A Traffic Flow Prediction Method Based on Residual Calibration Network and Multi-Scale Fusion Mechanism

Abstract

This invention discloses a traffic flow prediction method based on a residual calibration network and a multi-scale fusion mechanism. The method includes: acquiring traffic flow data and external feature data; preprocessing the traffic flow data and external feature data to obtain traffic image data to be measured; inputting the traffic image data to be measured into a pre-trained MS-ResCNet network model, and outputting the predicted traffic flow result. The network model traffic flow prediction method based on a residual calibration network and a multi-scale fusion mechanism proposed in this invention adds a calibration network after the fusion of multiple network paths on the basis of a deep residual network. This can calibrate the traffic flow characteristics of various regions in the urban traffic flow space, avoiding interference from irrelevant regions in the overall urban spatial features, and thus solving the problem of insufficient prediction accuracy due to inadequate extraction of spatiotemporal features.

Description

Technical Field

[0001] This invention relates to a traffic flow prediction method based on residual calibration network and multi-scale fusion mechanism, belonging to the field of traffic flow prediction technology. Background Technology

[0002] In intelligent transportation and scientific research, traffic flow data consists of multiple sets of random variables ordered spatiotemporally. Typically, it represents the results of observing a potential process within a specific spatial region at equal time intervals according to a certain sampling frequency. Traffic flow data essentially reflects the trend of one or more travel variables changing over time and space. Traffic flow prediction methods primarily involve extracting patterns from a series of flow data and using these patterns to estimate future data. With the increasing popularity of ride-hailing services like Uber and Didi Chuxing, and the continuous development of shared bicycles and the transportation industry, a vast amount of spatiotemporal image data can be continuously collected and generated, providing a solid data foundation for traffic flow prediction.

[0003] To address traffic flow prediction, researchers have proposed various methods, which can generally be categorized into two main types: model-driven methods and data-driven methods. Model-driven methods, also known as parametric methods, primarily rely on pre-defined theoretical assumptions. Based on these assumptions, model parameters are solved, and the resulting model is used to perform predictions. Examples include ARIMA (Autoregressive Integrated Moving Average) and VAR (Vector Autoregressive). VAR is a more advanced time series model that can capture pairwise relationships between all traffic flow sequences. However, these methods perform poorly in traffic flow prediction because the urban road network and trajectory data are often too complex, and the assumed variables cannot accurately represent the real, complex road network and trajectories.

[0004] Data-driven approaches can be divided into two categories: traditional machine learning methods and deep learning-based methods. Machine learning methods, such as KNN (K Nearest Neighbors) and SVM (Support Vector Machine), can model more complex data, but require detailed feature engineering. Traditional machine learning methods cannot handle high-dimensional data and struggle to describe complex traffic flow changes in road networks and trajectory data. Furthermore, the predictive performance of traditional machine learning largely depends on custom features, which are highly dependent on the problem and expert experience, thus limiting the general applicability of this approach. Deep learning methods can effectively build high-dimensional spatiotemporal data models and automatically discover complex features through hierarchical representations. Among these, RNN (Recurrent Neural Network) was first applied to traffic flow prediction, but due to weight sharing among neurons, it suffers from gradient explosion or vanishing gradient problems, making it difficult to achieve good results on longer time-series features. The hybrid model of CNN-LSTM (Convolutional Neural Networks-Long Short-Term Memory) provides a new approach to solving spatiotemporal sequence problems. RNN variants include GRU (Gated Recurrent Unit) and LSTM (Long Short-Term Memory). They primarily employ unique internal unit designs, introducing forget gates and update gates to selectively discard or update information from the previous state. This addresses the gradient explosion or vanishing problems of RNNs to some extent. However, because the internal units of GRU and LSTM are only connected to the internal units of the next time step, CNN-LSTM cannot effectively capture the periodic features of time. Residual convolutional neural networks can capture dependencies in spatiotemporal information, especially in processing complex spatiotemporal image data and predicting large-scale traffic flows, showing significant advantages. However, when acquiring urban spatial features, residual convolutional neural networks do not calibrate the flow characteristics of different areas in the urban traffic flow space, failing to avoid interference from irrelevant areas in the overall urban spatial features. Furthermore, the weak feature representation ability of lower-layer networks and the excessive information loss during downsampling of higher layers affect prediction accuracy. Summary of the Invention

[0005] The purpose of this invention is to provide a traffic flow prediction method based on residual calibration network and multi-scale fusion mechanism, so as to solve the problems of low prediction accuracy when using feature information of low-level network and excessive information loss when downsampling of high-level network.

