RSCN traffic flow prediction method based on time sequence convolutional network

Abstract

The invention discloses an RSCN traffic flow prediction method based on a time sequence convolutional network. A time sequence feature data preprocessing and normalization module, a feature data parallel processing module and an error judgment and correction module based on a residual network are included. The time sequence characteristic data preprocessing and normalization module is used for sorting the traffic flow data of the time sequence recorded by a plurality of transmitters on one road section into sequence data, and preprocessing and normalizing the traffic flow data; the characteristic data parallel processing module is used for distributing the preprocessed and normalized traffic flow data in an expansion convolution and radial basis network for parallel processing; and carrying out self-correction error analysis training on the residual network by taking the traffic flow data after parallel processing as the input of residual processing and taking the residual neuron as the output of residual processing. According to the invention, an RSCN traffic flow prediction algorithm based on a time sequence convolutional network is adopted, and based on a convolutional network layer, traffic flow data of a time sequence recorded by a plurality of transmitters on a road section are predicted and calculated by using parallel calculation and residual checking methods. A short-time traffic flow prediction model which is higher in prediction precision and capable of changing learning parameters in real time is constructed, and the problems that an original model is not high enough in precision and poor in real-time data processing capacity are solved.

Description

Technical Field

[0001] The present invention relates to the technical field of machine learning algorithms, and in particular to a RSCN traffic flow prediction method based on a temporal convolutional network. Background Art

[0002] Machine learning is the study of improving computer algorithms through experience. The main purpose of the research is to improve the performance and accuracy of computer algorithms.

[0003] The rapid development of social economy and urbanization has gradually increased the demand for traffic capacity. The increase in traffic capacity will also bring certain challenges in a certain space, such as the damage to the green economy. To solve these problems, the comprehensive application of multi-dimensional cross-technology, specifically the intelligent transportation system has become an important management model. Traffic flow prediction is a key component of the intelligent transportation system. It completes the prediction of traffic flow in a certain space in the future by analyzing historical traffic flow data and predicting real-time traffic volume. Among them, highways are important channels for the operation of the entire social logistics. However, affected by the regional economy and society, the traffic flow between different regions has great differences, which will lead to irrationality in regional resource allocation and energy use. Therefore, improving the accuracy and timeliness of traffic flow prediction models plays a vital role in promoting the development of regional transportation systems.

[0004] Traffic flow prediction is affected by many factors, which can be divided into internal and external factors. Traditional recurrent neural networks can capture long-term dependencies in time series well, but they are too dependent on the previous time step when computing in parallel, have low adaptability to time steps, and may face instability problems such as gradient explosion. In terms of sequence data modeling, convolutional networks can achieve better performance than recurrent networks because convolutional networks allow parallel computing of outputs. Therefore, temporal convolutional networks were proposed and began to be used for the calculation of sequence models. This convolutional network can compute in parallel and capture local dependencies, and can be well applied to sequence models in different contexts. Moreover, the sequential convolutional network can capture all two levels of information in a unified way, which makes it more stable in terms of gradient propagation. However, the modeling ability of temporal convolutional networks is limited, and it is difficult to maintain high-precision calculations when dealing with long-term dependencies. Proposing a model with higher accuracy and stronger adaptability to real-time time series data has become a technical problem that needs to be solved in predicting short-term traffic flow data. Summary of the invention

[0005] 1. Technical issues to be resolved

[0006] In view of the shortcomings of the prior art, the present invention provides a RSCN traffic flow prediction method based on a temporal convolutional network, which solves the problems of the original model being insufficiently accurate and having poor ability to process real-time data.

[0007] (II) Technical solution

[0008] To achieve the above-mentioned purpose, the present invention provides the following technical solutions: a RSCN traffic flow prediction method based on a time series convolutional network, which predicts and calculates the time series traffic flow data recorded by multiple transmitters on a road section based on parallel computing and residual verification methods, aiming to improve the prediction accuracy and enhance the stability of the prediction of time series data;

[0009] The module for preprocessing and normalizing time series feature data can sort out the data of sensors in a certain space one by one, and preprocess and normalize the traffic flow data;

[0010] The preprocessed and normalized data are used as the input of the parallel processing module, and the double-layer data are used as the output of the parallel processing module for processing by the residual module.

