CN118072518A - Traffic flow prediction method based on big data - Google Patents

Abstract

The application provides a traffic flow prediction method based on big data, which comprises the steps of dividing a monitoring area into M sample road sections by constructing a traffic flow prediction model, and determining historical traffic data (comprising corresponding traffic flow labels) corresponding to each sample road section by utilizing a historical traffic data set; determining real-time characteristics and global characteristics corresponding to each sample road section; performing feature fusion to obtain N training features corresponding to each sample road section, and forming a training data set containing MN training features; and constructing a GRU model, and training and testing to obtain a trained traffic flow prediction model. By combining the real-time features and the global features and adopting the GRU model for prediction, the accuracy of traffic flow prediction can be improved, and the prediction result is more reliable.

Description

Traffic flow prediction method based on big data

Technical Field

The present application relates to the field of big data technology, and in particular, to a traffic flow prediction method based on big data.

Background technique

With the continuous acceleration of urbanization, urban traffic congestion has become a problem that seriously affects people's daily life and economic development. Frequent traffic congestion leads to long commutes, aggravates vehicle exhaust emissions, and causes serious environmental pollution; at the same time, it also reduces people's quality of life and increases life pressure.

Traditional traffic management methods can usually only provide static information and lack the ability to respond to dynamically changing traffic conditions in a timely manner. Therefore, how to use modern information technology to improve traffic management (such as driving route planning) and improve the prediction accuracy of traffic flow has become an urgent problem to be solved in the current transportation field.

With the rapid development of big data technology, the transportation field has also begun to gradually apply big data technology to predict and manage traffic flow. With its powerful data processing and analysis capabilities, big data technology provides more diversified and real-time data support for traffic management departments. For example, by collecting vehicle trajectory data, road monitoring data, mobile terminal data, etc., traffic management departments can understand traffic conditions more accurately and take corresponding management measures in a timely manner. At the same time, big data technology can also discover the laws and trends of traffic flow changes through the analysis of massive data, providing a scientific basis for future traffic planning.

Although big data technology has brought new hope for traffic flow prediction, there are still problems in practical applications. The current big data processing and analysis technology has poor application effect in traffic flow prediction, and the prediction results are inaccurate, resulting in poor application effect (such as guidance on driving route planning and avoiding congestion in advance), which is not conducive to improving traffic congestion.

Summary of the invention

The purpose of the embodiments of the present application is to provide a traffic flow prediction method based on big data to improve the accuracy of traffic flow prediction, improve the applicability at the application level, and improve traffic congestion conditions.

In order to achieve the above purpose, the embodiments of the present application are implemented in the following ways:

In a first aspect, an embodiment of the present application provides a traffic flow prediction method based on big data, comprising: obtaining real-time traffic data of a target section, wherein the target section includes a monitoring section for flow prediction and an adjacent section connected to the monitoring section, and in each direction of travel of the target section, it is possible to start from the adjacent section at the starting point in the travel direction and then reach the adjacent section at the end point in the travel direction after passing through the monitoring section; preprocessing the real-time traffic data of the target section, and determining the real-time features corresponding to the target section; obtaining the global features corresponding to the target section, and performing feature fusion on the global features and real-time features corresponding to the target section to obtain the input features corresponding to the target section; inputting the input features corresponding to the target section into a preset traffic flow prediction model, and determining the traffic flow prediction result corresponding to the target section through the traffic flow prediction model.

In combination with the first aspect, in a first possible implementation method of the first aspect, before obtaining real-time traffic data, a traffic flow prediction model needs to be constructed, and the method of constructing the traffic flow prediction model is: obtaining a historical traffic data set, wherein the historical traffic data set contains historical traffic information collected in N consecutive time steps within the monitoring area; dividing the monitoring area to form M sample sections, and determining the historical traffic data corresponding to each sample section, wherein each piece of historical traffic data corresponding to each sample section contains a corresponding traffic flow label for revealing the traffic flow of the sample section at the time step, and each sample section contains a calibrated section and an adjacent section connected to the calibrated section, and in each direction of travel of the sample section, a traffic flow label can be obtained from the direction of travel. Starting from the adjacent road section at the starting point, passing through the calibrated road section, it arrives at the adjacent road section at the end point in the travel direction; preprocessing the historical traffic data corresponding to each sample road section in the historical traffic data set, and determining the real-time features and global features corresponding to each sample road section, and obtaining N real-time features corresponding to each sample road section and M global features corresponding to M sample road sections; feature fusion of the global features and real-time features corresponding to each sample road section, obtaining N training features corresponding to each sample road section, and forming a training data set containing MN training features; dividing the training data set into a training set and a test set; constructing a GRU model, and using the training set and the test set to train and test the GRU model, to obtain a trained traffic flow prediction model.

