Space-time convolution short-term traffic flow prediction method based on complementary ensemble empirical mode decomposition

Abstract

The application discloses a space-time convolution short-time traffic flow prediction method based on complementary ensemble empirical mode decomposition, and comprises the following steps: acquiring original traffic flow data and performing data preprocessing; and inputting the traffic flow data into a pre-constructed CEEMD-GAT-LSTM model to perform short-time traffic flow prediction. The prediction model comprises the following steps: performing signal decomposition on the traffic flow data by using CEEMD, fully mining the spatial features of the data by using GAT, and fully mining the time features of the data by using LSTM, so that the learning efficiency of the neural network is improved. The application can effectively improve the prediction accuracy of short-time traffic flow prediction and provide favorable reference value for the traffic department.

Description

Technical Field

[0001] This invention relates to a spatiotemporal convolutional short-term traffic flow prediction method based on complementary integrated empirical modulus decomposition, belonging to the field of intelligent transportation technology. Background Technology

[0002] Knowing the traffic conditions ahead is essential for proactive decision-making, and the effects of traffic control and guidance typically take time to become apparent. Depending on the time scale, traffic flow prediction is divided into long-term and short-term predictions, used to predict and analyze the number of vehicles passing through a specific road section within 5-30 minutes.

[0003] Improving the accuracy of traffic flow forecasting has become one of the major challenges in current traffic flow research. Accurate short-term traffic flow forecasting is currently the most fundamental and effective measure to solve urban road traffic congestion and reduce environmental pollution. With the continuous accumulation of traffic flow data and the continuous development of related technologies, forecasting methods are gradually evolving from the original classical statistical methods to a data-driven perspective.

[0004] However, traffic flow data is easily affected by environmental and external factors such as weather and communication, resulting in a large amount of noise in the collected data, making it highly nonlinear and complex. Currently, there is a lack of a short-term traffic flow prediction method that is both accurate and stable. Summary of the Invention

[0005] To address the shortcomings in the aforementioned background technology, this invention provides a spatiotemporal convolutional short-term traffic flow prediction method based on complementary integrated empirical modulus decomposition, with the aim of improving the accuracy of short-term traffic flow prediction.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] This invention provides a spatiotemporal convolutional short-time traffic flow prediction method based on complementary ensemble empirical modulus decomposition, comprising:

[0008] First, the raw traffic flow data is acquired and preprocessed to generate a dataset;

[0009] The traffic flow dataset is input into the pre-built CEEMD-GAT-LSTM model for short-term traffic flow prediction;

[0010] The construction process of the CEEMD-GAT-LSTM model includes:

[0011] Mathematical modeling of GAT: Calculate the hidden values ​​between each node and its neighboring nodes in the input traffic flow dataset, assign different weights, and combine multi-head attention mechanism to obtain more node feature information and improve the model's learning ability.

[0012] Mathematical modeling of LSTM: Based on RNN network, LSTM network adds sequentially connected input gate, forget gate and output gate, and realizes information protection and control through the three such "gate" structures;

[0013] Mathematical modeling of CEEMD-GAT-LSTM: Traffic flow data is decomposed using CEEMD, and the decomposed signals are sequentially input into the GAT model to capture the spatial features of the data and the LSTM model to capture the temporal features of the data. Feature fusion is performed through a fully connected layer to output the predicted value.

[0014] Furthermore, linear interpolation was used to handle missing values ​​in the data, the data was re-aggregated into a format where one record is generated every 20 minutes, and the min-max method was used to normalize the traffic flow data.

[0015] Furthermore, a traffic flow prediction model is constructed, and the traffic flow dataset is input into the traffic flow prediction model. The traffic flow prediction model is trained based on the input traffic flow dataset and outputs predicted values. The predicted values ​​are compared with the corresponding real values, and the loss value between the predicted values ​​and the real values ​​is calculated. Iterative training is performed based on the loss value until the preset maximum number of iterations is reached or the loss value converges.

