Urban traffic flow prediction method, system and product under complex road network conditions

Abstract

The application belongs to the field of traffic flow prediction, and provides a city traffic flow prediction method, system and product under complex road network conditions, and the technical scheme is as follows: time mixing features are extracted based on city road network traffic flow time series data, intersection information between different traffic features is captured based on the time mixing features, mixed feature representation is obtained by fusing the time mixing features and the intersection information between different traffic features; global dependence in a long sequence is captured by selective state space modeling based on the mixed feature representation, and dynamic time series features are output; after the dynamic time series features are enhanced, a trailing time dimension is projected to obtain prediction values of each feature variable of the traffic time series data. In city traffic flow prediction, data can be dynamically processed according to real-time traffic conditions, multivariate information can be effectively utilized, and efficient long sequence modeling can be performed, and the prediction accuracy of time series data is significantly improved.

Description

Technical Field

[0001] This invention belongs to the field of traffic flow prediction, and in particular relates to methods, systems and products for predicting urban traffic flow under complex road network conditions. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] For urban traffic flow forecasting, a typical time series forecasting scenario, existing technologies face challenges in practical applications. Urban traffic systems possess complex dynamic characteristics, encompassing multivariate time-series data such as intersection traffic flow, vehicle type, and speed. Traffic flow changes are influenced by multiple factors, including the periodicity of morning and evening peak hours, sudden event interference, and holiday pattern switching, exhibiting high volatility, non-periodicity, and non-stationarity. Traditional statistical time series forecasting models, such as ARIMA, face challenges in handling complex multivariate data. With the rapid development of deep learning, models based on CNNs, RNNs, and Transformers have been widely applied to time series forecasting tasks, achieving better predictive performance than traditional methods. Transformer-based models utilize self-attention mechanisms to capture long-term dependencies in sequences, achieving excellent performance; however, their quadratic complexity leads to high computational costs when processing long-sequence data generated by massive sensors in urban road networks. Recently, the Mamba model, based on the State-Space Model (SSM), has demonstrated performance comparable to Transformers in sequence data modeling. It combines the sequential reasoning capabilities of RNNs and the parallel training capabilities of CNNs, successfully addressing the low computational efficiency of Transformers when processing long-sequence data. Mamba introduces a selection mechanism on top of the SSM framework, enabling it to effectively focus on or ignore information in a data-dependent manner. Simultaneously, it employs a hardware-aware scanning algorithm for efficient parallel data processing, maintaining linear complexity even when modeling long sequences. Today, Mamba has demonstrated superior performance in language modeling, computer vision, genomics, and other fields, making it a strong competitor to the Transformer architecture and inspiring researchers to explore its applications in time series forecasting tasks.

[0004] Currently, simple linear models, such as DLinear, have been found to outperform most complex Transformer-based models in terms of performance and efficiency in time series forecasting tasks. The TSMixer model, a recently proposed multivariate time series forecasting architecture based on a full MLP, retains the ability of linear models to capture temporal dependencies while effectively utilizing cross-variable information to improve forecasting performance, making it highly competitive. However, in urban traffic flow forecasting, when non-periodic events such as sudden traffic congestion or abnormal weather occur, the model, using an MLP with static parameter weights to capture dependencies between variables, lacks a dynamic information filtering mechanism, making it difficult to distinguish between effective signals and noise. This limits its forecasting performance under complex road network conditions because it cannot effectively handle real-time changing traffic networks. Furthermore, MLP-based models rely only on local temporal patterns observed in historical windows for forecasting, which may ignore global temporal dependencies when processing long sequences generated by traffic flow sensors, leading to decreased forecast accuracy. Summary of the Invention

[0005] To address at least one of the technical problems mentioned above, this invention provides a method and system for predicting urban traffic flow under complex road network conditions. This method can dynamically process data based on real-time traffic conditions in urban traffic flow prediction, effectively utilize multivariate information, and perform efficient long-sequence modeling, thus exhibiting stronger expressive power and higher prediction accuracy.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] A first aspect of the present invention provides a method for predicting urban traffic flow under complex road network conditions, comprising the following steps:

[0008] Obtain time-series data of urban road network traffic flow;

[0009] Temporal mixed features are extracted from urban road network traffic flow time series data. Cross-information between different traffic features is captured based on temporal mixed features. The mixed feature representation is obtained by fusing temporal mixed features and cross-information between different traffic features.