In combination with the first possible implementation of the first aspect, in a second possible implementation of the first aspect, the monitoring area is divided to form M sample sections, and the historical traffic data corresponding to each sample section is determined, including: dividing the monitoring area into sections to obtain M sections; for each section, taking the current section as a calibration section, determining the calibration section and an adjacent section as a sample section, and obtaining a total of M sample sections, wherein each sample section includes a calibration section and an adjacent section connected to the calibration section, and in each travel direction of the sample section, it is possible to start from the adjacent section at the starting point in the travel direction and pass through the calibration section to reach the adjacent section at the end point in the travel direction; for each sample section, the information corresponding to each piece of historical traffic information in the current sample section is used as the historical traffic data corresponding to the sample section, and each sample section corresponds to N pieces of historical traffic data, wherein each piece of historical traffic data includes a traffic flow label, which is determined by the traffic flow of the marked section in the current sample section at the corresponding time step.

In combination with the first possible implementation method of the first aspect, in a third possible implementation method of the first aspect, the historical traffic data corresponding to each sample section in the historical traffic data set are preprocessed, and the real-time features and global features corresponding to each sample section are determined, including: performing data cleaning on the historical traffic data corresponding to each sample section in the historical traffic data set; standardizing the historical traffic data corresponding to each cleaned sample section; performing real-time feature extraction on the historical traffic data corresponding to each standardized sample section, and determining N real-time features corresponding to the sample section; performing global feature extraction on the historical traffic data of the same sample section at all time steps, and determining M global features corresponding to the M sample sections.

In combination with the first possible implementation method of the first aspect, in a fourth possible implementation method of the first aspect, feature fusion is performed on the global features and real-time features corresponding to each sample road section to obtain N training features corresponding to each sample road section, and a training data set containing MN training features is formed, including: for each sample road section: using a fully connected layer to feature embed the global features and each real-time feature corresponding to the current sample road section to obtain N groups of real-time embedded features and a group of global embedded features corresponding to the current sample road section; for each group of real-time embedded features: feature fusion is performed on the real-time embedded features and the corresponding global embedded features to obtain the corresponding training features; all training features corresponding to the M sample road sections are integrated to form a training data set containing MN training features.

In combination with the fourth possible implementation manner of the first aspect, in a fifth possible implementation manner of the first aspect, a global feature corresponding to the current sample section and each real-time feature are feature embedded using a fully connected layer to obtain N groups of real-time embedded features and a group of global embedded features corresponding to the current sample section, including: for the i-th real-time feature corresponding to the current sample section j The global feature Y corresponding to the current sample road section j is embedded using the following formula:

Among them, i∈[1,N],j∈[1,M], For real-time features/> Feature representation after feature embedding, is the feature representation after feature embedding of the global feature Y ^j ,/> For real-time features/> The weight matrix for feature embedding, /> is the weight matrix for embedding the global feature Y ^j ,/> For real-time features/> Bias term when embedding features, /> is the bias term when embedding the global feature ^Yj , and σ is the activation function.

In combination with the fifth possible implementation of the first aspect, in a sixth possible implementation of the first aspect, the real-time embedding feature is fused with the corresponding global embedding feature to obtain the corresponding training feature, including: for the real-time embedding feature and global embedding features/> The following formula is used for feature fusion:

in, is the i-th training feature corresponding to sample road section j, i.e., the real-time embedding feature/> and global embedding features/> The fused training features, α is the attention weight, /> For real-time embedding features/> Nonlinear transformation of For global embedding features/> Nonlinear transformation of Embedding features for real-time/> The corresponding weight parameter, /> is the global embedding feature/> The corresponding weight parameter, /> Embedding features for real-time/> The lag term at time step t, t∈[1,N],/> is the global embedding feature/> The lag term in the sample segment s.

In combination with the sixth possible implementation manner of the first aspect, in a seventh possible implementation manner of the first aspect, the attention weight α satisfies:

in, is the k-th dimension parameter of the attention weight α at time step t, F(·) is the attention function, h ^(t-1) is the hidden state of the previous time step, /> is the k-th dimension parameter in the i-th training feature corresponding to the sample road segment j, C is the total number of dimensions of the training feature, /> is the lth dimension parameter in the i-th training feature corresponding to sample road segment j.

In combination with the sixth possible implementation manner of the first aspect, in an eighth possible implementation manner of the first aspect, the weight parameter satisfy:

in, is the weight parameter.

In conjunction with the sixth possible implementation of the first aspect, in a ninth possible implementation of the first aspect, the weight parameter satisfy:

in, is the weight parameter.

