CN113205685B - Short-term traffic flow prediction method based on global-local residual error combination model - Google Patents

Abstract

The invention discloses a short-time traffic flow prediction method based on a global-local residual error combination model, which comprises the following steps: collecting traffic flow data of a city road network and transmitting the traffic flow data to a traffic big data cluster in real time, carrying out data preprocessing on original data to reduce data redundancy, and converting the preprocessed data into space-time raster data according to the longitude and latitude of the road network; carrying out standardization processing on the time-space grid data, and dividing the time-space grid data into a training set and a test set; constructing a global-local space-time residual error-based combined model; and predicting the space-time grid data of the urban road network at the next moment based on the global-local space-time residual error combination model which is trained and constructed. The invention analyzes the global and local spatial characteristics of the urban road network simultaneously on the basis of performing space-time analysis on the road network, thereby improving the capture of the global and local spatial characteristics and long-term time characteristics of the road network and further improving the short-time traffic flow prediction precision.

Description

Short-term traffic flow prediction method based on global-local residual error combination model

Technical Field

The invention belongs to the field of intelligent traffic and traffic flow prediction, and particularly relates to a short-time traffic flow prediction method based on a global-local residual error combination model.

Background

With the acceleration of the urbanization process in China, the construction of traffic infrastructure is gradually difficult to meet the increasing number of automobiles, so that the problem of traffic jam is increasingly serious. The short-time traffic flow prediction is one of the important components of the intelligent traffic subsystem, can provide traffic information at future time for traffic management departments, and helps the traffic management departments to make scientific and reasonable traffic guidance and traffic scheduling. The method has the advantages that the urban traffic flow is accurately acquired in real time, a traveler can be helped to make road planning in advance, traffic management departments can be helped to make corresponding traffic management methods for different traffic flows, the intelligent level of urban traffic is greatly improved, traffic jam is relieved, and travel cost is reduced. Therefore, how to effectively improve the accuracy of the traffic flow prediction model is important.

The models for traffic flow prediction are numerous, with the rise of deep learning, a deep neural network model taking a neural network as a core is popular, the urban road network traffic flow has complex space-time characteristics, the model prediction precision is not high only by considering a certain characteristic, many models are used for analyzing the space-time characteristics of the traffic flow at present, but the long-term time characteristic and the global space characteristic are not captured, so that the models cannot adapt to the traffic flow prediction of the urban road network macroscopic angle, and the global consideration is lacked.

Disclosure of Invention

The purpose of the invention is as follows: aiming at the problems, the invention introduces a short-time traffic flow prediction method based on a global-local residual error combination model. The method aims at the defect that the traditional space-time residual error model lacks capture of long-term characteristics of traffic flow when processing space-time grid data; meanwhile, the spatial features of the traffic network are affected by global and local features, and the traditional method for extracting the spatial features only analyzes one feature and ignores the effect of the other feature. Constructing a global spatial feature extraction component and a local spatial feature extraction component by utilizing a 2D convolution and residual error unit, and respectively capturing global spatial features and local spatial features of the urban road network; and a long-term time feature extraction component is constructed by utilizing the strong long-term time feature capturing capability of the GRU, so that the long-term time features of the urban road network are captured.

The technical scheme is as follows: a short-time traffic flow prediction method based on a global-local residual error combination model comprises the following steps:

step 1) collecting traffic flow data of a city road network and transmitting the traffic flow data to a traffic big data cluster in real time, carrying out data preprocessing on the traffic flow data of the original city road network to reduce data redundancy, and converting the preprocessed traffic flow data of the city road network into space-time grid data according to the longitude and latitude of the road network;

step 2) carrying out standardization processing on the time-space grid data, and dividing the time-space grid data into a training set and a test set;

step 3) constructing a global-local space-time residual error-based combined model, wherein the model consists of a global spatial feature extraction component, a local spatial feature extraction component and a long-term time feature extraction component;

and 4) training by using the training set obtained in the step 2) and predicting the space-time grid data of the city road network at the next moment based on the global-local space-time residual error combination model.