[0010] Based on hybrid feature representation, global dependencies in long sequences are captured by selective state space modeling, and dynamic temporal features are output.

[0011] By enhancing the dynamic temporal features and projecting them along the time dimension, the predicted values ​​of each feature variable in the traffic temporal data are obtained.

[0012] Furthermore, the temporal mixed features extracted from the urban road network traffic flow time series data include:

[0013] Perform two-dimensional batch normalization on time series data in both time and feature dimensions;

[0014] The normalized time series data is transposed, and a temporal mixture MLP is applied along the time dimension and shared among features to capture local temporal dependencies, thus obtaining temporal mixture features.

[0015] The temporal blending features are transposed again, and the dimensional order of the original data is restored before residual connection is performed to preserve the fluctuation characteristics of the original traffic, thus obtaining the final temporal blending features.

[0016] Furthermore, the method of capturing the cross-information between different traffic features based on temporal blending features includes:

[0017] Perform two-dimensional batch normalization on the temporal mixed feature sequence;

[0018] The normalized data is used in a feature mixing MLP along the feature domain and shared between time steps to capture the cross information between features, thus obtaining the cross information between different traffic features.

[0019] Furthermore, the method based on hybrid feature representation captures global dependencies in long sequences through selective state-space modeling and outputs dynamic temporal features, including:

[0020] The hybrid feature representation is transformed by a fully connected linear layer to map the original feature dimension to the hidden layer dimension of the Mamba block, thus encoding the high-dimensional semantics of traffic features.

[0021] The high-dimensional semantics of the linearly transformed traffic features are input into a Mamba block, which is then mapped to the hidden state space to capture global dependencies, model long-sequence dependencies, and output dynamic temporal features.

[0022] Furthermore, after obtaining the dynamic temporal features, the dynamic temporal features are mapped back to the original dimension through a fully connected linear layer, decoding the high-dimensional features into physically meaningful variables.

[0023] Furthermore, the enhancement of dynamic temporal features includes performing Dropout regularization on the dynamic temporal features, performing residual connection between the result and the hybrid feature representation, preserving the original features and skipping unnecessary transformations to obtain the enhanced feature representation.

[0024] Furthermore, when enhancing the projection of dynamic temporal features onto the trailing time dimension, a fully connected linear layer is used to project the historical window. Mapping to prediction length , to obtain the future Predicted values ​​of each characteristic variable in traffic time series data at each time step.

[0025] Furthermore, the method also includes preprocessing the time series data of urban road network traffic flow, and the preprocessed data is represented as follows: , This represents the preprocessed time-series traffic flow data continuously sampled from different sensors in the urban road network, where... Indicates time series The observed value at time, Indicates the length of the input time series data. This represents the number of traffic multivariate features contained in each time step.

[0026] A second aspect of the present invention provides an urban traffic flow prediction system under complex road network conditions, comprising:

[0027] The data acquisition module is used to acquire time-series data of urban road network traffic flow.

[0028] The hybrid feature extraction module is used to extract time-mixed features based on urban road network traffic flow time series data, capture the cross information between different traffic features based on the time-mixed features, and fuse the time-mixed features and the cross information between different traffic features to obtain a hybrid feature representation.

[0029] The traffic flow prediction module is used to capture global dependencies in long sequences based on hybrid feature representation and selective state space modeling, and output dynamic time-series features. The dynamic time-series features are enhanced by projection onto the trailing time dimension to obtain the predicted values ​​of each feature variable of the traffic time-series data.