[0016] Furthermore, the traffic flow dataset is decomposed using CEEMD by introducing paired positive and negative Gaussian white noise into the dataset, decomposing it into several IMF components and a residual component:

[0017]

[0018] In the formula, M1 represents the original sequence after adding positive noise, M2 represents the original sequence after adding negative noise, S represents the traffic flow dataset, and N represents the added Gaussian white noise.

[0019] EMD decomposition was performed on M1 and M2 to obtain 2n IMF components. The component of each component was denoted as follows: Calculate the average value for each group of IMF components:

[0020]

[0021] in, This represents the j-th IMF component decomposed from the i-th signal after adding positive noise; This represents the j-th IMF component decomposed from the i-th signal after adding negative noise; IMF jThis represents the j-th IMF component obtained after the signal is decomposed by CEEMD. In summary, the traffic flow dataset S is ultimately decomposed into multiple IMF components and a residual component res.

[0022]

[0023] Furthermore, spatial features of the traffic flow dataset are captured, and the attention coefficient e between node i and node j is calculated using GAT. ij The softmax function is used to normalize all neighboring nodes, and the attention coefficient is used to adopt a multi-head attention mechanism and update the feature vector of each node using a weighted summation method.

[0024] e ij =α(Wh i ,Wh j )

[0025]

[0026]

[0027] In the formula, α represents the function used to calculate the correlation, W represents a parameter matrix that performs a linear transformation on all node features, converting the input features into high-dimensional features, σ represents the activation function, and N... i Let || represent the neighboring nodes directly connected to node i, || represent the join operation, and K represent the number of attention heads. W represents the attention coefficient under the k-th attention head. k Let represent the linear transformation matrix under the k-th attention head.

[0028] Furthermore, to capture the temporal features of the traffic flow dataset, the forget gate will input vector x t and hidden state h t-1 The forgotten state f is obtained through the sigmoid function σ. t and set the unit state C t-1 and the forgotten state f t Multiply and output;

[0029] The input gate takes the input vector x t and hidden state h t-1 The input state i is obtained by using the sigmoid function σ and the tanh function respectively. t and unit state Input state i t and unit state Multiplying the result and adding it to the output of the forget gate yields the cell state C. t ;

[0030] The output gate will input vector x t and hidden state h t-1 The output state o is obtained by using the sigmoid function σ. t Set the unit state C t The result and output status are obtained through the tanh function. t Multiply to obtain the hidden state h t ;

[0031] f t =σ[W f ·(h t-1 ,x t )+b f ]

[0032] i t =σ[W i ·(h t-1 ,x t )+b i ]

[0033] o t =σ[W o ·(h t-1 ,x t )+b o ]

[0034]

[0035]

[0036] h t =o t ·tanh(C t )

[0037] Among them, W f For the weight of the forget gate, b f For the bias of the forget gate, W i b represents the weights of the input gate. i For the bias of the input gate, W o b represents the weight of the output gate. o For the output gate bias; W c b represents the weight of the memory unit. c This is the bias of the memory cell.

[0038] Furthermore, the evaluation metrics used to measure the loss between the predicted and actual values ​​include:

[0039]

[0040]

[0041]

[0042] Where N is the number of traffic flow data points, y i It is the experimental predicted value, x i It is the actual value of traffic flow data. It is the average of the actual values ​​of traffic flow data.

[0043] Furthermore, an apparatus includes:

[0044] One or more processors;

[0045] Memory, used to store one or more programs;

[0046] When one or more of the programs are executed by one or more of the processors, the one or more processors implement the spatiotemporal convolutional short-time traffic flow prediction method based on complementary ensemble empirical modulus decomposition as described above.

[0047] Furthermore, a storage medium containing computer-executable instructions, which, when executed by a computer processor, are used to perform the spatiotemporal convolutional short-time traffic flow prediction method based on complementary integrated empirical modulus decomposition as described above.