[0006] A traffic flow prediction method based on residual calibration network and multi-scale fusion mechanism, characterized in that the method includes:

[0007] Acquire traffic flow data and external feature data;

[0008] Traffic flow data and external feature data are preprocessed to obtain traffic image data to be measured;

[0009] The traffic image data to be tested is input into a pre-trained MS-ResCNet network model, which outputs the predicted traffic flow.

[0010] Furthermore, the preprocessing method includes:

[0011] Traffic flow data and external feature data are rasterized to generate traffic flow image data with spatiotemporal information;

[0012] The traffic flow image data is preprocessed using a batch normalization method to obtain the traffic image data to be measured.

[0013] Furthermore, the MS-ResCNet network model includes convolutional layers, residual calibration layers, and residual units. The convolutional layer consists of a single convolutional network and is connected to the residual calibration layer and residual units to extract the spatiotemporal features of urban traffic flow.

[0014] Furthermore, the residual calibration layer includes convolutional layers and calibration blocks. The calibration blocks are networks formed by fusing multiple convolutional layer branches, which are used to improve the extraction capability of spatiotemporal features.

[0015] Furthermore, the residual unit is composed of two convolutional blocks connected by residual connections. Each convolutional block consists of a batch normalization layer, an activation layer, and a convolutional layer. Through a multi-scale fusion mechanism, the inputs of the residual calibration layer and multiple residual units are used as the input feature scales of the traffic image. The multiple feature scales are spliced ​​and fused to output multi-feature scale data.

[0016] Furthermore, the convolutional layer is a convolutional network that converts output data at multiple feature scales into predicted target values ​​through convolution operations.

[0017] Furthermore, the training method for the MS-ResCNet network model includes:

[0018] The training samples are input into the MS-ResCNet network for training, and the weight parameters W in the MS-ResCNet network are continuously optimized and updated through a pre-constructed loss function;

[0019] The updated weight parameters W are fed into the MS-ResCNet network, and the loss function is calculated until the loss function result is minimized. The weight matrix parameters W and W' are then saved. b and bias parameter B b The trained MS-ResCNet network model is obtained.

[0020] Furthermore, before training the MS-ResCNet network model using the sample training, parameter initialization is first performed, using a mean of 0 and a standard deviation of . The Kaiming normal distribution is randomly initialized with re-parameters W and W'. b Bias parameter B b , where fan_in is the input dimension.

[0021] Furthermore, the loss function expression is as follows:

[0022]

[0023] Where RMSE is the loss function, n is the number of training samples, and obs is the number of training samples. i For the true value of the i-th training sample, pre i Let be the i-th predicted value output by the MS-ResCNet network from the first i-1 training samples.

[0024] Furthermore, updating the weight parameters W in the MS-ResCNet network includes: calculating the loss function result, performing backpropagation using the stochastic gradient descent algorithm, and updating the weight parameters.

[0025] Compared with the prior art, the beneficial effects achieved by the present invention are as follows:

[0026] (1) The network model traffic prediction method based on residual calibration network and multi-scale fusion mechanism proposed in this invention first performs normalization preprocessing on historical traffic flow spatiotemporal image data. Through normalization, the data is transformed to between 0 and 1, saving computing resources and accelerating the convergence speed of the neural network. The use of residual network deepens the number of network layers and solves the gradient vanishing problem.

[0027] (2) The network model traffic prediction method based on residual calibration network and multi-scale fusion mechanism proposed in this invention adds a calibration network after the fusion of multiple network paths on the basis of deep residual network. It can calibrate the traffic characteristics of each area in the urban traffic flow space, avoid the interference of irrelevant areas in the entire urban spatial characteristics to their respective areas, and solve the problem of insufficient prediction accuracy due to insufficient extraction of spatiotemporal features.

