CN115204445A - Travel demand prediction method and system for interaction between urban functional areas - Google Patents
Travel demand prediction method and system for interaction between urban functional areas Download PDFInfo
- Publication number
- CN115204445A CN115204445A CN202210541542.4A CN202210541542A CN115204445A CN 115204445 A CN115204445 A CN 115204445A CN 202210541542 A CN202210541542 A CN 202210541542A CN 115204445 A CN115204445 A CN 115204445A
- Authority
- CN
- China
- Prior art keywords
- matrix
- demand
- functional areas
- longitude
- latitude
- 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.)
- Pending
Links
Images
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
- G06Q10/00—Administration; Management
- G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
-
- 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)
- Business, Economics & Management (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Strategic Management (AREA)
- Human Resources & Organizations (AREA)
- Economics (AREA)
- Computational Linguistics (AREA)
- Development Economics (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Game Theory and Decision Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Health & Medical Sciences (AREA)
- Entrepreneurship & Innovation (AREA)
- Marketing (AREA)
- Operations Research (AREA)
- Quality & Reliability (AREA)
- Tourism & Hospitality (AREA)
- General Business, Economics & Management (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The embodiment of the invention discloses a travel demand prediction method and system for interaction between urban functional areas. The method specifically comprises the following steps: pre-dividing grids based on the irregularly distributed urban functional areas to obtain the longitude and latitude of grid edges; determining the distance between each area pair according to the longitude and latitude of the grid edge, and calculating an undirected graph adjacency matrix A based on the distance between each area pair; manufacturing a directed graph adjacency matrix TI based on the interaction type; manufacturing a matrix V based on the interactive demand among the functional areas; and training a travel demand prediction model according to the matrix V, the undirected graph adjacent matrix A and the directed graph adjacent matrix TI, and predicting demand information of travel demands interacted among the functional areas to be predicted in a preset time interval according to the travel demand prediction model. The problem of there is not enough research dynamics in interactive trip demand field between the present urban functional area is solved in this application.
Description
Technical Field
The invention belongs to the technical field of traffic prediction, and particularly relates to a method and a system for predicting travel demands interacted between urban functional areas.
Background
According to the statistics of Ministry of public Security, the number of motor vehicles in the country in 2021 is 3.95 hundred million. A large number of motor vehicles cause a series of problems of environmental pollution, energy consumption, congestion and diseases to road traffic. The travel demand prediction information based on the urban functional areas can help users to improve travel efficiency to a greater extent, provides a powerful traffic decision basis for traffic managers, and if accurate traffic conditions can be informed to the masses in advance, the traffic safety condition is improved, the carbon footprint is greatly reduced, and road network resources can be more fully utilized.
The deep learning method has strong big data representation and calculation capacity, has great success in a plurality of challenging works, and provides a new research direction for the learning task in the traffic field. Recurrent Neural Networks (RNNs) are mainly used for the prediction task of sequence data. As a variant of RNN, a Long Short-Term Memory network (LSTM) and a Gated Recursion Unit (GRU) relieve gradient disappearance or explosion phenomena existing in RNN and can effectively capture semantic association between Long sequences. However, LSTM is computationally intensive and difficult to train, GRU has relatively few parameters, and its prediction is not ideal in the case of large data size. Many researchers have employed the powerful feature extraction capabilities of Convolutional Neural Networks (CNNs) to model nonlinear spatial dependence. However, CNNs are for structured data, and cannot implement the capability of Graph Convolution Networks (GCNs) to handle non-euclidean structured data, and therefore cannot accurately extract the spatial correlation of functional inter-zone traffic in cities. And these methods above require the use of many layers to capture the longer sequences. The method combining CNN or GCN and RNN can jointly model the space-time information. Existing spatio-temporal models have further improved the ability to process data and capture spatio-temporal information. The prediction of travel demand based on only a single area at the starting place of a trip has more defects, for example, drivers in certain areas are limited by traffic control rules, only allowed to drive in a specific area or time, and for example, the drivers usually take orders after considering destinations according to personal conditions when accepting orders. When the problem of urban area division is solved, the uniform size division area grids cannot reflect the travel demand characteristics of the urban area, for example, the travel demand characteristics in the morning are that the demand from a living area to a school area and a business circle is large, the demand from the school area and the business circle to the living area is slightly large in the afternoon, and the demand from the living area to a leisure area is large on the weekend. Therefore, accurate urban travel demand prediction remains somewhat challenging.
In conclusion, the problem of insufficient research strength exists in the existing travel demand field of interaction between urban functional areas.
Disclosure of Invention
The embodiment of the invention aims to provide a travel demand prediction method and system for interaction between urban functional areas, which are used for solving the problem of insufficient research strength in the existing travel demand field of interaction between urban functional areas.
In order to achieve the above object, an embodiment of the present invention provides a method for predicting travel demands interacted between urban functional areas, where the method includes the following steps:
pre-dividing grids based on the irregularly distributed urban functional areas to obtain the longitude and latitude of grid edges;
determining the distance between each area pair according to the longitude and latitude of the grid edge, and calculating an undirected graph adjacency matrix A based on the distance between each area pair; manufacturing a directed graph adjacency matrix TI based on the interaction type; manufacturing a matrix V based on the interactive demand among the functional areas;
and training a travel demand prediction model according to the matrix V, the undirected graph adjacent matrix A and the directed graph adjacent matrix TI, and predicting demand information of travel demands interacted among the functional areas to be predicted in a preset time interval according to the travel demand prediction model.