Further, in step 1, the preprocessed data is converted into space-time grid data according to the road network longitude and latitude, and the method specifically comprises the following steps: dividing the city road network into I X J network regions according to longitude and latitude to form a network X ═ X_i,jEach position in the network is regularly distributed, and the distance between adjacent positions is constant; for each position (i, j), the traffic flow data, x, for the current position are recorded at fixed time Δ t intervals_t ^i,jRepresenting the traffic flow data counted in the position (I, J) at the time t, and representing the traffic flow data counted in the network area I J as tensor

Converting traffic flow prediction problem into given historical space-time grid data { X }_tI t 0, …, K, predicting the spatio-temporal raster data X at the time K + Δ t_K+ΔtAnd K is the total number of samples of the collected urban road network traffic flow.

Further, in the step 2, the space-time grid data is standardized by adopting z-score standardization, and the space-time grid data { X ] obtained in the step 2 is subjected to standardization_t| t ═ 0, …, K } by

Carrying out transformation; wherein

x_tIs the time-space grid data of a certain historical moment.

Further, in step 3, a global-local space-time residual combination-based model is constructed, and the model is composed of a global spatial feature extraction component, a local spatial feature extraction component and a long-term temporal feature extraction component, and the specific steps are as follows:

step 3-1: constructing a global spatial feature extraction component, which comprises the following specific steps:

step 3-1-1: determining a global spatial feature extraction component overall network architecture, extracting global spatial features of the time-space grid data, wherein the size of the time-space grid data is M, selecting the size of a convolution kernel as M, performing convolution operation on the time-space grid data, and setting the step length to be 0; the global spatial feature extraction component convolution operation is defined as:

wherein, X_s ^l-1Space-time grid data to be extracted global spatial features of the first layer of convolutional layer, W_sG1 ^lIs the convolution kernel parameter of the l-th layer global feature extraction convolution layer, the dimension and X of the parameter_s ^l-1Are equal in dimension X_sG1 ^lIs the spatial feature extracted by the l-th layer global feature extraction convolutional layer, b^l _sG1Is the bias term, L, of the L-th layer global feature extraction convolutional layer_sG1The number of layers of convolution required by the global spatial feature extraction component;

step 3-1-2: output X of global component_{s_G1} ^lInputting a residual unit, the operation being defined as:

wherein X_sG1 ^l-1Is the global spatial feature input, X, of the l-th layer residual unit_sG2 ^lIs the global spatial feature output of the l-th layer residual unit, θ_sG2 ^lIs the set of all learnable parameters in the layer i residual unit of the global component, F_GResidual mapping of the extraction component for global spatial features, L_sG2The number of layers of the global component needing residual error;

step 3-1-3: setting an activation function in the global spatial feature extraction component as a Relu function, determining the global convolution depth and the number of residual error units, and determining an initialized training step length lambda;

step 3-2: constructing a local spatial feature extraction component, which comprises the following specific steps:

step 3-2-1: determining a total network architecture of a local spatial feature extraction component, wherein the local spatial feature extraction component only performs convolution operation on the spatio-temporal raster data and does not perform pooling operation, and the size of a convolution kernel is smaller than the dimensionality of the data; the local spatial feature extraction component convolution operation is defined as:

wherein, X_s ^{l -1}Space-time grid data, W, of the first layer of convolutional layer to be extracted with local spatial features_sL1 ^lIs the convolution kernel parameter of the first layer local feature extraction convolution layer, the dimension of the parameter is less than X_s ^l-1Dimension of (2), X_sL1 ^lIs the spatial feature extracted by the first layer local feature extraction convolution layer, b^l _sL1Is the bias term, L, of the first layer local feature extraction convolution layer_sL1The number of layers required for convolution of the local spatial feature extraction component;

step 3-2-2: output X of partial assembly_{s_L1} ^lInputting a residual unit, the operation being defined as:

wherein, X_sL1 ^l-1Is the local spatial feature input, X, of the l-th layer residual unit_sL2 ^lIs the local spatial feature output of the layer I residual unit, θ_sL2 ^lIs the set of all learnable parameters in the local component layer I residual unit, F_LExtracting residual mappings of components for local spatial features, L_sL2The number of layers of residual errors required by local components;

step 3-2-3: setting an activation function in the local spatial feature extraction component as a Relu function, and determining the local convolution depth and the number of residual error units, the training step length and the training step length of the global component;

step 3-3: constructing a long-term time feature extraction component, which comprises the following specific steps:

step 3-3-1: determining the overall network architecture of the long-term time feature extraction component, and determining the neuron number and the hidden layer depth of an input layer and an output layer;

