CN114818984B - Refined urban ponding water level fitting method based on artificial intelligence - Google Patents

Refined urban ponding water level fitting method based on artificial intelligence Download PDF

Info

Publication number
CN114818984B
CN114818984B CN202210605679.1A CN202210605679A CN114818984B CN 114818984 B CN114818984 B CN 114818984B CN 202210605679 A CN202210605679 A CN 202210605679A CN 114818984 B CN114818984 B CN 114818984B
Authority
CN
China
Prior art keywords
information
water level
ponding
ponding water
hydrological
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.)
Active
Application number
CN202210605679.1A
Other languages
Chinese (zh)
Other versions
CN114818984A (en
Inventor
智协飞
吕阳
季焱
朱寿鹏
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.)
Nanjing University of Information Science and Technology
Original Assignee
Nanjing University of Information Science and Technology
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 Nanjing University of Information Science and Technology filed Critical Nanjing University of Information Science and Technology
Priority to CN202210605679.1A priority Critical patent/CN114818984B/en
Publication of CN114818984A publication Critical patent/CN114818984A/en
Application granted granted Critical
Publication of CN114818984B publication Critical patent/CN114818984B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/23—Clustering techniques
    • G06F18/232—Non-hierarchical techniques
    • G06F18/2323—Non-hierarchical techniques based on graph theory, e.g. minimum spanning trees [MST] or graph cuts
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/20—Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
    • G06F16/21—Design, administration or maintenance of databases
    • G06F16/215—Improving data quality; Data cleansing, e.g. de-duplication, removing invalid entries or correcting typographical errors
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/20—Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
    • G06F16/29—Geographical information databases
    • 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
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/048—Activation functions
    • 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
    • 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
    • G06Q50/00—Systems or methods specially adapted for specific business sectors, e.g. utilities or tourism
    • G06Q50/10—Services
    • G06Q50/26—Government or public services
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
    • Y02A10/00—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE at coastal zones; at river basins
    • Y02A10/40—Controlling or monitoring, e.g. of flood or hurricane; Forecasting, e.g. risk assessment or mapping

Abstract

The invention discloses a refined urban ponding water level fitting method based on artificial intelligence, which comprises the following steps: 1. constructing a dimensionless hydrological feature database; 2. developing cluster analysis based on a graph neural network method; 3. dividing the urban ponding water level subarea; 4. establishing a personalized urban hydrological concept model region by region based on a neural network; 5. according to the ponding water level monitoring information of the ponding monitoring station, combining longitude information, latitude information, time information and ground elevation information, the ponding water level of any position of each subarea is reversely derived. The model has strong applicability, the model integrates the geographic information and the time information of the ponding sites, the space-time distribution characteristic of the ponding water level is effectively extracted, and the fitting capability of the model is strong; the system can provide ponding water level products at any position of the whole area, realizes grid removal, can realize return of historical water level based on a historical synchronization database, and plays an important role in preventing urban waterlogging and reasonably planning cities.