Further, the method for obtaining the longitude and latitude of the grid edge by pre-dividing the grid based on the irregularly distributed urban functional area comprises the following steps:
selecting an area with the most dense trip demand data according to the distribution of the trip demand data in the city, and screening all trip demand data belonging to the area;
dividing four types of functional areas including a living area, a business circle, a learning area and a leisure area according to the screened areas;
dividing the city into grids with different sizes according to the attributes of the four types of functional areas;
wherein, the average value of the distance between the getting-on position and the getting-off position above the grid size is used as an approximate standard;
recording the longitude and latitude of the grid edge after the division, the functional attribute of the grid and the serial number;
and matching grids, functional area attributes and serial numbers of the getting-on position and the getting-off position according to the longitude and latitude of the grid edge of the original data and the longitude and latitude of the grid edge after the division is finished.
Furthermore, determining the longitude and latitude of the grid center points according to the longitude and latitude of the grid edges, wherein the distance between every two grid center points represents the distance between each area pair, and calculating an undirected graph adjacency matrix A based on the distance between each area pair;
determining an interaction type according to the functional attributes of two regions forming the interaction, and manufacturing a directed graph adjacency matrix TI based on the interaction type;
the requirements for interaction between each two functional zones within a time step form a matrix V.
Further, the method for training the travel demand prediction model according to the matrix V, the undirected graph adjacency matrix A and the directed graph adjacency matrix TI comprises the following steps:
extracting bottom layer characteristics of travel demand data input according to regional time sequence by a first layer time sequence convolution network of the travel demand prediction model, and inputting the bottom layer characteristics of the travel demand data, an undirected graph adjacent matrix A and a directed graph adjacent matrix TI to a second layer space convolution network;
the second layer of spatial convolution network receives the bottom layer characteristics of the first layer, processes the normalized undirected graph adjacent matrix A and the undirected graph adjacent matrix TI to acquire the topological characteristics and the spatial correlation of the road network, and inputs the time sequence bottom layer characteristics blended with the topological characteristics and the spatial correlation into the third layer of time sequence convolution network;
and the third layer of time sequence convolution network calculates the time dependency of the fusion data.
Further, predicting the demand information of the trip demands interacted between the functional areas to be predicted in a preset time interval according to the trip demand prediction model comprises the following steps:
because the change of the travel demand is influenced by the traffic state of the adjacent time period, a sequence prediction method in the travel demand prediction model is used;
since the travel demand change is affected by the periodic traffic state, a periodic prediction method in the travel demand prediction model is used.
Further, the method for extracting the bottom layer characteristics of the trip demand data input according to the regional time sequence by the time sequence convolution network comprises the following steps:
wherein f is i Represents a filter, K represents the size of the convolution kernel, X = (X) 1 ，x 2 ，……，x T ) Represents the input sequence, T represents the T-th implicit characteristic, d is a hole factor, which varies with an index of 2.
Further, the method for processing the normalized undirected graph adjacency matrix a and the normalized directed graph adjacency matrix TI to obtain the topological feature and the spatial correlation of the road network comprises the following steps:
wherein D ∈ R N×N Is a matrix of degrees, and is,
after the state of the neighbor node is considered, the adjacency matrix of the state of the self node is introduced G Representing a graph convolution operation, g θ For the convolution kernel, θ ∈ R K A vector representing the chebyshev coefficients, x representing the input data,
a matrix of laplacian values is represented,
λ max is the maximum eigenvalue, I, of the Laplace matrix n Is a unit matrix, T k (x)＝2xT k-1 (x)- T k-2 (x)，T 0 (x)＝1，T 1 (x)＝x。
A travel demand prediction system for interaction between urban functional areas, the system comprising:
the reading module is used for pre-dividing grids based on the urban functional areas distributed irregularly to obtain the longitude and latitude of the grid edges;
the analysis module is used for determining the distance between each area pair according to the longitude and latitude of the grid edge and calculating an undirected graph adjacency matrix A based on the distance between each area pair; manufacturing a directed graph adjacency matrix TI based on the interaction type; manufacturing a matrix V based on the interactive demand among the functional areas;
and the processing module is used for training a row demand prediction model according to the matrix V, the undirected graph adjacent matrix A and the directed graph adjacent matrix TI and predicting demand information of the trip demand interacted among the functional areas to be predicted in a preset time interval according to the row demand prediction model.
Furthermore, the reading module is used for pre-dividing grids based on the irregularly distributed urban functional areas to obtain the longitude and latitude of the grid edges;
selecting an area with the most dense trip demand data according to the distribution of the trip demand data in the city, and screening out all trip demand data belonging to the area;
dividing four types of functional areas including a living area, a business circle, a learning area and a leisure area according to the screened areas;
dividing the city into grids with different sizes according to the attributes of the four types of functional areas;
wherein, the average value of the distance between the upper vehicle position and the lower vehicle position above the grid size is used as an approximate standard;
recording the longitude and latitude of the grid edge after the division, the functional attribute of the grid and the serial number;
and matching grids, functional area attributes and serial numbers of the getting-on position and the getting-off position according to the longitude and latitude of the grid edge of the original data and the longitude and latitude of the grid edge after the division is finished.
Furthermore, determining the longitude and latitude of the grid center points according to the longitude and latitude of the grid edges, wherein the distance between every two grid center points represents the distance between each area pair, and calculating an undirected graph adjacency matrix A based on the distance between each area pair;
determining an interaction type according to the functional attributes of two regions forming the interaction, and manufacturing a directed graph adjacency matrix TI based on the interaction type;
the requirements for interaction between every two functional zones within a time step form a matrix V.
