A street space vitality prediction method based on deep learning

Abstract

The present invention provides a street space vitality prediction method based on deep learning, which relates to the technical field of regional pedestrian flow prediction, and aims to effectively combine fitting and deep learning to achieve complementary shortcomings, including dividing a year into N periodic submodules, establishing a fitting model for estimating the pedestrian flow of the street for each periodic submodule, constructing an objective function that minimizes the fitting error to obtain the parameters of the fitting model, and obtaining the first pedestrian flow estimation result of the target day through the fitting model; collecting a variety of real-time impact data and combining them with the first pedestrian flow estimation result, and then outputting the second pedestrian flow estimation result through the trained error correction model for prediction and correction, and the real-time impact data includes natural change impact data and artificial change impact data. The present invention has the advantages of strong adaptability, low computing power consumption and high precision in spatial vitality prediction.

Description

Technical Field

[0001] The present invention relates to the technical field of regional pedestrian flow prediction, and in particular to a street space vitality prediction method based on deep learning. Background Art

[0002] With the acceleration of urbanization, the study of street space vitality has gradually become a research hotspot in the fields of urban planning, public management and social economy.

[0003] However, street space vitality prediction faces many technical challenges. For example, traditional prediction methods usually rely on fitting techniques to establish prediction models. Although this method can provide certain prediction effects under small-scale data, as the amount of data increases, it is necessary to process a large amount of historical data, geographic data, and socioeconomic data. The computational complexity and model complexity increase sharply, and the type and parameters of the model are difficult to determine, resulting in poor adaptability and low accuracy of the model. Neural networks can solve the adaptability problem, but data preprocessing is complex, the model training cycle is long, and powerful computing resources are required. Especially when facing large-scale urban data, the computing power consumption is very high, which may lead to low computing efficiency or insufficient resources, thus affecting the real-time and accuracy of the prediction.

[0004] Therefore, it is necessary to optimize the prediction method of street space vitality, effectively combine fitting and deep learning to complement each other's shortcomings, and then obtain a spatial vitality prediction with strong adaptability, low computing power consumption and high accuracy. Summary of the invention

[0005] The purpose of the present invention is to provide a street space vitality prediction method based on deep learning, which effectively combines fitting and deep learning to achieve complementary shortcomings and thus obtain a spatial vitality prediction with strong adaptability, low computing power consumption and high accuracy.

[0006] The present invention is achieved through the following technical solutions:

[0007] A street space vitality prediction method based on deep learning, comprising the following steps:

[0008] Divide the year into periodic submodules, respectively establish a fitting model for estimating the pedestrian flow of the street for each of the periodic submodules, construct an objective function that minimizes the fitting error to obtain the parameters of the fitting model, and obtain the first pedestrian flow estimation result of the target day through the fitting model;

[0009] A variety of real-time impact data are collected and combined with the first pedestrian flow estimation result, and prediction correction is performed through the trained error correction model to output a second pedestrian flow estimation result. The real-time impact data includes natural change impact data and man-made change impact data.

[0010] Preferably, the method for establishing a fitting model for estimating the flow of people on a street is:

[0011] Divide a year into equal parts to obtain a plurality of periodic submodules, and number the periodic submodules in chronological order;

[0012] Based on the equal division, a fitting model for estimating the total flow of people under the periodic submodule is established.

[0013] Preferably, based on the total number of days in a year Get the period value :

[0014] ;

[0015] in, To round down;

[0016] Get the remaining days :

[0017] ;

[0018] in, To find the remainder function;

[0019] The initial value of the number of days for each of the periodic submodules is set to ,exist The periodic submodules are randomly selected And increase the number of days by one day respectively.

