CN117635216A - Shared bicycle demand prediction method based on improved GCN (generalized communication network) - Google Patents
Shared bicycle demand prediction method based on improved GCN (generalized communication network) Download PDFInfo
- Publication number
- CN117635216A CN117635216A CN202311775577.5A CN202311775577A CN117635216A CN 117635216 A CN117635216 A CN 117635216A CN 202311775577 A CN202311775577 A CN 202311775577A CN 117635216 A CN117635216 A CN 117635216A
- Authority
- CN
- China
- Prior art keywords
- time
- space
- shared bicycle
- graph
- improved
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 238000000034 method Methods 0.000 title claims abstract description 58
- 238000004891 communication Methods 0.000 title abstract description 3
- 239000011159 matrix material Substances 0.000 claims abstract description 84
- 230000004927 fusion Effects 0.000 claims abstract description 34
- 238000010586 diagram Methods 0.000 claims abstract description 26
- 238000007781 pre-processing Methods 0.000 claims abstract description 8
- 230000005540 biological transmission Effects 0.000 claims abstract description 7
- 238000012545 processing Methods 0.000 claims abstract description 7
- 238000004364 calculation method Methods 0.000 claims description 60
- 230000008569 process Effects 0.000 claims description 23
- 238000012549 training Methods 0.000 claims description 21
- 238000000605 extraction Methods 0.000 claims description 18
- 230000006870 function Effects 0.000 claims description 18
- 230000006872 improvement Effects 0.000 claims description 15
- 230000004913 activation Effects 0.000 claims description 14
- 230000003044 adaptive effect Effects 0.000 claims description 12
- 230000000737 periodic effect Effects 0.000 claims description 9
- 230000002776 aggregation Effects 0.000 claims description 5
- 238000004220 aggregation Methods 0.000 claims description 5
- 238000004140 cleaning Methods 0.000 claims description 4
- 238000010606 normalization Methods 0.000 claims description 4
- 230000008859 change Effects 0.000 claims description 3
- 238000005065 mining Methods 0.000 claims description 2
- 238000002759 z-score normalization Methods 0.000 claims description 2
- 230000000694 effects Effects 0.000 abstract description 5
- 238000013527 convolutional neural network Methods 0.000 description 4
- 238000013135 deep learning Methods 0.000 description 2
- 238000011161 development Methods 0.000 description 2
- 238000012360 testing method Methods 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 230000001419 dependent effect Effects 0.000 description 1
- 238000007499 fusion processing Methods 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 230000001932 seasonal effect Effects 0.000 description 1
- 238000000638 solvent extraction Methods 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
- 230000036962 time dependent Effects 0.000 description 1
- 238000010200 validation analysis Methods 0.000 description 1
- 238000012795 verification Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q30/00—Commerce
- G06Q30/02—Marketing; Price estimation or determination; Fundraising
- G06Q30/0201—Market modelling; Market analysis; Collecting market data
- G06Q30/0202—Market predictions or forecasting for commercial activities
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/042—Knowledge-based neural networks; Logical representations of neural networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/0464—Convolutional networks [CNN, ConvNet]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Business, Economics & Management (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Computational Linguistics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Biomedical Technology (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Biophysics (AREA)
- Health & Medical Sciences (AREA)
- Accounting & Taxation (AREA)
- Development Economics (AREA)
- Strategic Management (AREA)
- Finance (AREA)
- Entrepreneurship & Innovation (AREA)
- Game Theory and Decision Science (AREA)
- Economics (AREA)
- Marketing (AREA)
- General Business, Economics & Management (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a shared bicycle demand prediction method based on an improved GCN (general purpose communication network). The method comprises the steps of preprocessing shared bicycle demand data, calculating shared bicycle demand characteristics according to input data, generating a self-adaptive space-time diagram (namely an adjacency matrix of the diagram) with space-time dimension at the same time, inputting the bicycle demand characteristics and the self-adaptive space-time diagram into a space-time fusion backbone network, enabling the space-time fusion backbone network to conduct information transmission with space-time dimension at the same time, so as to obtain space-time fusion characteristics, and finally converting the space-time fusion characteristics into a shared bicycle demand prediction result by a prediction module. Based on the space-time data characteristics, the invention adopts the self-adaptive graph structure to learn more pertinent and real graph structure information, unifies the processing procedures of time and space, and thereby achieves better sharing bicycle demand prediction effect.
Description
Technical Field
The invention relates to the technical field of intelligent traffic, in particular to a shared bicycle demand prediction method based on an improved GCN (global positioning system) network.
Background
In the development and operation of smart cities and intelligent transportation systems (intelligent transportation systems, ITSs), sharing of bicycles has become a convenient and important city short trip mode, so accurate prediction research of demand of sharing of bicycles has important significance.