The embodiment of the invention has the following advantages:
the invention divides grids in advance based on the urban functional areas distributed irregularly to obtain the longitude and latitude of the grid edge; determining the distance between each area pair according to the longitude and latitude of the grid edge, and calculating an undirected graph adjacency matrix A based on the distance between each area pair; manufacturing a directed graph adjacency matrix TI based on the interaction type; manufacturing a matrix V based on the interactive demand among the functional areas; and training a travel demand prediction model according to the matrix V, the undirected graph adjacent matrix A and the directed graph adjacent matrix TI, and predicting demand information of travel demands interacted among the functional areas to be predicted in a preset time interval according to the travel demand prediction model. The problem of there is not enough research dynamics in interactive trip demand field between the current urban functional area is solved.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly described below. It should be apparent that the drawings in the following description are merely exemplary and that other implementation drawings may be derived from the provided drawings without inventive effort by those skilled in the art.
The structures, ratios, sizes, and the like shown in the present specification are only used for matching with the contents disclosed in the specification, so as to be understood and read by those skilled in the art, and are not used to limit the conditions of the present invention, so that the present invention has no technical significance, and any structural modifications, changes in the ratio relationship, or adjustments of the sizes should fall within the scope of the present invention without affecting the functions and the achievable purposes of the present invention.
Fig. 1 is a flowchart of a method for predicting travel demand for interaction between urban functional areas according to an embodiment of the present invention;
FIG. 2 is a schematic diagram of an undirected graph adjacency matrix that converts spatial locations into distance-based ones according to an embodiment of the present invention;
fig. 3 is a schematic diagram of converting data at a spatial location into a weighted directed graph and then into a matrix (a) according to an embodiment of the present invention;
fig. 4 is a schematic diagram of converting data at a spatial location into a weighted directed graph and then into a matrix (b) according to an embodiment of the present invention;
fig. 5 is a schematic diagram of converting data at a spatial location into a weighted directed graph and then into a matrix (c) according to an embodiment of the present invention;
fig. 6 is a schematic structural diagram of a travel demand prediction model for interaction between urban functional areas according to an embodiment of the present invention;
FIG. 7 is a schematic diagram of a hole causal convolution as provided by an embodiment of the present invention;
FIG. 8 is a diagram of a residual block according to an embodiment of the present invention;
fig. 9 is a schematic diagram of a spatial angle prediction result based on a sequential prediction method according to an embodiment of the present invention;
fig. 10 is a schematic diagram of a time-angle business circle-living area prediction result based on a sequence prediction method according to an embodiment of the present invention;
fig. 11 is a schematic diagram of a time-angle quotient circle-discipline area prediction result based on a sequential prediction method according to an embodiment of the present invention;
fig. 12 is a schematic diagram of a spatial angle prediction result based on a periodic prediction method according to an embodiment of the present invention;
fig. 13 is a schematic diagram of a time-angle leisure area-leisure area prediction result (a) based on a periodic prediction method according to an embodiment of the present invention;
fig. 14 is a schematic diagram of a time-angle leisure area-leisure area prediction result (b) based on a periodic prediction method according to an embodiment of the present invention;
fig. 15 is a schematic diagram of a time-angle leisure area-leisure area prediction result (c) based on a periodic prediction method according to an embodiment of the present invention;
fig. 16 is a schematic diagram of a time-based quotient circle-living area prediction result (a) based on a periodic prediction method according to an embodiment of the present invention;
fig. 17 is a schematic diagram of a time-based quotient circle-living area prediction result (b) based on a periodic prediction method according to an embodiment of the present invention;
fig. 18 is a schematic diagram of a time-based quotient circle-living area prediction result (c) based on a periodic prediction method according to an embodiment of the present invention;
fig. 19 is a schematic diagram of a time-angle quotient circle-discipline area prediction result (a) based on a periodic prediction method according to an embodiment of the present invention;
fig. 20 is a schematic diagram of a time-angle quotient circle-school zone prediction result (b) based on a periodic prediction method according to an embodiment of the present invention;
fig. 21 is a schematic diagram of a time-angle quotient circle-school zone prediction result (c) based on a periodic prediction method according to an embodiment of the present invention;
fig. 22 is a frame diagram of a system of a travel demand prediction system interacting between functional areas in cities according to an embodiment of the present invention.
Detailed Description
The present invention is described in terms of specific embodiments, and other advantages and benefits of the present invention will become apparent to those skilled in the art from the following disclosure. All other embodiments, which can be obtained by a person skilled in the art without making any creative effort based on the embodiments of the present invention, belong to the protection scope of the present invention.
Example 1
The embodiment is described with reference to fig. 1, and the travel demand prediction method for interaction between urban functional areas of the embodiment includes the following steps:
s1: pre-dividing grids based on the irregularly distributed urban functional areas to obtain the longitude and latitude of grid edges;
s2: determining the distance between each area pair according to the longitude and latitude of the grid edge, and calculating an undirected graph adjacency matrix A based on the distance between each area pair; manufacturing a directed graph adjacency matrix TI based on the interaction type; manufacturing a matrix V based on the interactive demand among the functional areas;
s3: and training a travel demand prediction model according to the matrix V, the undirected graph adjacent matrix A and the directed graph adjacent matrix TI, and predicting demand information of travel demands interacted among the functional areas to be predicted in a preset time interval according to the travel demand prediction model.
In the embodiment, the grids are divided in advance based on the urban functional areas distributed irregularly to obtain the longitude and latitude of the grid edges; determining the distance between each area pair according to the longitude and latitude of the grid edge, and calculating an undirected graph adjacency matrix A based on the distance between each area pair; manufacturing a directed graph adjacency matrix TI based on the interaction type; manufacturing a matrix V based on the interactive demand among the functional areas; and training a travel demand prediction model according to the matrix V, the undirected graph adjacent matrix A and the directed graph adjacent matrix TI, and predicting demand information of travel demands interacted among the functional areas to be predicted in a preset time interval according to the travel demand prediction model. The problem of there is not enough research dynamics in interactive trip demand field between the current urban functional area is solved.