[0048] The beneficial effects of this invention are:

[0049] This invention uses the CEEMD module to decompose traffic flow data based on noise frequency. By introducing positive and negative paired Gaussian white noise into the traffic flow dataset, the mixed signal is simultaneously decomposed into EMD. This not only overcomes the mode aliasing phenomenon present in the EMD module, but also, because complementary noise is added, the residual noise in the reconstructed data is almost negligible. To a certain extent, CEEMD can also use relatively fewer ensemble averaging operations, saving computation time while ensuring minimal residual noise interference.

[0050] This invention uses a graph attention network layer (GAT) and a long short-term memory network layer (LSTM) to build a prediction model. GAT calculates the attention coefficient between adjacent nodes, allowing adjacent nodes to obtain different weights based on their own characteristics, thereby better capturing the spatial characteristics of traffic flow datasets. LSTM adds an input gate, a forget gate, and an output gate to RNN, overcoming the problem that RNN is not good at capturing contextual dependencies in long-term sequences.

[0051] This invention provides a spatiotemporal convolutional short-term traffic flow prediction method based on complementary integrated empirical modulus decomposition. Accurate short-term traffic flow prediction has important practical significance in the field of intelligent transportation. It can provide travelers with more accurate road information, improve road traffic efficiency, solve traffic congestion problems, strengthen the connection between vehicles and roads, and is an important part of building smart cities. Attached Figure Description

[0052] The invention will now be further described with reference to the accompanying drawings.

[0053] Figure 1 This is a schematic diagram of the CEEMD structure provided in Embodiment 1 of the present invention;

[0054] Figure 2 This is a schematic diagram of the GAT structure provided in Embodiment 1 of the present invention;

[0055] Figure 3 This is a schematic diagram of the LSTM structure provided in Embodiment 1 of the present invention;

[0056] Figure 4 This is an overall flowchart provided for Embodiment 1 of the present invention; Detailed Implementation

[0057] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0058] Example 1:

[0059] This invention provides a spatiotemporal convolutional short-time traffic flow prediction method based on complementary ensemble empirical modulus decomposition, comprising the following steps:

[0060] 1. Acquire raw traffic flow data and preprocess it to generate a dataset;

[0061] 2. Input the traffic flow dataset into the pre-built CEEMD-GAT-LSTM model for short-term traffic flow prediction;

[0062] The following section mainly introduces the construction process of the CEEMD-GAT-LSTM model:

[0063] Mathematical modeling of GAT: Calculate the hidden values ​​between each node and its neighboring nodes in the input traffic flow dataset, assign different weights, and combine multi-head attention mechanism to obtain more node feature information and improve the model's learning ability.

[0064] like Figure 2 The diagram shown is a structural diagram of the GAT network;

[0065] The attention coefficient e between node i and node j is calculated using GAT. ijThe softmax function is used to normalize all neighboring nodes, and the attention coefficient is used to adopt a multi-head attention mechanism and update the feature vector of each node using a weighted summation method.

[0066] e ij =α(Wh i ,Wh j )

[0067]

[0068]

[0069] In the formula, α represents the function used to calculate the correlation, W represents a parameter matrix that performs a linear transformation on all node features, converting the input features into high-dimensional features, σ represents the activation function, and N... i Let || represent the neighboring nodes directly connected to node i, || represent the join operation, and K represent the number of attention heads. W represents the attention coefficient under the k-th attention head. k Let represent the linear transformation matrix under the k-th attention head.

[0070] Mathematical modeling of LSTM: Based on RNN network, LSTM network adds sequentially connected input gate, forget gate and output gate, and realizes information protection and control through the three such "gate" structures;

[0071] like Figure 3 The diagram shown is a structural diagram of an LSTM network.

