CN112862177A - Urban area concentration degree prediction method, equipment and medium based on deep neural network - Google Patents

Urban area concentration degree prediction method, equipment and medium based on deep neural network Download PDF

Info

Publication number
CN112862177A
CN112862177A CN202110142862.8A CN202110142862A CN112862177A CN 112862177 A CN112862177 A CN 112862177A CN 202110142862 A CN202110142862 A CN 202110142862A CN 112862177 A CN112862177 A CN 112862177A
Authority
CN
China
Prior art keywords
time
probability density
density distribution
stay
prediction
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
Application number
CN202110142862.8A
Other languages
Chinese (zh)
Other versions
CN112862177B (en
Inventor
肖竹
方辉
蔡成林
蒋洪波
陈红阳
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Hunan University
Original Assignee
Hunan University
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Hunan University filed Critical Hunan University
Priority to CN202110142862.8A priority Critical patent/CN112862177B/en
Publication of CN112862177A publication Critical patent/CN112862177A/en
Application granted granted Critical
Publication of CN112862177B publication Critical patent/CN112862177B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

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
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/044—Recurrent networks, e.g. Hopfield 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
    • G06N3/045—Combinations of 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
    • G06N3/08—Learning methods

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • General Health & Medical Sciences (AREA)
  • Evolutionary Computation (AREA)
  • Software Systems (AREA)
  • Mathematical Physics (AREA)
  • General Engineering & Computer Science (AREA)
  • Computing Systems (AREA)
  • Molecular Biology (AREA)
  • Business, Economics & Management (AREA)
  • Data Mining & Analysis (AREA)
  • Computational Linguistics (AREA)
  • Biophysics (AREA)
  • Biomedical Technology (AREA)
  • Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Human Resources & Organizations (AREA)
  • Economics (AREA)
  • Strategic Management (AREA)
  • Quality & Reliability (AREA)
  • Game Theory and Decision Science (AREA)
  • Tourism & Hospitality (AREA)
  • Development Economics (AREA)
  • Operations Research (AREA)
  • Marketing (AREA)
  • Entrepreneurship & Innovation (AREA)
  • General Business, Economics & Management (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)
  • Traffic Control Systems (AREA)

Abstract

The invention discloses a method, equipment and a medium for predicting urban area concentration degree based on a deep neural network, wherein the method comprises the following steps: extracting stay data of each time interval of holidays from private car tracks of cities, wherein the stay data comprise positions and stay durations of all stay points in the time interval; respectively calculating the spatial probability density distribution and the time probability density distribution of the relevant stay points in each time period according to the positions and stay durations of all the stay points in each time period; respectively constructing a training set and a prediction label set by using a spatial probability density distribution sequence and a temporal probability density distribution sequence which are correspondingly formed in the continuous time interval of the holiday, and training a pre-built neural network model to obtain an aggregation prediction model; and performing rolling prediction on the aggregation of the target prediction period by using an aggregation prediction model. The method is suitable for predicting the urban aggregation degree during holidays.