[0020] Preferably, the method for establishing a fitting model for estimating the total flow of people under the periodic submodule based on the equal division is:

[0021] ;

[0022] ;

[0023] ;

[0024] ;

[0025] ;

[0026] ;

[0027] ;

[0028] ;

[0029] in, For the first time in a year The crowd flow estimation result of the periodic submodule is , and Within one year, The first-order coefficient, the second-order coefficient and the exponential coefficient of the periodic submodule, For the first time in a year The observed value of the flow of people in the periodic submodule, For the previous year The pedestrian flow observation value of the periodic submodule, For the The number of days of the cycle submodule, is a constant representing the length of time involved in fitting, is a natural constant, U, V, C, D and E are intermediate parameters.

[0030] Preferably, the method for obtaining the parameters of the fitting model is:

[0031] Get daily passenger flow data for Y consecutive years from historical data;

[0032] Establish the objective function:

[0033] ;

[0034] in, and They are the first Year The pedestrian flow estimation results and observation values ​​of the periodic submodules;

[0035] Establish periodic submodule quantity constraints:

[0036] ;

[0037] Establish a constraint on the number of fitting terms:

[0038] ;

[0039] Based on the objective function and the plurality of constraints , and , .

[0040] Preferably, the error correction model receives the real-time impact data and the first pedestrian flow estimation result through an input layer, extracts feature parameters through a feature extraction layer based on the real-time impact data, and then nonlinearizes the feature parameters through a hidden layer, and finally outputs the second pedestrian flow estimation result through an output layer based on the feature parameters and the first pedestrian flow estimation result.

[0041] Preferably, the method for extracting feature parameters by the feature extraction layer is:

[0042] The collected natural change impact data include the temperature range, maximum wind force level and maximum rainfall of the target day obtained based on the weather forecast;

[0043] Extracting natural features based on the natural variation impact data , Represents the day number of the target day within a year. Represents the physical features of the target day;

[0044] The collected data on the impact of human changes include the locations of newly added commercial facilities outside the street within the threshold time and threshold range and the locations of commercial and entertainment activities held on the target day, as well as the number of newly added commercial facilities within the street within the threshold time and the number of commercial and entertainment activities held on the target day;

[0045] Extracting artificial features based on the artificial change impact data , Human characteristics for the target day;

[0046] The natural features and the man-made features of the target day are integrated to form the feature parameters of the target day. :

[0047] .

[0048] Preferably, the method for extracting the natural features according to the natural change impact data is:

[0049] ;

[0050] ;

[0051] ;

[0052] ;

[0053] ;

[0054] ;

[0055] in, Represents the serial number of the periodic submodule, is a truth function, , and The upper limit of the forecast temperature on the target day, The average temperature upper limit of the periodic submodules and the preset temperature upper limit threshold, , and The lower limit of the forecast temperature on the target day, The average temperature lower limit of each periodic submodule and the preset temperature lower limit threshold, , and They are the forecast maximum wind force level on the target day, The average maximum wind force level of the periodic submodules and the preset maximum wind force level threshold, , and They are the forecast maximum rainfall on the target day, The average maximum rainfall of the periodic submodules and the preset maximum rainfall, are the intermediate parameters respectively;

[0056] The method for extracting the human-made features according to the human-made change impact data is:

[0057] ;

[0058] ;

[0059] ;

[0060] ;

[0061] in, Added to the street The distance from the location of a commercial facility to the center of the street, For the first The angle between the two external tangent lines from the location of the commercial facility to the street range, For the street outside The distance from the location of a commercial and recreational activity to the center of the street, For the first The angle between the two external tangent lines from the location of the commercial and entertainment activities to the street range, and They are the total number of newly added commercial facilities outside the street and the total number of commercial and entertainment activities held. and are the experience weights, is the street area, The number of new commercial facilities in the street and the total number of commercial and entertainment activities held on the target day. are the intermediate parameters respectively.

[0062] Preferably, the output layer fuses multiple sigmoid functions to output:

[0063] ;

[0064] in, and are the first crowd flow estimation result and the second crowd flow estimation result of the target day, respectively. and are the natural features and the artificial features processed by the hidden layer, respectively, , and are the weights to be trained, , and are the biases to be trained respectively.