step 3-3-2: capturing temporal features of spatio-temporal raster data with a GRU, the temporal feature extraction component operation defined as: x_tem＝Re2(G(Re1(X_tem ^(m,n)) In a) of X), wherein X_tem ^(m,n)Is space-time grid data of dimension (m, n), Re1 is a matrix dimension-changing operation that changes matrix dimension from (m, n) to (1, m n), G is a GRU operation of the temporal feature extraction component, Re2 is a matrix dimension-changing operation that changes matrix dimension from (1, m n) to (m, n), X is_temThe final output of the temporal feature extraction component;

step 3-3-3: setting an activation function in the long-term time feature extraction component as a Tanh function, wherein the training step length is equal to the overall component;

step 3-4: constructing a global-local space-time residual error combination model, which comprises the following specific steps:

step 3-4-1: determining a global-local space-time residual error combination model overall network architecture;

step 3-4-2: outputting the global space characteristics X by adopting a parameter matrix fusion method_sGLocal spatial feature X input_sLOutbound and long term time feature output X_temCarrying out weighted fusion, wherein the formula is as follows: x_Fusion＝f(W_sG*X_sG+W_sL*X_sL+W_tem*X_tem) Wherein ". sup" represents the Hadamard product, W_sG、W_sLAnd W_temRespectively representing the proportion of the global space feature, the local space feature and the long-term time feature, f is an activation function, and the final output activation function is set to be a Sigmod function.

Further, in the step 4, the trained and constructed space-time grid data at the next moment of the urban road network is predicted based on the global-local space-time residual combination model, and the specific steps are as follows:

step 4-1: initializing model structure parameters including convolution kernel dimensionality, convolution step length, activation function, initial weight, hidden layer number, training step length, residual error unit number and the like;

step 4-2: inputting the training set into a global spatial feature extraction component to capture global road network spatial features to obtain global spatial features W of the traffic road network_sG；

Step 4-3: the training set is input into a partial spatial feature extraction component to capture partial road network spatial features to obtain partial spatial features W of the traffic road network_sL；

Step 4-4: inputting the training set into a long-term time feature extraction component to capture road network time features to obtain long-term time features W of the traffic road network_tem；

And 4-5: performing weighted fusion on the outputs from the step 4-2 to the step 4-4, and obtaining a final predicted value X through an activation function_Fusion；

And 4-6: setting iteration times, taking the root mean square error as an error evaluation standard of the model, calculating the error between a predicted value and an actual value of the model, and updating the weight value of the model by using a back propagation algorithm;

and 4-7: and after the model training is finished, testing the prediction precision of the model by using the test set.

Has the beneficial effects that: the traffic flow prediction method of the invention constructs a global space characteristic extraction component and a local space characteristic extraction component on the basis of the traditional space-time residual error model, and respectively captures global space characteristics and local space characteristics of an urban road network; the long-term time feature extraction component is constructed to capture the long-term space-time features of the urban road network, so that the capture of the global space features and the long-term time features by the model is improved, the training error is reduced, and the training precision is improved.

The method aims at the defect that the traditional space-time residual error model lacks capture of long-term characteristics of traffic flow when processing space-time grid data; meanwhile, the spatial features of the traffic network are affected by the global and local features, the traditional method for extracting the spatial features only analyzes one feature, ignores the function of the other feature, provides a global-local space-time residual error combination model, constructs a global spatial feature extraction component and a local spatial feature extraction component, and respectively captures the global spatial features and the local spatial features of the urban network; the long-term time feature extraction component is constructed to capture the long-term space-time features of the urban road network, so that the capture of the global space features and the long-term time features by the model is improved, the training error is reduced, and the training precision is improved.

Drawings

FIG. 1 is a schematic diagram of steps of a short-term traffic flow prediction method based on a global-local residual error combination model according to the present invention;

FIG. 2 is a flow chart of a short-term traffic flow prediction method based on a global-local residual error combination model according to the present invention;

FIG. 3 is a model structure diagram of a short-term traffic flow prediction method based on a global-local residual error combination model according to the present invention;

FIG. 4 is a global spatial feature extraction diagram of a short-term traffic flow prediction method based on a global-local residual error combination model according to the present invention;

FIG. 5 is a local spatial feature extraction diagram of a short-term traffic flow prediction method based on a global-local residual error combination model according to the present invention;

FIG. 6 is a comparison graph of real data and predicted data of a short-term traffic flow prediction method test set based on a global-local residual error combination model.

