model. It is worth noting that the training of the preset traffic index prediction model can be pre-performed based on the acquired training sample data set. Each time a traffic index prediction is required, the trained traffic index prediction model can be used to directly predict the traffic index without the need to perform network training again each time a traffic index prediction is performed. The acquired historical traffic index data is used as a training sample data set. For example, the constructed samples are divided into a training set, a validation set, and a test set in a ratio of 6:2:2 to train and verify the model.

[0111] During the training process, the mean absolute error, root mean square error, and mean absolute percentage error between the traffic indicator prediction results output by the traffic indicator prediction model and the actual traffic indicator results are calculated. The target loss function is optimized by the Adam optimizer, and the network parameters of the model are updated by back propagation and gradient descent. At the same time, the validation set is input into the traffic indicator prediction model, and the mean absolute error, root mean square error, and mean absolute percentage error between the traffic indicator prediction results output based on the validation set and the actual traffic indicator results are calculated to evaluate the model, so as to select the model with the best performance on the validation set as the final traffic indicator prediction model.

[0112] Among them, the calculations of mean absolute error MAPE, root mean square error MAE, and mean absolute percentage error RMSE are as follows:

[0113] ;

[0114] ;

[0115] ;

[0116] Where n is the number of samples, The real result of traffic indicators. Forecast results of traffic indicators.

[0117] like Fig. 9 A traffic index prediction device is shown, the device comprising:

[0118] An acquisition module 901 is used to acquire historical traffic index data of a target traffic area within multiple historical time windows;

[0119] A preprocessing module 902 is used to preprocess the historical traffic index data to obtain discrete fixed feature data and continuous time-varying feature data;

[0120] The feature extraction module 903 is used to input the discrete fixed feature data into the first embedding model in the traffic index prediction model to obtain the latent vector of the discrete feature, and input the continuous time-varying feature data into the second embedding model in the traffic index prediction model to obtain the high-dimensional latent vector of the continuous feature;

[0121] A feature fusion module 904 is used to use the feature fusion layer in the traffic index prediction model to fuse the latent vector of the discrete feature with the high-dimensional latent vector of the continuous feature to obtain the spatiotemporal feature;

[0122] The result prediction module 905 is used to predict the spatiotemporal characteristics using the prediction layer in the traffic index prediction model to obtain traffic index prediction data for the target traffic area in the future time window, wherein the historical time window and the future time window are continuous in time.

[0123] In this embodiment, the discrete fixed feature data and the continuous time-varying feature data are respectively encoded in parallel by the first embedding model and the second embedding model of the traffic index prediction model, and the feature vectors output by the two different embedding models are fused through the feature fusion layer, which makes full use of historical data information, efficiently extracts the implicit information in various features, improves the training speed of the model, reduces the complexity of the model, and effectively improves the prediction effect of the model.

[0124] Optionally, the preprocessing module 902 is specifically used for:

[0125] Based on the data attributes of historical traffic index data, it is determined that the data that can be exhaustively enumerated in the historical traffic index data is discrete fixed feature data, and the data that changes in real time in the historical traffic index data is continuous time-varying feature data, wherein the continuous time-varying feature data is at least one of the following: traffic data, speed data, density data, delay data, green-to-letter ratio data, saturation data, and travel time data.

[0126] Optionally, the feature extraction module 903 is specifically used for:

[0127] The facility data is input into the first embedding model to perform feature dimension reduction processing to obtain a latent vector of the facility data;

[0128] The timestamp information of each of the historical time windows is input into the first embedding model for feature dimensionality reduction processing to obtain a latent vector of the timestamp information.

[0129] Optionally, the first embedding model includes a multi-layer perceptron, and the feature extraction module 903 is specifically used for:

[0130] Encoding the facility data to obtain the facility code of the target traffic area;

[0131] Establish an index for the facility code, encode the facility code using one-hot encoding, and obtain the facility code vector;

[0132] For the timestamp information corresponding to each historical time window, each timestamp is cyclically encoded to determine the day cycle code and week cycle code corresponding to each timestamp;

[0133] Create an index for each day cycle code, and use one-hot encoding to obtain a day cycle code vector for each day cycle code;

[0134] Create an index for each weekly code, and use one-hot encoding to obtain a weekly code vector for each weekly code.

[0135] The facility coding vector, the day cycle coding vector and the period coding vector are input into the multilayer perceptron in parallel to obtain the latent vector of the facility coding, the latent vector of the day cycle coding and the latent vector of the week cycle coding.