[0065] The technical solution of the present invention has at least the following advantages and beneficial effects:

[0066] The present invention divides the prediction of passenger flow into two parts: global trend prediction and change prediction due to unexpected situations. The global trend prediction is performed by fitting and the change prediction due to unexpected situations is performed by neural network, which reduces the computational burden of neural network and makes up for the accuracy problem of fitting prediction.

[0067] When performing fitting, the present invention not only optimizes the fitting parameters, but also realizes automatic adaptation of the appropriate polynomial type by setting the fitting polynomial, the objective function and the constraint conditions for limiting the number of fitting items, thereby improving the adaptability and accuracy of the fitting to a certain extent;

[0068] The present invention uses a trained error correction model to learn the impact of natural change impact data and artificial change impact data on human flow. Since only fluctuations need to be learned, the learning burden of the model is reduced, which helps to improve computing efficiency and reduce computing resource consumption;

[0069] When performing error correction, the present invention takes into account the impact of the natural environment and the impact of changes in business behavior, further improving the adaptability and accuracy of the prediction and improving the accuracy of the prediction;

[0070] The invention is reasonably designed, easy to implement, and has a high cost-effectiveness ratio, and is convenient to be applied and promoted in the prediction of spatial vitality of various streets. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 A schematic flow chart of a street space vitality prediction method based on deep learning provided in Example 1 of the present invention. DETAILED DESCRIPTION

[0072] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations.

[0073] Example 1

[0074] This embodiment provides a street space vitality prediction method based on deep learning. Figure 1 , including the following steps:

[0075] Divide the year into periodic submodules, respectively establish a fitting model for estimating the pedestrian flow of the street for each of the periodic submodules, construct an objective function that minimizes the fitting error to obtain the parameters of the fitting model, and obtain the first pedestrian flow estimation result of the target day through the fitting model;

[0076] A variety of real-time impact data are collected and combined with the first pedestrian flow estimation result, and prediction correction is performed through the trained error correction model to output a second pedestrian flow estimation result. The real-time impact data includes natural change impact data and man-made change impact data.

[0077] Based on the above scheme, this embodiment divides the task of predicting the flow of people on the street into two main parts: global trend prediction and prediction of changes affected by emergencies. Global trend prediction is carried out through fitting methods. This part mainly focuses on stable trends and patterns over a long period of time. The traditional fitting method can quickly and accurately identify the overall change pattern of human flow. For changes caused by emergencies (such as extreme weather, large-scale events, etc.), trained models can be used for prediction. The model can flexibly respond to these nonlinear and complex emergency factors, and through the powerful fitting ability of the deep learning model, it can capture short-term changes and abnormal patterns in human flow fluctuations. Through this method, the burden of neural network calculation is effectively reduced, and the limitations of traditional fitting methods in prediction accuracy can be compensated, thereby improving the accuracy and efficiency of the overall prediction.

[0078] In this embodiment, the method for establishing a fitting model for estimating the pedestrian flow on a street is:

[0079] Divide a year into equal parts to obtain a plurality of periodic submodules, and number the periodic submodules in chronological order;

[0080] Based on the equal division, a fitting model for estimating the total flow of people under the periodic submodule is established.

[0081] Furthermore, based on the total number of days in a year Get the period value :

[0082] ;

[0083] in, To round down;

[0084] Get the remaining days :

[0085] ;

[0086] in, To find the remainder function;

[0087] The initial value of the number of days for each of the periodic submodules is set to ,exist The periodic submodules are randomly selected And increase the number of days by one day respectively.