[0030] A third aspect of the present invention provides a program product.

[0031] The program product is a computer program product, including a computer program. When the computer program is executed by a processor, it implements the steps in the urban traffic flow prediction method under complex road network conditions as described above.

[0032] Compared with the prior art, the beneficial effects of the present invention are:

[0033] 1. This invention can simultaneously capture the time dependence in time series, the cross information between features, and the global dependence in long series. In urban traffic flow prediction under complex road network conditions, it dynamically processes data based on real-time traffic conditions, effectively utilizes multivariate information, and performs efficient long series modeling, resulting in stronger expressive power and higher prediction accuracy.

[0034] 2. This invention effectively captures temporal dependencies and cross-information between features in time-series data by utilizing a hybrid layer. It fuses temporal hybrid features and cross-information between different traffic features to obtain a hybrid feature representation. Based on this hybrid feature representation, selective state-space modeling captures global dependencies in long sequences, outputting dynamic temporal features, reducing overfitting risk, and improving the model's generalization ability. The use of residual connections effectively alleviates the gradient vanishing problem and avoids information degradation in deep networks. In complex multivariate scenarios, this invention achieves higher prediction accuracy compared to the traditional TSMixer model.

[0035] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0036] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0037] Figure 1 This is a flowchart of the urban traffic flow prediction method under complex road network conditions provided in the embodiments of the present invention;

[0038] Figure 2 This is a structural diagram of the time series prediction model provided in the embodiments of the present invention. Detailed Implementation

[0039] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0040] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0041] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0042] As mentioned in the background section, in urban traffic flow prediction, when non-periodic events such as sudden traffic congestion or abnormal weather occur, the models, which use multilayer perceptrons (MLPs) with static parameter weights to capture dependencies between variables, lack dynamic information filtering mechanisms. This makes it difficult to distinguish between effective signals and noise, and they cannot effectively handle real-time changing traffic networks, thus limiting prediction performance under complex road network conditions. Furthermore, MLP-based models rely solely on local temporal patterns observed within historical windows for prediction, and may ignore global temporal dependencies when processing long sequences generated by traffic flow sensors, leading to decreased prediction accuracy.

[0043] This invention combines the hybrid layers of the TSMixer model with Mamba blocks sequentially. The hybrid layers effectively capture temporal dependencies and cross-cutting information between features in time-series data, while the dynamic selection capability of Mamba blocks filters noise, enhancing focus on key information and further capturing global dependencies in long sequences. Dropout and residual connections improve model training stability and generalization ability, and the prediction results are finally generated by a time projection layer. By integrating the advantages of TSMixer and Mamba, this invention can dynamically process data based on real-time traffic conditions in urban traffic flow prediction, effectively utilize multivariate information, and perform efficient long-series modeling, resulting in stronger expressive power and higher prediction accuracy.

[0044] Example 1

[0045] See Figure 1 and Figure 2 This embodiment provides a method for predicting urban traffic flow under complex road network conditions, including the following steps:

[0046] S101: Obtain and preprocess the time series data of urban road network traffic flow;

[0047] In this embodiment, time-series data of urban road network traffic flow is obtained and preprocessed to obtain... , This represents the preprocessed time-series traffic flow data continuously sampled from different sensors in the urban road network, where... Indicates time series The observed value at time, Indicates the length of the input time series data. This represents the number of traffic multivariate features contained in each time step.

[0048] S102: Based on the preprocessed urban road network traffic flow time series data, extract the cross information between time patterns and different features to generate a hybrid feature representation;

[0049] In this embodiment, time series data is input into the mixing layer of the TSMixer model to extract the cross information between time patterns and different features, and generate a mixed feature representation;

[0050] For inputting time series data into the hybrid layer of the TSMixer model, the model's input data is: ,in Indicates time series The observed value at time, Indicates the length of the input time series data. To represent the number of feature variables, the data is sequentially passed through a temporal mixing layer and a feature mixing layer in the mixing layer to capture local temporal dependencies and enhance feature interactions, generating a mixed representation. In form ,in This indicates time mixing in the time dimension. This indicates feature blending along the feature dimension.