Beneficial effects:

1. This scheme builds a traffic flow prediction model, divides the monitoring area into M sample sections, and uses the historical traffic data set to determine the historical traffic data corresponding to each sample section (including the corresponding traffic flow label); preprocesses the historical traffic data and determines the real-time features and global features corresponding to each sample section; further performs feature fusion to obtain N training features corresponding to each sample section, forming a training data set containing MN training features; builds a GRU model, trains and tests, and obtains a trained traffic flow prediction model. By combining real-time features and global features and using the GRU model for prediction, the accuracy of traffic flow prediction can be improved and the prediction results can be made more reliable. Using real-time traffic data for prediction can better reflect the dynamic changes of traffic flow. The fusion of global features and real-time features can make full use of historical and current data, improve the model's perception of complex traffic conditions, and thus take into account the impact of emergencies such as traffic congestion and accidents, and improve the accuracy of traffic flow prediction.

2. Designing feature embedding and feature fusion ideas can enable the model to better utilize the relationship between global and real-time features, improve the effect of feature expression, and thus enhance the model's ability to comprehensively consider different factors. By weighted fusion of real-time embedded features and global embedded features, the model can simultaneously consider the real-time information at the current moment and the historical global information, thereby comprehensively utilizing the advantages of both. This helps to improve the model's ability to comprehensively grasp the characteristics of traffic data, allowing the model to understand road conditions more comprehensively. Introducing attention weights, dynamically learning the importance between real-time embedded features and global embedded features, and performing weighted fusion accordingly, this allows the model to adaptively focus on feature information that is more critical to the current prediction task, and improve the model's adaptability and generalization ability in different scenarios. When nonlinear transformations are performed on real-time embedded features and global embedded features, more nonlinear factors are introduced, enhancing the ability to model complex relationships between features, helping to improve the model's expression ability, so that it can better fit the nonlinear characteristics of the data and improve the accuracy of predictions. According to the formation process of training features and the way of regional division (the division of a sample road section includes the calibration road section and the adjacent road section), the lag term of the real-time embedded features and the lag term of the global embedded features are designed accordingly, which can help the model capture the time dependency and long-term memory of the data, better understand the time series characteristics of the data, and help improve the model's understanding and prediction capabilities of the time-varying trends of traffic data. The weight parameters of the real-time embedded features and the global embedded features can be learned through training data (the weight distribution scheme designed by this scheme can also be used), so that the weight parameters can be automatically adjusted according to the characteristics of the data, making the feature fusion process more flexible and adaptable to different data distribution situations.

3. The calculation formula of attention weight can enable the model to dynamically adjust the weight according to the importance of different features, and better focus on the information that is more critical to the current prediction task. By calculating the hidden state and training features, the model can determine the importance of each feature based on the relationship between historical information and current features, which helps to improve the model's sensitivity to data features and enhance the accuracy of prediction. Weight parameter The design uses the tanh function. As the difference between time step t and feature number i (training feature corresponding to the i-th time step) increases, the weight gradually decreases, which can make the model pay more attention to recent feature information and reduce its dependence on long-term features. This helps the model better capture the short-term change trend of data and improve its sensitivity to real-time information. Weight parameter/> Different values are taken according to the position s of the sample road section j in the overall sequence, where M is the total number of sample road sections. This enables the model to assign different weights to road sections at different positions (mainly to distinguish the influence of the preceding adjacent road sections and the succeeding adjacent road sections based on the travel direction), and can better distinguish the contribution of different road sections to the prediction task. It also enables the model to focus on the road section information that is more important for the overall prediction, improve the generalization ability of the model and the understanding of global information, and ultimately improve the accuracy of traffic flow prediction, improve the applicability at the application level (such as driving route planning, guiding intelligent traffic management design, etc.), and improve traffic congestion.

In order to make the above-mentioned objects, features and advantages of the present application more obvious and easy to understand, preferred embodiments are specifically cited below and described in detail with reference to the attached drawings.

BRIEF DESCRIPTION OF THE DRAWINGS

In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments of the present application will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

Figure 1 is a flow chart of building a traffic flow prediction model.

FIG2 is a flow chart of a traffic flow prediction method based on big data provided in an embodiment of the present application.

Detailed ways

The technical solutions in the embodiments of the present application will be described below in conjunction with the drawings in the embodiments of the present application.

Since the traffic flow prediction method based on big data of this scheme mainly relies on the traffic flow prediction model constructed by this scheme, in order to facilitate the understanding of this scheme, the process of constructing the traffic flow prediction model is first introduced here.

Please refer to Figure 1, which is a flow chart of constructing a traffic flow prediction model. In this embodiment, constructing a traffic flow prediction model may include steps S11, S12, S13, S14, S15, and S16.