[0088] As a preferred solution of this embodiment, a method for establishing a fitting model for estimating the total flow of people under the periodic submodule based on the equal division is:

[0089] ;

[0090] ;

[0091] ;

[0092] ;

[0093] ;

[0094] ;

[0095] ;

[0096] ;

[0097] in, For the first time in a year The crowd flow estimation result of the periodic submodule is , and Within one year, The first-order coefficient, the second-order coefficient and the exponential coefficient of the periodic submodule, For the first time in a year The observed value of the flow of people in the periodic submodule, For the previous year The pedestrian flow observation value of the periodic submodule, For the The number of days of the cycle submodule, is a constant representing the length of time involved in fitting, is a natural constant, U, V, C, D and E are intermediate parameters.

[0098] Furthermore, the method for obtaining the parameters of the fitting model may be:

[0099] Get daily passenger flow data for Y consecutive years from historical data;

[0100] Establish the objective function:

[0101] ;

[0102] in, and They are the first Year The pedestrian flow estimation results and observation values ​​of the periodic submodules;

[0103] Establish periodic submodule quantity constraints:

[0104] ;

[0105] Establish a constraint on the number of fitting terms:

[0106] ;

[0107] Based on the objective function and the plurality of constraints , and , .

[0108] When fitting in this embodiment, the constraints ensure that there will be no overfitting or underfitting in the fitting process, so that the fitting results are smoother and have better generalization ability. First, the fitting of this embodiment divides the whole year into periods. When obtaining the global trend, different fitting formulas are established for the local area, which helps to improve the fitting accuracy. The size of the period division can be determined according to the computing power and demand. For example, each period can preferably include 1-7 days. The less time the period includes, the more accurate it is. Under the condition of computing power operation, 1 day is selected for the best accuracy. At the same time, the design of the fitting polynomial including first-order fitting, second-order fitting and exponential fitting and the setting of the objective function can automatically adapt to the changing trend of the data, ensuring that the model can automatically adjust to the appropriate polynomial type, thereby improving the adaptability and accuracy of the fitting. The fitting design method adopted in this embodiment not only optimizes the fitting parameters, but also can automatically and flexibly respond to different data environments, thereby improving the robustness of the model. It is particularly noted that the fitting model can be periodically updated based on the objective function and constraints to ensure immediate adaptability. When obtaining the first person flow estimation result of a specific target day, it can be obtained The value is then divided by the number of days in the submodule of the cycle.

[0109] On the other hand, the error correction model receives the real-time impact data and the first pedestrian flow estimation result through the input layer, extracts feature parameters based on the real-time impact data through the feature extraction layer, and then nonlinearizes the feature parameters through the hidden layer, and finally outputs the second pedestrian flow estimation result through the output layer based on the feature parameters and the first pedestrian flow estimation result.

[0110] Specifically, the method for extracting feature parameters in the feature extraction layer is:

[0111] The collected natural change impact data include the temperature range, maximum wind force level and maximum rainfall of the target day obtained based on the weather forecast;

[0112] Extracting natural features based on the natural variation impact data , Represents the day number of the target day within a year. Represents the physical features of the target day;

[0113] The collected data on the impact of human changes include the locations of newly added commercial facilities outside the street within the threshold time and threshold range and the locations of commercial and entertainment activities held on the target day, as well as the number of newly added commercial facilities within the street within the threshold time and the number of commercial and entertainment activities held on the target day;

[0114] Extracting artificial features based on the artificial change impact data , Human characteristics for the target day;

[0115] The natural features and the man-made features of the target day are integrated to form the feature parameters of the target day. :

[0116] .