In a preferred embodiment, the method for obtaining the longitude and latitude of the grid edge by pre-dividing the grid based on the irregularly distributed urban functional areas in the embodiment includes:
selecting an area with the most dense trip demand data according to the distribution of the trip demand data in the city, and screening all trip demand data belonging to the area;
dividing four types of functional areas including a living area, a business circle, a learning area and a leisure area according to the screened areas;
dividing the city into grids with different sizes according to the attributes of the four types of functional areas;
wherein, the average value of the distance between the upper vehicle position and the lower vehicle position above the grid size is used as an approximate standard;
recording the longitude and latitude of the grid edge after the division, the functional attribute of the grid and the serial number;
and matching grids, functional area attributes and serial numbers of the getting-on position and the getting-off position according to the longitude and latitude of the grid edge of the original data and the longitude and latitude of the grid edge after the division is finished.
In the present embodiment, the taxi order demand in haikou city, hainan province and metropolis, sichuan province will be described as an example. Firstly, drawing a data scatter diagram, selecting an area with the most dense trip demand data in a city, and screening all data belonging to the area.
Four types of functional areas are divided according to living areas, business circles, school areas and leisure areas. Cities are divided into grids of different sizes according to the property of the functional areas. In the determination of the grid size, the average value of the distances between the getting-on and getting-off positions is used as an approximate standard of the grid size. When the grid is too large, the total number of the grids is reduced, and the number of the interaction data discarded in the same grid is increased; when the grid is too small, the total number of grids increases and the feature matrix becomes too sparse. And finally, recording the longitude and latitude of the grid edge, the functional attribute and the serial number of the grid, and matching the grid, the functional area attribute and the serial number of the getting-on and getting-off positions according to the longitude and latitude of the original data and the longitude and latitude of the grid edge.
In a preferred embodiment, the longitude and latitude of the grid center points are determined according to the longitude and latitude of the grid edges, the distance between every two grid center points represents the distance between each area pair, and the undirected graph adjacency matrix a is calculated based on the distance between each area pair;
determining an interaction type according to the functional attributes of two regions forming the interaction, and manufacturing a directed graph adjacency matrix TI based on the interaction type;
the requirements for interaction between each two functional zones within a time step form a matrix V.
In this embodiment, a is an undirected graph adjacency matrix based on the distance between two regions, and represents the distance between different grids. In the calculation of the matrix a, the distance between the center points of the meshes is used to represent the distance between each pair of regions. Fig. 2 exemplarily shows a schematic diagram of the construction of the matrix a according to the embodiment of the present disclosure, as shown in fig. 2, the left side is the real space area, and the right side is the calculated distance matrix a. Since the city is divided into grids of different sizes, it is impossible to obtain all the coordinates of the center point by only one calculation. Obtaining coordinates of each central point according to the obtained edge coordinates of the grid, and using the following formula:
where d represents the distance between two regions, φ represents latitude, λ represents longitude, and r is the earth radius.
The directed graph adjacency matrix TI is composed of values of interaction types determined by functional attributes of two regions constituting the interaction, for example, four types of functional regions have 16 interactions in total, and thus 16 values are used to distinguish interaction categories.
The V matrix consists of the demand for interaction between every two grids in a time step. Since the V matrix is time-varying, a plurality of V matrices are formed. Fig. 3 exemplarily shows a schematic diagram of the matrix V construction according to the embodiment of the present disclosure, and as shown in fig. 3, data in a spatial position is abstracted as shown in fig. 3 into a weighted directed graph as shown in fig. 4, and then is converted into a matrix as shown in fig. 5. In fig. 5, the first row stores the total interactive demand in a time step from area 1 to the rest of areas, the second row stores the total interactive demand in the same time step from area 2 to the rest of grids, and so on to construct the V matrix. Where each value in the V matrix represents not a demand point value for an interaction within a single region but a demand edge value for an interaction between regions. We can see that the demand for interaction from zone 2 to zone 4 is 1 and the demand for interaction from zone 4 to zone 2 is 2.
In a preferred embodiment, the method for training the travel demand prediction model according to the matrix V, the undirected graph adjacency matrix a and the directed graph adjacency matrix TI in this embodiment includes:
extracting bottom layer characteristics of trip demand data input according to regional time sequence by a first layer time sequence convolution network of the trip demand prediction model, and inputting the bottom layer characteristics of the trip demand data, an undirected graph adjacent matrix A and a directed graph adjacent matrix TI into a second layer space convolution network;
the second-layer spatial convolution network receives the bottom-layer characteristics of the first layer, processes the normalized undirected graph adjacency matrix A and the normalized directed graph adjacency matrix TI to obtain the topological characteristics and the spatial correlation of the road network, and inputs the time sequence bottom-layer characteristics fused with the topological characteristics and the spatial correlation into the third-layer time sequence convolution network;
and the third layer of time sequence convolution network calculates the time dependency of the fusion data.
The travel demand prediction model in the embodiment is based on two layers of time sequence convolution networks and one layer of space convolution network, and is used for predicting interactive travel demand between functional areas.
Fig. 6 exemplarily shows an architecture diagram of a travel demand prediction model for interaction between city functional areas, as shown in fig. 6, a left-most column of fig. 6 represents input data, i.e., V input to a first time-sequence convolutional layer, and output result joint matrices a and TI thereof are input to a space convolutional layer.
As shown in fig. 4, the right part of the graph is from top to bottom, the first layer of time-series convolutional network extracts the bottom layer features of the input data, the second layer of space convolutional layer enhances the relative spatial relationship of the data features, the other layer of time-series convolutional network realizes the calculation of the high layer features on the time series, and finally, the final predicted values are output after being respectively normalized and fully connected.