In order to construct a traffic flow prediction model, step S11 may be executed first.

Step S11: Acquire a historical traffic data set, wherein the historical traffic data set includes historical traffic information collected in N consecutive time steps within the monitoring area.

In this embodiment, historical traffic information of N consecutive time steps in the monitoring area can be collected. Here, N consecutive time steps can be 1 minute, 10 minutes, etc. as a time step, without limitation. In theory, in order to maintain a high prediction accuracy, the time span of data collection is preferably 1 year or more. However, since the amount of data is too large, this embodiment takes a time span of 1 month as an example, and gradually improves the long-term features through continuous learning and updating of the model during use (that is, the global features can be updated and optimized as the model is used, or the global features are designed to be updated regularly). The historical traffic information includes data collected by sensors (such as traffic cameras, geomagnetic sensors, radars), GPS data (data sent by GPS devices on vehicles), geographic information system (GIS) data, weather data (rainfall, wind speed, visibility, etc.), event data (traffic accidents, construction projects, special activities, etc.), vehicle data (vehicle operation status and driving data, etc.), etc.

After obtaining the historical traffic information, step S12 may be executed.

Step S12: Divide the monitoring area to form M sample sections, and determine the historical traffic data corresponding to each sample section, wherein each piece of historical traffic data corresponding to each sample section contains a corresponding traffic flow label, which is used to reveal the traffic flow of the sample section at the time step. Each sample section includes a calibrated section and an adjacent section connected to the calibrated section. In each travel direction of the sample section, it is possible to start from the adjacent section at the starting point in the travel direction and then pass through the calibrated section to reach the adjacent section at the end point in the travel direction.

In this embodiment, in order to realize the flow prediction of the local road section in the monitoring area, it can be divided into M sample sections, and the historical traffic data corresponding to each sample section can be determined.

Exemplarily, the monitoring area may be divided into sections to obtain M sections.

For each road section, the current road section can be used as a calibration section, and the calibration section and adjacent sections (for example, in this embodiment, the section directly connected to the calibration section is used as the adjacent section. In other embodiments, in order to further improve the prediction accuracy, sections in a wider range can also be considered as adjacent sections, such as the three nearest sections connected to the calibration section in each travel direction as adjacent sections) are determined as a sample section, and a total of M sample sections are obtained, wherein each sample section includes a calibration section and adjacent sections connected to the calibration section. In each travel direction of the sample section, it is possible to start from the adjacent section at the starting point in the travel direction and then pass through the calibration section to reach the adjacent section at the end point in the travel direction.

Then, for each sample road section, the information corresponding to each piece of historical traffic information in the current sample road section is used as the historical traffic data corresponding to the sample road section, and each sample road section corresponds to N pieces of historical traffic data, wherein each piece of historical traffic data contains a traffic flow label, which is determined by the traffic flow of the marked section in the current sample road section at the corresponding time step. Thus, MN pieces of historical traffic data corresponding to M sample road sections can be obtained (each sample road section corresponds to N pieces of historical traffic data).

After the historical traffic data corresponding to each sample road section is determined, step S13 may be executed.

Step S13: pre-process the historical traffic data corresponding to each sample section in the historical traffic data set, and determine the real-time features and global features corresponding to each sample section, to obtain N real-time features corresponding to each sample section and M global features corresponding to M sample sections.

In this embodiment, data cleaning may be performed on the historical traffic data corresponding to each sample road section in the historical traffic data set, such as processing missing values, identifying and removing outliers, etc., to ensure the reliability and accuracy of data quality.

Then, the historical traffic data corresponding to each cleaned sample road section is standardized, such as logarithmic transformation, normalization and other operations.

After completing data standardization, real-time feature extraction can be performed on the historical traffic data corresponding to each standardized sample road section to determine N real-time features corresponding to the sample road section, and a total of MN real-time features can be obtained. Feature extraction is implemented here, such as aggregation, extraction of statistical features (such as mean, variance, etc.), and construction of spatiotemporal features (jointly constructing spatiotemporal features by combining adjacent sections and calibration sections in each direction of travel, and each direction of travel forms a part of the feature parameters), so as to integrate and obtain the real-time features corresponding to the sample road section.

Furthermore, global features can be extracted from the historical traffic data of the same sample road section at all time steps to determine the M global features corresponding to the M sample road sections. In the training phase, this embodiment takes the data with a time span of 1-3 months as an example to extract the global features of this part. After the model training is completed and put into use, the global features of each sample road section can be updated as the time span of the data is lengthened, thereby further improving the accuracy of the model as time goes by.

After obtaining N real-time features corresponding to each sample road section and M global features corresponding to M sample road sections, step S14 may be executed.