[0117] When performing feature extraction, the method for extracting the natural features according to the natural change impact data is:

[0118] ;

[0119] ;

[0120] ;

[0121] ;

[0122] ;

[0123] ;

[0124] in, Represents the serial number of the periodic submodule, is a truth function, , and The upper limit of the forecast temperature on the target day, The average temperature upper limit of the periodic submodules and the preset temperature upper limit threshold, , and The lower limit of the forecast temperature on the target day, The average temperature lower limit of each periodic submodule and the preset temperature lower limit threshold, , and They are the forecast maximum wind force level on the target day, The average maximum wind force level of the periodic submodules and the preset maximum wind force level threshold, , and They are the forecast maximum rainfall on the target day, The average maximum rainfall of the periodic submodules and the preset maximum rainfall, are the intermediate parameters, representing the relevant characteristics of the upper and lower limits of temperature, wind force, and rainfall, respectively;

[0125] The method for extracting the human-made features according to the human-made change impact data is:

[0126] ;

[0127] ;

[0128] ;

[0129] ;

[0130] in, Added to the street The distance from the location of a commercial facility to the center of the street, For the first The angle between the two external tangent lines from the location of the commercial facility to the street range, For the street outside The distance from the location of a commercial and recreational activity to the center of the street, For the first The angle between the two external tangent lines from the location of the commercial and entertainment activities to the street range, and They are the total number of newly added commercial facilities outside the street and the total number of commercial and entertainment activities held. and are the experience weights, is the street area, The number of new commercial facilities in the street and the total number of commercial and entertainment activities held on the target day. are the intermediate parameters, They represent the relevant characteristics of newly added commercial facilities and commercial entertainment activities. It represents a combination of street area and related characteristics of new commercial facilities and the total number of commercial and recreational activities, as street area also affects the flow of people.

[0131] In the above formula, and Represents the radiation range, and The radiation range is corrected based on the distance factor. and It can be set according to the degree of correction required.

[0132] Finally, the method of outputting the output layer by fusing multiple sigmoid functions is preferably:

[0133] ;

[0134] in, and are the first crowd flow estimation result and the second crowd flow estimation result of the target day, respectively. and are the natural features and the artificial features processed by the hidden layer, respectively, , and are the weights to be trained, , and are the biases to be trained respectively.

[0135] This embodiment mainly uses an error correction model to learn and correct the impact of changes in the natural environment and changes in human factors on the flow of people. The natural environment change mainly extracts the weather forecast information of the target day, and then compares it with the actual weather observation data of the same period in previous years and the preset threshold. The preset threshold can be set according to the actual weather characteristics of the corresponding area, and is set based on the boundary of the local comfortable weather. Changes in human factors take into account the additional flow of people brought by new commercial facilities and commercial activities in and around the streets. Based on the above two factors, the error correction model performs feature learning, captures the relationship between these features and the percentage increase or decrease in the flow of people, and then achieves error correction of the fitted prediction with a certain degree of accuracy, and finally obtains a prediction result that is more in line with the characteristics of the target day.

[0136] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A street space vitality prediction method based on deep learning, characterized in that: The following steps are involved: Divide the year into periodic submodules, respectively establish a fitting model for estimating the pedestrian flow of the street for each of the periodic submodules, construct an objective function that minimizes the fitting error to obtain the parameters of the fitting model, and obtain the first pedestrian flow estimation result of the target day through the fitting model; Collecting a variety of real-time impact data and combining them with the first pedestrian flow estimation result, performing prediction correction through the trained error correction model and then outputting a second pedestrian flow estimation result, wherein the real-time impact data includes natural change impact data and man-made change impact data; The error correction model receives the real-time impact data and the first crowd flow estimation result through an input layer, extracts feature parameters through a feature extraction layer based on the real-time impact data, and then nonlinearizes the feature parameters through a hidden layer, and finally outputs the second crowd flow estimation result through an output layer according to the feature parameters and the first crowd flow estimation result; The method for extracting feature parameters in the feature extraction layer is: The collected natural change impact data include the temperature range, maximum wind force level and maximum rainfall of the target day obtained based on the weather forecast; Extracting natural features based on the natural variation impact data , Represents the day number of the target day within a year. Represents the physical features of the target day; The collected data on the impact of human changes include the locations of newly added commercial facilities outside the street within the threshold time and threshold range and the locations of commercial and entertainment activities held on the target day, as well as the number of newly added commercial facilities within the street within the threshold time and the number of commercial and entertainment activities held on the target day; Extracting artificial features based on the artificial change impact data , Human characteristics for the target day; The natural features and the man-made features of the target day are integrated to form the feature parameters of the target day. : 。 2. According to claim 1, a street space activity prediction method based on deep learning is characterized in that: The method for establishing a fitting model for estimating the flow of people on a street is: Divide a year into equal parts to obtain a plurality of periodic submodules, and number the periodic submodules in chronological order; Based on the equal division, a fitting model for estimating the total flow of people under the periodic submodule is established.