The shared bicycle demand prediction problem is more challenging than other time series prediction problems because it involves high-dimensional large data volumes, as well as a variety of variations including emergency situations. The bicycle demand status at a particular location is spatially dependent, and may be affected not only by the bicycle station, but also by the remote bicycle station, where an increase in bicycle station demand may result in an increase in the demand at another bicycle station. Furthermore, the bicycle site demand is time dependent, may be seasonal, and may be affected by holidays. Traditional linear time series models, such as autoregressive and integrated moving average (ARIMA) models, do not effectively address such spatio-temporal prediction problems. Machine learning and deep learning techniques have been introduced in this field to improve prediction accuracy, for example by modeling the bicycle sites of the entire city as a grid and applying Convolutional Neural Networks (CNNs) for processing. CNN can handle euclidean data such as images well, however, the method of CNN is not optimal in the face of the problem of the demand of a bicycle site, which is a form of a graph.
With the development of deep learning, a novel network structure GCN is proposed, the aggregation of feature aggregation sharing bicycle site information can be completed on the basis of a graph, the method is stronger in interpretation, a model capable of processing the sharing bicycle demand is modified on the basis of the method, the prediction effect can be improved, and the model is more interpretable.
However, in the space-time data prediction based on the GCN network, there are some problems, such as that the predefined graph has a large deviation from the actual situation, and the predefined graph is difficult to reflect the real spatial relationship; the information is not fully utilized, the operation in the time dimension and the operation in the space dimension are independent, and unified modeling is difficult to perform; the GCN network computation has an overcomplete problem, resulting in loss of valid information in the data.
Disclosure of Invention
The invention aims to overcome the defects and the shortcomings of the prior art, and provides a shared bicycle demand prediction method based on an improved GCN (global control system) network, which adopts a self-adaptive graph structure to learn more pertinent and real graph structure information, unifies the processing procedures of time and space, and achieves a better shared bicycle demand prediction effect.
In order to achieve the above purpose, the technical scheme provided by the invention is as follows: the method is based on an improved GCN (global control system) network to realize accurate prediction of the demand of the shared bicycle, wherein the improvement of the GCN network comprises improvement of input characteristics, an adjacency matrix of a graph and a calculation layer; the improvement of the input features is that a shared bicycle demand feature module is newly added, the shared bicycle demand feature module decomposes an input time sequence signal into a trend signal and a periodic signal through a moving average method and is used for improving the learning of a GCN network on complex signals, the shared bicycle demand feature module can learn stable and independent shared node features in a training set for the ultra-long time sequence signal, and the node features can enable the GCN network to better distinguish different bicycle stations; the improvement of the adjacent matrix of the graph is that an adaptive space-time graph structure learning module is newly added, the adaptive space-time graph structure learning module changes an N multiplied by N matrix which is originally used as the adjacent matrix of the graph into a TN multiplied by TN matrix, wherein N is the number of bicycle stations, and T is the length of a time window, so that the adjacent matrix of the graph is not only related to the bicycle stations but also related to time, the adjacent matrix of the graph contains information of space dimension and time dimension, and the adjacent matrix of the graph changes along with the training process, so that the adjacent matrix of the graph is more suitable for a real scene; the improvement of the calculation layer is to improve the message transmission calculation mode of the calculation layer from the original calculation mode which only has a space relation to the calculation mode which simultaneously comprises two dimensions of space and time according to the structural change of the adjacent matrix of the graph, so that the GCN network can simultaneously aggregate the characteristics of the space dimension and the time dimension, and the problem of overcorrection of the GCN network is relieved by adding a residual structure;
the specific implementation of the shared bicycle demand prediction method comprises the following steps:
acquiring and preprocessing the shared bicycle demand data to obtain the shared bicycle demand data added with time information;
the trained improved GCN network is utilized to carry out the following processing on the shared bicycle demand data added with time information:
deeply mining the input data by a shared bicycle demand characteristic module to obtain a time sequence signal, a node characteristic and a time characteristic, and mixing the time sequence signal, the node characteristic and the time characteristic to form a shared bicycle demand characteristic E, namely the shared bicycle demand characteristic E comprises the time sequence characteristic, the node characteristic and the time characteristic;
generating an adjacency matrix A of the graph simultaneously containing space dimension and time dimension by a self-adaptive space-time graph structure learning module according to the time window length T and the number N of bicycle stations;
inputting the adjacent matrix A of the shared bicycle demand quantity feature E and the graph into a space-time fusion backbone network of the improved GCN network to obtain a space-time fusion feature H, and inputting the space-time fusion feature H into a prediction module of the improved GCN network to obtain a final prediction result; wherein the spatio-temporal fusion backbone network consists of an improved computational layer and residual structure.
Further, the preprocessing includes the following operations:
data cleaning: in the process of collecting the shared bicycle demand data, invalid or erroneous data exist, so that the data need to be cleaned, and the invalid or erroneous data are removed to ensure the quality and consistency of the data;
data normalization: because the range and the size of the data of different data sets can be different, in order to enable the model to have better universality, the data is normalized by using the z-score normalization;
adding time information: time information is added according to the collection time interval, and the time information comprises the data of the time of day and the day of week in the week.