[0011] Preferably, the preprocessing and normalization module includes:

[0012] (1) Preprocessing: Through the traditional outlier method, the preprocessing operation of isolation forest is used. We determine whether each data is an outlier by separating the data points. After that, the outlier data is replaced by calculating the average path length of all numbers.

[0013] (2) Data normalization: Using the entropy method (homogenization of heterogeneous indicators), normalize the indicator data from different populations or with different characteristics, and standardize the data to the range of [0, 1], so that the algorithm can more easily calculate the relevant parameters.

[0014] Preferably, the relevant neuron parameters of the dilated convolution are obtained by comparing the error data under different parameters after testing the network with a test set; the error parameters are specifically the root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE) and determination coefficient (R 2 ).

[0015] Preferably, the parallel processing module of feature data includes:

[0016] (1) Parameter determination: Through the test set test analysis, the test ability of the dilated convolution layer with different parameters is determined to determine the optimal parameters;

[0017] (2) Dilated convolution layer: The hybrid dilated convolution method is used to enable the model to expand the receptive field and make high use of the input information;

[0018] (3) Radial basis network layer: The data is processed using a feed-forward neural network radial basis network, which enables the model to adaptively change learning parameters when processing real-time data.

[0019] (4) Flattening layer: Flattens the two layers of data into one-dimensional data, making it suitable for dimensional convolution processing.

[0020] Preferably, a total of three layers of error analysis models are constructed in the error judgment and correction module based on the residual network, including dimensionality-raising error processing, residual network processing and monitoring layer processing layers.

[0021] Preferably, the error judgment and correction module based on the residual network includes:

[0022] (1) Dimensionality increase error processing: The flattened data is increased in dimension, and the convolution layer performs a convolution operation on the multi-dimensional data to further process the time information and extract multi-dimensional effective information.

[0023] (2) Residual network processing: The residual network layer is composed of several residual blocks and a forward network, and uses skip connections to pass temporal information through the residual blocks more quickly and efficiently.

[0024] (3) Monitoring layer and processing layer: randomly extract a time series segment without any preceding or following time logic (the randomly extracted number set is 0.8%-1.5% of the total number set), perform prediction processing on this part of the data, and compare the data errors of the previous and subsequent predictions of the same time series segment. If the prediction results are within the allowable error range, we believe that the prediction of this time series is accurate and there is no problem such as gradient explosion.

[0025] The RSCN traffic flow prediction method based on a temporal convolutional network of the present invention has the following beneficial effects:

[0026] (1) A parallel analysis module was constructed. The system is based on a convolutional neural network and uses the parallel computing capabilities of the time series convolutional network, the ability to expand the receptive field, and the ability of the radial basis function to self-adjust learning parameters in a single operation. A traffic flow prediction model that can be calculated in real time was constructed, which solves the problem that traditional machine learning has low accuracy in capturing temporal features and weak real-time prediction capabilities in traffic flow prediction problems.

[0027] (2) Construct an error judgment and correction module. This system is based on a residual network and focuses on the unstable problems such as gradient explosion that the model may face when processing big data. It innovates the monitoring set solution to monitor the stability of the entire prediction process and solves many problems such as the inability of machine learning to self-repair and gradient explosion.

[0028] (3) The research proposed in the present invention was verified using real highway data. Based on a real truck data set of multiple detection points for 365 days in a year, the average prediction accuracy of traffic flow reached 92.12%. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 is the result of the first training in Example 1, representing the error data of different dilated convolution parameters;

[0030] Figure 2 is the result of the second training in Example 1, representing the error data of the training set;

[0031] Figure 3 is the result of the second training in Example 1, indicating the visual error result of the training set;

[0032] Figure 4 is the result of the first monitoring set in Example 1, indicating the error data of the monitoring set;

[0033] Figure 5 is the result of the first monitoring set in Example 1, indicating the visual error result of the monitoring set; DETAILED DESCRIPTION

[0034] In order to make the purpose, technical scheme and advantages disclosed in the embodiments of the present invention clearer, the embodiments of the present invention are further described in detail in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the embodiments of the present invention and are not intended to limit the embodiments of the present invention. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application. Examples of the embodiments are shown in the accompanying drawings, where the same or similar reference numerals throughout represent the same or similar elements or elements with the same or similar functions.

[0035] It should be noted that the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or server that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units that are not explicitly listed or inherent to these processes, methods, products or devices.