[0028] (3) The network model traffic prediction method based on residual calibration network and multi-scale fusion mechanism proposed in this invention uses a multi-scale fusion mechanism to fuse urban spatial feature information of high and low layers in MS-ResCNet network. The traffic flow image input features of each layer in MS-ResCNet network are used, and multiple input features are spliced ​​and fused using CNN. This solves the problems of low prediction accuracy when using spatial feature information of low layer network and excessive information loss when downsampling of high layer network. Attached Figure Description

[0029] Figure 1 This is a flowchart illustrating an embodiment of the method of the present invention;

[0030] Figure 2 This is a schematic diagram of the structure of an embodiment of the method of the present invention;

[0031] Figure 3 This is a schematic diagram of the residual unit described in an embodiment of the method of the present invention;

[0032] Figure 4 This is an internal structure diagram of the residual calibration network described in the embodiment of the method of the present invention. Detailed Implementation

[0033] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0034] like Figure 1 As shown, this invention discloses a traffic flow prediction method based on a residual calibration network and a multi-scale fusion mechanism. The steps of this method are as follows:

[0035] Acquire traffic flow data and external feature data;

[0036] Traffic flow data and external feature data are preprocessed to obtain traffic image data to be measured;

[0037] The traffic image data to be tested is input into a pre-trained MS-ResCNet network model, which outputs the predicted traffic flow.

[0038] The preprocessing method includes:

[0039] Traffic flow data and external feature data are rasterized to generate traffic flow image data with spatiotemporal information;

[0040] The traffic flow image data is preprocessed using a batch normalization method to obtain the traffic image data to be measured.

[0041] The structural improvements of the MS-ResCNet network model in this invention are as follows:

[0042] The MS-ResCNet network mainly consists of convolutional layers, residual calibration layers, and residual units. The convolutional layer is composed of a single convolutional network with a kernel size of 3. The convolutional layer is connected to the residual calibration layer and residual units to extract the spatiotemporal features of urban traffic flow.

[0043] Furthermore, the residual calibration layer consists of convolutional layers and calibration blocks. The convolutional layers employ a convolutional network with a kernel size of 3. The calibration block is a network formed by fusing multiple convolutional layer branches. It integrates the interactive feature information and nearby spatial feature information of each region in the spatial characteristics of urban traffic flow across its temporal channels, thereby improving the extraction capability of spatiotemporal features. This calibration block has five branches. Branch 1 consists of two convolutional blocks connected by residual connections. Each convolutional block consists of a batch normalization layer, an activation layer, and a convolutional layer. Branch 2 uses an average pooling layer connected to a convolutional layer and a normalization layer, with a pooling rate of 2. Branches 3, 4, and 5 each consist of convolutional layers and normalization layers. Finally, branch 5 fuses the two outputs of the four branches and performs residual connection processing. The residual calibration layer outputs feature image information. All convolutional kernels are 3 in size, and the number of input channels is 64, with the calibration block having 64 channels.

[0044] Furthermore, each residual unit consists of two convolutional blocks connected by residual connections. Each convolutional block comprises a batch normalization layer, an activation layer, and a convolutional layer. Multiple residual units are connected, with the input and output of each unit linked together, to address the gradient vanishing problem while capturing spatiotemporal features. The number of residual units used is 12, all convolutional kernels are 3 in size, and the number of channels is 64.

[0045] Furthermore, through a multi-scale fusion mechanism, the inputs of the residual calibration layer and multiple residual units are used as the input feature scales of the flow image;

[0046] Furthermore, multiple feature scales are spliced ​​and fused to output multi-feature scale data;

[0047] Furthermore, the convolutional layer is a convolutional network with 896 input channels and 2 output channels. This layer converts multi-feature scale output data into predicted target values ​​through convolution operations.

[0048] refer to Figure 1 This paper describes the specific methods for training and predicting the MS-ResCNet network model.