Further, the improved GCN network includes:
the sharing bicycle demand quantity feature module is divided into a time sequence extraction feature module, a node feature extraction module, a word embedding and feature fusion linear layer of time information;
the self-adaptive space-time diagram structure learning module generates an adjacent matrix of the diagram through a matrix initialized randomly, and controls the time and space relation of the diagram through a mask matrix;
the space-time fusion backbone network consists of 3 improved calculation layers and residual error structures;
the prediction module consists of a linear layer, a relu activation function, a residual error structure and Dropout, and the Dropout is used for effectively slowing down the overfitting and training network parameters with more generalization.
Further, the time sequence extraction feature module decomposes an input time sequence signal into a trend signal and a periodic signal through a moving average method, obtains the trend feature and the periodic feature of the two signals through a linear layer, and adds the two features to obtain a final time sequence feature Z, so that a network can better distinguish the two signals, and the capability of learning complex time sequence features is improved;
the node characteristic extraction module acquires an ultra-long time sequence signal from the training set, obtains corresponding time sequence characteristics from the ultra-long time sequence signal through a plurality of one-dimensional expansion convolutions, batchNorm and residual structures, expands the time sequence characteristics, and finally obtains node characteristics D through linear layers, dropout and relu activation functions; the node characteristic extraction module can acquire the ultra-long time sequence characteristic by using one-dimensional expansion convolution and a small amount of parameters, and acquire a stable and independent node characteristic by using a training set ultra-long time sequence signal, wherein the node characteristic can enable a GCN network to better distinguish different bicycle stations;
the word embedding of the time information is to randomly initialize the corresponding word embedding parameter E according to the data of the time signal in one day and the day of the week tid And E is diw Word embedding parameters are passed through the webTraining of collaterals is continuously learned and optimized; the corresponding time characteristic T is obtained by embedding the time signal and the word of the input data tid And T diw ；
The feature fusion linear layer is Z, D, T tid And T diw And splicing to form the shared bicycle demand characteristic E.
Further, the adaptive space-time diagram structure learning module randomly initializes two learnable matrixes W first 1 And W is 2 The two matrixes are TN multiplied by c matrixes, c is a super parameter and is used for controlling the complexity of generating the matrixes, then, a mask matrix with the dimension TN multiplied by TN is generated and is used for controlling the information transmission calculation of the improved GCN network, the space feature aggregation is considered only on the last time step, the real scene is more closely related, and finally, the matrix W is used for passing through 1 、W 2 And a mask matrix generating adjacent matrix A of the graph, wherein the process is as shown in (1):
A＝mask⊙(SoftMax(relu(W 1 ·transpose(W 2 )))) (1)
wherein, the formula is that the mask is a 0-1 matrix of TN multiplied by TN, the relu activation function is used for eliminating weak connection, the softMax function is used for normalizing the adaptive adjacent matrix, and the adaptive space-time diagram matrix generated by the process, namely the adjacent matrix A of the diagram, not only contains time dimension information, but also is more similar to the real situation.
Further, in the whole space-time fusion backbone network, the shared bicycle demand feature E is a total input, the output of each improved calculation layer is taken as the input of the next improved calculation layer, and the output of each improved calculation layer is added to be taken as the output of the space-time fusion backbone network; for each improved calculation layer, padding the input to control the scale before and after the input, obtaining the hidden characteristic of each node by using a linear layer, carrying out graph convolution along the time dimension, finally splicing the result of the graph convolution in the time dimension, and adding the output of the calculation layer with improved residual error formed by the input, wherein the calculation process of the improved calculation layer is shown in the formulas (2), (3) and (4):
H l ＝padding(G l ) (2)
the calculation process of the first +1th modified calculation layer is shown in formulas (2), (3) and (4), wherein G l Is the input of layer 1, G l+1 Is the output of layer l+1, in formula (2), H l Is G l The result after padding is performed, equation (3) is the convolution calculation process at the t-th time step of layer 1, where σ is the activation function, W l+1 Is the linear layer parameter of the 1 st calculation layer,is H l Part of the time step from T-T to T, -/->Is the result of the graph convolution calculation of the (1+1) th layer of the (t) th time step, A is the adjacency matrix A of the graph generated by the self-adaptive space-time graph structure learning module, P is the time step of the input data in the formula (4), and->Is the result of the graph convolution calculation for the t+p time step at layer l+1.
Compared with the prior art, the invention has the following advantages and beneficial effects:
1. compared with the common shared bicycle demand prediction, the method disclosed by the invention has the advantages that the space-time sequence prediction modeling is adopted instead of the simple linear time sequence modeling, the bicycle site itself and factors among bicycle sites are comprehensively considered for carrying out the shared bicycle demand prediction, and the accuracy and the reliability of the model are improved.