[0136] Optionally, the first embedding model includes a facility encoder, and the feature extraction module 903 is specifically used to:

[0137] Based on the facility data, encoding processing is performed through a facility encoder to obtain the facility code of the target traffic area, wherein the facility data is at least one of the following: intersection data, road section data, lane data, direction data, lane type data, and the facility code is at least one of the following: intersection number, road section number, lane number, direction number, lane type number.

[0138] Optionally, the first embedding model includes a day cycle encoder and a week cycle encoder, and the feature extraction module 903 is specifically used for:

[0139] The day cycle encoder determines the total amount of time windows of one day based on the duration of the historical time window, determines the day number corresponding to each timestamp based on the timestamp corresponding to each historical time window, and cyclically encodes each timestamp according to each day number and the total amount of time windows to determine the day cycle code corresponding to each timestamp;

[0140] The weekly cycle encoder determines the week number corresponding to each timestamp based on the timestamp corresponding to each historical time window, and cyclically encodes each timestamp according to each week number and the total number of days in a week to determine the weekly cycle code corresponding to each timestamp.

[0141] Optionally, the second embedding model adopts a WaveNet neural network model, the WaveNet neural network model includes a causal convolution layer and a stacked dilated convolution layer, each dilated convolution layer includes a dilated convolution, a gating unit, a skip layer connection and a residual connection, and the feature extraction module 903 is specifically used for:

[0142] The continuous time-varying feature data is processed by the causal convolution layer to obtain time series data, which is then input into the stacked dilated convolution layer.

[0143] For each dilated convolution layer, time features of time series data are extracted through dilated convolution. The gating unit inputs the output of the dilated convolution into the hyperbolic tangent function and the S-type function respectively, and completes the skip-layer connection after multiplying the output results of the two functions element by element. The skip-layer connection and the one-dimensional output of the causal convolution layer are added element by element to form a residual. The residual result is input into the next dilated convolution layer. The output results of each dilated convolution layer are all skip-layer connected and then subjected to activation function and convolution operation to obtain a high-dimensional latent vector of continuous features.

[0144] Optionally, the feature fusion layer includes L stacked convolutional layers, each convolutional layer includes a fully connected layer and a residual connection, and the feature fusion module 904 is specifically used for:

[0145] Stack the latent vector of discrete features with the high-dimensional latent vector of continuous features to obtain a fused feature vector;

[0146] For each convolution layer, the fused feature vector is linearly mapped through the fully connected layer to obtain the output vector. The output vector is summed and fused with the vector calculated by the residual connection and mapped to the next convolution layer. The output vector and the vectors output by each convolution layer are connected through a skip layer and then passed through the activation function and convolution operation to obtain the spatiotemporal features.

[0147] An embodiment of the present application also provides a computer-readable storage medium, in which a computer program is stored. When the computer program is executed by a processor, any one of the above-mentioned traffic index prediction methods is implemented.

[0148] Fig.10 It is a schematic diagram of the hardware structure of a computer device provided in an embodiment of the present application. Fig.10 The computer device shown includes: a processor 1001, a communication interface 1002, a memory 1003 and a communication bus 1004. The processor 1001, the communication interface 1002 and the memory 1003 communicate with each other via the communication bus 1004. Fig.10 The connection method between the processor 1001, the communication interface 1002, and the memory 1003 shown is merely exemplary. During implementation, the processor 1001, the communication interface 1002, and the memory 1003 may also be connected to each other in communication with each other using other connection methods besides the communication bus 1004.

[0149] The memory 1003 can be used to store a computer program, which may include instructions and data to implement the steps of any of the above traffic index prediction methods. In the embodiment of the present application, the memory 1003 may be various types of storage media, such as random access memory (RAM), read only memory (ROM), non-volatile RAM (NVRAM), programmable ROM (PROM), erasable PROM (EPROM), electrically erasable PROM (EEPROM), flash memory, optical storage and registers, etc. The memory 1003 may include a hard disk and / or a memory.

[0150] The processor 1001 may be a general-purpose processor, which may be a processor that performs specific steps and / or operations by reading and executing a computer program (e.g., a computer program) stored in a memory (e.g., memory 1003). The general-purpose processor may use data stored in the memory (e.g., memory 1003) in the process of performing the steps and / or operations. The general-purpose processor may be, for example, but not limited to, a central processing unit (CPU). In addition, the processor 1001 may also be a special-purpose processor, which may be a processor specially designed to perform specific steps and / or operations. The special-purpose processor may be, for example, but not limited to, an ASIC and an FPGA. In addition, the processor 1001 may also be a combination of multiple processors, such as a multi-core processor.