Step S14: performing feature fusion on the global features and real-time features corresponding to each sample road section to obtain N training features corresponding to each sample road section, thereby forming a training data set containing MN training features.

In this embodiment, for each sample road section:

The fully connected layer can be used to embed the global features and each real-time feature corresponding to the current sample road section, so as to obtain N groups of real-time embedded features and a group of global embedded features corresponding to the current sample road section.

For example, for the i-th real-time feature corresponding to the current sample road segment j The global feature Y corresponding to the current sample road section j can be embedded using the following formula:

Among them, i∈[1,N],j∈[1,M], For real-time features/> Feature representation after feature embedding, is the feature representation after feature embedding of the global feature Y ^j ,/> For real-time features/> The weight matrix for feature embedding, /> is the weight matrix for embedding the global feature Y ^j ,/> For real-time features/> Bias term when embedding features, /> is the bias term when embedding the global feature ^Yj , and σ is the activation function (the Sigmoid function is used as an example in this embodiment).

For each set of real-time embedding features: the real-time embedding features can be fused with the corresponding global embedding features to obtain the corresponding training features.

Specifically, for real-time embedding features and global embedding features/> The following formula can be used for feature fusion:

in, is the i-th training feature corresponding to sample road section j, i.e., the real-time embedding feature/> and global embedding features/> The fused training features, α is the attention weight, /> For real-time embedding features/> Nonlinear transformation of For global embedding features/> Nonlinear transformation of Embedding features for real-time/> The corresponding weight parameter, /> is the global embedding feature/> The corresponding weight parameter, /> Embedding features for real-time/> The lag term at time step t, t∈[1,N],/> is the global embedding feature/> The lag term in the sample segment s.

The attention weight α satisfies:

in, is the k-th dimension parameter of the attention weight α at time step t, F(·) is the attention function, h ^(t-1) is the hidden state of the previous time step, /> is the k-th dimension parameter in the i-th training feature corresponding to the sample road segment j, C is the total number of dimensions of the training feature, /> is the lth dimension parameter in the i-th training feature corresponding to sample road segment j.

Weight parameters satisfy:

Weight parameters satisfy:

Then, all training features corresponding to the M sample road sections may be integrated to form a training data set containing MN training features.

After obtaining the training data set, step S15 may be executed.

Step S15: Divide the training data set into a training set and a test set.

In this embodiment, the training data set is divided into a training set (accounting for 85%) and a test set (accounting for 15%) in an 8.5:1.5 division method. When dividing, it is necessary to divide it in equal proportion according to the sample sections to facilitate subsequent training and testing of each sample section in the monitoring area to avoid sample imbalance.

After obtaining the training set and the test set, step S16 may be executed.

Step S16: construct a GRU model, and use the training set and the test set to train and test the GRU model to obtain a trained traffic flow prediction model.

In this embodiment, a GRU model can be constructed. GRU (Gated Recurrent Unit) is a recurrent neural network model suitable for processing time series data. The update formulas of each gate in the model are as follows:

Reset the gate:

Where r ^(t) is the reset gate at time step t, σ is the activation function (also Sigmoid function), _Wr is the weight of the reset gate, h ^(t-1) is the hidden state at time step t-1, is the training feature at time step t corresponding to the sample road segment j (the training feature obtained in the previous article/> Rewritten as/> Parameter description of the adapted GRU model), _br is the bias of the reset gate.

Update Gate:

Where z ^(t) is the update gate at time step t, σ is the activation function (also Sigmoid function), _Wz is the weight of the update gate, h ^(t-1) is the hidden state at time step t-1, is the training feature at time step t corresponding to sample road segment j, and b _z is the bias of the update gate.

Candidate hidden states:

in, is the candidate hidden state at time step t, _Wk is the weight corresponding to the candidate hidden state, ⊙ is the element-by-element multiplication operation, and _bk is the bias corresponding to the candidate hidden state.

Hide status updates:

Where h ^(t) is the hidden state at time step t.

The mean square error is selected as the loss function of the GRU model.

Finally, the constructed GRU model is trained using the training set, and the GRU model is tested using the test set, and finally a trained traffic flow prediction model is obtained.

After obtaining the trained traffic flow prediction model, the trained traffic flow prediction model can be mounted on a server, and the running program of the traffic flow prediction method based on big data can be placed in the server, and the traffic flow prediction method based on big data can be run by the server.

Please refer to Figure 2, which is a flow chart of a traffic flow prediction method based on big data provided in an embodiment of the present application. In this embodiment, the traffic flow prediction method based on big data may include steps S21, S22, S23, and S24.

First, the server may execute step S21.