[0051] Specifically, the steps include the following:

[0052] S201. Time series data are subjected to two-dimensional batch normalization in both time and feature dimensions. This ensures data consistency across two dimensions and eliminates dimensional differences between different sensors.

[0053] S202. Transpose the normalized input, apply a temporal mixture MLP along the time dimension and share it among features to capture local temporal dependencies, obtaining temporal mixture features, such as the periodicity of traffic during peak hours. ;in, , , It is an activation function. It is Perform a transpose operation;

[0054] S203. Transpose the time-mixed result again, restore the dimensional order of the original data, and then perform a residual join to preserve the fluctuation characteristics of the original traffic. The temporal blending feature representation of the output of the temporal blending layer is obtained. ;in, This indicates a Dropout operation;

[0055] S204, Sequence Two-dimensional batch normalization was performed again to obtain The normalized data is applied along the feature domain using a feature mixing MLP and shared across time steps to capture cross-information between features, such as the dynamic relationship between vehicle speed and traffic flow. , ;in, The first layer output of the feature mixing MLP represents the intermediate result after performing a nonlinear transformation on the normalized input. The output of the second layer of the feature fusion MLP represents the final result of the feature fusion MLP obtained by linearly transforming the intermediate results of the first layer. , , , , To feature blending of the dimensions of hidden layers in an MLP;

[0056] S205. After residual connection and fusion of time pattern and cross-variable information, The output of the hybrid layer is obtained. .

[0057] S103: Based on hybrid feature representation, it captures global dependencies in long sequences through selective state space modeling and outputs dynamic temporal features;

[0058] In this embodiment, the output data of the hybrid layer is input into the Mamba block, and global dependencies in long sequences are captured through selective state space modeling, outputting dynamic temporal features;

[0059] Specifically, the steps include the following:

[0060] S301, Transfer data The feature dimension is transformed through a fully connected linear layer, from the original feature dimension. Mapped to hidden layer dimensions of Mamba blocks Encoding high-dimensional semantics of traffic features enhances the model's expressive power, as shown below: The weight matrix Bias term Data after linear transformation .

[0061] S302, Time Series After Linear Transformation Enter into the Mamba block, This is mapped to a hidden state space. By selectively ignoring unnecessary mixing operations in the state space, relevant patterns crucial for prediction are extracted, and global dependencies that the mixing layer might overlook are captured. Long-sequence dependencies are modeled, and dynamic temporal features are output, formally represented as... , .

[0062] In this embodiment, the Mamba block dynamically processes data based on real-time traffic conditions through a selective state space, ignoring unnecessary mixing operations and extracting relevant patterns that are crucial for prediction.

[0063] When faced with sudden events such as abnormal weather, the weight of historical periodicity is reduced, prioritizing the fitting of short-term fluctuations in weather impacts. Simultaneously, global dependencies that might be overlooked by the mixing layer are captured, and the dependencies of long sequences, such as historical data spanning multiple weeks, are modeled, ultimately outputting dynamic time-series features. .

[0064] S303, Process the output data obtained from the Mamba block. By mapping back to the original dimension through a fully connected linear layer, high-dimensional features are decoded into physically meaningful variables, specifically represented as follows: The weight matrix Bias term The output of this layer after linear transformation is .

[0065] S104: Enhance the dynamic time series features to obtain an enhanced dynamic time series feature representation;

[0066] Specifically, the output of the Mamba block is subjected to Dropout regularization, and the result is residually connected with the hybrid feature representation. The original features are preserved and unnecessary transformations are skipped to obtain the enhanced feature representation.

[0067] In this embodiment, when performing Dropout regularization on the output of the Mamba block, some neurons are randomly dropped with a probability of 0.1 to 0.9. , By reducing the interdependence between variables, overfitting of the model to fixed road patterns is prevented. Probability values ​​are dynamically adjusted during training to adapt to different data distributions, thus improving the model's generalization ability.

[0068] S105: The enhanced dynamic temporal feature representation is mapped to obtain the final traffic flow prediction result;

[0069] In this embodiment, the result after residual connection is input into the time projection layer of the TSMixer model and mapped to the final prediction result.

[0070] Specifically, when inputting the result after residual concatenation into the time projection layer of the TSMixer model, the residual output is first... The transpose is then applied along the time dimension using a fully connected linear layer for linear projection, and shared among features. This layer learns temporal patterns and extracts the time series from a historical window. Mapping to prediction length Specifically, it is expressed as The weight matrix Bias term To obtain the final prediction result Including the future Predicted values ​​of each characteristic variable in traffic time series data at each time step.

[0071] Effect verification

[0072] The predictive performance of this invention is evaluated using mean squared error (MSE) and mean absolute error (MAE), with smaller errors indicating better performance. The model constructed in this embodiment was tested on a publicly available real traffic dataset, with input historical data lengths of 512 and prediction lengths of 96, 192, 336, and 720. Table 1 shows the experimental results comparing the errors of the method of this invention and the TSMixer model on the publicly available dataset.

[0073] Table 1. Experimental results comparing prediction model errors.

[0074]

[0075] As shown in Table 1, the present invention performs well in prediction tasks of different lengths, with mean squared error (MSE) and mean absolute error (MAE) values ​​both being smaller than those of the TSMixer model, proving the effectiveness of the proposed model for predicting urban traffic flow under complex road network conditions.

[0076] Example 2

[0077] This embodiment provides an urban traffic flow prediction system under complex road network conditions, including:

[0078] The data acquisition module is used to acquire time-series data of urban road network traffic flow.

[0079] The hybrid feature extraction module is used to extract time-mixed features based on urban road network traffic flow time series data, capture the cross information between different traffic features based on the time-mixed features, and fuse the time-mixed features and the cross information between different traffic features to obtain a hybrid feature representation.

[0080] The traffic flow prediction module is used to capture global dependencies in long sequences based on hybrid feature representation and selective state space modeling, and output dynamic time-series features. The dynamic time-series features are enhanced by projection onto the trailing time dimension to obtain the predicted values ​​of each feature variable of the traffic time-series data.

[0081] It should be noted that the specific implementation of the urban traffic flow prediction system under complex road network conditions in this embodiment of the invention is similar to the specific implementation of the urban traffic flow prediction method under complex road network conditions in this embodiment of the invention. Please refer to the description in the method section for details. In order to reduce redundancy, it will not be repeated here.

[0082] Example 3

[0083] This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps in the urban traffic flow prediction method under complex road network conditions as described above.

[0084] Example 4

[0085] This embodiment provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps in the urban traffic flow prediction method under complex road network conditions as described above.

[0086] Example 5

[0087] This embodiment provides a program product, which is a computer program product, including a computer program. When the computer program is executed by a processor, it implements the steps in the urban traffic flow prediction method under complex road network conditions as described above.

[0088] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting urban traffic flow under complex road network conditions, characterized in that, Includes the following steps: Obtain time-series data of urban road network traffic flow; Temporal mixed features are extracted from urban road network traffic flow time series data. Cross-information between different traffic features is captured based on temporal mixed features. The mixed feature representation is obtained by fusing temporal mixed features and cross-information between different traffic features. Based on hybrid feature representation, global dependencies in long sequences are captured by selective state space modeling, and dynamic temporal features are output. By projecting the enhanced dynamic temporal features along the time dimension, the predicted values ​​of each feature variable in the traffic time series data are obtained. The temporal hybrid features extracted from urban road network traffic flow time series data include: Time series data Two-dimensional batch normalization is performed on both time and feature dimensions to obtain... ; The normalized time series data is transposed, and a temporal mixture MLP is applied along the time dimension and shared among features to capture local temporal dependencies, thus obtaining temporal mixture features. This includes the periodicity of traffic flow during peak hours; The temporal blending features are transposed again, and the dimensional order of the original data is restored before residual joins are performed to preserve the original traffic fluctuation characteristics, resulting in the final temporal blending features. ,in, This represents the Dropout operation; wherein, the capture of cross-information between different traffic features based on temporal blending features includes: Temporal blending feature sequences Two-dimensional batch normalization is performed to obtain ; The normalized data is processed along the feature domain using a feature mixing MLP and shared across time steps to capture cross-information between features, thus obtaining cross-information between different traffic features. , is represented as: , ;in, The first layer output of the feature mixing MLP represents the intermediate result after performing a nonlinear transformation on the normalized input. The output of the second layer of the feature fusion MLP represents the final result of the feature fusion MLP obtained by linearly transforming the intermediate results of the first layer. , , , , To feature blending, the dimension of the hidden layer in the MLP is used; For activation function, This represents the number of traffic multivariate features contained in each time step; The hybrid feature representation is obtained by fusing temporal mixed features and cross-information between different traffic features. , is represented as: ; The method based on hybrid feature representation, which captures global dependencies in long sequences through selective state-space modeling and outputs dynamic temporal features, includes: Representing hybrid features The feature dimensions are transformed using a fully connected linear layer, mapping them from the original feature dimensions to the hidden layer dimensions of the Mamba block. Encoding yields high-dimensional semantics of traffic features, represented as The weight matrix Bias term ; High-dimensional semantics of traffic features after linear transformation The input Mamba block is mapped to the hidden state space, global dependencies are captured, long-sequence dependencies are modeled, and dynamic temporal features are output.

2. The urban traffic flow prediction method under complex road network conditions as described in claim 1, characterized in that, After obtaining the dynamic temporal features, the dynamic temporal features are mapped back to the original dimension through a fully connected linear layer, and the high-dimensional features are decoded into physically meaningful variables.

3. The urban traffic flow prediction method under complex road network conditions as described in claim 1, characterized in that, Enhancement of dynamic temporal features includes performing Dropout regularization on the dynamic temporal features, connecting the result with the hybrid feature representation using residuals, preserving the original features, and obtaining the enhanced feature representation.

4. The method for predicting urban traffic flow under complex road network conditions as described in claim 1, characterized in that, When enhancing the projection of dynamic temporal features onto the trailing time dimension, a fully connected linear layer is used to project the historical window. Mapping to prediction length To obtain the future Predicted values ​​of each characteristic variable in traffic time series data at each time step.

5. The urban traffic flow prediction method under complex road network conditions as described in claim 1, characterized in that, The method further includes preprocessing the urban road network traffic flow time series data, and the preprocessed data is represented as follows: , This represents the preprocessed time-series traffic flow data continuously sampled from different sensors in the urban road network, where... Indicates time series The observed value at time, Indicates the length of the input time series data. This represents the number of traffic multivariate features contained in each time step.

6. A system for predicting urban traffic flow under complex road network conditions, characterized in that, The method for predicting urban traffic flow under complex road network conditions as described in any one of claims 1-5 includes: The data acquisition module is used to acquire time-series data of urban road network traffic flow. The hybrid feature extraction module is used to extract time-mixed features based on urban road network traffic flow time series data, capture the cross information between different traffic features based on the time-mixed features, and fuse the time-mixed features and the cross information between different traffic features to obtain a hybrid feature representation. The traffic flow prediction module is used to capture global dependencies in long sequences based on hybrid feature representation and selective state space modeling, and output dynamic time-series features. The dynamic time-series features are enhanced by projection onto the trailing time dimension to obtain the predicted values ​​of each feature variable of the traffic time-series data.

7. A program product, said program product being a computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps in the urban traffic flow prediction method under complex road network conditions as described in any one of claims 1-5.

Images