[0072] The forget gate will input vector x t and hidden state h t-1 The forgotten state f is obtained through the sigmoid function σ. t and set the unit state C t-1 and the forgotten state f t Multiply and output;

[0073] The input gate takes the input vector x t and hidden state h t-1 The input state i is obtained by using the sigmoid function σ and the tanh function respectively. t and unit state Input state i t and unit state Multiplying the result and adding it to the output of the forget gate yields the cell state C. t ;

[0074] The output gate will input vector x t and hidden state h t-1The output state o is obtained by using the sigmoid function σ. t Set the unit state C t The result and output status are obtained through the tanh function. t Multiply to obtain the hidden state h t ;

[0075] f t =σ[W f ·(h t-1 ,x t )+b f ]

[0076] i t =σ[W i ·(h t-1 ,x t )+b i ]

[0077] o t =σ[W o ·(h t-1 ,x t )+b o ]

[0078]

[0079]

[0080] h t =o t ·tanh(C t )

[0081] Among them, W f For the weight of the forget gate, b f For the bias of the forget gate, W i b represents the weights of the input gate. i For the bias of the input gate, W o b represents the weight of the output gate. o For the output gate bias; W c b represents the weight of the memory unit. c This is the bias of the memory cell.

[0082] Mathematical modeling of CEEMD-GAT-LSTM: Traffic flow data is decomposed using CEEMD, and the decomposed signals are sequentially input into the GAT model to capture the spatial features of the data and the LSTM model to capture the temporal features of the data. Feature fusion is performed through a fully connected layer to output the predicted value.

[0083] The traffic flow dataset is decomposed using CEEMD by introducing paired positive and negative Gaussian white noise into the dataset, which decomposes it into several IMF components and a residual component:

[0084]

[0085] In the formula, M1 represents the original sequence after adding positive noise, M2 represents the original sequence after adding negative noise, S represents the traffic flow dataset, and N represents the added Gaussian white noise.

[0086] EMD decomposition was performed on M1 and M2 to obtain 2n IMF components. The component of each component was denoted as follows: Calculate the average value for each group of IMF components:

[0087]

[0088] in, This represents the j-th IMF component decomposed from the i-th signal after adding positive noise; This represents the j-th IMF component decomposed from the i-th signal after adding negative noise; IMF j This represents the j-th IMF component obtained after the signal is decomposed by CEEMD. In summary, the traffic flow dataset S is ultimately decomposed into multiple IMF components and a residual component res.

[0089]

[0090] A traffic flow prediction model is constructed, and a traffic flow dataset is input into the traffic flow prediction model. The traffic flow prediction model is trained based on the input traffic flow dataset and outputs predicted values. The predicted values ​​are compared with the corresponding real values, and the loss value between the predicted values ​​and the real values ​​is calculated. Iterative training is performed based on the loss value until the preset maximum number of iterations is reached or the loss value converges.

[0091] like Figure 4 The following is a flowchart of the overall process of CEEMD-GAT-LSTM.

[0092] Linear interpolation was used to handle missing values ​​in the data, the data was re-aggregated into a format where one record is generated every 20 minutes, and the min-max method was used to normalize the traffic flow data.

[0093] This paper uses traffic flow and weather data from a highway in a certain province, provided by KDD CUP 2017. The traffic flow data comes from three toll stations: toll stations 1 and 3 allow two-way traffic, while toll station 2 only allows entry and not exit. Toll station 1 is selected as the central node, and toll stations 2 and 3 are selected as neighboring nodes. The weather data includes seven features: air pressure, sea pressure, wind direction, wind speed, temperature, humidity, and rainfall, recorded every 3 hours. The time period for both traffic flow and weather data is from September 19, 2016 to October 17, 2016. Based on the data characteristics, the first 90% is used as the training set, and the last 10% as the test set, with a sliding window of 7.

[0094] This paper describes the model construction and training using the PyTorch neural network framework in a Python 3.6 environment. Auxiliary white noise with an amplitude of 0.2 was selected to perform CEEMD decomposition on the traffic flow dataset. The number of hidden units in the GAT network layer was set to 128, and the number of attention heads K was set to 3. The LSTM network had 2 layers with 64 and 32 hidden units respectively. The Adam optimizer was used to optimize the model, with a learning rate of 0.001, a batch size of 32, a dropout rate of 0.2, and 50 iterations.