Description

Refined urban ponding water level fitting method based on artificial intelligence
Technical Field
The invention relates to a fitting method for urban accumulated water level, in particular to a refined urban accumulated water level fitting method based on artificial intelligence.
Background
Under the global warming background, extreme events occur frequently, and urban waterlogging is serious day by day, so that the life and property safety of people is seriously influenced. Under the background, high-quality urban ponding water level data has important significance for disaster prevention and reduction and urban future reasonable planning. At present, a part of cities are provided with urban ponding information monitoring systems, ponding information monitoring of a part of important road sections is realized, but in consideration of the cost of a ponding monitoring station and the actual conditions of all road sections, ponding monitoring of all areas of the cities is difficult to realize. In addition, the newly-added ponding monitoring station can not monitor the historical ponding water level at this position. Therefore, fitting is carried out on the water level of the whole area of the water tank based on the water tank monitoring information of the existing water tank monitoring station, and the return of the historical water level is realized.
The method is characterized in that the accumulated water level is fitted based on a simple linear interpolation method, namely linear interpolation is carried out according to the accumulated water levels of a plurality of accumulated water monitoring stations around a target position, but the method neglects the nonlinear characteristic of the accumulated water level distribution, does not consider the space-time distribution characteristic of the accumulated water level, and is difficult to generate high-quality accumulated water products. In addition, the existing method realizes the fitting of the urban accumulated water level based on the same model, neglects the difference of hydrological characteristics of different areas of the city, and has poor pertinence of the model. Meanwhile, regular gridding products are generated by the conventional method, effective fitting of ponding products in any position of the whole area is difficult to realize, and the application scene is limited.
Disclosure of Invention
The purpose of the invention is as follows: the invention aims to provide a refined urban ponding water level fitting method based on artificial intelligence.
The technical scheme is as follows: the invention discloses an artificial intelligence-based fitting method for refining urban ponding water level, which comprises the following steps:
step 1, collecting ponding water level monitoring information, ground elevation information, longitude information, latitude information and time information of an urban ponding monitoring station, eliminating abnormal values, then carrying out standardization processing on data, and constructing a dimensionless hydrological feature database;
step 2, performing cluster analysis on the database in the step 1 based on a Graph Neural Network method, and clustering sites with similar hydrological characteristics together;
step 3, dividing the city into subareas based on the clustering result in the step 2, and dividing the subareas of the ponding water level of the city;
step 4, taking ground elevation information, longitude information, latitude information and time information of all the sites and ponding water level monitoring information outside the target site as input variables, taking the ponding water level of the target site as an output variable, and constructing a personalized hydrological conceptual model region by region;
and 5, based on the hydrological conceptual model constructed region by region in the step 4, utilizing the ponding water level monitoring information, the ground elevation information, the longitude information, the latitude information and the time information of the ponding monitoring stations in the sub-regions and the ground elevation information, the longitude information, the latitude information and the time information of the target positions to reversely derive the ponding water level of any position of each sub-region.
Further, in step 1, the calculation formula of the normalization process is:
Figure 978251DEST_PATH_IMAGE001
in the formula (I), the compound is shown in the specification,
Figure 400005DEST_PATH_IMAGE002
for the purpose of the result after the normalization,
Figure 524956DEST_PATH_IMAGE003
represent
Figure 415683DEST_PATH_IMAGE004
Is determined by the average value of (a) of (b),
Figure 854754DEST_PATH_IMAGE005
to represent
Figure 131015DEST_PATH_IMAGE004
Standard deviation of (d); after all variables are standardized, hydrological feature vectors of all ponding sites are formed:
Figure 630129DEST_PATH_IMAGE006
in the formula (I), the compound is shown in the specification,
Figure 8152DEST_PATH_IMAGE007
representing the hydrological feature vector for the kth ponding site,
Figure 47652DEST_PATH_IMAGE008
and respectively representing the standardized ponding water level monitoring information, the ground elevation information, the longitude information, the latitude information and the time information, thereby forming a dimensionless hydrological feature database.
Further, step 2 specifically comprises:
step 2.1, presume study area
Figure 709578DEST_PATH_IMAGE009
Exist of
Figure 864747DEST_PATH_IMAGE010
A site, constructing a topology network based on site distribution
Figure 713754DEST_PATH_IMAGE011
Topological networks
Figure 760208DEST_PATH_IMAGE011
Mainly consists of two parts which are respectively nodes
Figure 276640DEST_PATH_IMAGE012
And an edge
Figure 665027DEST_PATH_IMAGE013
Wherein the node
Figure 1330DEST_PATH_IMAGE012
Representing the stations in the research area, wherein the characteristic vector of each station is shown as formula (2); edge of a container
Figure 585895DEST_PATH_IMAGE013
Representing the interrelationship between the various sites;
step 2.2, network based topology
Figure 441987DEST_PATH_IMAGE011
Building a corresponding input matrix
Figure 453805DEST_PATH_IMAGE004
Input matrix
Figure 542984DEST_PATH_IMAGE004
Topology network
Figure 665661DEST_PATH_IMAGE011
Imaging, and combining n sites
Figure 704155DEST_PATH_IMAGE014
The two-dimensional matrix represents the upstream and downstream influences existing among different nodes by using the weight;
step 2.3, constructing Graph Auto Encode, which comprises compiling original input matrix by using encoder Encode
Figure 886875DEST_PATH_IMAGE004
And reconstructing the original network structure by using a decoder; first, the input matrix in step 2.2 is encoded using an encoder
Figure 197770DEST_PATH_IMAGE004
Feature vector of
Figure 874871DEST_PATH_IMAGE015
Is projected to
Figure 220401DEST_PATH_IMAGE016
Coding space of dimension
Figure 574022DEST_PATH_IMAGE017
I.e. by
Figure 106635DEST_PATH_IMAGE018
(ii) a At this point, consider the use of a mechanism with attentionThe encoder performs weighted average on the neighborhood nodes, and the calculation formula is written as:
Figure 587426DEST_PATH_IMAGE019
wherein the content of the first and second substances,
Figure 584201DEST_PATH_IMAGE020
is composed of
Figure 108723DEST_PATH_IMAGE021
The function of the function is that of the function,
Figure 144943DEST_PATH_IMAGE022
、
Figure 678693DEST_PATH_IMAGE023
are respectively nodes
Figure 733236DEST_PATH_IMAGE024
In the encoder
Figure 428660DEST_PATH_IMAGE025
Layer and the first
Figure 686597DEST_PATH_IMAGE026
The information of the layer(s) is,
Figure 758458DEST_PATH_IMAGE027
is a node
Figure 667508DEST_PATH_IMAGE024
Of the neighborhood of the node in the cluster,
Figure 346882DEST_PATH_IMAGE028
is a node
Figure 341383DEST_PATH_IMAGE024
And between the neighboring nodes
Figure 216935DEST_PATH_IMAGE029
The attention weight of (1);
Figure 980492DEST_PATH_IMAGE021
the calculation formula of the function is:
Figure 34030DEST_PATH_IMAGE030
then, the coding space is calculated
Figure 250248DEST_PATH_IMAGE017
Obtaining the inner product between the inner node pairs to obtain the reconstructed original network
Figure 663911DEST_PATH_IMAGE031
(ii) a At this time, the error of the graph self-encoder reconstruction network is represented as:
Figure 95024DEST_PATH_IMAGE032
step 2.4, constructing a self-training clustering network, and assuming that a clustering center is
Figure 568730DEST_PATH_IMAGE033
Then node
Figure 272244DEST_PATH_IMAGE024
Belong to
Figure 489599DEST_PATH_IMAGE034
Probability of class
Figure 978480DEST_PATH_IMAGE035
Expressed as:
Figure 419826DEST_PATH_IMAGE036
wherein the content of the first and second substances,
Figure 876215DEST_PATH_IMAGE037
is as followskThe center of the class is classified into a class center,
Figure 382414DEST_PATH_IMAGE038
is a node
Figure 240648DEST_PATH_IMAGE024
And cluster center
Figure 993841DEST_PATH_IMAGE033
The Euclidean distance between the nodes and the similarity between the nodes are represented, in the clustering problem, each node and the corresponding clustering center are forced to tend to the minimum intra-class distance and the maximum inter-class distance, and the target distribution is realized
Figure 734264DEST_PATH_IMAGE039
Is defined as:
Figure 44153DEST_PATH_IMAGE040
at this time, the KL divergence is used to characterize the predicted cluster distribution
Figure 756895DEST_PATH_IMAGE035
And target distribution
Figure 477726DEST_PATH_IMAGE039
Error between, defining a self-trained loss function:
Figure 643128DEST_PATH_IMAGE041
the final loss function
Figure 491129DEST_PATH_IMAGE042
Writing as follows:
Figure 120694DEST_PATH_IMAGE043
Figure 278006DEST_PATH_IMAGE044
in order to be a weight coefficient of the image,
Figure 681436DEST_PATH_IMAGE045
reconstructing errors of the network for the image self-encoder by minimizing
Figure 582396DEST_PATH_IMAGE042
Training the model according to the time
Figure 4150DEST_PATH_IMAGE035
Judging node
Figure 879834DEST_PATH_IMAGE024
The class to which it most likely belongs.
Further, step 4 specifically includes:
under the same subregion, taking the ground elevation information, longitude information, latitude information and time information of all the sites and the ponding water level monitoring information outside the target site as input variables, taking the ponding water level of the target site as output variables, constructing an individualized hydrological conceptual model region by region based on a neural network, wherein the input variables can be expressed as:
Figure 19828DEST_PATH_IMAGE046
wherein
Figure 458899DEST_PATH_IMAGE008
Respectively represents the ponding water level monitoring information, the ground elevation information, the longitude information, the latitude information and the time information,
Figure 739DEST_PATH_IMAGE010
representing the number of water accumulation sites in the sub-area, assuming each site has
Figure 250586DEST_PATH_IMAGE047
Data of individual accumulated water, all are
Figure 612297DEST_PATH_IMAGE048
The model is trained by each sample, the model effectively extracts the spatial distribution characteristics of the water level of the ponding water by adding the ground elevation information, the longitude information and the latitude information, the seasonal change and the daily change characteristics of the ponding water level are favorably captured by adding the time information, the fitting capacity of the model is enhanced, and the calculation formula of the neural network is as follows:
Figure 855060DEST_PATH_IMAGE049
Figure 251406DEST_PATH_IMAGE050
，
Figure 492329DEST_PATH_IMAGE051
，
Figure 606916DEST_PATH_IMAGE052
representing input layer, hidden layer and output layer vectors respectively,
Figure 653369DEST_PATH_IMAGE010
in order to hide the number of layers,
Figure 654955DEST_PATH_IMAGE053
and
Figure 495872DEST_PATH_IMAGE054
for each of the training parameters of the layers,
Figure 97754DEST_PATH_IMAGE055
representing the activation function, which is also a non-linear source of the neural network, the activation function to be adopted by the method is a ReLU function:
Figure 682319DEST_PATH_IMAGE056
。
has the advantages that: compared with the prior art, the invention has the following remarkable advantages:
(1) the invention carries out cluster analysis on the hydrological characteristics of the ponding monitoring station based on the graph neural network method, divides the city into a plurality of sub-areas, and constructs an individual hydrological model for the actual hydrological characteristics of each sub-area by area, and the model has stronger applicability.
(2) The method is based on the neural network method for modeling, can effectively extract the nonlinear characteristics of the ponding water level, integrates the geographic information (ground elevation information, longitude and latitude information) and time information of the ponding sites, is favorable for extracting the spatial-temporal distribution characteristics of the ponding water level, and improves the fitting capability of the model.
(3) The invention can provide ponding water level products at any position in the whole area, realizes gridding removal, can realize return of historical water level based on the historical synchronization database, has strong application value, and plays an important role in preventing urban waterlogging and reasonably planning cities.
Drawings
FIG. 1 is a general flow chart of the present invention;
FIG. 2 is a schematic diagram of a Graph Neural Network (GNN) according to the present invention;
FIG. 3 is a schematic diagram of a city ponding information site clustering;
figure 4 is a schematic view of a neural network,
Figure 803990DEST_PATH_IMAGE057
，
Figure 815809DEST_PATH_IMAGE058
，
Figure 904987DEST_PATH_IMAGE059
representing input layer, hidden layer and output layer vectors, respectively.
Detailed Description
The technical scheme of the invention is further explained by combining the attached drawings.
The flow of the fitting method for refining the urban ponding water level based on artificial intelligence is shown in figure 1.
Step 1, collecting ponding water level monitoring information, ground elevation information, longitude information, latitude information and time information of an urban ponding monitoring station, and carrying out standardized processing on data after abnormal values are eliminated, wherein the calculation formula is as follows:
Figure 762085DEST_PATH_IMAGE001
in the formula (I), the compound is shown in the specification,
Figure 738262DEST_PATH_IMAGE060
for the purpose of the normalized result, the results,
Figure 186561DEST_PATH_IMAGE061
to represent
Figure 497457DEST_PATH_IMAGE062
Is determined by the average value of (a) of (b),
Figure 971295DEST_PATH_IMAGE063
to represent
Figure 316825DEST_PATH_IMAGE062
Standard deviation of (2). After all variables are standardized, hydrological feature vectors of all ponding sites are formed:
Figure 670446DEST_PATH_IMAGE006
in the formula (I), the compound is shown in the specification,
Figure 468638DEST_PATH_IMAGE007
representing the hydrological feature vector for the kth ponding site,
Figure 949429DEST_PATH_IMAGE008
and respectively representing the standardized ponding water level monitoring information, the ground elevation information, the longitude information, the latitude information and the time information, thereby forming a dimensionless hydrological feature database.
And 2, performing cluster analysis based on a Graph Neural Network (GNN) method, and clustering sites with similar hydrological features. The method specifically comprises the following steps:
step 2.1 postulating the region of investigation
Figure 149466DEST_PATH_IMAGE009
Exist of
Figure 408409DEST_PATH_IMAGE010
Each site, constructing a topology network based on site distribution
Figure 693897DEST_PATH_IMAGE011
As shown in fig. 2 (a), fig. 2 (a) is mainly composed of two parts, which are nodes respectively
Figure 712800DEST_PATH_IMAGE012
And an edge
Figure 767343DEST_PATH_IMAGE013
Wherein the node
Figure 728346DEST_PATH_IMAGE012
Representing the stations in the research area, wherein the characteristic vector of each station is shown as formula (2); edge of a container
Figure 501130DEST_PATH_IMAGE013
Representing the interrelationship between the various sites; if the underground drainage network is shared, the terrain is high or low, and the like. Note that the edges
Figure 58145DEST_PATH_IMAGE013
The vector for the belt direction, as in fig. 2 (b), may represent a unidirectional effect, and may also characterize a bidirectional effect. At this time, the network is topological
Figure 232774DEST_PATH_IMAGE011
Can use a feature matrix
Figure 99099DEST_PATH_IMAGE064
（
Figure 359179DEST_PATH_IMAGE065
A station,
Figure 516622DEST_PATH_IMAGE066
A feature) and an edge
Figure 545758DEST_PATH_IMAGE013
And (4) performing representation.
Step 2.2, known from step 2.1, the node
Figure 582984DEST_PATH_IMAGE012
The neighborhood influence is controlled by the constructed topology, and the influence networks of different nodes are very different. Network according to topology
Figure 815513DEST_PATH_IMAGE011
Constructing a corresponding regular input matrix
Figure 229177DEST_PATH_IMAGE067
As in fig. 2 (c). Input matrix
Figure 909557DEST_PATH_IMAGE067
Topology network
Figure 133996DEST_PATH_IMAGE011
Is like an elephant and is composed of
Figure 103089DEST_PATH_IMAGE068
The two-dimensional matrix of (2) represents the upstream and downstream influences existing among different nodes by using the weight.
Step 2.3, constructing Graph Auto encoder (Graph Auto encoder) which includes compiling original input matrix by encoder (encoder)
Figure 54865DEST_PATH_IMAGE067
And reconstructing the original network structure using an inverse encoder (Decode). First, the input matrix in step 2.2 is encoded using an encoder
Figure 58593DEST_PATH_IMAGE067
Feature vector of
Figure 188354DEST_PATH_IMAGE064
Is projected to
Figure 644743DEST_PATH_IMAGE069
Coding space of dimension
Figure 400209DEST_PATH_IMAGE070
I.e. by
Figure 992865DEST_PATH_IMAGE071
. At this time, considering that an encoder with attention mechanism is used to perform weighted average on the neighborhood nodes, the calculation formula can be written as:
Figure 559106DEST_PATH_IMAGE019
wherein the content of the first and second substances,
Figure 502792DEST_PATH_IMAGE020
is composed of
Figure 61949DEST_PATH_IMAGE021
The function of the function is that of the function,
Figure 774690DEST_PATH_IMAGE022
、
Figure 246254DEST_PATH_IMAGE023
are respectively nodes
Figure 473973DEST_PATH_IMAGE024
In the encoder it is
Figure 571242DEST_PATH_IMAGE025
Layer and the first
Figure 889222DEST_PATH_IMAGE026
The information of the layer(s) is,
Figure 780954DEST_PATH_IMAGE027
is a node
Figure 699232DEST_PATH_IMAGE024
The set of neighborhood nodes of (a) is,
Figure 600192DEST_PATH_IMAGE028
is a node
Figure 772678DEST_PATH_IMAGE024
And between the neighboring nodes
Figure 897629DEST_PATH_IMAGE029
The attention weight of (1);
Figure 37623DEST_PATH_IMAGE021
the calculation formula of the function is:
Figure 493007DEST_PATH_IMAGE030
then, the coding space is calculated
Figure 769267DEST_PATH_IMAGE017
Obtaining the inner product between the inner node pairs to obtain the reconstructed original network
Figure 268382DEST_PATH_IMAGE031
(ii) a At this time, the error of the graph self-encoder reconstructing the network is represented as:
Figure 895672DEST_PATH_IMAGE032
step 2.4, constructing a self-training clustering network, and assuming that a clustering center is
Figure 889167DEST_PATH_IMAGE033
Then node
Figure 285513DEST_PATH_IMAGE024
Belong to
Figure 752267DEST_PATH_IMAGE034
Probability of class
Figure 899477DEST_PATH_IMAGE035
Expressed as:
Figure 680351DEST_PATH_IMAGE036
wherein the content of the first and second substances,
Figure 196783DEST_PATH_IMAGE037
is as followskThe center of the class is classified into a class center,
Figure 788432DEST_PATH_IMAGE038
is a node
Figure 390315DEST_PATH_IMAGE024
And cluster center
Figure 709301DEST_PATH_IMAGE033
The Euclidean distance between the nodes and the similarity between the nodes are represented, in the clustering problem, each node and the corresponding clustering center are forced to tend to the minimum intra-class distance and the maximum inter-class distance, and the target distribution is realized
Figure 80239DEST_PATH_IMAGE039
Is defined as:
Figure 842790DEST_PATH_IMAGE040
at this time, the KL divergence is used to characterize the predicted cluster distribution
Figure 463127DEST_PATH_IMAGE035
And target distribution
Figure 851383DEST_PATH_IMAGE039
Error between, defining loss of self-trainingFunction:
Figure 827561DEST_PATH_IMAGE041
the final loss function
Figure 10280DEST_PATH_IMAGE042
Write as:
Figure 586755DEST_PATH_IMAGE043
Figure 247544DEST_PATH_IMAGE044
in order to be the weight coefficient,
Figure 78227DEST_PATH_IMAGE045
reconstructing errors of the network for the image self-encoder by minimizing
Figure 697428DEST_PATH_IMAGE042
Training the model according to the time
Figure 292357DEST_PATH_IMAGE035
Judging node
Figure 773148DEST_PATH_IMAGE024
The class to which it most likely belongs. Fig. 3 is a schematic diagram of a clustering result, in which a city is taken as an example and is divided into 3 sub-regions.
And 3, dividing the sites with similar hydrological characteristics into the same sub-area based on the clustering result in the step 2, and then dividing the whole city into k sub-areas (the number k of the sub-areas depends on the clustering result).
And 4, constructing the personalized hydrological concept model region by region.
Under the same subregion, the ground elevation information, longitude information, latitude information, time information of all the sites and the ponding water level monitoring information outside the target site are used as input variables, the ponding water level of the target site is used as an output variable, a personalized hydrological conceptual model is constructed region by region based on a neural network, and the schematic diagram of the neural network is shown in fig. 4. The input variables may be represented as:
Figure 707606DEST_PATH_IMAGE046
wherein
Figure 497707DEST_PATH_IMAGE008
Respectively represents the ponding water level monitoring information, the ground elevation information, the longitude information, the latitude information and the time information,
Figure 517616DEST_PATH_IMAGE010
representing the number of water accumulation sites in the sub-area, assuming each site has
Figure 536519DEST_PATH_IMAGE047
Data of water accumulation, all are
Figure 591062DEST_PATH_IMAGE048
The model is trained by each sample, the model effectively extracts the spatial distribution characteristics of the water level of the ponding water by adding the ground elevation information, the longitude information and the latitude information, the seasonal change and the daily change characteristics of the ponding water level are favorably captured by adding the time information, the fitting capacity of the model is enhanced, and the calculation formula of the neural network is as follows:
Figure 348803DEST_PATH_IMAGE049
Figure 606740DEST_PATH_IMAGE050
，
Figure 678601DEST_PATH_IMAGE051
，
Figure 853231DEST_PATH_IMAGE052
are respectively provided withRepresenting the input layer, hidden layer and output layer vectors,
Figure 719555DEST_PATH_IMAGE010
in order to hide the number of layers,
Figure 730368DEST_PATH_IMAGE053
and
Figure 74762DEST_PATH_IMAGE054
for each of the training parameters of the layers,
Figure 651368DEST_PATH_IMAGE055
representing the activation function, which is also a non-linear source of the neural network, the activation function to be adopted by the method is a ReLU function:
Figure 439326DEST_PATH_IMAGE056
。
and 5, according to the ponding information of the ponding monitoring station, the ponding information of each subarea is reversely derived.
And (4) inputting longitude information, latitude information, ground elevation information and time information of any target position based on the personalized hydrological model trained region by region in the step 4, and matching the ponding water level monitoring information, the ground elevation information, the longitude information, the latitude information and the time information of all the ponding monitoring stations in the sub-region at the same period to obtain the ponding water level information of the position. The scheme can obtain ponding water level information of all areas at any positions, and can provide high-resolution refined ponding products. In addition, the scheme can realize the return of the water level of the whole area based on the historical contemporaneous water level data.

Claims (2)

1. A refined urban ponding water level fitting method based on artificial intelligence is characterized by comprising the following steps:
step 1, collecting ponding water level monitoring information, ground elevation information, longitude information, latitude information and time information of an urban ponding monitoring station, eliminating abnormal values, then carrying out standardization processing on data, and constructing a dimensionless hydrological feature database;
step 2, performing cluster analysis on the database in the step 1 based on a Graph Neural Network method, and clustering sites with similar hydrological characteristics together, wherein the method comprises the following steps:
step 2.1, presume study area
Figure 892816DEST_PATH_IMAGE001
Exist of
Figure 799592DEST_PATH_IMAGE002
Each site, constructing a topology network based on site distribution
Figure 938318DEST_PATH_IMAGE003
Topological networks
Figure 879729DEST_PATH_IMAGE003
Mainly consists of two parts which are respectively nodes
Figure 462020DEST_PATH_IMAGE004
And an edge
Figure 121672DEST_PATH_IMAGE005
Wherein the node
Figure 814821DEST_PATH_IMAGE004
Representing the stations in the research area, wherein the characteristic vector of each station is shown as formula (2); edge of a container
Figure 594427DEST_PATH_IMAGE005
Representing the interrelationship between the various sites;
step 2.2, according to the topological network
Figure 347620DEST_PATH_IMAGE003
Building a corresponding input matrix
Figure 228988DEST_PATH_IMAGE006
Input matrix
Figure 725828DEST_PATH_IMAGE006
Topology network
Figure 891099DEST_PATH_IMAGE003
Imaging, and combining n sites
Figure 549614DEST_PATH_IMAGE007
The two-dimensional matrix of (2) represents the upstream and downstream influences existing among different nodes by using the weight;
step 2.3, constructing Graph Auto Encode, which comprises compiling original input matrix by using encoder Encode
Figure 918278DEST_PATH_IMAGE006
And reconstructing the original network structure by using a decoder; first, the input matrix in step 2.2 is encoded using an encoder
Figure 953230DEST_PATH_IMAGE006
Feature vector of
Figure 989320DEST_PATH_IMAGE008
Is projected to
Figure 68003DEST_PATH_IMAGE009
Coding space of dimension
Figure 923963DEST_PATH_IMAGE010
I.e. by
Figure 762606DEST_PATH_IMAGE011
(ii) a At this time, considering that an encoder with attention mechanism is used to perform weighted average on the neighborhood nodes, the calculation formula is written as:
Figure 122044DEST_PATH_IMAGE012
wherein, the first and the second end of the pipe are connected with each other,
Figure 637207DEST_PATH_IMAGE013
is composed of
Figure 714885DEST_PATH_IMAGE014
The function of the function is that of the function,
Figure 622798DEST_PATH_IMAGE015
、
Figure 102321DEST_PATH_IMAGE016
are respectively nodes
Figure 788386DEST_PATH_IMAGE017
In the encoder
Figure 618939DEST_PATH_IMAGE018
Layer and the first
Figure 799384DEST_PATH_IMAGE019
The information of the layer(s) is,
Figure 133414DEST_PATH_IMAGE020
is a node
Figure 741113DEST_PATH_IMAGE017
The set of neighborhood nodes of (a) is,
Figure 308229DEST_PATH_IMAGE021
is a node
Figure 26786DEST_PATH_IMAGE017
And between the neighboring nodes
Figure 480901DEST_PATH_IMAGE022
Attention weight of (1);
Figure 259502DEST_PATH_IMAGE014
the calculation formula of the function is:
Figure 64647DEST_PATH_IMAGE023
then, the coding space is calculated
Figure 836162DEST_PATH_IMAGE010
Obtaining the inner product between the inner node pairs to obtain the reconstructed original network
Figure 144784DEST_PATH_IMAGE024
(ii) a At this time, the error of the graph self-encoder reconstructing the network is represented as:
Figure 359865DEST_PATH_IMAGE025
step 2.4, constructing a self-training clustering network, and assuming that a clustering center is
Figure 121147DEST_PATH_IMAGE026
Then node
Figure 899616DEST_PATH_IMAGE017
Belong to
Figure 328324DEST_PATH_IMAGE027
Probability of class
Figure 448726DEST_PATH_IMAGE028
Expressed as:
Figure 212152DEST_PATH_IMAGE029
wherein the content of the first and second substances,
Figure 76203DEST_PATH_IMAGE030
is as followskThe center of the class is classified into a class center,
Figure 93837DEST_PATH_IMAGE031
is a node
Figure 916300DEST_PATH_IMAGE017
And cluster center
Figure 652175DEST_PATH_IMAGE026
The Euclidean distance between the nodes and the similarity between the nodes are represented, in the clustering problem, each node and the corresponding clustering center are forced to tend to the minimum intra-class distance and the maximum inter-class distance, and the target distribution is realized
Figure 569184DEST_PATH_IMAGE032
Is defined as:
Figure 706904DEST_PATH_IMAGE033
at this time, the KL divergence is used to characterize the predicted cluster distribution
Figure 434689DEST_PATH_IMAGE028
And target distribution
Figure 657860DEST_PATH_IMAGE032
Error between, defining a self-trained loss function:
Figure 112981DEST_PATH_IMAGE034
the final loss function
Figure 105207DEST_PATH_IMAGE035
Write as:
Figure 269472DEST_PATH_IMAGE036
Figure 714360DEST_PATH_IMAGE037
in order to be the weight coefficient,
Figure 723905DEST_PATH_IMAGE038
reconstructing errors of the network for the image self-encoder by minimizing
Figure 351064DEST_PATH_IMAGE035
Training the model according to the time
Figure 155072DEST_PATH_IMAGE028
Judging node
Figure 352835DEST_PATH_IMAGE017
The category to which it most likely belongs;
step 3, dividing the city into subareas based on the clustering result in the step 2, and dividing the subareas of the ponding water level of the city;
step 4, taking the ground elevation information, longitude information, latitude information and time information of all the sites and the monitoring information of the ponding water level outside the target site as input variables, taking the ponding water level of the target site as output variables, and constructing an individualized hydrological conceptual model region by region, wherein the method comprises the following steps:
under the same subregion, taking ground elevation information, longitude information, latitude information and time information of all stations and ponding water level monitoring information outside a target station as input variables, taking the ponding water level of the target station as an output variable, and constructing an individualized hydrological conceptual model region by region based on a neural network, wherein the input variables can be expressed as follows:
Figure 166070DEST_PATH_IMAGE039
wherein
Figure 132889DEST_PATH_IMAGE040
Respectively represents the ponding water level monitoring information, the ground elevation information, the longitude information, the latitude information and the time information,
Figure 380504DEST_PATH_IMAGE002
representing the number of water accumulation sites in the sub-area, assuming each site has
Figure 65563DEST_PATH_IMAGE041
Data of individual accumulated water, all are
Figure 603860DEST_PATH_IMAGE042
The model is trained by each sample, the model effectively extracts the spatial distribution characteristics of the water level of the ponding water by adding the ground elevation information, the longitude information and the latitude information, the seasonal change and the daily change characteristics of the ponding water level are favorably captured by adding the time information, the fitting capacity of the model is enhanced, and the calculation formula of the neural network is as follows:
Figure 425186DEST_PATH_IMAGE043
Figure 836576DEST_PATH_IMAGE044
，
Figure 743352DEST_PATH_IMAGE045
，
Figure 882078DEST_PATH_IMAGE046
representing input layer, hidden layer and output layer vectors respectively,
Figure 823489DEST_PATH_IMAGE002
in order to hide the number of layers,
Figure 874622DEST_PATH_IMAGE047
and
Figure 534273DEST_PATH_IMAGE048
for each of the training parameters of the layers,
Figure 211111DEST_PATH_IMAGE049
representing the activation function, which is also a non-linear source of the neural network, the activation function to be adopted by the method is a ReLU function:
Figure 7029DEST_PATH_IMAGE050
；
and 5, based on the hydrological conceptual model constructed region by region in the step 4, utilizing the ponding water level monitoring information, the ground elevation information, the longitude information, the latitude information and the time information of the ponding monitoring stations in the sub-regions and the ground elevation information, the longitude information, the latitude information and the time information of the target positions to reversely derive the ponding water level of any position of each sub-region.
2. The method for fitting the water level of the urban ponding water based on the artificial intelligence refinement of the claim 1, wherein in the step 1, the calculation formula of the standardization processing is as follows:
Figure 25800DEST_PATH_IMAGE051
in the formula (I), the compound is shown in the specification,
Figure 907169DEST_PATH_IMAGE052
for the purpose of the normalized result, the results,
Figure 387697DEST_PATH_IMAGE053
to represent
Figure 303701DEST_PATH_IMAGE006
Is determined by the average value of (a) of (b),
Figure 227795DEST_PATH_IMAGE054
represent
Figure 596459DEST_PATH_IMAGE006
Standard deviation of (d); after all variables are standardized, hydrological feature vectors of all ponding sites are formed:
Figure 880679DEST_PATH_IMAGE055
in the formula (I), the compound is shown in the specification,
Figure 651189DEST_PATH_IMAGE056
representing the hydrological feature vector for the kth ponding site,
Figure 746184DEST_PATH_IMAGE040
and respectively representing the standardized ponding water level monitoring information, the ground elevation information, the longitude information, the latitude information and the time information, thereby forming a dimensionless hydrological feature database.
CN202210605679.1A 2022-05-31 2022-05-31 Refined urban ponding water level fitting method based on artificial intelligence Active CN114818984B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202210605679.1A CN114818984B (en) 2022-05-31 2022-05-31 Refined urban ponding water level fitting method based on artificial intelligence