Detailed Description

The technical method of the present invention will be further described in detail with reference to the accompanying drawings.

As shown in fig. 1-3, a method for predicting short-term traffic flow of an urban road network by using a global-local space-time residual error combination model comprises the following steps:

step 1) collecting traffic flow data of a city road network and transmitting the traffic flow data to a traffic big data cluster in real time, carrying out data preprocessing on the traffic flow data of the original city road network to reduce data redundancy, and converting the preprocessed traffic flow data of the city road network into space-time grid data according to the longitude and latitude of the road network; (ii) a

In the step 1, the preprocessed data are converted into space-time grid data according to the longitude and latitude of the road network, and the method specifically comprises the following steps: according to longitude and latitude, the city road network is constructedNetwork region divided into I and J, forming network X ═ { X_i,jEvery position in the network is regularly distributed, and the distance between adjacent positions is constant; for each position (i, j), the traffic flow data, x, for the current position are recorded at fixed time Δ t intervals_t ^i,jRepresenting the traffic flow data counted in the position (I, J) at the time t, and representing the traffic flow data counted in the network area I J as tensor

Converting traffic flow prediction problem into given historical space-time grid data { X }_tI t 0, …, K, predicting the spatio-temporal raster data X at the time K + Δ t_K+ΔtAnd K is the total number of samples of the collected urban road network traffic flow.

This embodiment adopts metropolis taxi data to carry out the training and the verification of model, and the data is derived from public data set, and this experimental data record metropolis 6 am taxi locating data to 12 o' clock evening, and the date is 2014 8 month 3 days to 2014 8 month 23 days, according to the taxi distribution situation, use metropolis 3 ring to turn into traffic grid data with the original data into as the boundary, use 5 minutes as the sampling interval, 3888 traffic grid data in total, regard the data of the previous 14 days as the training set, the data of the last 7 days as the test set.

Step 2) carrying out standardization processing on the time-space grid data, and dividing the time-space grid data into a training set and a testing set according to a certain proportion;

in the step 2, the Z-score standardization is adopted to carry out standardization processing on the space-time raster data, and the space-time raster data { X ] obtained in the step 2 is subjected to standardization processing_t| t ═ 0, …, K } by

Carrying out transformation; wherein

x_tIs the time-space grid data of a certain historical moment.

Step 3) constructing a global-local space-time residual error-based combined model, wherein the model consists of a global spatial feature extraction component, a local spatial feature extraction component and a long-term time feature extraction component;

in the step 3, a global-local space-time residual error-based combined model is constructed, and the model consists of a global space feature extraction component, a local space feature extraction component and a long-term time feature extraction component, and the specific steps are as follows:

step 3-1: constructing a global spatial feature extraction component, which comprises the following specific steps:

step 3-1-1: determining a global spatial feature extraction component overall network architecture, extracting global spatial features of the time-space grid data, wherein the size of the time-space grid data is M, selecting the convolution kernel size of M to perform convolution operation on the time-space grid data, and setting the step length to be 0; the global spatial feature extraction component convolution operation is defined as:

wherein, X_s ^l-1Space-time grid data to be extracted global spatial features of the first layer of convolutional layer, W_sG1 ^lIs the convolution kernel parameter of the l-th layer global feature extraction convolution layer, the dimension and X of the parameter_s ^l-1Are equal in dimension X_sG1 ^lIs the spatial feature extracted by the l-th layer global feature extraction convolutional layer, b^l _sG1Is the bias term, L, of the L-th layer global feature extraction convolutional layer_sG1The number of layers of convolution required by the global spatial feature extraction component;

step 3-1-2: output X of global component_{s_G1} ^lInputting a residual unit, the operation being defined as:

wherein, X_sG1 ^l-1Is the global spatial feature input, X, of the l-th layer residual unit_sG2 ^lIs the global spatial feature output of the l-th layer residual unit, θ_sG2 ^lIs the set of all learnable parameters in the l-th layer residual unit of the global component, F_GResidual mapping of the extraction component for global spatial features, L_sG2The number of layers of the global component needing residual error;

step 3-1-3: setting an activation function in the global spatial feature extraction component as a Relu function, determining the global convolution depth and the number of residual error units, and determining an initialized training step length lambda;

step 3-2: constructing a local spatial feature extraction component, which comprises the following specific steps:

step 3-2-1: determining a total network architecture of a local spatial feature extraction component, wherein the local spatial feature extraction component only performs convolution operation on the spatio-temporal raster data and does not perform pooling operation, and the size of a convolution kernel is smaller than the dimensionality of the data; the local spatial feature extraction component convolution operation is defined as:

wherein, X_s ^l-1Space-time grid data, W, of the first layer of convolutional layer to be extracted with local spatial features_sL1 ^lIs the convolution kernel parameter of the first layer of local feature extraction convolution layer, the dimension of the parameter is less than X_s ^l-1Dimension of (2), X_sL1 ^lIs the spatial feature extracted by the first layer local feature extraction convolution layer, b^l _sL1Is the bias term, L, of the first layer local feature extraction convolution layer_sL1The number of layers required for convolution of the local spatial feature extraction component;

step 3-2-2: output X of local component_{s_L1} ^lInputting a residual unit, the operation being defined as:

wherein, X_sL1 ^l-1Is the local spatial feature input, X, of the l-th layer residual unit_sL2 ^lIs the local spatial feature output of the l-th layer residual unit, θ_sL2 ^lIs the set of all learnable parameters in the local component layer I residual unit, F_LExtracting residual mappings of components for local spatial features, L_sL2The number of layers of residual errors required by local components;

step 3-2-3: setting an activation function in the local spatial feature extraction component as a Relu function, and determining the local convolution depth and the number of residual error units, the training step length and the training step length of the global component;

step 3-3: constructing a long-term time feature extraction component, which comprises the following specific steps:

step 3-3-1: determining the overall network architecture of the long-term time feature extraction component, and determining the neuron number and the hidden layer depth of an input layer and an output layer;

step 3-3-2: capturing temporal features of spatio-temporal raster data with a GRU, the temporal feature extraction component operation defined as: x_tem＝Re2(G(Re1(X_tem ^(m,n)) In a) of X), wherein X_tem ^(m,n)Is space-time grid data of dimension (m, n), Re1 is a matrix dimension-changing operation that changes matrix dimension from (m, n) to (1, m n), G is a GRU operation of the temporal feature extraction component, Re2 is a matrix dimension-changing operation that changes matrix dimension from (1, m n) to (m, n), X is_temThe final output of the temporal feature extraction component;

step 3-3-3: setting an activation function in the long-term time feature extraction component as a Tanh function, wherein the training step length is equal to the overall component;

step 3-4: constructing a global-local space-time residual error combination model, which comprises the following specific steps:

step 3-4-1: determining a global-local space-time residual error combination model overall network architecture;

step 3-4-2: outputting the global space characteristics X by adopting a parameter matrix fusion method_sGLocal spatial feature X input_sLOutbound and long term time feature output X_temCarrying out weighted fusion, wherein the formula is as follows: x_Fusion＝f(W_sG*X_sG+W_sL*X_sL+W_tem*X_tem) Wherein ". sup" represents the Hadamard product, W_sG、W_sLAnd W_temRespectively representing the proportion of the global space feature, the local space feature and the long-term time feature, f is an activation function, and the final output activation function is set to be a Sigmod function.

And 4) training by using the training set obtained in the step 2) and predicting the space-time grid data of the city road network at the next moment based on the global-local space-time residual error combination model.

In the step 4, the trained and constructed space-time grid data at the next moment of the urban road network is predicted based on the global-local space-time residual error combination model, and the specific steps are as follows:

step 4-1: initializing model structure parameters including convolution kernel dimensionality, convolution step length, activation function, initial weight, hidden layer number, training step length, residual error unit number and the like;

step 4-2: inputting historical space-time grid data into a global spatial feature extraction component to capture global road network spatial features to obtain global spatial features W of the traffic road network_sG；

Step 4-3: inputting historical space-time grid data into a local spatial feature extraction component to capture local road network spatial features to obtain local spatial features W of the traffic road network_sL；

Step 4-4: inputting historical space-time grid data into a long-term time feature extraction component to capture road network time features to obtain long-term time features W of a traffic road network_tem；

And 4-5: performing weighted fusion on the outputs from the step 4-2 to the step 4-4, and obtaining a final predicted value X through an activation function_Fusion；

And 4-6: setting iteration times, taking the root mean square error as an error evaluation standard of the model, calculating the error between a predicted value and an actual value of the model, and updating the weight value of the model by using a back propagation algorithm;

and 4-7: after the model training is completed, the prediction precision of the model is tested by using the test set, and the prediction result in the test set in a random day is shown in fig. 6.