[0095] 1. To verify the predictive performance of the model in this paper on traffic flow data, root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), and coefficient of determination (R²) were used. 2 As an evaluation indicator:

[0096]

[0097]

[0098]

[0099] Where N is the number of traffic flow data points, y i It is the experimental predicted value, x i It is the actual value of traffic flow data. It is the average of the actual values ​​of traffic flow data.

[0100] The model of this invention was compared with five other models on three metrics on a public dataset, and the results are shown in Table 1.

[0101] Model Structure RMSE MAE MAPE LSTM 8.3143 6.1280 0.6313 GAT 8.2009 6.0067 0.5810 CNN-LSTM 5.6031 4.5063 0.3145 GAT-LSTM 5.1649 4.1247 0.3040 CEEMD-GAT-LSTM 4.5116 3.4888 0.2049

[0102] Table 1

[0103] As shown in Table 2, the traffic flow prediction model proposed in this embodiment of the invention achieves good results in terms of evaluation indicators RMSE, MAE, and MAPE. Compared with other models, the prediction accuracy is significantly improved, and it can more accurately reflect changes in road traffic flow, providing valuable reference for travelers.

[0104] Based on the same inventive concept, this invention also provides a computer device, comprising: one or more processors, and a memory for storing one or more computer programs; the programs include program instructions, and the processor executes the program instructions stored in the memory. The processor may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, used to implement one or more instructions, specifically for loading and executing one or more instructions stored in a computer storage medium to implement the above-described method.

[0105] It should be further explained that, based on the same inventive concept, the present invention also provides a computer storage medium storing a computer program, which, when executed by a processor, performs the above-described method. This storage medium can be any combination of one or more computer-readable media. The computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. The computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In the present invention, the computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0106] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0107] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.