[0151] The communication interface 1002 may include an input / output (I / O) interface, a physical interface, and a logical interface for interconnecting devices within the network device, as well as an interface for interconnecting the network device with other devices (e.g., network devices). The communication network may be Ethernet, a radio access network (RAN), a wireless local area network (WLAN), etc. The communication interface 1002 may be a module, a circuit, a transceiver, or any device capable of implementing communication.

[0152] In the implementation process, each step of the above method can be completed by an integrated logic circuit of hardware in the processor 1301 or an instruction in the form of software. The method disclosed in conjunction with the embodiment of the present application can be directly embodied as a hardware processor for execution, or a combination of hardware and software modules in the processor for execution. The software module can be located in a mature storage medium in the field such as a random access memory flash memory, a read-only memory, a programmable read-only memory or an electrically erasable programmable memory, a register, etc. The storage medium is located in the memory 1003, and the processor 1001 reads the information in the memory 1003 and completes the steps of the above method in conjunction with its hardware. To avoid repetition, it is not described in detail here.

[0153] Although the preferred embodiments of the present application have been disclosed for illustrative purposes, those skilled in the art will appreciate that various modifications, additions and substitutions are possible, without departing from the scope and spirit of the application as disclosed in the accompanying claims.

Claims

1. A traffic index prediction method, characterized in that: The method comprises: Obtain historical traffic index data of the target traffic area within multiple historical time windows; Preprocessing the historical traffic index data to obtain discrete fixed feature data and continuous time-varying feature data; Inputting the discrete fixed feature data into a first embedding model in a traffic index prediction model to obtain a latent vector of the discrete feature, and inputting the continuous time-varying feature data into a second embedding model in the traffic index prediction model to obtain a high-dimensional latent vector of the continuous feature; Using the feature fusion layer in the traffic index prediction model, the latent vector of the discrete feature is fused with the high-dimensional latent vector of the continuous feature to obtain a spatiotemporal feature; The prediction layer in the traffic index prediction model is used to predict the spatiotemporal characteristics to obtain traffic index prediction data of the target traffic area in a future time window, wherein the historical time window and the future time window are continuous in time; The discrete fixed feature data includes the facility data of the target traffic area and the timestamp information of each of the historical time windows, the first embedding model includes a multi-layer perceptron, and the discrete fixed feature data is input into the first embedding model in the traffic index prediction model to obtain the latent vector of the discrete feature, including: Performing encoding processing on the facility data to obtain the facility code of the target traffic area; Establishing an index for the facility code, encoding the facility code using one-hot encoding to obtain a facility code vector; For the timestamp information corresponding to each historical time window, each timestamp is cyclically encoded to determine the day cycle code and week cycle code corresponding to each timestamp; Create an index for each day cycle code, and use one-hot encoding to obtain a day cycle code vector for each day cycle code; Create an index for each weekly code, and use one-hot encoding to obtain a weekly code vector for each weekly code. The facility coding vector, the day cycle coding vector and the period coding vector are input into the multilayer perceptron in parallel to obtain a latent vector of the facility coding, a latent vector of the day cycle coding and a latent vector of the week cycle coding.

2. The traffic index prediction method according to claim 1, characterized in that: The first embedding model includes a facility encoder, and the encoding process of the facility data to obtain the facility code of the target traffic area includes: Based on the facility data, encoding processing is performed through the facility encoder to obtain the facility code of the target traffic area, wherein the facility data is at least one of the following: intersection data, road section data, lane data, direction data, lane type data, and the facility code is at least one of the following: intersection number, road section number, lane number, direction number, lane type number.

3. The traffic index prediction method according to claim 1, characterized in that: The first embedding model includes a day cycle encoder and a week cycle encoder. For the timestamp information corresponding to each historical time window, each timestamp is cyclically encoded to determine the day cycle code and the week cycle code corresponding to each timestamp, including: The day cycle encoder determines the total amount of time windows of one day based on the duration of the historical time window, determines the day number corresponding to each time stamp based on the timestamp corresponding to each historical time window, performs cyclic encoding on each time stamp according to each day number and the total amount of time windows, and determines the day cycle code corresponding to each time stamp; The weekly cycle encoder determines the week number corresponding to each timestamp based on the timestamp corresponding to each historical time window, and cyclically encodes each timestamp according to each week number and the total number of days in a week to determine the weekly cycle code corresponding to each timestamp.