[0049] Step 1: Obtain historical traffic flow spatiotemporal image data x t ={x1,...,x n-1}, where n-1 are the timestamps of the traffic data, and we need to predict X for the nth timestamp. n The value of X. t This represents the observed flow rate value at the t-th (0 ≤ t ≤ n-1) timestamp in the flow rate data. Each x... t ={x t in,i,j ,x t out,i,j Traffic data is represented as traffic in the target city divided into i×j grid types, s=(s 0,0 ,....s i,j )∈R represents the spatial characteristics of urban traffic flow, S i,j This represents the traffic data in the i-th row and j-th column of the city traffic data;

[0050] in:

[0051]

[0052] Step 2: To map the data to a range of 0 to 1, thereby accelerating the training speed of the neural network, the traffic data X is processed. t Normalization is performed to obtain data x t '={x0',...,x' n-1 For each observation X t Its normalization method is as follows:

[0053]

[0054] Among them, X max and X min These correspond to the maximum and minimum values ​​among the observed values, x t 'This is the result after normalization.'

[0055] Step 3: Divide the normalized traffic image dataset into a 9:1 ratio to obtain the training set and the test set;

[0056] Step 4: Using the training sample data from Step 3 as input to the network model, design the MS-ResCNet network. Figure 2 shows a schematic diagram of the overall structure of the MS-ResCNet network implemented by the method of this invention. The flow matrix data first enters the convolutional layer to extract the short-term dependencies of the data in the spatiotemporal dimension. After one convolutional layer, the feature matrix Conv is obtained.

[0057] Step 5: Further, the feature matrix Conv is input into the residual calibration layer for calculation. The residual calibration layer consists of two convolutional layers and a calibration block. Conv is processed by convolutional layers, ReLU (Rectified Linear Activation Function), and max pooling layers, and then input into the calibration block. The calibration block first uses convolutional layers, ReLU, and normalization layers to divide Conv into two input branches, input1 and input2. Then, input1 is input into branch 1, which consists of two convolutional layers, and mainly outputs the short-term dependency Conv2 in the spatial dimension. Input2 is simultaneously input into branches 2 and 3. After branch 2 performs downsampling processing on the spatial and channel dimensions, it is multiplied with the output of branch 3 to output the feature matrix Conv1. Conv1 is fused through branch 4 to output the spatiotemporal matrix, thereby obtaining the attention features Conv3 in the channel and spatial dimensions. The output Conv3 and Conv2 are concatenated and fused through branch 5 to output Conv4 in the channel dimension. It is further residually connected with Conv to output the calibration feature matrix. This allows each region in space to integrate information from nearby areas and interactive information from channels, avoiding interference from irrelevant regions in the overall global information. This addresses the problem of insufficient prediction accuracy due to inadequate extraction of spatiotemporal features.

[0058] The internal structure of the residual calibration layer is as follows: Figure 4 As shown, now combined with Figure 4 A detailed explanation of the residual calibration layer and the result output process:

[0059] (1) The characteristic matrix Conv is x = {x1,...,x} c}∈R c×h×w After convolution processing through convolutional layers:

[0060] T1 = Relu(BN(Conv(x))) (5)

[0061] Where T1 is the output of the convolution operation, ReLU is the linear rectified activation function, BN is the batch normalization layer, and Conv is a 3*3 convolution.

[0062] (2) Input T1 into the calibration block and output the calibration feature matrix diagram C1. The process is as follows:

[0063] input1=Relu(BN(Conv_a(T1))) (6)

[0064] input2=Relu(BN(Conv_b(T1))) (7)

[0065] Y1=Conv(Relu(BN(input1))) (8)

[0066] K2=BN(Conv(AvgPool(input2))) (9)

[0067] K3 = BN(Conv(input2)) (10)

[0068] Y2=K3·σ(input2+K2) (11)

[0069] Y2=BN(Conv(Y2)) (12)

[0070] C1=Relu(BN(Conv(Cat(Y1,Y2)))) (13)

[0071] In this architecture, all Conv, Conv_a, and Conv_b are 3x3 convolutional kernels, ReLU is the activation function, and BN is the batch normalization layer. The pooling rate of AvgPool is 2, and both the convolutional kernel and stride are set to 2. Downsampling is performed using AvgPool, and upsampling is performed using the Up function to achieve the same resolution and number of channels as input2. σ maps the values ​​to the interval [0,1] using the sigmoid activation function. · represents the multiplication symbol; Cat concatenates the two outputs Y1 and Y2 along the channel dimension.