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202210605679.1A CN114818984B (en) 2022-05-31 2022-05-31 Refined urban ponding water level fitting method based on artificial intelligence

Publications (2)

Publication Number Publication Date
CN114818984A CN114818984A (en) 2022-07-29
CN114818984B true CN114818984B (en) 2022-09-23

Family

ID=82519839

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202210605679.1A Active CN114818984B (en) 2022-05-31 2022-05-31 Refined urban ponding water level fitting method based on artificial intelligence

Country Status (1)

Country Link
CN (1) CN114818984B (en)

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116955964B (en) * 2023-09-19 2023-12-12 江苏省气象服务中心 Continuous pavement meteorological condition analysis and deduction method

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CA2133877A1 (en) * 1993-10-25 1995-04-26 Mark A. Malamud Quick info windows and audio information cursors
CN111949749A (en) * 2020-07-30 2020-11-17 中国科学技术大学 High-order graph convolution network-based air quality monitoring station position recommendation method
CN113222328A (en) * 2021-03-25 2021-08-06 中国科学技术大学先进技术研究院 Air quality monitoring equipment point arrangement and site selection method based on road section pollution similarity

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CA2133877A1 (en) * 1993-10-25 1995-04-26 Mark A. Malamud Quick info windows and audio information cursors
CN111949749A (en) * 2020-07-30 2020-11-17 中国科学技术大学 High-order graph convolution network-based air quality monitoring station position recommendation method
CN113222328A (en) * 2021-03-25 2021-08-06 中国科学技术大学先进技术研究院 Air quality monitoring equipment point arrangement and site selection method based on road section pollution similarity

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
区域回归法对无资料地区设计洪水的估算;周芬等;《水力发电》;20040712(第07期);第10页第2栏最后1段-第13页第2栏最后1段 *
周芬等.区域回归法对无资料地区设计洪水的估算.《水力发电》.2004,(第07期),第10页第2栏最后1段-第13页第2栏最后1段. *
图神经网络时代的深度聚类;Houye;《https://zhuanlan.zhihu.com/p/114452245》;20200330;第2页第3段-第3页最后1段 *

Also Published As

Publication number Publication date
CN114818984A (en) 2022-07-29

Similar Documents

Publication Publication Date Title
CN111223301B (en) Traffic flow prediction method based on graph attention convolution network
Shirmohammadi et al. Application of several data-driven techniques for predicting groundwater level
Shirmohammadi et al. Forecasting of meteorological drought using Wavelet-ANFIS hybrid model for different time steps (case study: southeastern part of east Azerbaijan province, Iran)
CN106650767B (en) Flood forecasting method based on cluster analysis and real-time correction
Sehgal et al. Wavelet bootstrap multiple linear regression based hybrid modeling for daily river discharge forecasting
Elhenawy et al. Dynamic travel time prediction using data clustering and genetic programming
Shen et al. Research on traffic speed prediction by temporal clustering analysis and convolutional neural network with deformable kernels (May, 2018)
Sarkar et al. Modeling spatial determinants of urban expansion of Siliguri a metropolitan city of India using logistic regression
CN114818984B (en) Refined urban ponding water level fitting method based on artificial intelligence
Li et al. Deep spatio-temporal adaptive 3d convolutional neural networks for traffic flow prediction
Nunes Carvalho et al. Urban water demand modeling using machine learning techniques: Case study of Fortaleza, Brazil
Mohammad-Azari et al. State-of-art of genetic programming applications in water-resources systems analysis
CN115859801A (en) Landslide space-time risk assessment method combined with effective rainfall model
Zhang et al. Social media meets big urban data: A case study of urban waterlogging analysis
CN112561132A (en) Water flow prediction model based on neural network
Rubio et al. Adaptive non-parametric identification of dense areas using cell phone records for urban analysis
Xu et al. Traffic state data imputation: An efficient generating method based on the graph aggregator
Pal et al. Modelling water richness in riparian flood plain wetland using bivariate statistics and machine learning algorithms and figuring out the role of damming
CN109829583B (en) Mountain fire risk prediction method based on probability programming technology
Yang et al. Dual temporal gated multi-graph convolution network for taxi demand prediction
Jiang et al. Predicting citywide crowd dynamics at big events: A deep learning system
Zhang et al. Off-deployment traffic estimation—a traffic generative adversarial networks approach
WO2023155546A1 (en) Structure data generation method and apparatus, device, medium, and program product
CN107729941B (en) Air pollutant concentration estimation method based on deep terrain features
CN115204304A (en) Strip mine road network generation method based on two-way graph convolutional 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