4. The traffic index prediction method according to claim 1, characterized in that: The second embedding model adopts a WaveNet neural network model, and the WaveNet neural network model includes a causal convolution layer and a stacked dilated convolution layer, each dilated convolution layer includes a dilated convolution, a gate unit, a skip layer connection and a residual connection, wherein the continuous time-varying feature data is input into the second embedding model of the traffic index prediction model to obtain a high-dimensional latent vector of the continuous feature, including: The continuous time-varying feature data is processed by a causal convolution layer to obtain time series data, and the time series data is input into a stacked dilated convolution layer; For each dilated convolution layer, time features of the time series data are extracted by dilated convolution, the gating unit inputs the output of the dilated convolution into the hyperbolic tangent function and the S-type function respectively, and completes the skip-layer connection after multiplying the output results of the two functions element by element. The skip-layer connection and the one-dimensional output of the causal convolution layer are added element by element to form a residual, and the residual result is input into the next dilated convolution layer. The output results of each dilated convolution layer are all connected through the skip-layer connection and then through the activation function and convolution operation to obtain a high-dimensional latent vector of continuous features.

5. The traffic index prediction method according to claim 1, characterized in that: The feature fusion layer includes L stacked convolutional layers, each of which includes a fully connected layer and a residual connection, wherein the latent vector of the discrete feature is fused with the high-dimensional latent vector of the continuous feature to obtain the spatiotemporal feature, including: Stacking the latent vector of the discrete feature with the high-dimensional latent vector of the continuous feature to obtain a fused feature vector; For each convolution layer, the fused feature vector is linearly mapped through the fully connected layer to obtain an output vector, the output vector is summed and fused with the vector calculated by the residual connection and mapped to the next convolution layer, the output vector and the vectors output by each convolution layer are skip-connected and then passed through the activation function and convolution operation to obtain the spatiotemporal features.

6. The traffic index prediction method according to claim 1, characterized in that: The preprocessing of the historical traffic index data to obtain discrete fixed feature data and continuous time-varying feature data includes: Based on the data attributes of the historical traffic index data, it is determined that the data that can be exhaustively enumerated in the historical traffic index data is discrete fixed feature data, and it is determined that the data that changes in real time in the historical traffic index data is continuous time-varying feature data, wherein the continuous time-varying feature data is at least one of the following: traffic data, speed data, density data, delay data, green-to-signal ratio data, saturation data, and travel time data.

7. A traffic index prediction device, characterized in that: The device comprises: An acquisition module, used to acquire historical traffic index data of a target traffic area within multiple historical time windows; A preprocessing module, used for preprocessing the historical traffic index data to obtain discrete fixed feature data and continuous time-varying feature data; A feature extraction module, used for inputting the discrete fixed feature data into a first embedding model in the traffic index prediction model to obtain a latent vector of the discrete feature, and inputting the continuous time-varying feature data into a second embedding model in the traffic index prediction model to obtain a high-dimensional latent vector of the continuous feature; A feature fusion module, used to use the feature fusion layer in the traffic index prediction model to fuse the latent vector of the discrete feature with the high-dimensional latent vector of the continuous feature to obtain a spatiotemporal feature; A result prediction module, used to predict the spatiotemporal features using the prediction layer in the traffic index prediction model, to obtain traffic index prediction data of the target traffic area in a future time window, wherein the historical time window and the future time window are continuous in time; The discrete fixed feature data includes the facility data of the target traffic area and the timestamp information of each of the historical time windows, the first embedding model includes a multi-layer perceptron, and the feature extraction module is specifically used for: Performing encoding processing on the facility data to obtain the facility code of the target traffic area; Establishing an index for the facility code, encoding the facility code using one-hot encoding to obtain a facility code vector; For the timestamp information corresponding to each historical time window, each timestamp is cyclically encoded to determine the day cycle code and week cycle code corresponding to each timestamp; Create an index for each day cycle code, and use one-hot encoding to obtain a day cycle code vector for each day cycle code; Create an index for each weekly code, and use one-hot encoding to obtain a weekly code vector for each weekly code. The facility coding vector, the day cycle coding vector and the period coding vector are input into the multilayer perceptron in parallel to obtain a latent vector of the facility coding, a latent vector of the day cycle coding and a latent vector of the week cycle coding.

8. A computer device, characterized in that: include: A processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other via the communication bus; The memory is used to store computer programs; The processor is used to implement the steps of the traffic index prediction method as described in any one of claims 1-6 when executing the computer program stored in the memory.

Images