[0072] Step 6: Input the output C1 of the residual calibration layer into the residual unit layer. Each residual unit consists of two convolutional blocks connected by residual connections. Multiple residual units are connected, with the input and output of each unit linked together. This is used to capture spatiotemporal features while solving the gradient vanishing problem, thereby increasing the number of network layers.

[0073] The internal structure of the residual convolutional unit layer is as follows: Figure 3 As shown, now combined with Figure 3 A detailed explanation of the residual convolutional unit layer and the process of outputting the results:

[0074] (1) Input the calibration feature matrix C1 into multiple residual cells, and output matrix R. i The forward propagation process of each layer of the network is as follows:

[0075] R i =Conv(Relu(BN(R) i )))(1≤i≤12) (14)

[0076] R i+1 =Conv(Relu(BN(R) i ')))+R i (1≤i≤12) (15)

[0077] Where R0 is the output result C1 of the previous residual calibration layer, and each residual unit is processed by the input R of the current unit. i Perform residual connection to output the input R of the next residual unit. i+1 The number of residual units is set to 12, and the number of channels in each unit is 64.

[0078] Step 7: Finally, a multi-scale fusion mechanism is used to concatenate and fuse the outputs of the convolutional layer, residual calibration layer, and multiple residual convolutional units as scale outputs, producing multi-scale data. A convolutional layer is used to convert the multi-scale output data into predicted target values. This layer has 64*14 input channels and 2 output channels. ReLU (Rectified Linear Function) is used as the activation function.

[0079] The specific structure of the multi-scale fusion mechanism is as follows: Figure 2 As shown, now combined with Figure 2 A detailed explanation of the multi-scale fusion mechanism and the process of outputting results:

[0080] (1) The output Conv of the convolutional layer, the output C1 of the residual calibration layer, and the output R of multiple residual units are combined. i (1≤i≤14) are concatenated and fused to output multi-scale data. The process is as follows:

[0081] M' = Cat(Conv,C1,R1,...R) i (16)

[0082] (2) The fusion result M' is input into the convolutional layer to convert the multi-scale output data into the predicted target value. The process is as follows:

[0083] M = Conv(Relu(M')) (17)

[0084] Step 8: Construct the loss function, train the MS-ResCNet network model using the training samples, and input the test samples into the trained network to obtain predicted traffic data. The process includes:

[0085] First, the root mean square error (RMSE) is used as the loss function, and its definition is as follows:

[0086]

[0087] Where RMSE is the loss function, n is the number of training samples, and obs is the number of training samples. i For the true value of the i-th training sample, pre i The i-th predicted value is generated by the MS-ResCNet network from the first i-1 training samples.

[0088] Furthermore, the MS-ResCNet network was trained using a method with a mean of 0 and a standard deviation of . The weight matrix parameters of each stage of the network are randomly initialized using a Kaiming normal distribution (where fan_in is the input dimension).

[0089] Furthermore, after initialization, the SGD (Stochastic Gradient Descent) optimizer in PyTorch is used to continuously optimize and update the weight matrix parameters using stochastic gradient descent until the loss function is minimized. The update rule follows the formula below:

[0090]

[0091] In the formula, α is the learning rate, and W is the weight matrix parameter. The partial derivatives in the above formula can be obtained through backpropagation.

[0092] The main steps of backpropagation are as follows: First, the given samples are forward-propagated to obtain the final output value of the network. Then, the error value is calculated, and the partial derivative of the error value with respect to each node is obtained to determine the influence of each node on the final output. Therefore, the complete network training model is as follows: Initialize the weight parameters of the network at each stage; for each sample, first use forward propagation to calculate the total error, then use backpropagation to calculate the partial derivatives of each weight parameter, and finally update the weight parameters according to the above formula. Repeat this step until the loss function reaches its minimum. At this point, the MS-ResCNet network converges, and the network model and model parameters are saved. Input the test samples into the trained network model to obtain the predicted traffic data P.