[0036] Like reference numerals and letters denote similar items in the following drawings, and thus, once an item is defined in one drawing, further definition and explanation thereof is not required in subsequent drawings.

[0037] Embodiment 1

[0038] The data used in the present invention is collected from the highway section between Yandun Toll Station and Erbao Toll Station of Xinjiang G30. The traffic flow data is obtained through sensor equipment, and the data transmitted by the sensor is split to calculate the short-term traffic flow data every 15 minutes.

[0039] By excluding abnormal data such as mutation data that does not conform to the time series logic in the original data, this embodiment finally obtains 35040 valid time series for subsequent processing based on 365 days in a year on some sections of the highway.

[0040] According to the actual situation, multiple vehicle detection sites will be recorded, and the traffic flow results are recorded as follows:

[0041] Q={p1,p2,p3…,p t ,…p T}

[0042] Among them, p t is the traffic flow input corresponding to the current time step t. After flattening, it can be regarded as the target time series data, and T is the total sampling time step.

[0043] The establishment process is as follows:

[0044] (1) Data collection, preprocessing and normalization

[0045] 1) Data collection: Use multiple vehicle detection stations on a road section to collect the real value of traffic flow data pt. Traffic flow data on the same road section is recorded from 00:00 every day, with an average of one recording node every 15 minutes. There are about 96 data points per day, totaling about 35,040 sets of data.

[0046] 2) Noise and missing data filling: Use the isolation forest preprocessing operation to separate data points to determine whether each data is an outlier, delete the outlier samples and fill them in, and use the mean interpolation method to smooth the noise data.

[0047] The mean value interpolation method is described as follows:

[0048] Among them, x represents the average value of successive adjacent moments, and t is an arbitrary time series moment.

[0049] 3) Data normalization: Using the entropy method (homogenization of heterogeneous indicators), normalize the indicator data from different populations or with different characteristics, standardize the data to the range of [0,1], and improve the accuracy of the experiment.

[0050] The normalization description of the entropy method is as follows:

[0051] Among them, x is the original attribute value, y is the normalized attribute value, and x max Indicates the maximum value among the original attribute values, x min Indicates the minimum value among the original attribute values.

[0052] Dataset division: Divide the normalized data set into a training set and a test set. For this part of the experimental data, use 80% of the data as the training set and 20% of the data as the test set.

[0053] (2) Parallel processing module for feature data

[0054] The radial time series convolutional network (RSCN) is used to process short-term traffic flow data. Based on the traditional time series convolutional network, RSCN introduces a radial basis network structure layer as a parallel processing structure of the dilated convolution. By adding a forward radial network layer, the learning parameters of the radial network can be corrected by real-time processing of weight distribution. Combined with the dilated convolution layer with strong ability to extract short-term information, it is more suitable for solving time series problems in the same space, and is used to learn and describe the changing trend of traffic flow over time on the road section to better evaluate and allocate road section resources.

[0055] A three-step neural network is constructed in the feature data parallel processing module based on the temporal convolutional network, including the dilated convolution layer, the radial basis network layer and the tiling layer.

[0056] The dilated convolution layer processes time series data and extracts time series features. The one-dimensional convolution layer parameter is set to 5. The radial basis network layer is used to adaptively change the learning parameters according to the real-time input data and assist in extracting time series features. Here, the radial diffusion speed is set to 1000 to increase the radial analysis speed. The tiling layer is used for full connection, connecting the parallel analysis data and the residual network section.

[0057] (3) Error judgment and correction module based on residual network

[0058] An error judgment and correction module based on a residual network is used to detect and process errors in the data after parallel analysis. The monitoring layer is introduced and compared with the basic model to help verify the stability and accuracy of the model prediction. The monitoring set can monitor the stability of the data fitting process in real time, and can also respond to the role of the residual network in stabilizing the model.

[0059] Three error analysis and processing steps are constructed in the feature data parallel processing module based on the temporal convolutional network, including dimensionality increase error processing, residual network processing and monitoring layer processing layer.

[0060] For dimensionality-raising error processing, for data with processing errors or large errors, the one-dimensional convolution neuron size is set to 16 to ensure the convolution speed and preliminary calculation; the residual network processing layer is used to ensure the stability of the model. Here, the jump residual block parameter is set to 2 to improve the stability and the model processing speed; the monitoring layer processing layer is used to check whether the model can calculate the results stably. The randomly extracted number set is set to 0.8%-1.5% of the total number set to ensure randomness while also monitoring the model.