2. The factors influencing the demand of the shared bicycle are analyzed and mined in detail by combining the data set condition, a shared bicycle demand characteristic module is designed, and the interpretation and accuracy of the model are improved.
3. The graph convolution and graph structure are expanded to the space-time double dimension, so that unification of time dimension operation and space dimension operation is realized, and the model can better fuse time-space information.
4. By adopting the self-adaptive graph structure, more pertinent and real time space information can be learned, and a mask matrix is designed for controlling and improving the time space fusion process of the message transmission calculation of the GCN network, so that a better prediction effect is achieved.
5. A residual structure is added in the GCN network calculation process to effectively relieve the problem of excessive smoothness of the GCN.
Drawings
FIG. 1 is a block diagram of a model of the present invention; in the figure, X is the input of a model, E is the output shared bicycle demand characteristic of the shared bicycle demand characteristic module, A is the adjacency matrix of the space-time diagram generated by the self-adaptive space-time diagram structure learning module, and H is the output space-time fusion characteristic of a space-time fusion backbone network (which can be called a backbone).
FIG. 2 is a schematic flow chart of the method of the present invention shown in FIG. 1.
FIG. 3 is a block diagram of a shared bicycle demand feature module; in the figure, Z is a time sequence feature, D is a node feature extracted by a node feature extraction module, and T is tid And T diw Representing the time signal data of the day and the time of week of the week, respectively.
Fig. 4 is a block diagram of a space-time fusion backbone network (which may be referred to as a backhaul).
Detailed Description
The present invention will be described in further detail with reference to examples and drawings, but embodiments of the present invention are not limited thereto.
The embodiment discloses a shared bicycle demand prediction method based on an improved GCN (global control network) network, which is used for realizing accurate prediction of the shared bicycle demand based on the improved GCN network, wherein the improvement of the GCN network comprises improvement of input characteristics, an adjacent matrix of a graph and a calculation layer; the improvement of the input features is that a shared bicycle demand feature module is newly added, the shared bicycle demand feature module decomposes an input time sequence signal into a trend signal and a periodic signal through a moving average method and is used for improving the learning of a GCN network on complex signals, the shared bicycle demand feature module can learn stable and independent shared node features in a training set for the ultra-long time sequence signal, and the node features can enable the GCN network to better distinguish different bicycle stations; the improvement of the adjacent matrix of the graph is that an adaptive space-time graph structure learning module is newly added, the adaptive space-time graph structure learning module changes an N multiplied by N matrix which is originally used as the adjacent matrix of the graph into a TN multiplied by TN matrix, wherein N is the number of bicycle stations, and T is the length of a time window, so that the adjacent matrix of the graph is not only related to the bicycle stations but also related to time, the adjacent matrix of the graph contains information of space dimension and time dimension, and the adjacent matrix of the graph changes along with the training process, so that the adjacent matrix of the graph is more suitable for a real scene; the improvement of the calculation layer is to improve the message transmission calculation mode of the calculation layer from the original calculation mode which only has a space relation to the calculation mode which simultaneously comprises two dimensions of space and time according to the structural change of the adjacent matrix of the graph, so that the GCN network can simultaneously aggregate the characteristics of the space dimension and the time dimension, and the problem of overcorrection of the GCN network is relieved by adding a residual structure.
As shown in fig. 1, the improved GCN network architecture includes:
the sharing bicycle demand quantity feature module is divided into a time sequence extraction feature module, a node feature extraction module, a word embedding and feature fusion linear layer of time information;
the self-adaptive space-time diagram structure learning module generates an adjacent matrix of the diagram through a matrix initialized randomly, and controls the time and space relation of the diagram through a mask matrix;
a spatio-temporal fusion backbone network (which may be referred to as a backhaul) consisting of 3 improved computational layers and a residual structure;
the prediction module consists of a linear layer, a relu activation function, a residual structure and Dropout.