[0093] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A traffic flow prediction method based on residual calibration network and multi-scale fusion mechanism, characterized in that, The method comprises: acquiring traffic flow data and external feature data; preprocessing the traffic flow data and the external feature data to obtain to-be-measured traffic image data; inputting the to-be-measured traffic image data into a pre-trained MS-ResCNet network model to output a predicted result of the traffic flow; the MS-ResCNet network model comprises a convolution layer, a residual calibration layer and a residual unit, the convolution layer is composed of one convolution network, the convolution layer is connected with the residual calibration layer and the residual unit, and is used for extracting the spatiotemporal features of urban traffic flow; the training method of the MS-ResCNet network model comprises: The training sample is input into the MS-ResCNet network for training, and the weight parameters in the MS-ResCNet network are constantly optimized and updated through a pre-constructed loss function ; The updated weight parameters into the MS-ResCNet network, and calculate the loss function until the loss function result is minimum, save the weight parameters at this time , parameters and bias parameters , to obtain the trained MS-ResCNet network model; The updated weight parameters into the MS-ResCNet network, and calculate the loss function until the loss function result is minimum, save the weight matrix parameters at this time , parameters and bias parameters , to obtain the trained MS-ResCNet network model; the residual calibration layer is composed of a convolution layer and a calibration block, the calibration block is a network fused by multiple convolution layer branches, the calibration block has five branches in total, branch 1 adopts two convolution blocks and is connected through residual connection. Each convolution block is composed of a batch normalization layer, an activation layer and a convolution layer, branch 2 adopts an average pooling layer, a convolution layer and a normalization layer and is connected, the pooling rate is 2, branch 3, branch 4 and branch 5 are all composed of a convolution layer and a normalization layer, and finally the two outputs of the four branches are fused and processed through residual connection by branch 5.

2. The traffic flow prediction method based on residual calibration network and multi-scale fusion mechanism according to claim 1, characterized in that, the preprocessing method comprises: rasterizing the traffic flow data and the external feature data to generate traffic flow image data with spatiotemporal information; using a batch normalization method to preprocess the traffic flow image data to obtain the to-be-measured traffic image data.

3. The traffic flow prediction method based on residual calibration network and multi-scale fusion mechanism according to claim 1, characterized in that, The residual calibration layer comprises a convolution layer and a calibration block, the calibration block is a network fused by multiple convolution layer branches, and is used for improving the extraction capability of spatiotemporal features.

4. The traffic flow prediction method based on residual calibration network and multi-scale fusion mechanism according to claim 1, characterized in that, The residual unit is composed of two convolution blocks and is connected through residual connection, each convolution block is composed of a batch normalization layer, an activation layer and a convolution layer, and through a multi-scale fusion mechanism, each input of the residual calibration layer and the multiple residual units is taken as a traffic image input feature scale, multiple feature scales are spliced and fused, and multi-feature scale data is output.

5. The traffic flow prediction method based on residual calibration network and multi-scale fusion mechanism according to claim 4, characterized in that, The convolution layer is a convolution network, and through convolution operation, the output data of the multiple feature scales are converted into predicted target values.

6. The traffic flow prediction method based on residual calibration network and multi-scale fusion mechanism according to claim 1, characterized in that, The sample training first performs parameter initialization before training the MS-ResCNet network model, and the parameters are randomly initialized with a mean value of 0 and a standard deviation of Normal distribution random initialization reparameterization , parameters and bias parameters , wherein is the input dimension.

7. The traffic flow prediction method based on residual calibration network and multi-scale fusion mechanism according to claim 1, characterized in that, The loss function expression is as follows: ; wherein RMSE is a loss function, is the number of training samples, is the true value of the th training sample, is the th predicted value of the th training sample output by the MS-ResCNet network from the first previous training samples.

8. The traffic flow prediction method based on residual calibration network and multi-scale fusion mechanism according to claim 1, characterized in that, updating the weight parameters in the MS-ResCNet network comprises: calculating a loss function result, performing back propagation by using a stochastic gradient descent algorithm, and updating the weight parameters.

Images