Description

Urban area concentration degree prediction method, equipment and medium based on deep neural network
Technical Field
The invention belongs to the field of intelligent traffic, and particularly relates to a method, equipment and medium for predicting urban regional concentration based on a deep neural network.
Background
With the development of urbanization in recent years, the number of road automobiles, especially private cars, is increasing, and accounts for a large proportion of urban vehicles. According to a report of the Chinese statistical office, the number of private cars exceeds 2.2 hundred million and accounts for 88.68% of all vehicles by 2019, more and more people choose to drive the private cars to meet daily needs of the people, and particularly, the driving behaviors of the private cars have clear driving routes and destinations and often stay for a period of time to complete required activities after the private cars reach the destinations. Each area in the city is provided with private cars coming from different places at any time, the staying action is a key characteristic for generating the urban area gathering effect, and the length of the staying time is an important time characteristic for the gathering effect.
Understanding the effects of private car aggregation may provide a wide range of applications from location-based services to intelligent traffic management, including traffic condition monitoring, advertising and destination recommendation, and city planning, among others, for researchers and decision makers.
Existing studies focus primarily on urban mobility and its relationship to aggregation effects. It is generally accepted that private cars are more predictable on weekdays than on weekends because people tend to exhibit a regular spatio-temporal flow pattern when traveling on weekdays, meaning that the aggregations of regions and associated features do not vary much, and thus the weekdays have a strong correlation and predictability. However, on weekends, predicting the effects of private car aggregation on weekends is more challenging due to the diversity and randomness of people's travel.
In the prior art, a proper space-time modeling method is lacked, and the aggregation effect distribution change caused by human random movement during holidays cannot be predicted, so that the invention designs a method for predicting the aggregation effect of private cars in urban areas around the weekends based on a neural network framework.
Disclosure of Invention
The invention provides a method, equipment and a medium for predicting urban regional concentration based on a deep neural network, which are suitable for predicting urban concentration during holidays.
In order to achieve the technical purpose, the invention adopts the following technical scheme:
a city region aggregation degree prediction method based on a deep neural network comprises the following steps:
step 1, extracting stay data of each time interval of holidays from private car tracks of cities, wherein the stay data comprise positions and stay durations of all stay points in the time interval;
step 2, respectively calculating the spatial probability density distribution and the time probability density distribution of the relevant stay points in each time period according to the positions and stay durations of all the stay points in each time period;
step 3, taking T time intervals as prediction step lengths, taking a space probability density distribution sequence and a time probability density distribution sequence which are correspondingly formed by N + T-1 continuous time intervals on a holiday as a training set, taking a space probability density distribution sequence and a time probability density distribution sequence which are correspondingly formed by N continuous time intervals which are delayed by the training set as prediction label sets, and training a pre-built neural network model to obtain an aggregation prediction model;
and 4, acquiring the spatial probability density distribution and the temporal probability density distribution of the relevant stay points of the continuous T periods before the target prediction period, performing rolling prediction by using an aggregation degree prediction model, and acquiring the spatial probability density distribution and the temporal probability density distribution of the relevant stay points of the target prediction period, namely the aggregation degree of the target prediction period.
In a more preferred embodiment, the training set further includes weather data formed by a prediction period.
In a more preferred embodiment, the formula for calculating the spatial probability density distribution of the stop point is:
Figure BDA0002929921570000021
in the formula, piThe cluster center of the ith sub-region obtained by dividing the city is 1,2, …, n, n represents that the city is divided into n sub-regions; kσ(. cndot.) is a Gaussian kernel with a bandwidth of σ,
Figure BDA0002929921570000026
the stopping point p obtained by the k iteration calculation belongs to the clustering center piAnd is weighted, and
Figure BDA0002929921570000022
in a more optimal technical scheme, the maximum likelihood is adopted to estimate the clustering center p of the cityiAt this time, the calculation formula of the spatial probability density distribution can be expressed as:
Figure BDA0002929921570000023
in the formula, theta and psi (p)i) Respectively representing the cluster centers piAnd the dwell point p is mapped to the regenerated kernel Hilbert space, i.e.
Figure BDA0002929921570000024
The method is a robust loss function, and particularly adopts a Log-Cosh loss function and a kernel iterative reweighted least square method to analyze and solve the time-space probability density distribution.
In a more preferred embodiment, the formula for calculating the time probability density distribution of the dwell point is:
Figure BDA0002929921570000025
in the formula, TiAnd lambda is a hyperparameter of inverse Gaussian distribution, and the empirical value lambda is 1.
In a more preferred technical scheme, the neural network model comprises a space-time attention mechanism, an LSTM, an MLP, a cascade layer, a neural algorithm logic unit and a full connection layer;
the space-time attention mechanism comprises the same number of time attention mechanisms, space attention mechanisms and fusion gates, the input of the time attention mechanisms is time probability density distribution, the input of the space attention mechanisms is space probability density distribution, and the fusion gates perform weighted fusion on the time representation output by the corresponding time attention mechanisms and the space representation output by the space attention mechanisms and output the time representation and the space representation to the LSTM;
the input of the MLP is weather data, the cascade layer cascades the output of the LSTM and the MLP and then outputs the output to the neural algorithm logic unit, the full-connection layer is used for full-connecting the output data of the neural algorithm logic unit, and finally the full-connection layer outputs prediction data.
In a more preferable technical scheme, the fully-connected layer comprises two dense layers and a dropout module, and the ratio of the dropout module is 0.5.
An apparatus comprising a processor and a memory; wherein: the memory is to store computer instructions; the processor is configured to execute the computer instructions stored in the memory, and specifically, to perform the method according to any of the above technical solutions.
A computer storage medium storing a program for implementing the method of any one of the preceding claims when executed.
Advantageous effects
1. The invention provides a feature extraction method for effectively capturing space-time features of an aggregation effect, which comprises the steps of improving a nuclear density estimator for modeling space features by using a Log-Cosh loss function, and representing nonlinear time correlation by using the residence time of a vehicle.
2. The invention designs a space-time attention mechanism to give different weights to space-time characteristics so as to better capture the space-time characteristics, and adds holiday weather data as external characteristics.
3. According to the invention, a STATET-NALU framework is designed to generalize a model, and the generalization capability of the model is enhanced by adding a NALU network supporting data extrapolation, so that more accurate urban area aggregation effect prediction is realized, and the prediction has numerical extrapolation capability to adapt to the nonlinear change of urban area aggregation degree during holidays.
4. The invention is applied to a real private car data experiment, and the result verifies the effectiveness and superiority of the invention, thereby providing a new idea for further exploring human mobility.
Drawings
Fig. 1 is an overview of the prediction method for holiday aggregation in urban areas proposed by the present invention.
FIG. 2 is a diagram of a spatiotemporal attention mechanism.
Fig. 3 is a visualization result of the region aggregation effect modeling.
FIG. 4 is a neural network learning rate curve for different hyper-parameters.
Fig. 5 shows the results of prediction of the degree of aggregation of three-dimensional regions, and (a), (b), (c), (d), (e), and (f) respectively show the results of prediction by the method proposed in the present invention, the true value, the MLR predicted value, the STARIMA predicted value, the DCNN predicted value, and the Multi-LSTM predicted value.
Fig. 6 shows the results of prediction of the regional concentrations of the planes, (a), (b), (c), (d), (e), (f) respectively show the results of true values, MLR predicted values, STARIMA predicted values, DCNN predicted values, Multi-LSTM predicted values and prediction according to the method proposed by the present invention.
Detailed Description
The following describes embodiments of the present invention in detail, which are developed based on the technical solutions of the present invention, and give detailed implementation manners and specific operation procedures to further explain the technical solutions of the present invention.
The embodiment provides a city region concentration degree prediction method based on a deep neural network, which is shown in fig. 1 and includes the following steps:
step 1, extracting stay data of each time interval of holidays from private car tracks of cities, wherein the stay data comprise positions and stay durations of all stay points in the time interval.
In the embodiment, track data of holidays of private cars in cities are collected by using a GPS and an Internet of things communication technology, and the track data are stored in a data set according to a trip track (trip) of a vehicle ID, wherein the track data comprise the parking time, the driving-away time and the parking position of each private car. The step 1 mainly finishes the extraction of the stay data from the track data of the holidays of private car festivals in cities: firstly, cleaning collected travel track data, wherein the collected travel track data comprises abnormal track data such as ID information loss, abnormal movement within a short time and parking less than three minutes; then, dividing the holidays into 24 time periods on average, wherein each time period is 1 hour; and acquiring the positions and the stay time of all stay points in each time interval in the city by taking the parking time as a standard.
And 2, respectively calculating the spatial probability density distribution and the temporal probability density distribution of the relevant stay points in each time interval according to the positions and stay durations of all the stay points in each time interval.
In the invention, the distribution of the staying points is difficult to calculate directly, and a Kernel Density Estimation (KDE) method is adopted, because the data distribution of prior knowledge is not needed, the automatic clustering can be carried out according to the relative positions of the staying points.
Specifically, a city is divided into n small regions, the staying points in each small region have the characteristic of independent and same distribution, and the clustering centers of all the small regions are sequentially set as p1,p2,...,pnThen, for the dwell point in each cell, the spatial kernel density value can be expressed as:
Figure BDA0002929921570000041
wherein p represents the position of the stop point, and specifically comprises longitude and latitude information of the stop point, KσIs a kernel function with bandwidth of sigma and is defined as a sample point p to an aggregation center p in spaceiA monotonic function of the euclidean distance between.
Since there are some outliers in the data set that are far from the cluster center, this may be due to errors in the vehicle positioning system or information upload process. The classical KDE method is particularly sensitive to outliers, which are treated as isolated clusters. Thus, even a normal dwell point, the classic KDE treats it as an independent cluster center, greatly impacting the aggregate prediction results. Therefore, in a preferred embodiment, the adopted KDE method is expanded:
first, the objective function of the kernel density estimate is defined in the form of a weighted sum of:
Figure BDA0002929921570000051
here, the number of the first and second electrodes,
Figure BDA00029299215700000513
belonging to a cluster center p for a dwell point piAnd is weighted, and
Figure BDA0002929921570000052
in this embodiment, a gaussian kernel function is selected, and the general form is:
Figure BDA0002929921570000053
the key to solve the problem is to find the aggregation center, and the existing method is used for maximum likelihood estimation in which the kernel density estimation function
Figure BDA0002929921570000054
Can be expressed as:
Figure BDA0002929921570000055
for computational convenience, we will refer to p and piMapping to a regenerative nuclear hilbert space
Figure BDA0002929921570000056
In, i.e.
Figure BDA0002929921570000057
And
Figure BDA0002929921570000058
respectively represents piAnd a mapping function for p, and p,
Figure BDA0002929921570000059
the invention selects Log-Cosh loss, adopts hyperbolic cosine to calculate, has all the advantages of Huber loss, is twice-derivable at each point and is convenient to calculate:
Figure BDA00029299215700000510
since the above formula does not have a closed solution, the present invention adopts a kernel iterative reweighted least square method which must initialize the weight matrix first
Figure BDA00029299215700000514
Then a sequence representation is generated as
Figure BDA00029299215700000511
Figure BDA00029299215700000512
Figure BDA0002929921570000061
Then, after k iterations, the function is judged
Figure BDA0002929921570000062
Whether or not to converge, the final expression being
Figure BDA0002929921570000063
In short, in this embodiment, the distribution of the stopping points is characterized by the expanded kernel density estimation, and the spatial probability density distribution of the stopping points of the private car is obtained
Figure BDA0002929921570000064
The urban aggregation can be quantified in spatial dimension, and the method has adaptive robustness. The larger the spatial probability density distribution value, the stronger the aggregation.
The stay time of the stay point in the invention is an important index reflecting the urban aggregation effect on the time dimension. For example, people may take their children to an amusement park on weekends, often taking two or more hours, or they may take an hour to buy commodities in supermarkets. Clearly, the aggregate effect of amusement parks differs from that of supermarkets. The invention uses the probability density distribution of the residence time to characterize the time characteristic of urban aggregativity, and the probability density distribution is approximately inverse Gaussian distribution:
Figure BDA0002929921570000065
Tilambda is a super parameter of inverse Gaussian distribution, and is the time difference between the driving-off time and the parking time. According to experience, the residence time distribution can be calculated by substituting 1 for λ.
And 3, training a pre-built neural network model by taking the T time intervals as prediction step lengths, taking a space probability density distribution sequence and a time probability density distribution sequence which are correspondingly formed by N + T-1 continuous time intervals of the holidays as training sets, and taking a space probability density distribution sequence and a time probability density distribution sequence which are correspondingly formed by N continuous time intervals of the T time intervals delayed by the training sets as prediction label sets, so as to obtain an aggregation prediction model.
The neural network model in the present embodiment includes a spatiotemporal attention mechanism, LSTM, MLP, cascade layer, neural algorithm logic unit, and full-link layer.
After the spatial probability density distribution and the temporal probability density distribution of the stopping points are extracted as the spatial feature and the temporal feature of the urban aggregation in step 2, the ST-attention block, namely the empty attention mechanism, is designed in the neural network model to model the dynamic spatial correlation and the nonlinear temporal correlation.
The spatiotemporal attention mechanisms include the same number of temporal and spatial attention mechanisms and fusion gates. As shown in fig. 2, the spatio-temporal attention mechanism, on the one hand, the spatial embedding layer provides a geographical representation of the aggregation effect, in particular a probability density estimation based on historical ASL behavior; on the other hand, the distribution of the residence time is taken as a time characteristic of the embedded network. The input of the time attention mechanism is time probability density distribution, the input of the space attention mechanism is space probability density distribution, and the fusion gate performs weighted fusion on the time representation output by the corresponding time attention mechanism and the space representation output by the space attention mechanism and outputs the time representation and the space representation to the LSTM.
The spatial attention mechanism is as follows: in the spatial dimension, the probability density distributions of stop-and-wait points at different positions influence each other, and for this purpose, a spatial attention mechanism is deduced to capture the relevance between the spaces of the aggregation effect, according to the division of the area into a plurality of small areas and according to the spatial distribution f of the stop-and-wait points of the ith small areaiCorrelation score with true distribution f
Figure BDA0002929921570000071
Different weights are dynamically assigned:
Figure BDA0002929921570000072
Figure BDA0002929921570000073
wherein | | is a join operation,
Figure BDA0002929921570000074
is a probability density distribution fiAt tjHidden states in block 1 at time l,
Figure BDA0002929921570000075
two different non-linear projections in k parallel attention mechanisms are shown.
Figure BDA0002929921570000076
As output of the spatial attention layer.
The time attention mechanism is as follows: in the time dimension, the regional concentration degree distribution of different time periods also has correlation, the invention uses a time attention mechanism to capture the nonlinear correlation of the residence time in observation, and defines
Figure BDA0002929921570000077
To measure the average stay prediction time t of the jth time periodjThe correlation with the true mean residence time t can be expressed as
Figure BDA0002929921570000078
Figure BDA0002929921570000079
Figure BDA00029299215700000710
As output of the temporal layer.
ST feature fusion: namely the fusion gate designed by the invention, the spatial representation generated by the k parallel attention mechanisms of the adaptive fusion attention layer
Figure BDA00029299215700000711
And time characterization
Figure BDA00029299215700000712
Figure BDA00029299215700000713
And q is a gate activated by a sigmoid function, and the fusion mechanism adaptively controls the output of each spatial feature and each temporal feature.
While classical LSTM has the ability to infer a function, it often fails in the learning process, especially when the data is out of range during the training process, the results fail to reach the expected level. The invention adds a digital extrapolation framework of the drum excitation system, namely a neural algorithm logic unit (namely a NALU layer) behind the LSTM, and enhances the numerical extrapolation capability, namely generalization capability, of the model aiming at the condition that people have high randomness during holidays. The neural algorithm logic unit in this embodiment is operated by a combination of two weighted Neural Accumulators (NACs). One standard NAC is used to implement addition and subtraction operations, and the other is used to learn more complex arithmetic operations in logarithmic space, such as multiplication, division, power functions, etc., which can be expressed as:
Figure BDA0002929921570000081
since NAC does not contain bias parameters, and non-linear compression is not performed at the hidden layer, NAC is difficult to learn by standard neural networks. Thus two parameters that can be randomly initialized are used to obtain the weight values, i.e.
Figure BDA0002929921570000082
Let
Figure BDA0002929921570000083
Representing the hidden layer state of the previous LSTM structure, can be obtained
Figure BDA0002929921570000084
Figure BDA0002929921570000085
Wherein G ═ σ (G)s)，GsFor the gain function, o (y) represents an infinitesimal value.
The invention names the overall neural network architecture as STANet-NALU, which is able to learn various types of functions, rather than simple numerical characteristics, which allows better generalization within and outside the numerical range, and therefore has a good expected effect on the prediction of holiday city aggregations.
In a more preferred embodiment, weather also affects the urban aggregation effect to a certain extent, so that when the urban aggregation effect is predicted, the weather condition is taken as an external feature to improve the prediction accuracy, and therefore the multi-layer perceptron MLP and the cascade layer are arranged. The MLP is used for inputting weather data subjected to one-hot encoding and extracting characteristic data from the weather data; and then the output of the LSTM and the output of the MLP are cascaded by the cascade layer and then output to the neural algorithm logic unit, the output data of the neural algorithm logic unit are fully connected by the full-connection layer, and finally the prediction data are output by the full-connection layer.
The fully-connected layer comprises two dense layers and a dropout module, the activation function of the dense layers is 'linear', the optimizer adopts 'Adam', and the ratio of the dropout module is 0.5. In addition, the activation functions of the LSTM layer are all "Sigmoid".
In the training set of this embodiment, for example, any holiday in the data set is 24 hours, the spatial probability density distribution sequence and the temporal probability density distribution sequence of 23 consecutive periods in the process from 0 point to 23 point of a holiday are taken as the training set, the spatial probability density distribution sequence and the temporal probability density distribution sequence of 21 consecutive periods in the process from 4 point to 24 point of a holiday are taken as the prediction label set, and the prediction step is taken as 3 hours, then the spatial probability density distribution sequence and the temporal probability density distribution sequence of 3 consecutive periods in the process from 0 point to 3 point are taken as 1 training sample, the spatial probability density distribution and the temporal probability density distribution in the period from 3 point to 4 point are taken as the corresponding prediction labels, … …, and finally 21 training samples and corresponding prediction labels can be obtained in 1 day holiday for training the pre-built neural network model, and obtaining the urban festival and holiday concentration degree prediction model.
And 4, acquiring the spatial probability density distribution and the temporal probability density distribution of the relevant stay points of the continuous T periods before the target prediction period, performing rolling prediction by using an aggregation degree prediction model, and acquiring the spatial probability density distribution and the temporal probability density distribution of the relevant stay points of the target prediction period, namely the aggregation degree of the target prediction period.
The method selects a private car track data set collected in a Roche region (Dongchong 114: 04-114: 21 and North latitude 22: 50-22: 65) of Shenzhen city in China from 7/2017 to 7/29/2018, and the selected region comprises a plurality of functional regions such as an entertainment region, a residential region and a business region. The data set contained 5000 more private cars, nearly 10 ten thousand travel tracks and 2125 urban roads. The present invention evaluates the change in concentration based on the distribution of the dwell point data. Table 1 shows the cleaned partial parking equal point data, T1 and T2 respectively indicate the arrival time and departure time of the vehicle, and StopLon and StopLat respectively indicate the parking latitude and longitude, that is, the positions of the parking points.
TABLE 1
ObjectID StartTime StopTime StopLon StopLat
118091 2018/5/1 8:51 2018/5/1 12:15 114.094 22.57
118091 2018/5/1 12:24 2018/5/1 15:32 114.078 22.5734
118091 2018/5/1 16:11 2018/5/2 7:59 114.075 22.5736
First, 112 × 24 hours is taken as a training set, all training sets are delayed for 1 hour to be taken as a verification set, and finally 37 × 24 hours is taken as a test set, specifically, 7920 training samples, 2640 verification samples and 2520 test samples are included. As shown in the first drawing, the relevant parameters of the neural network are set to be m-64, n-16, the optimizer is "Adam", the loss function is "MSE", the metric standard is "accuracy", the activation functions of the LSTM layer are all "Sigmoid", and the activation function of the dense layer is "linear".
To show the superiority of the present invention, the test was conducted by selecting a Time period in which the city aggregation suddenly changed, and taking the aggregation effect of 13:00-14:00 in 29-week days in 2018 as an example, the test was compared with MLR (Multiple Linear Regression), STARIMA (Space-Time Autoregressive and Moving Average mode), DCNN (Deep Convolutional Neural Network), Multi-LSTM (Multiple Long Short Term) using five evaluation criteria of MSE (mean Square error), RMSE (mean Square error), MAE (mean Square error), KL (absolute Square error), KL-Square error, and KL-Square error (KL-2).
Table 2 below is the Root Mean Square Error (RMSE) between predicted and actual values for predictions taken for different numbers of NALU layers. Table 3 shows that the measures of the market holiday aggregation prediction model of the embodiment of the invention on Shenzhen dataset include MSE, RMSE, MAE, KL and R2 compared with the errors of MLR, STARIMA, DCNN and LSTM.
TABLE 2
Method Row Column Matrix
NALU(3)-LSTM 0.10258 0.1002 0,1237
NALU(2)-LSTM 0.10351 0.10325 0.12843
NALU(1)-LSTM 0.10264 0.10235 0.11962
Muti-LSTM 0.10186 0.10257 0.10366
TABLE 3
Method MSE RMSE MAE KL R2
MLR 0.0145 0.1205 0.0910 0.0638 0.7639
STARIMA 0.0832 0.2885 0.2092 inf -0.1738
DCNN 0.0172 0.1312 0.1033 0.0615 0.7202
LSTM 0.0377 0.1930 0.1796 0.1834 0.3947
Muti-LSTM 0.0129 0.1138 0.0940 0.0671 0.7897
STANet-NALU(3) 0.0113 0.1062 0.0863 0.0645 0.8168
STANet-NALU(2) 0.0109 0.1041 0.0855 0.0696 0.8238
STANet-NALU(1) 0.0159 0.1259 0.1042 0.0685 0.7423
The invention observes the performance of the urban area stop equal point distribution model from the angle of a plane. FIG. 3(a) shows the distribution of 13:00-14:00 selected regional stops on 7/month 29. As shown in fig. 3(b), the KDE function based on Huber loss is sensitive to outlier anomalies, particularly the distribution of the tie-points in low density regions. In fig. 3(c), the present invention proposes that the KDE model based on Log-cosh loss is hardly affected by outliers, generating a relatively true distribution. In order to effectively learn the spatio-temporal correlation of the training data, find out the optimal batch size, learning rate and NALU number of layers, the present invention has performed a series of experiments. In the experiment, the learning rate is set as invariant, the batch size is set as independent variable, and the Root Mean Square Error (RMSE) is set as dependent variable, and the results are shown in table 2. Studies have found that selecting smaller batches results in less error, but results in greater time consumption at the same training load. Thus, 135 is taken as the optimal batch size parameter. Fig. 4 shows the learning curves for different hyper-parameters. The learning curve shows that they eventually converge regardless of the learning rate, but 0.001 is the best learning rate parameter because it converges faster than the other settings. In addition, experiments also evaluated the impact of different NALU layers on learning. The results show that the number of layers has no substantial effect on the convergence of the algorithm. On the basis, the optimal parameters of the experiment are applied to the scheme provided by the invention.
To intuitively explain city aggregation predictions, the experiment contrasts with other baseline methods, whose hyper-parameter settings and feature selection are as follows. (1) MLR: temporal and spatial features are arguments that build multiple inputs. And optimizing parameter selection by using a Bayesian bridge algorithm. (2) STARIMA: the hysteresis order, difference index, moving average window sizes are set to 2,1, respectively. (3) DCNN. The filter size of the convolutional layer was 32, the kernel size was 2x2, Adam was used as the optimizer and the loss was MSE. (4) LSTM has the same structure and parameter settings as Multi-LSTM. In addition, Multi-LSTM adds weather as an external feature.
Fig. 5 shows the prediction results of the three-dimensional view. The experimental result shows that compared with the ground truth value, the method provided by the invention obtains better performance. Comparing the performance with the results of fig. 5 and 6, a significant difference from the true values is observed in fig. 6(a), especially in the rectangular region, because MLR can only capture linear relations, while the randomness of people vacating rows brings about a non-linear clustering effect of the distribution of adjacent time segments. In the green circle region in fig. 6(c), STARIMA attenuates the results of this region. It can also be seen in fig. 5(c) that it generates an erroneous cluster center. The DCNN method performed better than MLR and ARIMA, but some incorrect focal regions appeared in the red rectangular and green circular regions of fig. 6 (d). The results of FIG. 6(e) show the bias of the rectangular concentration center relative to FIG. 6 (a). That is, the Multi-LSTM method produces a bias in representing the aggregate effect. In conclusion, the method STANet-NALU model provided by the invention well captures the change of the aggregation effect and obtains the optimal visualization performance. The method not only accurately predicts the position of the peak value (namely the aggregation center) of the stop-equal point distribution, but also gives a relatively correct aggregation degree compared with a reference, and compares the traditional LSTM model on a series of indexes of MSE, RMSE and MAE.
The above embodiments are preferred embodiments of the present application, and those skilled in the art can make various changes or modifications without departing from the general concept of the present application, and such changes or modifications should fall within the scope of the protection claimed in the present application.

Claims (9)

1. A city region aggregation degree prediction method based on a deep neural network is characterized by comprising the following steps:
step 1, extracting stay data of each time interval of holidays from private car tracks of cities, wherein the stay data comprise positions and stay durations of all stay points in the time interval;
step 2, respectively calculating the spatial probability density distribution and the time probability density distribution of the relevant stay points in each time period according to the positions and stay durations of all the stay points in each time period;
step 3, taking T time intervals as prediction step lengths, taking a space probability density distribution sequence and a time probability density distribution sequence which are correspondingly formed by N + T-1 continuous time intervals on a holiday as training sets, taking the space probability density distribution sequence and the time probability density distribution sequence which are correspondingly formed by N continuous time intervals delayed by the training sets as prediction label sets, and training a pre-built neural network model to obtain an aggregation prediction model;
and 4, acquiring the spatial probability density distribution and the temporal probability density distribution of the relevant stay points of the continuous T periods before the target prediction period, performing rolling prediction by using an aggregation degree prediction model, and acquiring the spatial probability density distribution and the temporal probability density distribution of the relevant stay points of the target prediction period, namely the aggregation degree of the target prediction period.
2. The method of claim 1, wherein the training set further comprises weather data comprising predicted time periods.
3. The method of claim 1, wherein the spatial probability density distribution for the dwell point is calculated as:
Figure FDA0002929921560000011
in the formula, piThe cluster center of the ith sub-region obtained by dividing the city is 1,2, …, n, n represents that the city is divided into n sub-regions; kσ(. cndot.) is a Gaussian kernel with a bandwidth of σ,
Figure FDA0002929921560000012
the stopping point p obtained by the k iteration calculation belongs to the clustering centerpiAnd is weighted, and
Figure FDA0002929921560000013
4. method according to claim 3, characterized in that the maximum likelihood is used to estimate the cluster center p of a cityiAt this time, the calculation formula of the spatial probability density distribution can be expressed as:
Figure FDA0002929921560000014
in the formula, theta and psi (p)i) Respectively representing the cluster centers piAnd the dwell point p is mapped to the regenerated kernel Hilbert space, i.e.
Figure FDA0002929921560000015
Figure FDA0002929921560000016
The method is a robust loss function, and specifically adopts a Log-Cosh loss function and a nuclear iteration reweighted least square method to analyze and solve space-time probability density distribution.
5. The method of claim 1, wherein the time probability density distribution for the dwell point is calculated by:
Figure FDA0002929921560000021
in the formula, TiAnd lambda is a hyperparameter of inverse Gaussian distribution, and the empirical value lambda is 1.
6. The method of claim 2, wherein the neural network model comprises a spatiotemporal attention mechanism, LSTM, MLP, cascade layer, neural algorithmic logic unit, and full connectivity layer;
the space-time attention mechanism comprises the same number of time attention mechanisms, space attention mechanisms and fusion gates, the input of the time attention mechanisms is time probability density distribution, the input of the space attention mechanisms is space probability density distribution, and the fusion gates perform weighted fusion on the time representation output by the corresponding time attention mechanisms and the space representation output by the space attention mechanisms and output the time representation and the space representation to the LSTM;
the input of the MLP is weather data, the cascade layer cascades the output of the LSTM and the MLP and then outputs the output to the neural algorithm logic unit, the full-connection layer is used for full-connecting the output data of the neural algorithm logic unit, and finally the full-connection layer outputs prediction data.
7. The method of claim 6, wherein the fully connected layer comprises two dense layers and a dropout module, and the ratio of the dropout module is 0.5.
8. An apparatus comprising a processor and a memory; wherein: the memory is to store computer instructions; the processor is configured to execute the computer instructions stored by the memory, in particular to perform the method according to any one of claims 1 to 7.
9. A computer storage medium storing a program which, when executed, performs the method of any one of claims 1 to 7.
CN202110142862.8A 2021-02-02 2021-02-02 Urban area aggregation degree prediction method, device and medium based on deep neural network Active CN112862177B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110142862.8A CN112862177B (en) 2021-02-02 2021-02-02 Urban area aggregation degree prediction method, device and medium based on deep neural network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110142862.8A CN112862177B (en) 2021-02-02 2021-02-02 Urban area aggregation degree prediction method, device and medium based on deep neural network

Publications (2)

Publication Number Publication Date
CN112862177A true CN112862177A (en) 2021-05-28
CN112862177B CN112862177B (en) 2024-01-19

Family

ID=75986190

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110142862.8A Active CN112862177B (en) 2021-02-02 2021-02-02 Urban area aggregation degree prediction method, device and medium based on deep neural network

Country Status (1)

Country Link
CN (1) CN112862177B (en)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113724504A (en) * 2021-08-06 2021-11-30 之江实验室 Urban area traffic prediction system and method oriented to vehicle track big data
CN115933409A (en) * 2023-01-06 2023-04-07 青岛哈尔滨工程大学创新发展中心 Method for adaptively adjusting time step based on multi-physical field weak coupling
CN116151489A (en) * 2023-04-19 2023-05-23 之江实验室 Vehicle stay behavior analysis and prediction method and system based on multitask learning

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20030187720A1 (en) * 2002-03-28 2003-10-02 Fujitsu Limited Vehicle allocating method, system and program
CN110147904A (en) * 2019-04-23 2019-08-20 深圳先进技术研究院 A kind of urban agglomerations event prediction and localization method and device
CN110223510A (en) * 2019-04-24 2019-09-10 长安大学 A kind of multifactor short-term vehicle flowrate prediction technique based on neural network LSTM
CN111242352A (en) * 2020-01-02 2020-06-05 深圳市麦谷科技有限公司 Parking aggregation effect prediction method based on vehicle track
CN111275962A (en) * 2019-12-30 2020-06-12 深圳市麦谷科技有限公司 Vehicle track data aggregation effect prediction method and device

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20030187720A1 (en) * 2002-03-28 2003-10-02 Fujitsu Limited Vehicle allocating method, system and program
CN110147904A (en) * 2019-04-23 2019-08-20 深圳先进技术研究院 A kind of urban agglomerations event prediction and localization method and device
CN110223510A (en) * 2019-04-24 2019-09-10 长安大学 A kind of multifactor short-term vehicle flowrate prediction technique based on neural network LSTM
CN111275962A (en) * 2019-12-30 2020-06-12 深圳市麦谷科技有限公司 Vehicle track data aggregation effect prediction method and device
CN111242352A (en) * 2020-01-02 2020-06-05 深圳市麦谷科技有限公司 Parking aggregation effect prediction method based on vehicle track

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
范娇娇: "基于私家车轨迹的停等聚集效应分析研究", 工程科技Ⅱ辑, no. 07, pages 034 - 418 *

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113724504A (en) * 2021-08-06 2021-11-30 之江实验室 Urban area traffic prediction system and method oriented to vehicle track big data
CN115933409A (en) * 2023-01-06 2023-04-07 青岛哈尔滨工程大学创新发展中心 Method for adaptively adjusting time step based on multi-physical field weak coupling
CN115933409B (en) * 2023-01-06 2023-05-05 青岛哈尔滨工程大学创新发展中心 Self-adaptive time step adjustment method based on multi-physical field weak coupling
CN116151489A (en) * 2023-04-19 2023-05-23 之江实验室 Vehicle stay behavior analysis and prediction method and system based on multitask learning
CN116151489B (en) * 2023-04-19 2023-07-21 之江实验室 Vehicle stay behavior analysis and prediction method and system based on multitask learning

Also Published As

Publication number Publication date
CN112862177B (en) 2024-01-19

Similar Documents

Publication Publication Date Title
CN110827544B (en) Short-term traffic flow control method based on graph convolution recurrent neural network
Zhang et al. Trafficgan: Network-scale deep traffic prediction with generative adversarial nets
Xiao et al. Understanding private car aggregation effect via spatio-temporal analysis of trajectory data
Liu et al. Dynamic spatial-temporal representation learning for traffic flow prediction
CN115240425B (en) Traffic prediction method based on multi-scale space-time fusion graph network
CN114220271B (en) Traffic flow prediction method, equipment and storage medium based on dynamic space-time diagram convolution circulation network
CN112862177A (en) Urban area concentration degree prediction method, equipment and medium based on deep neural network
Yuan et al. An effective joint prediction model for travel demands and traffic flows
CN115578852B (en) DSTGCN-based traffic prediction method
Zhang et al. Curb-gan: Conditional urban traffic estimation through spatio-temporal generative adversarial networks
CN113591380B (en) Traffic flow prediction method, medium and equipment based on graph Gaussian process
CN113570859B (en) Traffic flow prediction method based on asynchronous space-time expansion graph convolution network
CN114944053B (en) Traffic flow prediction method based on space-time hypergraph neural network
CN112488185B (en) Method and system for predicting vehicle operating parameters including spatiotemporal characteristics
Shen et al. Research on traffic speed prediction by temporal clustering analysis and convolutional neural network with deformable kernels (May, 2018)
CN112598165B (en) Urban functional area transfer flow prediction method and device based on private car data
CN110164129A (en) Single Intersection multi-lane traffic flow amount prediction technique based on GERNN
Ouyang et al. Spatial-temporal dynamic graph convolution neural network for air quality prediction
James Citywide estimation of travel time distributions with Bayesian deep graph learning
CN114565187A (en) Traffic network data prediction method based on graph space-time self-coding network
CN115206092A (en) Traffic prediction method of BiLSTM and LightGBM model based on attention mechanism
Shuai et al. Relationship analysis of short-term origin–destination prediction performance and spatiotemporal characteristics in urban rail transit
Zhang et al. Boosted trajectory calibration for traffic state estimation
CN117271899A (en) Interest point recommendation method based on space-time perception
CN117131991A (en) Urban rainfall prediction method and platform based on hybrid neural network

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