As shown in fig. 2, the specific implementation of the method for predicting the demand of the shared bicycle includes:
1) Preprocessing shared bicycle demand data in a data set, including: data cleaning, data normalization, adding time information and data set division;
data cleaning: in the process of collecting the shared bicycle demand data, invalid or erroneous data exist, so that the data need to be cleaned, and the invalid or erroneous data are removed to ensure the quality and consistency of the data;
data normalization: because the range and the size of the data of different data sets can be different, in order to enable the model to have better universality, the data is normalized by using z-score standardization, the method is to calculate standard deviation and average value in a training set, and the whole data is standardized according to the standard deviation and the average value;
adding time information: increasing time information according to the collection time interval, wherein the time information comprises data of the day and the day of the week, for example, the collection time interval of a selected data set is 30 minutes, then the day is divided into 48 serial numbers to mark the data of the day, and the week is divided into 7 serial numbers to mark the day of the week, and the day of the week is divided into 7 serial numbers;
data set partitioning: the method comprises the following steps of: 2:2 dividing the data set into a training set, a validation set, a test set, and ensuring that each data set appears in only one data set. The training set is used for network training, the verification set is used for selecting the best model parameters, and the test set is used for verifying the effect of the model;
after data preprocessing, the whole sharing bicycle demand prediction problem can be defined as formula (1):
in the method, in the process of the invention,shared bicycle with t representing time stepThe demand condition, N is the number of bicycle stations, and C is the data dimension of the shared bicycle demand condition, including the data dimension collected by the data set and the time information dimension added by the data preprocessing; a is a matrix of a graph structure, and in the embodiment, A is a matrix generated by a self-adaptive space-time graph structure learning module; the problem of shared bicycle demand prediction aims at learning a function f, and predicting the shared bicycle demand of Q time steps in the future through the shared bicycle demand conditions of P time steps in the past and the matrix A; />Also representing the demand status of the shared bicycle with time step t, but containing only demand data,/->Is f according to [ X ] t-P+1 ,...,X t ；A]Is a predicted result of (a).
2) Inputting the preprocessed data into a trained improved GCN network to perform the following processing:
the shared bicycle demand characteristic is formed by a shared bicycle demand characteristic module, and is composed of a time sequence characteristic, a space characteristic and a time characteristic, the specific structure of the shared bicycle demand characteristic module is shown in fig. 3, wherein the shared bicycle demand characteristic module can be divided into a time sequence extraction characteristic module, a node characteristic extraction module, a word embedding and characteristic fusion linear layer of time information, and the specific structure is as follows:
and a time sequence extraction feature module: the input time sequence signal is decomposed into two signals, namely a trend signal and a periodic signal, the two signals are subjected to linear layer to obtain two characteristics, namely a trend characteristic and a periodic characteristic, and the two characteristics are added to obtain a final time sequence characteristic Z, so that the network can better distinguish the two signals, and the capability of learning complex time sequence characteristics is improved;
the node characteristic extraction module: acquiring an ultra-long time sequence signal from a training set, acquiring corresponding time sequence features of the ultra-long time sequence signal through a plurality of one-dimensional expansion convolutions, batchNorm and residual structures, expanding the time sequence features, and finally acquiring node features D through linear layers, dropout and relu activation functions; the node characteristic extraction module can acquire the ultra-long time sequence characteristic by using one-dimensional expansion convolution and a small amount of parameters, and acquire a stable and independent node characteristic by using a training set ultra-long time sequence signal, wherein the node characteristic can enable a GCN network to better distinguish different bicycle stations;
word embedding of time information: randomly initializing a corresponding word embedding parameter E according to the data of the added time signal in one day and the day of the week tid And E is diw The word embedding parameters are continuously learned and optimized through the training of the network; the corresponding time characteristic T is obtained by embedding the time signal and the word of the input data tid And T diw ；
Feature fusion linear layer: will Z, D, T tid And T diw And splicing to form a shared bicycle demand characteristic E, wherein the shared bicycle demand characteristic E is fused with various characteristics and can better reflect the condition of input data.
Generating an adjacency matrix A of the graph simultaneously containing space dimension and time dimension by a self-adaptive space-time graph structure learning module according to the time window length T and the number N of bicycle stations; the self-adaptive space-time diagram structure learning module can learn a space-time diagram matrix which contains time dimension information and is closer to the real situation, and the matrix is used for preparing for the calculation of a subsequent improved calculation layer; the self-adaptive space-time diagram structure learning module randomly initializes two learnable matrixes W firstly 1 And W is 2 Both matrices are TN×c matrices, c is a hyper-parameter for controlling the complexity of generating the matrix, then generating a mask matrix of dimension TN×TN, the mask matrix being represented by equation (2),
where i is a row of the mask matrix, j is a column of the mask matrix, t=t, i.e. the value in the time dimension is equal to the time windowAt T; at this time, the mask matrix can be used to control the message passing computation of the improved computation layer, only consider the space feature aggregation on the last time step, and more closely fit the real scene, and finally pass through the matrix W 1 、W 2 And a mask matrix generating adjacent matrix A of the graph, wherein the process is as shown in (3):
A＝mask⊙(SoftMax(relu(W 1 ·transpose(W 2 )))) (3)
wherein, the formula is that the formula is Hadamard product, the mask is TN multiplied by TN's 0-1 matrix, the relu activation function is used for eliminating weak connection, the softMax function is used for standardizing the self-adaptive adjacent matrix, and the adjacent matrix A of the graph generated by the process not only comprises time dimension information, but also is more close to the real situation.