Claims

1. A spatiotemporal convolutional short-term traffic flow prediction method based on complementary ensemble empirical modulus decomposition, characterized in that, Includes the following steps: Obtain raw traffic flow data and perform data preprocessing on the raw traffic flow data to generate a traffic flow dataset; The traffic flow dataset is input into a pre-built CEEMD-GAT-LSTM model for short-term traffic flow prediction. The construction process of the CEEMD-GAT-LSTM model includes: Mathematical modeling of CEEMD: CEEMD is a decomposition algorithm based on EMD and improved upon EEMD. It processes nonlinear signals according to noise frequency by introducing positive and negative pairs of Gaussian white noise into the traffic flow dataset, and simultaneously decomposes the mixed signals using EMD. Mathematical modeling of GAT: Calculate the hidden values ​​between each node and its neighboring nodes in the input traffic flow dataset, assign different weights, and combine multi-head attention mechanism to obtain more node feature information and improve the model's learning ability. Mathematical modeling of LSTM: Based on RNN network, LSTM network adds sequentially connected input gate, forget gate and output gate to realize information protection and control through input gate, forget gate and output gate; Mathematical modeling of CEEMD-GAT-LSTM: Traffic flow data is decomposed using CEEMD, and the decomposed signals are sequentially input into the GAT model to capture the spatial features of the data and the LSTM model to capture the temporal features of the data. Feature fusion is performed through a fully connected layer to output the predicted value. The data preprocessing operations specifically include: handling missing values ​​in the data using linear interpolation, re-aggregating the data into a format where one record is generated every 20 minutes, and normalizing the traffic flow data using the min-max method; The EMD can decompose a complex signal into a finite number of IMF components. Each IMF component contains local feature signals of the original signal at different time scales, that is, decompose the original sequence into eigenmode functions and residual components. in, m The number of IMF components; For the first i One IMF component; For the first m One residual component; The EEMD calculates the lumped average of the IMF components obtained after multiple EMD decompositions by adding auxiliary white noise; the CEEMD decomposes the traffic flow dataset by introducing paired positive and negative Gaussian white noise into it, decomposing it into several IMF components and a residual component. The signal decomposition process includes: In the formula, This represents the original sequence after adding positive noise. This represents the original sequence after adding negative noise. Represents a traffic flow dataset. This indicates the addition of Gaussian white noise; right , Perform EMD decomposition to obtain 2n Group IMF components, each component is denoted as... , Calculate the average value for each group of IMF components: in, Indicates the first i The first signal, after adding positive noise, is decomposed into the... j One IMF component; Indicates the first i The first signal, after adding negative noise, is decomposed into the... j One IMF component; This indicates the signal obtained after CEEMD decomposition. j One IMF component; in summary, the traffic flow dataset Ultimately, it is decomposed into multiple IMF components and a residual component res in data form. The GAT computing node i and nodes j Attention coefficient between and utilize softmax The function normalizes all neighboring nodes and uses attention coefficients, employs a multi-head attention mechanism, and updates the feature vector of each node using a weighted summation method. In the formula, Representative node i The feature set, Representative node j The feature set, This represents the function used to calculate the correlation. W This represents a parameter matrix that performs a linear transformation on all node features, converting the input features into high-dimensional features. `exp` represents the transformation of the input features into high-dimensional features. e An exponential function with base 0. This represents the activation function. Represents nodes i Directly connected neighbor nodes, K Indicates the number of attention heads. Indicates the first k Attention coefficient under each attention head Indicates the first k A linear transformation matrix under an attention head; The RNN is a type of neural network used to process sequential data; the calculation process of an RNN is as follows: in, y For output, x For input, W These are the weighting coefficients. f For activation functions; The forget gate will input vector and hidden state pass sigmoid function Get the forgotten state and the cell state and the state of forgetting Multiply and output; The input gate will input vector and hidden state Through respectively sigmoid function and the tanh function to get the input state and unit state , input status and unit state Multiply the result and add it to the output of the forget gate to obtain the cell state. ; The output gate will input vector and hidden state Through respectively sigmoid function Get output status , will the cell state Obtain the result and output status using the tanh function. Multiplication yields the hidden state ; in, As for the weight of the Forgotten Gate, For the offset of the forget gate, The weights of the input gates, This is the bias of the input gate. The weights of the output gates, For the output gate bias; The weights of the memory units, This is the bias of the memory cell.

2. The spatiotemporal convolutional short-time traffic flow prediction method based on complementary ensemble empirical modulus decomposition according to claim 1, characterized in that, The training process of the CEEMD-GAT-LSTM model includes: constructing a traffic flow prediction model, inputting a traffic flow dataset into the traffic flow prediction model, training the traffic flow prediction model based on the input traffic flow dataset and outputting predicted values, comparing the predicted values ​​with the corresponding true values, calculating the loss value between the predicted values ​​and the true values, and iteratively training based on the loss value until the preset maximum number of iterations is reached or the loss value converges.

3. The spatiotemporal convolutional short-time traffic flow prediction method based on complementary ensemble empirical modulus decomposition according to claim 2, characterized in that, The evaluation metrics for the loss value between the predicted value and the actual value include: Where N is the number of traffic flow data points, It is the experimental predicted value, It is the actual value of traffic flow data. It is the average of the actual values ​​of traffic flow data.

4. A device, characterized in that, The device includes: One or more processors; Memory, used to store one or more programs. When the one or more programs are executed by the one or more processors, the one or more processors perform the spatiotemporal convolutional short-time traffic flow prediction method based on complementary ensemble empirical modulus decomposition as described in any one of claims 1-3.

5. A computer-readable storage medium storing a computer program, characterized in that, When executed by the processor, the program implements the spatiotemporal convolutional short-time traffic flow prediction method based on complementary integrated empirical modulus decomposition as described in any one of claims 1-3.

Images