Step S21: Acquire real-time traffic data of the target road section, wherein the target road section includes a monitoring section for traffic prediction and an adjacent road section connected to the monitoring section. In each travel direction of the target road section, it is possible to start from the adjacent road section at the starting point in the travel direction and then reach the adjacent road section at the end point in the travel direction after passing through the monitoring section.

In this embodiment, in order to realize the traffic flow prediction of the monitored section, it is necessary to obtain the real-time traffic data of the target section (including the monitored section for flow prediction and the adjacent section connected to the monitored section. In each travel direction of the target section, it is possible to start from the adjacent section at the starting point in the travel direction and then reach the adjacent section at the end point in the travel direction after passing through the monitored section). The real-time traffic data includes data collected by sensors (such as traffic cameras, geomagnetic sensors, radars), GPS data (data sent by the GPS device on the vehicle), geographic information system (GIS) data, weather data (rainfall, wind speed, visibility, etc.), event data (traffic accidents, construction projects, special activities, etc.), vehicle data (vehicle operating status and driving data, etc.), etc., which correspond to the historical traffic data of the sample section in the previous text.

After obtaining the real-time traffic data of the target road section, the server may execute step S22.

Step S22: pre-processing the real-time traffic data of the target road section, and determining the real-time features corresponding to the target road section.

In this embodiment, the server may pre-process the real-time traffic data of the target road section, such as data cleaning and standardization, which will not be described in detail here, and then perform real-time feature extraction to obtain the real-time features corresponding to the target road section.

After obtaining the real-time features corresponding to the target road section, the server may execute step S23.

Step S23: obtaining the global features corresponding to the target road section, and performing feature fusion on the global features and real-time features corresponding to the target road section to obtain the input features corresponding to the target road section.

In this embodiment, the server may obtain the global features corresponding to the target road section. Since the target road section corresponds to a sample road section in the monitoring area, the global features corresponding to the sample road section may be directly obtained as the global features of the target road section.

After obtaining the global features corresponding to the target road section, the global features and real-time features corresponding to the target road section can be fused to obtain the input features corresponding to the target road section. The feature fusion process of the global features and the real-time features can be specifically referred to the process shown in step S14 above, which will not be repeated here.

After obtaining the input features, the server may execute step S24.

Step S24: inputting the input features corresponding to the target road section into a preset traffic flow prediction model, and determining the traffic flow prediction result corresponding to the target road section through the traffic flow prediction model.

In this embodiment, the input features can be input into a preset traffic flow prediction model, and the traffic flow prediction result corresponding to the target road section can be determined by the traffic flow prediction model. Of course, in actual application, if it is just beginning to be used (i.e., there are no previous prediction results and input data), it is best to perform continuous predictions for multiple times in order to obtain more accurate prediction results.

After obtaining the traffic flow prediction results, the traffic flow prediction results can be put into use in the application layer, such as route planning based on the traffic flow prediction results, or route congestion prompts, traffic management guidance, etc., which will not be further explained here.

In summary, the embodiment of the present application provides a traffic flow prediction method based on big data, by constructing a traffic flow prediction model, dividing the monitoring area into M sample sections, and using the historical traffic data set to determine the historical traffic data corresponding to each sample section (including the corresponding traffic flow label); preprocessing the historical traffic data, and determining the real-time features and global features corresponding to each sample section; further performing feature fusion to obtain N training features corresponding to each sample section, forming a training data set containing MN training features; constructing a GRU model, training and testing, and obtaining a trained traffic flow prediction model. By combining real-time features and global features, and using the GRU model for prediction, the accuracy of traffic flow prediction can be improved, and the prediction results can be made more reliable. Using real-time traffic data for prediction can better reflect the dynamic changes of traffic flow. The fusion of global features and real-time features can make full use of historical and current data, improve the model's perception of complex traffic conditions, and thus take into account the impact of emergencies such as traffic congestion and accidents, and improve the accuracy of traffic flow prediction.

Designing feature embedding and feature fusion ideas can enable the model to better utilize the relationship between global and real-time features, improve the effect of feature expression, and thus enhance the model's ability to comprehensively consider different factors. By weighted fusion of real-time embedded features and global embedded features, the model can simultaneously consider the real-time information at the current moment and the historical global information, thereby comprehensively utilizing the advantages of both. This helps to improve the model's ability to comprehensively grasp the characteristics of traffic data, allowing the model to understand road conditions more comprehensively. Introducing attention weights, dynamically learning the importance between real-time embedded features and global embedded features, and performing weighted fusion accordingly, this allows the model to adaptively focus on feature information that is more critical to the current prediction task, improving the model's adaptability and generalization ability in different scenarios. When nonlinear transformations are performed on real-time embedded features and global embedded features, more nonlinear factors are introduced, enhancing the ability to model complex relationships between features, helping to improve the model's expression ability, enabling it to better fit the nonlinear characteristics of data and improve the accuracy of predictions. According to the formation process of training features and the way of regional division (the division of a sample road section includes the calibration road section and the adjacent road section), the lag term of the real-time embedded features and the lag term of the global embedded features are designed accordingly, which can help the model capture the time dependency and long-term memory of the data, better understand the time series characteristics of the data, and help improve the model's understanding and prediction capabilities of the time-varying trends of traffic data. The weight parameters of the real-time embedded features and the global embedded features can be learned through training data (the weight distribution scheme designed by this scheme can also be used), so that the weight parameters can be automatically adjusted according to the characteristics of the data, making the feature fusion process more flexible and adaptable to different data distribution situations.