The space-time fusion backbone network is shown in fig. 4, and consists of 3 improved calculation layers and residual structures; the output of each improved computation layer serves as the input of the next improved computation layer, and the output of each improved computation layer is added to serve as the output of the space-time fusion backbone network; for each improved calculation layer, padding the input to control the scale before and after the input, obtaining the hidden characteristic of each node by using a linear layer, carrying out graph convolution along the time dimension, and finally, splicing the result of the graph convolution in the time dimension and adding the output of the calculation layer with improved residual error formed by the input, wherein the calculation process of the improved calculation layer is shown in the formulas (4), (5) and (6):
H l ＝padding(G l ) (4)
the calculation process of the first +1th modified calculation layer is shown in formulas (4), (5) and (6), wherein G l Is the input of layer 1, G l+1 Is the output of layer l+1, in formula (4), H l Is G l The result after padding is performed, equation (5) is the convolution calculation process at the t-th time step of layer 1, where σ is the activation function, and in the present method, the relu activation function, W l+1 Is the linear layer parameter of the 1 st calculation layer,is H l Part of the time step from T-T to T, -/->Is the result of the graph convolution calculation of the (1+1) th layer of the (t) th time step, A is the adjacency matrix A of the graph generated by the self-adaptive space-time graph structure learning module, P is the time step of the input data in the formula (6), and->Is the graph convolution calculation result of the (l+1) th layer T+P time step;
and inputting the generated shared bicycle demand quantity feature E and the adjacency matrix A of the graph into a space-time fusion backbone network, and obtaining a space-time fusion feature H after the space-time unified graph convolution.
The prediction module consists of a linear layer, a relu activation function, a residual error structure and Dropout, and the Dropout is used for effectively slowing down the overfitting, training network parameters with more generalization, and inputting the space-time fusion characteristic H into the prediction module to obtain a final prediction result.
The above examples are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above examples, and any other changes, modifications, substitutions, combinations, and simplifications that do not depart from the spirit and principle of the present invention should be made in the equivalent manner, and the embodiments are included in the protection scope of the present invention.
Claims (6)
1. The method is characterized in that the method is based on an improved GCN to realize accurate prediction of the shared bicycle demand, and the improvement of the GCN comprises improvement of input characteristics, an adjacency matrix of a graph and a calculation layer; the improvement of the input features is that a shared bicycle demand feature module is newly added, the shared bicycle demand feature module decomposes an input time sequence signal into a trend signal and a periodic signal through a moving average method and is used for improving the learning of a GCN network on complex signals, the shared bicycle demand feature module can learn stable and independent shared node features in a training set for the ultra-long time sequence signal, and the node features can enable the GCN network to better distinguish different bicycle stations; the improvement of the adjacent matrix of the graph is that an adaptive space-time graph structure learning module is newly added, the adaptive space-time graph structure learning module changes an N multiplied by N matrix which is originally used as the adjacent matrix of the graph into a TN multiplied by TN matrix, wherein N is the number of bicycle stations, and T is the length of a time window, so that the adjacent matrix of the graph is not only related to the bicycle stations but also related to time, the adjacent matrix of the graph contains information of space dimension and time dimension, and the adjacent matrix of the graph changes along with the training process, so that the adjacent matrix of the graph is more suitable for a real scene; the improvement of the calculation layer is to improve the message transmission calculation mode of the calculation layer from the original calculation mode which only has a space relation to the calculation mode which simultaneously comprises two dimensions of space and time according to the structural change of the adjacent matrix of the graph, so that the GCN network can simultaneously aggregate the characteristics of the space dimension and the time dimension, and the problem of overcorrection of the GCN network is relieved by adding a residual structure;
the specific implementation of the shared bicycle demand prediction method comprises the following steps:
acquiring and preprocessing the shared bicycle demand data to obtain the shared bicycle demand data added with time information;
the trained improved GCN network is utilized to carry out the following processing on the shared bicycle demand data added with time information:
deeply mining the input data by a shared bicycle demand characteristic module to obtain a time sequence signal, a node characteristic and a time characteristic, and mixing the time sequence signal, the node characteristic and the time characteristic to form a shared bicycle demand characteristic E, namely the shared bicycle demand characteristic E comprises the time sequence characteristic, the node characteristic and the time characteristic;
generating an adjacency matrix A of the graph simultaneously containing space dimension and time dimension by a self-adaptive space-time graph structure learning module according to the time window length T and the number N of bicycle stations;
inputting the adjacent matrix A of the shared bicycle demand quantity feature E and the graph into a space-time fusion backbone network of the improved GCN network to obtain a space-time fusion feature H, and inputting the space-time fusion feature H into a prediction module of the improved GCN network to obtain a final prediction result; wherein the spatio-temporal fusion backbone network consists of an improved computational layer and residual structure.
2. The method for predicting shared bicycle demand based on an improved GCN network of claim 1, wherein said preprocessing comprises the operations of:
data cleaning: in the process of collecting the shared bicycle demand data, invalid or erroneous data exist, so that the data need to be cleaned, and the invalid or erroneous data are removed to ensure the quality and consistency of the data;
data normalization: because the range and the size of the data of different data sets can be different, in order to enable the model to have better universality, the data is normalized by using the z-score normalization;
adding time information: time information is added according to the collection time interval, and the time information comprises the data of the time of day and the day of week in the week.