In a preferred embodiment, the predicting, according to the travel demand prediction model, demand information of travel demands interacted between functional areas to be predicted in a preset time interval includes:
because the change of the travel demand is influenced by the traffic state of the adjacent time period, a sequence prediction method in the travel demand prediction model is used;
since the travel demand change is influenced by the periodic traffic state, a periodic prediction method in the travel demand prediction model is used.
In the embodiment, the accuracy of the travel demand prediction model is checked by using data of Hainan Haikou city and Sichuan metropolis.
Considering that the change in demand may be affected by recent traffic conditions, the order prediction method, i.e., streaming data for the first 12 time steps, is first used as an input to predict the demand for the next 4 time steps in the disclosed embodiment. The formula of the input data S is as follows:
wherein t is the current time slice, d is the d day, and n is the number of time slices required.
The evaluation indicators in the results of this experiment are Mean Absolute Error (MAE) and Root Mean Square Error (RMSE).
The calculation formula is as follows:
wherein the content of the first and second substances,
for prediction, y = { y = { [ y ] 1 ，y 2 ，……，y n The real value, n represents the total amount of data.
As can be seen from Table 1, we can see that the error of the model of the present application is the smallest among all models.
Table 1: comparison results of the model and other models under the sequence prediction method
9-11 schematically illustrate diagrams of verification of accuracy of a prediction task by demand data on a sequential prediction method according to an embodiment of the present disclosure.
From a spatial perspective, to verify that the demand conditions for different interaction types in the same time period are different, the average demand of 16 interaction situations in two specific time periods is selected for comparison, as shown in fig. 7. The abscissa represents the business district-business district, business district-living district, etc. 16 interaction, and the ordinate represents the demand for the same time period in a day.
From the time sequence perspective, in order to check that the demand prediction effect of the model of the present application in different time periods of the same interaction type is the best, the predicted value of the model with the more prominent effect is compared with the true value in a certain continuous time period, as shown in fig. 10 and 11. It can be seen that the fitting effect of the model is the best under the condition of large demand or small demand.
Considering that the demand change is influenced by the periodic traffic state, a periodic prediction method is added in the embodiment of the disclosure, that is, the demand data of a certain same time step in each day of the first seven days is used as an input to predict the demand of the same time step in the eighth day. The formula of the input data P is as follows:
wherein m is the number of days required.
From table 2, we can see that the error of the model of the present application is the smallest among all models, and the overall accuracy is improved compared with the result of the sequence prediction method.
Table 2: comparison result of model and other models under periodic prediction method
The results predicted by the model of the present application and other methods on both data sets are shown in table 2. The model of the present application achieves the lowest MAE and RMSE in this prediction task. In the Haiko dataset, the MAE and RMSE for the model of the present application were 3.3% and 3.6% lower than STGCN, and 27.0% and 27.1% lower than GRU.
12-21 illustratively show diagrams of verification of accuracy of a prediction task by demand data on a periodic prediction method by embodiments of the present disclosure.
From a spatial perspective, to verify that the demand conditions for different interaction types in the same time period are different, the average demand of 16 interaction situations in two specific time periods is selected for comparison, as shown in fig. 12. The abscissa indicates 16 interaction types of business district-business district, business district-living district, and the like, and the ordinate indicates the amount of demand for the same time period in a day.
From the time sequence perspective, in order to test that the demand prediction effect of the model is the best in the same interaction type and different time periods, the prediction values of all the models are compared with the real values in groups within a certain period of continuous time. Fig. 13-21 are predicted results of a large demand from business circles to living areas, a medium demand from business circles to school areas, and a small demand from leisure areas to leisure areas, respectively. It can be seen that the prediction trend and the fitting effect of the model are the best no matter the demand is large. The demand from the leisure area to the leisure area is not much in fig. 19 to 21, so that it is difficult for the model to obtain the regularity of the data. When the demand value changes long around 1, the model has difficulty predicting sudden peaks, resulting in an inadequate fit.
As can be seen from fig. 13 to 21, since only the time variation of the demand state is considered and the calculation of the spatial correlation is omitted, the conventional time series analysis method has a poor prediction effect and a limited capability of modeling nonlinear complex traffic data. On the basis of the LSTM model, CNN is added to form the STDN model for realizing space-time information processing, so that the prediction precision is improved. However, since CNN cannot process graph structures, its accuracy still remains to be improved. GCN is suitable for computing the spatial relationships of graph structures, but its prediction is not optimal due to its lack of temporal feature dependencies. The T-GCN respectively uses GCN and recurrent neural network to calculate the space-time relationship, and the experimental result is superior to LSTM, GRU, STDN and GCN. The STGCN, the ASTGCN and the model of the application are completely formed by convolution, so that the parameters are smaller, and the performance is higher than that of a recurrent neural network model. Compared with STGCN and ASTGCN, the predicted value of the model is closer to the true value when the local peak value is caused by rapid change.
In the prediction of the two methods, the overall performance of the periodic prediction result is better than that of the sequential prediction. The MAE and RMSE of the model under the periodic method are respectively 4.9% and 6.9% lower than those of the model under the sequence method, and the model achieves the best performance when compared with other models in the two methods.
In addition to the predictive performance of the model, the computational cost also needs to be considered. The average calculated time per round is shown in table 3.
Table 3: run time per round (seconds) for all models
| STDN | T-GCN | LSTM | GRU | ASTGCN | STGCN | Text model | GCN |
| 5.817 | 5.426 | 5.030 | 4.927 | 2.426 | 0.158 | 0.321 | 0.081 |
The time-series models LSTM and GRU are time-series models, and t-time depends on information of t-1 and cannot be calculated in parallel, so that the running time is long. The T-GCN and the STDN introduce GCN and CNN respectively on the basis of a recurrent neural network, so that parameters and calculation time are increased. The ASTGCN, STGCN and the model of the present application consist entirely of convolutions. However, as the ASTGCN designs a multi-path network consisting of three time periods and three are not parallel computation, the computation time is relatively long. The application model can realize parallel computation, thereby greatly reducing the computation time.
In a preferred embodiment, the method for extracting the bottom layer features of the trip demand data input according to the regional time sequence by using the time sequence convolutional network in the embodiment includes:
wherein f is i Represents a filter, K represents the size of the convolution kernel, X = (X) 1 ，x 2 ，……，x T ) Represents the input sequence, T represents the T-th implicit characteristic, d is a hole factor, which varies with an index of 2.
In this embodiment, according to the time sequence requirement change of the region, a time sequence convolution layer is constructed by using a residual error network to extract time-related feature information, where the time-related feature information includes causal convolution and hole convolution, and fig. 5 exemplarily illustrates a hole causal convolution implementation process in an embodiment of the present disclosure.
To ensure that historical data is not missed, causal convolutions are introduced. Causal convolution is a one-way time constraint model, i.e. only historical information is considered, not future information. The causal convolution is calculated as follows:
wherein f is i Represents a filter, K represents the size of the convolution kernel, X = (X) 1 ，x 2 ，……，x T ) Represents the input sequence, T represents the T-th implicit characteristic, d is a hole factor, which varies with an index of 2.
If we consider the variable x long before, the convolutional layer must be increased, but the increase of the convolutional layer causes the problems of gradient disappearance, complex training, poor fitting result and the like. The hole convolution can effectively solve the problems, and the receptive field can be enlarged in the convolution process. The calculation formula of the hole convolution is as follows:
where d is the void factor, which varies by an index of 2.
The addition of the residual structure can realize deepening of the network while reducing complexity and ensuring accuracy, and the time sequence convolution layer has more generalization capability. Fig. 8 is a schematic diagram illustrating an implementation process of a residual block according to an embodiment of the present disclosure, where as shown in fig. 8, the residual block mainly includes two layers of hole causal convolution, tanh activation, weight normalization, dropout regularization, and nonlinear mapping. The residual block is mainly used to compute features within a larger receptive field and to track historical information over a longer period.
In a preferred embodiment, the method for processing the normalized undirected graph adjacency matrix a and the normalized directed graph adjacency matrix TI to obtain the topological feature and the spatial correlation of the road network in the embodiment includes:
wherein D ∈ R N×N Is a matrix of the degrees, and the degree matrix,
after the states of the neighbor nodes are considered, the adjacency matrix of the state of the self node is introduced G Representing a graph convolution operation, g θ For the convolution kernel, θ ∈ R K A vector representing the chebyshev coefficients, x represents the input data,
a matrix of laplacian values is represented,
λ max is the maximum eigenvalue, I, of the Laplace matrix n Is a unit matrix, T k (x)＝2xT k-1 (x)- T k-2 (x)，T 0 (x)＝1，T 1 (x)＝x。
In the present embodiment, a layer of GCN for graph structure data is used to extract a highly valuable spatial correlation between regions according to spatial position information of the regions.
The core idea of GCN is to aggregate nodes v in a graph i Characteristic x of i With its neighbour characteristics x j Generating a node v i Is shown. I.e. the node information is aggregated with the information of the edges to generate a new node representation. The N vertices in the graph will change their state continuously until equilibrium is reached, influenced by points that are closer or farther away.
Since convolution calculations are simple in the fourier domain, the disclosed embodiment uses eigenvectors of the laplace matrix for fourier transformation on the graph. Laplacian is the basic tool of spectral analysis, and Laplacian matrix can define the derivative, and the smoothness of the signal on the drawing. In order to introduce an own degree matrix and solve the self-propagation problem, a modified Laplace matrix is used, and the formula is as follows:
wherein D ∈ R N×N Is a matrix of degrees, and is,
the adjacency matrix of the self node state is introduced on the basis of considering the neighbor node state.
However, eigendecomposition of the laplacian matrix is less efficient in the decomposition process in large graph structures.
Therefore, the computational complexity of the laplacian matrix can be reduced by using the chebyshev polynomial approximation. The chebyshev graph convolution based on a chebyshev polynomial of order K is defined as follows:
wherein G Representing a graph convolution operation, g θ For the convolution kernel, θ ∈ R K Vector representing the chebyshev coefficient, x representing the input data,
a matrix of laplacian values is represented,
λ max is the maximum eigenvalue of the Laplace matrix, I n Is a unit matrix, T k (x)＝2xT k-1 (x)-T k-2 (x)， T 0 (x)＝1，T 1 (x)＝x。
Example 2
Referring to fig. 22, the present embodiment is described, and a travel demand prediction system for interaction between functional urban areas of the present embodiment includes:
the reading module 10 is used for pre-dividing grids based on the urban functional areas distributed irregularly to obtain the longitude and latitude of the grid edges;
the analysis module 20 is configured to determine a distance between each pair of regions according to the longitude and latitude of the grid edge, and calculate an undirected graph adjacency matrix a based on the distance between each pair of regions; manufacturing a directed graph adjacency matrix TI based on the interaction type; manufacturing a matrix V based on the interactive demand among the functional areas;
and the processing module 30 is configured to train a travel demand prediction model according to the matrix V, the undirected graph adjacency matrix a and the directed graph adjacency matrix TI, and predict, according to the travel demand prediction model, demand information of travel demands interacted among the functional areas to be predicted within a preset time interval.
In a preferred embodiment, the reading module 10 of this embodiment is configured to divide a grid in advance based on an irregularly distributed urban functional area to obtain a longitude and a latitude of a grid edge;
selecting an area with the most dense trip demand data according to the distribution of the trip demand data in the city, and screening all trip demand data belonging to the area;
dividing four types of functional areas including a living area, a business circle, a learning area and a leisure area according to the screened areas;
dividing the city into grids with different sizes according to the attributes of the four types of functional areas;
wherein, the average value of the distance between the getting-on position and the getting-off position above the grid size is used as an approximate standard;
recording the longitude and latitude of the grid edge after the division, the functional attribute of the grid and the serial number;
and matching grids, functional area attributes and serial numbers of the getting-on position and the getting-off position according to the longitude and latitude of the grid edge of the original data and the longitude and latitude of the grid edge after the division is finished.
In a preferred embodiment, the longitude and latitude of the grid center points are determined according to the longitude and latitude of the grid edges, the distance between every two grid center points represents the distance between each area pair, and the undirected graph adjacency matrix a is calculated based on the distance between each area pair;
determining an interaction type according to the functional attributes of two regions forming the interaction, and manufacturing a directed graph adjacency matrix TI based on the interaction type;
the requirements for interaction between every two functional zones within a time step form a matrix V.
Although the invention has been described in detail with respect to the general description and the specific embodiments, it will be apparent to those skilled in the art that modifications and improvements may be made based on the invention. Accordingly, such modifications and improvements are intended to be within the scope of the invention as claimed.
Claims (10)
1. A travel demand forecasting method for interaction between city functional areas is characterized by comprising the following steps:
pre-dividing grids based on the irregularly distributed urban functional areas to obtain the longitude and latitude of grid edges;
determining the distance between each area pair according to the longitude and latitude of the grid edge, and calculating an undirected graph adjacency matrix A based on the distance between each area pair; manufacturing a directed graph adjacency matrix TI based on the interaction type; manufacturing a matrix V based on the interactive demand among the functional areas;
and training a travel demand prediction model according to the matrix V, the undirected graph adjacent matrix A and the directed graph adjacent matrix TI, and predicting demand information of travel demands interacted among the functional areas to be predicted in a preset time interval according to the travel demand prediction model.
2. The method for predicting travel demand interacting between functional urban areas according to claim 1, wherein grids are pre-divided based on the urban functional areas distributed irregularly, and the method for obtaining the longitude and latitude of the grid edge comprises:
selecting an area with the most dense trip demand data according to the distribution of the trip demand data in the city, and screening out all trip demand data belonging to the area;
dividing four types of functional areas including a living area, a business circle, a learning area and a leisure area according to the screened areas;
dividing the city into grids with different sizes according to the attributes of the four types of functional areas;
wherein, the average value of the distance between the getting-on position and the getting-off position above the grid size is used as an approximate standard;
recording the longitude and latitude of the grid edge after the division, the functional attribute of the grid and the serial number;
and matching grids, functional area attributes and serial numbers of the getting-on position and the getting-off position according to the longitude and latitude of the grid edge of the original data and the longitude and latitude of the grid edge after the division is finished.
3. The method of predicting travel demand for interaction between urban functional areas according to claim 1, wherein the longitude and latitude of the grid center points are determined according to the longitude and latitude of the grid edges, the distance between every two grid center points represents the distance between each pair of regions, and the undirected graph adjacency matrix a is calculated based on the distance between each pair of regions;
determining an interaction type by the functional attributes of two regions forming the interaction, and making a directed graph adjacency matrix TI based on the interaction type;
the requirements for interaction between each two functional zones within a time step form a matrix V.
4. The method for predicting travel demand interacted among urban functional areas according to claim 1, wherein the method for training the travel demand prediction model according to the matrix V, the undirected graph adjacency matrix A and the directed graph adjacency matrix TI comprises the following steps:
extracting bottom layer characteristics of the trip demand data input according to the regional time sequence by a first layer time sequence convolution network of the trip demand prediction model, and inputting the bottom layer characteristics of the trip demand data, an undirected graph adjacent matrix A and a directed graph adjacent matrix TI to a second layer space convolution network;
the second-layer spatial convolution network receives the bottom-layer characteristics of the first layer, processes the normalized undirected graph adjacency matrix A and the normalized directed graph adjacency matrix TI to obtain the topological characteristics and the spatial correlation of the road network, and inputs the time sequence bottom-layer characteristics blended with the topological characteristics and the spatial correlation into the time sequence convolution network of the third layer;
and the third layer of time sequence convolution network calculates the time dependency of the fusion data.
5. The method for predicting travel demands interacted among urban functional areas according to claim 1, wherein predicting demand information of travel demands interacted among functional areas to be predicted in a preset time interval according to a travel demand prediction model comprises:
because the change of the travel demand is influenced by the traffic state of the adjacent time period, a sequence prediction method in the travel demand prediction model is used;
since the travel demand change is affected by the periodic traffic state, a periodic prediction method in the travel demand prediction model is used.
6. The method for predicting travel demand interacted among urban functional areas according to claim 4, wherein the method for extracting the bottom-layer characteristics of the travel demand data input according to the regional time sequence by the time sequence convolutional network comprises the following steps of:
wherein f is i Represents a filter, K represents the size of the convolution kernel, X = (X) 1 ，x 2 ，......，x T ) Represents the input sequence, T represents the T-th implicit characteristic, d is a hole factor, and d is changed with an exponent of 2.
7. The method for predicting travel demand interacted among urban functional areas according to claim 4, wherein the method for processing the normalized undirected graph adjacency matrix A and the normalized directed graph adjacency matrix TI to obtain the topological characteristics and the spatial correlation of the road network comprises the following steps of:
wherein D ∈ R N×N Is a matrix of the degrees, and the degree matrix,
after the state of the neighbor node is considered, the adjacency matrix of the state of the self node is introduced G Representing a graph convolution operation, g θ For the convolution kernel, θ ∈ R K A vector representing the chebyshev coefficients, x represents the input data,
a matrix of laplacian values is represented,
λ max is the maximum eigenvalue, I, of the Laplace matrix n Is a unit matrix, T k (x)＝2xT k-1 (x)-T k-2 (x)，T 0 (x)＝1，T 1 (x)＝x。
8. A travel demand prediction system for interaction between urban functional areas, the system comprising:
the reading module is used for pre-dividing grids based on the urban functional areas distributed irregularly to obtain the longitude and latitude of the grid edges;
the analysis module is used for determining the distance between each area pair according to the longitude and latitude of the grid edge and calculating an undirected graph adjacency matrix A based on the distance between each area pair; manufacturing a directed graph adjacency matrix TI based on the interaction type; manufacturing a matrix V based on the interactive demand among the functional areas;
and the processing module is used for training a travel demand prediction model according to the matrix V, the undirected graph adjacent matrix A and the directed graph adjacent matrix TI and predicting demand information of travel demands interacted among the functional areas to be predicted in a preset time interval according to the travel demand prediction model.
9. The system of claim 8, wherein the reading module is configured to pre-partition grids based on the urban functional areas distributed irregularly, and obtain a longitude and a latitude of a grid edge;
selecting an area with the most dense trip demand data according to the distribution of the trip demand data in the city, and screening out all trip demand data belonging to the area;
dividing four types of functional areas including a living area, a business circle, a learning area and a leisure area according to the screened areas;
dividing the city into grids with different sizes according to the attributes of the four types of functional areas;
wherein, the average value of the distance between the getting-on position and the getting-off position above the grid size is used as an approximate standard;
recording the longitude and latitude of the grid edge after the division, the functional attribute of the grid and the serial number;
and matching grids, functional area attributes and serial numbers of the getting-on position and the getting-off position according to the longitude and latitude of the grid edge of the original data and the longitude and latitude of the grid edge after the division is finished.
10. The system of claim 8, wherein the longitude and latitude of the grid center points are determined according to the longitude and latitude of the grid edges, the distance between each two grid center points represents the distance between each pair of regions, and the undirected graph adjacency matrix a is calculated based on the distance between each pair of regions;
determining an interaction type by the functional attributes of two regions forming the interaction, and making a directed graph adjacency matrix TI based on the interaction type;
the requirements for interaction between every two functional zones within a time step form a matrix V.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210541542.4A CN115204445A (en) | 2022-05-17 | 2022-05-17 | Travel demand prediction method and system for interaction between urban functional areas |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210541542.4A CN115204445A (en) | 2022-05-17 | 2022-05-17 | Travel demand prediction method and system for interaction between urban functional areas |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115204445A true CN115204445A (en) | 2022-10-18 |
Family
ID=83574923
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210541542.4A Pending CN115204445A (en) | 2022-05-17 | 2022-05-17 | Travel demand prediction method and system for interaction between urban functional areas |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115204445A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117271959A (en) * | 2023-11-21 | 2023-12-22 | 中南大学 | Uncertainty evaluation method and equipment for PM2.5 concentration prediction result |
-
2022
- 2022-05-17 CN CN202210541542.4A patent/CN115204445A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117271959A (en) * | 2023-11-21 | 2023-12-22 | 中南大学 | Uncertainty evaluation method and equipment for PM2.5 concentration prediction result |
| CN117271959B (en) * | 2023-11-21 | 2024-02-20 | 中南大学 | Uncertainty evaluation method and equipment for PM2.5 concentration prediction result |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Tekouabou et al. | Improving parking availability prediction in smart cities with IoT and ensemble-based model | |
| CN110633871B (en) | Regional traffic demand prediction method based on convolution long-term and short-term memory network | |
| CN115240425B (en) | Traffic prediction method based on multi-scale space-time fusion graph network | |
| CN113792929B (en) | Traffic flow prediction method, electronic device and storage medium | |
| CN113159364A (en) | Passenger flow prediction method and system for large-scale traffic station | |
| CN113570859B (en) | Traffic flow prediction method based on asynchronous space-time expansion graph convolution network | |
| Lu et al. | Lane-level traffic speed forecasting: A novel mixed deep learning model | |
| CN112270355A (en) | Active safety prediction method based on big data technology and SAE-GRU | |
| CN115578852A (en) | Traffic prediction method based on DSTGCN | |
| CN112862177B (en) | Urban area aggregation degree prediction method, device and medium based on deep neural network | |
| Zhao et al. | Prediction in traffic accident duration based on heterogeneous ensemble learning | |
| Chen et al. | Discrimination and prediction of traffic congestion states of urban road network based on spatio-temporal correlation | |
| CN114973678B (en) | Traffic prediction method based on graph attention neural network and space-time big data | |
| CN108320027B (en) | Big data processing method based on quantum computation | |
| CN111047078A (en) | Traffic characteristic prediction method, system and storage medium | |
| James | Citywide estimation of travel time distributions with Bayesian deep graph learning | |
| CN115204445A (en) | Travel demand prediction method and system for interaction between urban functional areas | |
| Dai et al. | Attention based simplified deep residual network for citywide crowd flows prediction | |
| Bao et al. | PKET-GCN: prior knowledge enhanced time-varying graph convolution network for traffic flow prediction | |
| Rico et al. | Graph neural networks for traffic forecasting | |
| Feng et al. | Multi-step ahead traffic speed prediction based on gated temporal graph convolution network | |
| CN117456730A (en) | Traffic flow prediction method based on self-adaptive partial attention spread graph convolutional neural network | |
| CN117131991A (en) | Urban rainfall prediction method and platform based on hybrid neural network | |
| CN116665448A (en) | Traffic speed real-time prediction method and system based on graph convolution network | |
| Li et al. | Multigraph aggregation spatiotemporal graph convolution network for ride-hailing pick-up region prediction |
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 |