3. The street space activity prediction method based on deep learning according to claim 2 is characterized in that: The method for performing equal division is: Based on the total number of days in a year Get the period value : ； in, To round down; Get the remaining days : ； in, To find the remainder function; The initial value of the number of days for each of the periodic submodules is set to ,exist The periodic submodules are randomly selected And increase the number of days by one day respectively.

4. The street space activity prediction method based on deep learning according to claim 3 is characterized in that: The method for establishing a fitting model for estimating the total flow of people under the periodic submodule based on the equal division is: ； ； ； ； ； ； ； ； in, For the first time in a year The crowd flow estimation result of the periodic submodule is , and Within one year, The first-order coefficient, the second-order coefficient and the exponential coefficient of the periodic submodule, For the first time in a year The observed value of the flow of people in the periodic submodule, For the previous year The pedestrian flow observation value of the periodic submodule, For the The number of days of the cycle submodule, is a constant representing the length of time involved in fitting, is a natural constant, U, V, C, D and E are intermediate parameters.

5. The street space activity prediction method based on deep learning according to claim 4 is characterized in that: The method for obtaining the parameters of the fitting model is: Get daily passenger flow data for Y consecutive years from historical data; Establish the objective function: ； in, and They are the first Year The pedestrian flow estimation results and observation values ​​of the periodic submodules; Establish the periodic submodule quantity constraint: ； Establish a constraint on the number of fitting terms: ； Based on the objective function and the plurality of constraints , and , .

6. The street space activity prediction method based on deep learning according to claim 1 is characterized in that: The method for extracting the natural features according to the natural change impact data is: ； ； ； ； ； ； in, Represents the serial number of the periodic submodule, is a truth function, , and The upper limit of the forecast temperature on the target day, The average temperature upper limit of the periodic submodules and the preset temperature upper limit threshold, , and The lower limit of the forecast temperature on the target day, The average temperature lower limit of each periodic submodule and the preset temperature lower limit threshold, , and They are the forecast maximum wind force level on the target day, The average maximum wind force level of the periodic submodules and the preset maximum wind force level threshold, , and They are the forecast maximum rainfall on the target day, The average maximum rainfall of the periodic submodules and the preset maximum rainfall, are the intermediate parameters respectively; The method for extracting the human-made features according to the human-made change impact data is: ； ； ； ； in, Added to the street The distance from the location of a commercial facility to the street center, For the new The angle between the two external tangent lines from the location of a commercial facility to the street range, For the street outside The distance from the location of a commercial and recreational activity to the center of the street, For the first The angle between the two external tangent lines from the location of the commercial and entertainment activities to the street range, and They are the total number of newly added commercial facilities outside the street and the total number of commercial and entertainment activities held. and are the experience weights, is the street area, The number of new commercial facilities in the street and the total number of commercial and entertainment activities held on the target day. are the intermediate parameters respectively.

7. The street space activity prediction method based on deep learning according to claim 6 is characterized in that: The output layer fuses multiple sigmoid functions to output: ; in, and are the first crowd flow estimation result and the second crowd flow estimation result of the target day, respectively. and are the natural features and the artificial features processed by the hidden layer, respectively, , and are the weights to be trained, , and are the biases to be trained respectively.

Images