3. The method for predicting shared bicycle demand based on an improved GCN network of claim 2, wherein said improved GCN network comprises:
the sharing bicycle demand quantity feature module is divided into a time sequence extraction feature module, a node feature extraction module, a word embedding and feature fusion linear layer of time information;
the self-adaptive space-time diagram structure learning module generates an adjacent matrix of the diagram through a matrix initialized randomly, and controls the time and space relation of the diagram through a mask matrix;
the space-time fusion backbone network consists of 3 improved calculation layers and residual error structures;
the prediction module consists of a linear layer, a relu activation function, a residual error structure and Dropout, and the Dropout is used for effectively slowing down the overfitting and training network parameters with more generalization.
4. The method for predicting the demand of a shared bicycle based on an improved GCN network according to claim 3, wherein: the time sequence extraction feature module decomposes an input time sequence signal into a trend signal and a periodic signal through a moving average method, obtains the trend feature and the periodic feature of the two signals through a linear layer, and adds the two features to obtain a final time sequence feature Z, so that a network can better distinguish the two signals to improve the capability of learning complex time sequence features;
the node characteristic extraction module acquires an ultra-long time sequence signal from the training set, obtains corresponding time sequence characteristics from the ultra-long time sequence signal through a plurality of one-dimensional expansion convolutions, batchNorm and residual structures, expands the time sequence characteristics, and finally obtains node characteristics D through linear layers, dropout and relu activation functions; the node characteristic extraction module can acquire the ultra-long time sequence characteristic by using one-dimensional expansion convolution and a small amount of parameters, and acquire a stable and independent node characteristic by using a training set ultra-long time sequence signal, wherein the node characteristic can enable a GCN network to better distinguish different bicycle stations;
the word embedding of the time information is to randomly initialize the corresponding word embedding parameter E according to the data of the time signal in one day and the day of the week tid And E is diw The word embedding parameters are continuously learned and optimized through the training of the network; the corresponding time characteristic T is obtained by embedding the time signal and the word of the input data tid And T diw ；
The feature fusion linear layer is Z, D, T tid And T diw And splicing to form the shared bicycle demand characteristic E.
5. The method for predicting the demand of a shared bicycle based on an improved GCN network as claimed in claim 3, wherein said adaptive space-time diagram structure learning module randomly initializes two learnable matrices W first 1 And W is 2 The two matrixes are TN multiplied by c matrixes, c is a super parameter and is used for controlling the complexity of generating the matrixes, then, a mask matrix with the dimension TN multiplied by TN is generated and is used for controlling the information transmission calculation of the improved GCN network, the space feature aggregation is considered only on the last time step, the real scene is more closely related, and finally, the matrix W is used for passing through 1 、W 2 And a mask matrix generating adjacent matrix A of the graph, wherein the process is as shown in (1):
A＝mask⊙(SoftMax(relu(W 1 ·transpose(W 2 )))) (1)
wherein, the formula is that the mask is a 0-1 matrix of TN multiplied by TN, the relu activation function is used for eliminating weak connection, the softMax function is used for normalizing the adaptive adjacent matrix, and the adaptive space-time diagram matrix generated by the process, namely the adjacent matrix A of the diagram, not only contains time dimension information, but also is more similar to the real situation.
6. A method of predicting shared bicycle demand based on an improved GCN network according to claim 3, wherein the shared bicycle demand feature E is the total input throughout the spatio-temporal fusion backbone network, the output of each improved computational layer being the input of the next improved computational layer, the output of each improved computational layer being added as the output of the spatio-temporal fusion backbone network; for each improved calculation layer, padding the input to control the scale before and after the input, obtaining the hidden characteristic of each node by using a linear layer, carrying out graph convolution along the time dimension, finally splicing the result of the graph convolution in the time dimension, and adding the output of the calculation layer with improved residual error formed by the input, wherein the calculation process of the improved calculation layer is shown in the formulas (2), (3) and (4):
H l ＝padding(G l ) (2)