The calculation formula of attention weights allows the model to dynamically adjust weights according to the importance of different features, and better focus on information that is more critical to the current prediction task. By calculating the hidden state and training features, the model can determine the importance of each feature based on the relationship between historical information and current features, which helps to improve the model's sensitivity to data features and enhance the accuracy of predictions. Weight parameters The design uses the tanh function. As the difference between time step t and feature number i (training feature corresponding to the i-th time step) increases, the weight gradually decreases, which can make the model pay more attention to recent feature information and reduce its dependence on long-term features. This helps the model better capture the short-term change trend of data and improve its sensitivity to real-time information. Weight parameter/> Different values are taken according to the position s of the sample road section j in the overall sequence, where M is the total number of sample road sections. This enables the model to assign different weights to road sections at different positions (mainly to distinguish the influence of the preceding adjacent road sections and the succeeding adjacent road sections based on the travel direction), and can better distinguish the contribution of different road sections to the prediction task. It also enables the model to focus on the road section information that is more important for the overall prediction, improve the generalization ability of the model and the understanding of global information, and ultimately improve the accuracy of traffic flow prediction, improve the applicability at the application level (such as driving route planning, guiding intelligent traffic management design, etc.), and improve traffic congestion.

In this document, relational terms such as first and second, etc. are used merely to distinguish one entity or operation from another entity or operation, but do not necessarily require or imply any such actual relationship or order between these entities or operations.

The above description is only an embodiment of the present application and is not intended to limit the scope of protection of the present application. For those skilled in the art, the present application may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the scope of protection of the present application.

Claims (10)

1. A traffic flow prediction method based on big data, comprising:

Acquiring real-time traffic data of a target road section, wherein the target road section comprises a monitoring road section for flow prediction and an adjacent road section communicated with the monitoring road section, and in each passing direction of the target road section, the adjacent road section positioned at a starting point in the passing direction can start from the adjacent road section positioned at the starting point in the passing direction, passes through the monitoring road section and then reaches the adjacent road section positioned at a terminal point in the passing direction;

preprocessing real-time traffic data of a target road section, and determining real-time characteristics corresponding to the target road section;

acquiring global features corresponding to the target road segments, and carrying out feature fusion on the global features corresponding to the target road segments and the real-time features to obtain input features corresponding to the target road segments;

And inputting the input characteristics corresponding to the target road section into a preset traffic flow prediction model, and determining a traffic flow prediction result corresponding to the target road section through the traffic flow prediction model.

2. The traffic flow prediction method based on big data according to claim 1, wherein before acquiring real-time traffic data, a traffic flow prediction model is constructed by:

Acquiring a historical traffic data set, wherein the historical traffic data set comprises historical traffic information acquired by N continuous time steps in a monitoring area;

Dividing a monitoring area to form M sample road sections, and determining historical traffic data corresponding to each sample road section, wherein each historical traffic data corresponding to each sample road section comprises a corresponding traffic flow label used for revealing the traffic flow of the sample road section in the time step, each sample road section comprises a calibration road section and an adjacent road section communicated with the calibration road section, and in each passing direction of the sample road section, the adjacent road section which is positioned at a starting point in the passing direction can be started from the adjacent road section which is positioned at the starting point in the passing direction, passes through the calibration road section and then reaches the adjacent road section which is positioned at an end point in the passing direction;

preprocessing historical traffic data corresponding to each sample road section in the historical traffic data set, and determining real-time features and global features corresponding to each sample road section to obtain N real-time features corresponding to each sample road section and M global features corresponding to M sample road sections;

Feature fusion is carried out on the global features and the real-time features corresponding to each sample road section, N training features corresponding to each sample road section are obtained, and a training data set containing MN training features is formed;

dividing the training data set into a training set and a testing set;

And constructing a GRU model, and training and testing the GRU model by using a training set and a testing set to obtain a trained traffic flow prediction model.

3. The traffic flow prediction method based on big data according to claim 2, wherein dividing the monitoring area to form M sample segments, determining historical traffic data corresponding to each sample segment, includes:

Carrying out road section division on the monitoring area to obtain M road sections;

For each road section, taking the current road section as a calibration road section, determining that the calibration road section and the adjacent road sections are one sample road section, and obtaining M sample road sections in total, wherein each sample road section comprises one calibration road section and the adjacent road sections communicated with the calibration road section, and in each passing direction of the sample road section, the adjacent road sections positioned at the starting point in the passing direction can reach the adjacent road sections positioned at the end point in the passing direction after passing through the calibration road sections from the adjacent road sections positioned at the starting point in the passing direction;

And aiming at each sample road section, taking the information corresponding to each piece of historical traffic information in the current sample road section as historical traffic data corresponding to the sample road section, wherein each sample road section corresponds to N pieces of historical traffic data, each piece of historical traffic data comprises a traffic flow label, and the traffic flow of the marked road section in the current sample road section under the corresponding time step is determined.

4. The big data based traffic flow prediction method according to claim 2, wherein preprocessing the historical traffic data corresponding to each sample section in the historical traffic data set and determining the real-time feature and the global feature corresponding to each sample section comprises:

Carrying out data cleaning on the historical traffic data corresponding to each sample road section in the historical traffic data set;

Normalizing the historical traffic data corresponding to each cleaned sample road section;

Carrying out real-time feature extraction on the standardized historical traffic data corresponding to each sample road section, and determining N real-time features corresponding to the sample road sections;

and carrying out global feature extraction on the historical traffic data of the same sample road section in all time steps, and determining M global features corresponding to the M sample road sections.

5. The traffic flow prediction method based on big data according to claim 2, wherein the feature fusion is performed on the global feature and the real-time feature corresponding to each sample road segment to obtain N training features corresponding to each sample road segment, and a training data set including MN training features is formed, and the method includes:

for each sample segment: performing feature embedding on the global features corresponding to the current sample road section and each real-time feature by using a full connection layer to obtain N groups of real-time embedded features and a group of global embedded features corresponding to the current sample road section;

For each set of real-time embedded features: performing feature fusion on the real-time embedded features and the corresponding global embedded features to obtain corresponding training features;

And integrating all training features corresponding to the M sample road sections to form a training data set containing the MN training features.

6. The traffic flow prediction method based on big data according to claim 5, wherein feature embedding is performed on global features and each real-time feature corresponding to a current sample section by using a full connection layer, so as to obtain N sets of real-time embedded features and a set of global embedded features corresponding to the current sample section, and the method comprises:

ith real-time feature corresponding to current sample segment j The global feature Y corresponding to the current sample section j is embedded by adopting the following formula:

Wherein i is E [1, N ], j is E [1, M ], To real-time feature/>Feature representation after feature embedding,/>For feature representation after feature embedding of global feature Y ^j,/>To real-time feature/>Weight matrix during feature embedding,/>For the weight matrix when the global feature Y ^j is embedded with the features,/>To real-time feature/>Bias term in feature embeddingFor bias terms when feature embedding is performed on global feature Y ^j, σ is the activation function.

7. The traffic flow prediction method based on big data according to claim 6, wherein feature fusion is performed between the real-time embedded feature and the corresponding global embedded feature to obtain the corresponding training feature, comprising:

For real-time embedded features And globally embedded features/>The following formula is used for feature fusion:

wherein, For the ith training feature corresponding to the sample section j, namely the real-time embedded feature/>And globally embedded features/>Fused training features, alpha is the attention weight,/>To embed features/>, in real timeIs used for the non-linear transformation of (a),For the purpose of global embedded features/>Nonlinear transformation of,/>For embedding features in real time/>The corresponding weight parameter is used to determine the weight of the object,For global embedded features/>Corresponding weight parameter,/>For embedding features in real time/>Hysteresis term at time step t, t.epsilon.1, N,/>For global embedded features/>Hysteresis term at sample segment s.

8. The big data based traffic flow prediction method according to claim 7, wherein the attention weight α satisfies:

wherein, The kth dimension parameter of the attention weight alpha at the time step t is F (·) is the attention function, h ^(t-1) is the hidden state of the last time step,/>For the kth dimension parameter in the ith training feature corresponding to the sample section j, C is the total number of dimensions of the training feature,/>And the first dimension parameter in the ith training feature corresponding to the sample section j.

9. The traffic flow prediction method based on big data according to claim 7, wherein the weight parameter isThe method meets the following conditions:

wherein, Is a weight parameter.

10. The traffic flow prediction method based on big data according to claim 7, wherein the weight parameter isThe method meets the following conditions:

wherein, Is a weight parameter.