Claims (4)

1. A short-time traffic flow prediction method based on a global-local residual error combination model is characterized by comprising the following steps: the method comprises the following steps:

step 1) collecting traffic flow data of an urban road network and transmitting the traffic flow data to a traffic big data cluster in real time, carrying out data preprocessing on the traffic flow data of the urban road network to reduce data redundancy, and converting the preprocessed traffic flow data of the urban road network into space-time grid data according to the longitude and latitude of the road network;

step 2) carrying out standardization processing on the time-space grid data, and dividing the time-space grid data into a training set and a test set;

step 3) constructing a global-local space-time residual error-based combined model, wherein the model consists of a global spatial feature extraction component, a local spatial feature extraction component and a long-term time feature extraction component;

step 4) training by using the training set obtained in the step 2) and predicting space-time grid data of the city road network at the next moment based on a global-local space-time residual error combination model;

in the step 3), a global-local space-time residual error-based combined model is constructed, and the model consists of a global space feature extraction component, a local space feature extraction component and a long-term time feature extraction component, and the specific steps are as follows:

step 3-1: constructing a global spatial feature extraction component, which comprises the following specific steps:

step 3-1-1: determining a global spatial feature extraction component overall network architecture, extracting global spatial features of the time-space grid data, wherein the size of the time-space grid data is M, selecting the convolution kernel size of M to perform convolution operation on the time-space grid data, and setting the step length to be 0; the global spatial feature extraction component convolution operation is defined as:

wherein, X_s ^l-1Space-time grid data to be extracted global spatial features of the first layer of convolutional layer, W_sG1 ^lIs the convolution kernel parameter of the l-th layer global feature extraction convolution layer, the dimension and X of the parameter_s ^l-1Are equal in dimension X_sG1 ^lIs the spatial feature extracted by the l-th layer global feature extraction convolutional layer, b^l _sG1Is the bias term, L, of the L-th layer global feature extraction convolutional layer_sG1The number of layers of convolution required by the global spatial feature extraction component;

step 3-1-2: extracting output X of global space characteristic extraction component_{s_G1} ^lInput residual unit, operation is defined as:

wherein, X_sG1 ^l-1Is the global spatial feature input, X, of the l-th layer residual unit_sG2 ^lIs the global spatial feature output of the l-th layer residual unit, θ_sG2 ^lIs the set of all learnable parameters in the residual unit of the l level of the global spatial feature extraction module, F_GResidual mapping of the extraction component for global spatial features, L_sG2The number of layers of residual errors required by the global spatial feature extraction component is set;

step 3-1-3: setting an activation function in the global spatial feature extraction component as a Relu function, determining the global convolution depth and the number of residual error units, and determining the initialized training step length lambda;

step 3-2: constructing a local spatial feature extraction component, which comprises the following specific steps:

step 3-2-1: determining a total network architecture of a local spatial feature extraction component, wherein the local spatial feature extraction component only performs convolution operation on the spatio-temporal raster data and does not perform pooling operation, and the size of a convolution kernel is smaller than the dimensionality of the data; the local spatial feature extraction component convolution operation is defined as:

wherein, X_s ^l-1Space-time grid data, W, of the first layer of convolutional layer to be extracted with local spatial features_sL1 ^lIs the convolution kernel parameter of the first layer local feature extraction convolution layer, the dimension of the parameter is less than X_s ^l-1Dimension of (2), X_sL1 ^lIs the spatial feature extracted by the first layer local feature extraction convolution layer, b^l _sL1Is the bias term, L, of the first layer local feature extraction convolution layer_sL1The number of layers required for convolution of the local spatial feature extraction component;

step 3-2-2: extracting output X of local spatial feature extraction component_{s_L1} ^lInput residual unit, operation is defined as:

wherein, X_sL1 ^l-1Is the local spatial feature input, X, of the l-th layer residual unit_sL2 ^lIs the local spatial feature output of the l-th layer residual unit, θ_sL2 ^lIs the set of all learnable parameters in the residual unit of the l layer of the local spatial feature extraction module, F_LExtracting residual mappings of components for local spatial features, L_sL2The number of layers of residual errors required by the local spatial feature extraction component;

step 3-2-3: setting an activation function in the local spatial feature extraction component as a Relu function, determining the number of local convolution depths and residual error units, wherein the training step length is equal to that of the global spatial feature extraction component;

step 3-3: constructing a long-term time feature extraction component, which comprises the following specific steps:

step 3-3-1: determining the overall network architecture of the long-term time feature extraction component, and determining the neuron number and the hidden layer depth of an input layer and an output layer;

step 3-3-2: the GRU is used for capturing the time characteristics of the space-time grid data, and the operation of the long-term time characteristic extraction component is defined as follows: x_tem＝Re2(G(Re1(X_tem ^(m,n)) In a) of X), wherein X_tem ^(m,n)Is space-time grid data of dimension (m, n), Re1 is a matrix dimension-changing operation that changes the matrix dimension from (m, n) to (1, m n), G is a GRU operation of the long-term temporal feature extraction component, Re2 is a matrix dimension-changing operation that changes the matrix dimension from (1, m n) to (m, n), X2_temA final output for the long-term temporal feature extraction component;

step 3-3-3: setting an activation function in the long-term time feature extraction component as a Tanh function, wherein the training step length is equal to that of the global component;

step 3-4: constructing a global-local space-time residual error combination model, which comprises the following specific steps:

step 3-4-1: determining a global-local space-time residual error combination model overall network architecture;

step 3-4-2: outputting the global space characteristics X by adopting a parameter matrix fusion method_sGLocal spatial feature X input_sLOutput and long term time characteristicsOut of X_temCarrying out weighted fusion, wherein the formula is as follows: x_Fusion＝f(W_sG*X_sG+W_sL*X_sL+W_tem*X_tem) Wherein ". sup" represents the Hadamard product, W_sG、W_sLAnd W_temRespectively representing the proportion of the global space feature, the local space feature and the long-term time feature, f is an activation function, and the final output activation function is set to be a Sigmod function.

2. The short-time traffic flow prediction method based on the global-local residual error combination model according to claim 1, characterized in that: in the step 1), the preprocessed urban road network traffic flow data is converted into space-time grid data according to the longitude and latitude of the road network, and the method specifically comprises the following steps: dividing the city road network into I X J network regions according to longitude and latitude to form a network X ═ X_i,jEvery position in the network is regularly distributed, and the distance between adjacent positions is constant; for each position (i, j), the traffic flow data, x, for the current position are recorded at fixed time Δ t intervals_t ^i,jRepresenting the traffic flow data counted in the position (I, J) at the time t, and representing the traffic flow data counted in the network area I J as tensor

Converting traffic flow prediction problem into given historical space-time grid data { X }_tI t 0, …, K, predicting the spatio-temporal raster data X at the time K + Δ t_K+ΔtAnd K is the total number of samples of the collected urban road network traffic flow.

3. The short-time traffic flow prediction method based on the global-local residual error combination model according to claim 2, characterized in that: in the step 2), the space-time raster data is standardized by adopting z-score standardization, and the space-time raster data { X) obtained in the step 2) is subjected to standardization treatment_tI t | (0, …, K) } through

Carrying out transformation; wherein

x_tIs the time-space grid data of a certain historical moment.

4. The short-time traffic flow prediction method based on the global-local residual error combination model according to claim 1, characterized in that: in the step 4), the training and construction of the global-local space-time residual error combination model based on prediction of the space-time grid data of the city road network at the next moment specifically comprises the following steps:

step 4-1: initializing model structure parameters including convolution kernel dimensionality, convolution step length, activation function, initial weight, hidden layer number, training step length and residual error unit number;

step 4-2: inputting the training set into a global spatial feature extraction component to capture global road network spatial features to obtain global spatial features W of the traffic road network_sG；

Step 4-3: the training set is input into a partial spatial feature extraction component to capture partial road network spatial features to obtain partial spatial features W of the traffic road network_sL；

Step 4-4: inputting the training set into a long-term time feature extraction component to capture road network time features to obtain long-term time features W of the traffic road network_tem；

And 4-5: performing weighted fusion on the outputs from the step 4-2 to the step 4-4, and obtaining a final predicted value X through an activation function_Fusion；

And 4-6: setting iteration times, taking the root mean square error as an error evaluation standard of the model, calculating the error between a predicted value and an actual value of the model, and updating the weight value of the model by using a back propagation algorithm;

and 4-7: and after the model training is finished, testing the prediction precision of the model by using the test set.

Images