[0061] (4) Prediction result data visualization

[0062] According to Example 1, the prediction results of the training set and test set models are visualized. According to the prediction data of the test set, we find that the prediction effect is best when the time step is 10 and the delay step is 0. Then, for the same data set, the RSCN traffic flow prediction method based on the temporal convolutional network outperforms the other four traditional prediction models under the comparison of four error parameters.

[0063] Therefore, the present invention adopts the above-mentioned RSCN traffic flow prediction method based on time series convolutional network. Based on the convolutional network layer, it uses parallel computing and residual verification methods to predict and calculate the time series traffic flow data recorded by multiple transmitters on a road section, and constructs a short-term traffic flow prediction model with higher prediction accuracy and the ability to change learning parameters in real time, which solves the problem that the original model is not accurate enough and has poor ability to process real-time data.

Claims

1. A RSCN traffic flow prediction method based on a temporal convolutional network, characterized by: A preprocessing and normalization module for time series feature data, a parallel processing module for feature data, and an error judgment and correction module based on a residual network; The module for preprocessing and normalizing time series characteristic data organizes the time series traffic flow data recorded by multiple transmitters on a road section into sequence data, and preprocesses and normalizes the traffic flow data; The parallel processing module of feature data divides the pre-processed and normalized traffic flow data into dilated convolution and radial basis network for parallel processing; The traffic flow data after parallel processing is used as the input of residual processing, and the residual neurons are used as the output of residual processing to carry out self-correction error analysis training of the residual network.

2. According to claim 1, a RSCN traffic flow prediction method based on a temporal convolutional network is characterized in that: The preprocessing and normalization modules of time series feature data include: (1) Preprocessing: Through the traditional outlier method, the preprocessing operation of isolation forest is used. We determine whether each data is an outlier by separating the data points. After that, the outlier data is replaced by calculating the average path length of all numbers. (2) Data normalization: Using the entropy method (homogenization of heterogeneous indicators), normalize the indicator data from different populations or with different characteristics, and standardize the data to the range of [0, 1], so that the algorithm can more easily calculate the relevant parameters.

3. The RSCN traffic flow prediction method based on a temporal convolutional network according to claim 1 is characterized in that: The relevant neuron parameters of the dilated convolution are obtained by testing the network with a test set and comparing the error data under different parameters; The error parameters are root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE) and coefficient of determination (R 2 ).

4. According to claim 2, a RSCN traffic flow prediction method based on a temporal convolutional network is characterized in that: The parallel processing modules of feature data include: (1) Parameter determination: Through the test set test analysis, the test ability of the dilated convolution layer with different parameters is determined to determine the optimal parameters; (2) Dilated convolution layer: The hybrid dilated convolution method is used to enable the model to expand the receptive field and make high use of the input information; (3) Radial basis network layer: The data is processed using a feed-forward neural network radial basis network, which enables the model to adaptively change learning parameters when processing real-time data. (4) Flattening layer: Flattens the two layers of data into one-dimensional data, making it suitable for dimensionality-enhancing convolution processing.

5. The RSCN traffic flow prediction method based on a temporal convolutional network according to claim 1 is characterized in that: A total of three layers of error analysis models are constructed in the error judgment and correction module based on the residual network, including dimensionality-enhanced error processing, residual network processing and monitoring layer processing layer.

6. The RSCN traffic flow prediction method based on a temporal convolutional network according to claim 5 is characterized by: The error judgment and correction module based on the residual network includes: (1) Dimensionality increase error processing: The flattened data is increased in dimension, and the convolution layer performs a convolution operation on the multi-dimensional data to further process the time information and extract multi-dimensional effective information. (2) Residual network processing: The residual network layer is composed of several residual blocks and a forward network, and uses skip connections to pass temporal information through the residual blocks more quickly and efficiently. (3) Monitoring layer and processing layer: randomly extract a time series segment without any preceding or following time logic, perform prediction processing on this part of the data, and compare the data errors of the previous and following predictions of the same time series segment. If the prediction results are within the allowable error range, we consider the prediction of this time series to be accurate and there is no problem such as gradient explosion.

7. An RSCN traffic flow prediction method based on a temporal convolutional network as described in any one of claims 1 to 6 is applied in predicting short-term traffic flow in a certain space.

Images