the calculation process of the first +1th modified calculation layer is shown in formulas (2), (3) and (4), wherein G l Is the input of layer 1, G l+1 Is the output of layer l+1, in formula (2), H l Is G l The result after padding is performed, equation (3) is the convolution calculation process at the t-th time step of layer 1, where σ is the activation function, W l+1 Is the linear layer parameter of the 1 st calculation layer,is H l Part of the time step from T-T to T, -/->Is the result of the graph convolution calculation of the (1+1) th layer of the (t) th time step, A is the adjacency matrix A of the graph generated by the self-adaptive space-time graph structure learning module, P is the time step of the input data in the formula (4), and->Is the result of the graph convolution calculation for the t+p time step at layer l+1.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311775577.5A CN117635216B (en) | 2023-12-22 | 2023-12-22 | Shared bicycle demand prediction method based on improved GCN (generalized communication network) |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311775577.5A CN117635216B (en) | 2023-12-22 | 2023-12-22 | Shared bicycle demand prediction method based on improved GCN (generalized communication network) |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN117635216A true CN117635216A (en) | 2024-03-01 |
| CN117635216B CN117635216B (en) | 2024-04-30 |
Family
ID=90035680
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202311775577.5A Active CN117635216B (en) | 2023-12-22 | 2023-12-22 | Shared bicycle demand prediction method based on improved GCN (generalized communication network) |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN117635216B (en) |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20210391723A1 (en) * | 2020-06-12 | 2021-12-16 | Tsinghua University | Method for monitoring short-term voltage stability of power system |
| CN114757418A (en) * | 2022-04-18 | 2022-07-15 | 广州市香港科大霍英东研究院 | Training method, system and equipment for sharing bicycle demand prediction model |
| CN115271833A (en) * | 2022-09-28 | 2022-11-01 | 湖北省楚天云有限公司 | Shared bicycle demand prediction method and prediction system |
| US20230334981A1 (en) * | 2022-04-19 | 2023-10-19 | East China Jiaotong University | Traffic flow forecasting method based on multi-mode dynamic residual graph convolution network |
| CN116993391A (en) * | 2023-06-15 | 2023-11-03 | 东南大学 | Site type shared bicycle system use demand prediction method |
-
2023
- 2023-12-22 CN CN202311775577.5A patent/CN117635216B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20210391723A1 (en) * | 2020-06-12 | 2021-12-16 | Tsinghua University | Method for monitoring short-term voltage stability of power system |
| CN114757418A (en) * | 2022-04-18 | 2022-07-15 | 广州市香港科大霍英东研究院 | Training method, system and equipment for sharing bicycle demand prediction model |
| US20230334981A1 (en) * | 2022-04-19 | 2023-10-19 | East China Jiaotong University | Traffic flow forecasting method based on multi-mode dynamic residual graph convolution network |
| CN115271833A (en) * | 2022-09-28 | 2022-11-01 | 湖北省楚天云有限公司 | Shared bicycle demand prediction method and prediction system |
| CN116993391A (en) * | 2023-06-15 | 2023-11-03 | 东南大学 | Site type shared bicycle system use demand prediction method |
Also Published As
| Publication number | Publication date |
|---|---|
| CN117635216B (en) | 2024-04-30 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN113905391A (en) | Ensemble learning network traffic prediction method, system, device, terminal, and medium | |
| CN116187555A (en) | Traffic flow prediction model construction method and prediction method based on self-adaptive dynamic diagram | |
| CN112910690A (en) | Network traffic prediction method, device and equipment based on neural network model | |
| CN112766600B (en) | Urban area crowd flow prediction method and system | |
| CN112910711A (en) | Wireless service flow prediction method, device and medium based on self-attention convolutional network | |
| CN115762147B (en) | Traffic flow prediction method based on self-adaptive graph meaning neural network | |
| CN113920379B (en) | Zero sample image classification method based on knowledge assistance | |
| CN112382081A (en) | Traffic flow prediction method based on multiple tasks | |
| CN116383616B (en) | Track GPS coordinate recovery method and frame based on track similarity and deep learning | |
| CN115376317A (en) | Traffic flow prediction method based on dynamic graph convolution and time sequence convolution network | |
| CN116844041A (en) | Cultivated land extraction method based on bidirectional convolution time self-attention mechanism | |
| CN112418939A (en) | Method for mining space-time correlation of house price based on neural network to predict house price | |
| CN113327417A (en) | Traffic flow prediction method based on 3D dynamic space-time residual convolution associated network | |
| CN116052419A (en) | Deep learning-based graph neural network traffic flow prediction method | |
| CN117852588A (en) | AI chip evaluation parameter determination method based on deep residual neural network | |
| CN117635216B (en) | Shared bicycle demand prediction method based on improved GCN (generalized communication network) | |
| CN117236666A (en) | Emergency material demand analysis method and system | |
| CN117910814A (en) | Traffic management and control strategy optimization method and device based on digital twin and storage medium | |
| CN116246465A (en) | Traffic flow prediction method based on generated space-time evolution modeling | |
| CN115713155A (en) | Traffic sequence prediction method based on multivariate time sequence data analysis | |
| CN115938112A (en) | Traffic demand prediction method, system, electronic device, and computer storage medium | |
| CN111260121B (en) | Urban-range pedestrian flow prediction method based on deep bottleneck residual error network | |
| CN113515890A (en) | Renewable energy day-ahead scene generation method based on federal learning | |
| Cheng et al. | Traffic flow prediction based on information aggregation and comprehensive temporal-spatial synchronous graph neural network | |
| CN117496699B (en) | Traffic flow prediction method for multi-element traffic flow time-space data information |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |