CN117875190A - Space-time prediction method for traffic network function state in extreme storm event - Google Patents
Space-time prediction method for traffic network function state in extreme storm event Download PDFInfo
- Publication number
- CN117875190A CN117875190A CN202410264225.1A CN202410264225A CN117875190A CN 117875190 A CN117875190 A CN 117875190A CN 202410264225 A CN202410264225 A CN 202410264225A CN 117875190 A CN117875190 A CN 117875190A
- Authority
- CN
- China
- Prior art keywords
- time
- traffic network
- data
- rainfall
- network function
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 238000000034 method Methods 0.000 title claims abstract description 101
- 238000004088 simulation Methods 0.000 claims abstract description 23
- 238000012549 training Methods 0.000 claims abstract description 20
- 230000007246 mechanism Effects 0.000 claims abstract description 12
- 238000012360 testing method Methods 0.000 claims abstract description 10
- 238000007781 pre-processing Methods 0.000 claims abstract description 7
- 238000007499 fusion processing Methods 0.000 claims abstract description 5
- 239000011159 matrix material Substances 0.000 claims description 38
- 239000003795 chemical substances by application Substances 0.000 claims description 21
- 238000010586 diagram Methods 0.000 claims description 19
- 238000012545 processing Methods 0.000 claims description 16
- 238000011176 pooling Methods 0.000 claims description 15
- 238000004364 calculation method Methods 0.000 claims description 7
- 238000005516 engineering process Methods 0.000 claims description 7
- 238000010230 functional analysis Methods 0.000 claims description 7
- 238000012614 Monte-Carlo sampling Methods 0.000 claims description 5
- 238000004458 analytical method Methods 0.000 claims description 5
- 238000002790 cross-validation Methods 0.000 claims description 5
- 238000013528 artificial neural network Methods 0.000 claims description 4
- 230000008859 change Effects 0.000 claims description 4
- 238000013461 design Methods 0.000 claims description 4
- 238000013507 mapping Methods 0.000 claims description 3
- XLYOFNOQVPJJNP-UHFFFAOYSA-N water Substances O XLYOFNOQVPJJNP-UHFFFAOYSA-N 0.000 claims description 3
- 230000004927 fusion Effects 0.000 claims description 2
- 238000001556 precipitation Methods 0.000 claims 1
- 230000006870 function Effects 0.000 abstract description 70
- 230000001186 cumulative effect Effects 0.000 description 6
- 230000008569 process Effects 0.000 description 6
- 230000004044 response Effects 0.000 description 6
- 230000002123 temporal effect Effects 0.000 description 5
- 230000000694 effects Effects 0.000 description 4
- 238000011160 research Methods 0.000 description 4
- 230000007423 decrease Effects 0.000 description 3
- 238000011156 evaluation Methods 0.000 description 3
- 230000003993 interaction Effects 0.000 description 3
- 230000009471 action Effects 0.000 description 2
- 230000004913 activation Effects 0.000 description 2
- 239000002131 composite material Substances 0.000 description 2
- 238000013135 deep learning Methods 0.000 description 2
- 230000007547 defect Effects 0.000 description 2
- 238000012417 linear regression Methods 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 238000012935 Averaging Methods 0.000 description 1
- 238000006424 Flood reaction Methods 0.000 description 1
- 238000009825 accumulation Methods 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000001413 cellular effect Effects 0.000 description 1
- 230000000295 complement effect Effects 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 230000001934 delay Effects 0.000 description 1
- 230000003111 delayed effect Effects 0.000 description 1
- 230000007613 environmental effect Effects 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 230000004907 flux Effects 0.000 description 1
- 238000007654 immersion Methods 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 238000007726 management method Methods 0.000 description 1
- 238000003012 network analysis Methods 0.000 description 1
- 230000037361 pathway Effects 0.000 description 1
- 238000013349 risk mitigation Methods 0.000 description 1
- 238000005070 sampling Methods 0.000 description 1
- 238000012731 temporal analysis Methods 0.000 description 1
- 238000000700 time series analysis Methods 0.000 description 1
Classifications
-
- 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
Landscapes
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a space-time prediction method of traffic network function states in extreme storm events, which comprises the following steps: generating a physical simulation data set and preprocessing; a data driving agent model is built based on the TA-STGCN, a physical simulation data set is randomly divided into a training set and a testing set, and prediction of future traffic is completed by fusion processing of input data and capturing of disaster characteristics and space-time characteristics of the input data. The method uses TA-STGCN to capture the space-time characteristic of the dynamic network function state; by establishing a multi-mode input mechanism, the two-state input of rainfall-related data and traffic network data is realized, so that the space-time influence of extreme storm events on traffic network functions is captured; based on a self-attention mechanism and space-time convolution operation, stable multi-step traffic function prediction under the condition of non-stationary rainfall is realized.
Description
Technical Field
The invention relates to the technical field of traffic network function evaluation, in particular to a space-time prediction method of traffic network function states in extreme storm events.
Background
Extreme stormwater events can lead to severe floods or urban inland inundations and have a significant impact on the service function of the traffic network, including at least reduced driving speeds, increased accident risk and even road damage. Under the influence of extreme storms, emergency response actions in a disaster can be severely delayed. For example, a decrease in traffic network functionality would result in significant delays in search and rescue actions, and emergency supplies and medical personnel may not be reachable in time. Predicting functional status of a traffic network during extreme stormwater events can provide valuable support for real-time risk mitigation decisions to minimize potential loss of life and property. However, there is considerable uncertainty in the weather forecast of the current extreme storms. Thus, throughout the emergency response process, it is necessary for the decision maker to re-evaluate the spatio-temporal dynamics of the traffic network functional status with the latest meteorological data and modify the in-disaster emergency response plan accordingly. This presents a significant challenge to the computational speed of the traffic network function prediction model.
In the current research, there are two main technical frameworks for predicting the function of traffic networks during extreme storm events, a prediction method based on physical simulation and a prediction method based on data, respectively. Prediction methods based on physical simulation typically utilize a physical-based model to build and solve a set of mathematical equations describing disaster evolution, including rainfall runoff, flood inundation, and then perform traffic network function analysis based on the simulation results. While these methods are likely to provide highly accurate predictions, they are computationally demanding due to the complexity of flood and waterlogging simulations. Therefore, the prediction method based on physical simulation is generally suitable for a pre-construction or pre-disaster planning stage, but is not suitable for regional decision making of real-time disaster risk relief. With advances in intelligent computing technology, some research has developed data-based predictive methods for studying changes in traffic network functions during extreme weather events. The data-based prediction method provides faster computation speed and requires less computation resources than the physical simulation-based prediction method. However, due to data source limitations, these studies have focused mainly on functional assessment based on user service data, and cannot predict the inherent functions of the traffic network, including whether roads are destroyed or submerged by flooding. Thus, due to unpredictable fluctuations in daily traffic demand, these methods cannot directly support emergency response planning during disaster events, let alone when many extreme stormwater events occur during periods of insufficient traffic data, such as nights or early morning. Some studies have utilized conventional rainfall events as a complement to extreme stormwater events to augment their training data set, however, this approach weakens the non-stationary nature of the extreme stormwater sequence to some extent, thus resulting in poor prediction accuracy of the model.
Data driving methods based on real data are limited to a large number of high-quality historical disaster data sets and cannot be widely popularized. However, the data-driven proxy model is developed based on the data set generated by the physical simulation, so that the defect of the physical model in terms of calculation speed and the dilemma of the actual data-driven model in terms of insufficient training data can be effectively overcome. In recent years, various studies have employed machine learning techniques to construct data-driven proxy models for rainfall flood prediction, or developed proxy models for traffic network analysis. However, none of these models establishes a direct spatiotemporal correlation between rainfall forecasts and traffic network functional status to support real-time emergency response planning.
Notably, developing an end-to-end traffic network function prediction model for extreme stormwater events faces three challenges, as it is a typical multivariable spatiotemporal dynamic prediction task. First, such data-based proxy models must capture the dynamic effects of extreme storms on traffic network functions, which represent complex interactions of multiple physical processes, as compared to mature traffic prediction studies. Second, in the time dimension, extreme stormwater sequences often exhibit significant non-stationary characteristics, with different temporal characteristics between different extreme stormwater events, which makes it difficult for conventional time series analysis methods to model them effectively. Third, in the spatial dimension, the functional state of the traffic network during extreme stormwater events presents significant non-local characteristics, which are not only related to euclidean distance, influenced by geography, weather, hydrologic factors, traffic network topology, flood control infrastructure, etc., and traditional data driven methods have difficulty in accurately understanding such complex spatio-temporal relationships, even in the recent field of deep learning research, a significant challenge.
Disclosure of Invention
The invention provides a space-time prediction method for traffic network function states in extreme storm events in order to overcome the defects of the technology. The method uses a time-space diagram convolutional network (TA-STGCN) based on time attention to capture the space-time characteristics of the dynamic network function state; by establishing a multi-mode input mechanism, the two-state input of rainfall-related data and traffic network data is realized, so that the space-time influence of extreme storm events on traffic network functions is captured; based on a self-attention mechanism and space-time convolution operation, stable multi-step traffic function prediction under the condition of non-stationary rainfall is realized.
Term interpretation:
1. IPW: edge-independent pathway, independent paths, do not share paths between pairs of road nodes at any common edge (i.e., road). The independent paths can represent network redundancy and are the most important topological features of the traffic network.
2. TA-STGCN: temporal Attention-Spatiotemporal Graph Convolutional Network, a time space diagram of time attention convolves the network.
3. LISFLOOD-FP: a grid-based flood inundation model designed for research purposes at the university of bristol.
4. Info Works ICM: info Works Integrated Catchment Management, city comprehensive drainage model business software developed by Sekerman corporation.
5. MIKE flood: flood simulation tools developed by danish hydraulic engineering and environmental simulation company (DHI) include complete one-dimensional and two-dimensional flood simulation engines and the like.
6. K-fold cross validation method: the training data D is divided into K parts, the model is trained by the K-1 data, the rest 1 data are used for evaluating the quality of the model, the process is sequentially circulated on the K parts of data, and the obtained K evaluation results are combined, such as averaging or voting.
The technical scheme adopted for overcoming the technical problems is as follows:
a space-time prediction method for traffic network function states in extreme storm events comprises the following steps:
s1, generating a physical simulation data set:
firstly, generating a rainfall sequence aiming at a target area, then simulating an urban flood inundation scene, carrying out functional analysis on a traffic network to obtain a traffic network functional state, and respectively carrying out pretreatment on a space dimension and a time dimension on an original data set formed by the rainfall sequence and the traffic network functional state;
s2, constructing a data driving agent model and training:
constructing a data driving agent model based on the TA-STGCN, wherein the data driving agent model at least comprises an input data processing module, a disaster feature capturing module, a space-time feature capturing module and a future traffic prediction module; randomly dividing a physical simulation data set into a training set and a testing set, and training network parameters of a data driving agent model by using a K-fold cross validation method; and the future traffic prediction is completed by carrying out fusion processing on the input data and capturing disaster characteristics and space-time characteristics of the input data.
Further, in step S1, the generating a rainfall sequence specifically includes:
the method comprises the steps of generating a historical record of a design rain type and an extreme storm event based on a target area by adopting a Monte Carlo sampling methodBar resolution 1 hour, duration +.>An hourly rainfall sequence, wherein +.>，/>And->And->The rainfall is positive integer, the peak rainfall of each rainfall sequence is 20 mm-80 mm, and the accumulated rainfall is 45.5-652.5 mm。
Further, in step S1, the model simulation of the urban flood inundation scene specifically includes:
simulating a flood inundation scene of the city by a hydrological hydrodynamic analysis method based on the geographic and hydrological conditions of the target area;
after obtaining the flood inundation scene of the city, mapping the flood inundation scene of the city onto a traffic network by utilizing ArcGIS software to extract the road inundation state, and evaluating the influence of the city flood on the traffic network function, wherein the influence at least comprises the road passing capability and the maximum safe driving speed, so that the service level of all roads in the traffic network under the extreme heavy rain scene is obtained.
Further, in step S1, the performing functional analysis on the traffic network specifically includes:
the traffic accessibility index based on IPW evaluates the functional state of the traffic network, and the traffic network is set as a directed graph Wherein->Is a node set representing road nodes, +.>Representing roads forming a traffic network as an edge set; />And->Are all positive integers, & gt>The value of (2) is determined by the actual edges and nodes in the traffic network;
is provided withRepresenting road node->And (2) road node->The total number of IPWs is set to->Representing road node->And (2) road node->The%>IPWs each of which is a series of ordered edges, wherein +.>And->The method comprises the steps of carrying out a first treatment on the surface of the Therefore, the road node->At time->Is defined as the road node +.>Mean value of reliability of all IPWs +.>The method comprises the following steps:
（1）
in the formula (1),for road node->And (2) road node->Between (1)>Bar IPW at time->Is used as a weight factor of (1),for road node->And (2) road node->Between (1)>Bar IPW at time->The formula is:（2）
（3）
in the formula (2) and the formula (3),and->The number representing the road and the corresponding time of the road are respectively +.>Is a service level of (2); />For road->In kilometers.
Further, in step S1, the preprocessing of the spatial dimension of the original data set composed of the rainfall sequence and the traffic network function state specifically includes:
the rainfall sequence is structured grid data obtained by a meteorological observation and Monte Carlo sampling method, the traffic network functional state is unstructured graph data consistent with network topology, namely, the rainfall sequence is different from the spatial characteristics of two types of data of the traffic network functional state, a Thiessen polygon method is adopted to divide a target area, the average rainfall of polygons where each road node is positioned is calculated, and the rainfall sequence is converted from the structured grid data into graph data with the same topological structure as the traffic network functional state, which is that Wherein->For the rainfall at time 1, +.>For time->Is to (1) rainfall, 18)>For time->Is to (1) rainfall, 18)>，/>Represents a set of real numbers,representing matrix->Is composed of->Go->A matrix of column real numbers;
and (3) withThe corresponding traffic network function state in the time dimension is thatWherein->For the traffic network functional state at time 1, +.>For time->Traffic network function status of->For time->Traffic network function status of (c).
Further, in step S1, the preprocessing of the time dimension of the original data set composed of the rainfall sequence and the traffic network function state specifically includes:
usingTime step window +.>And->Processing and movingStep size of->Hour, wherein->Representing historical data time step,/->Representing the predicted time step, +.>Is a positive integer; obtainingGroup rainfall sequence data->Andgroup traffic network functional status->；
Splitting the traffic network function state to obtain the pastTime-step traffic network function statusAnd future->Time-step traffic network function statusThe method comprises the steps of carrying out a first treatment on the surface of the Thus (S)>And->Input constituting a data driven proxy modelData set->Is a standard label for the output variables of the data driven proxy model.
Further, in step S2,
the input data processing module is used for processing input data so as to realize consistency of time and characteristic channel dimensions;
the disaster feature capturing module is used for extracting a nonlinear relation between a rainfall sequence and a traffic network functional state and capturing time information features of the rainfall sequence and the traffic network functional state;
the space-time characteristic capturing module captures space-time characteristics of the traffic network function state change in the extreme storm event by adopting a space-time diagram convolution method;
the predictive future traffic module decodes the information using the fully connected neural network and predicts a future traffic network functional status.
Further, in step S2, the processing of the input data specifically includes:
the input to the data driven proxy model includes a rainfall sequenceAnd pass->Time-step traffic network function status->；
In the pastTime-step traffic network function status->Is increased by a time dimension with the same structureFilling tensors of time steps to preserve the time dimension of both rainfall sequences and traffic network functional statesKeeping consistent; all elements in the fill tensor are 0, which acts as placeholders so that the two input variables have the same length in the time dimension;
Then, the two input tensors of rainfall sequence and traffic network function state are operated by connection on characteristic channel dimension ""operation is concatenated into a hidden tensor->The following are provided:
（4）
in the formula (4) of the present invention,for the hidden variable corresponding to time 1, +.>For time->Corresponding hidden variable,/->For time->Corresponding hidden variable,/->For time->Corresponding hidden variables.
Further, in step S2, capturing disaster characteristics of the input data specifically includes:
to hidden tensorsUsing inverse convolution kernelsExtracting channel characteristics to obtain an intermediate tensor with +.>A channel dimension; fusing the features of all channels using pooling operations to extract key features from the channel dimensions, each pooling operation reducing the channel size, and performing several rounds of pooling operations to obtain a hidden tensor ++1 for the number of channels>；
Then, a self-attention mechanism is introduced to find out the influence on the futureInformation of the traffic network function status of the time step and assigning different weights to them; the self-attention mechanism uses dot product operation to calculate the attention distribution matrix, specifically as follows:
first, three weight matrices are set、/>And->Hide tensor->Multiplying the weight matrix ∈ - >Sum weight matrixObtaining query tensors respectively>And Key tensor->By +.>And Key tensor->Dot product is performed divided by +.>Obtaining an attention distribution matrix of information at different moments in the whole time sequence and normalizing attention coefficients by using a softmax function, namely: />（5）
In the formula (5) of the present invention,for query tensor->And Key tensor->Dimension of->Transpose symbols representing the matrix;
then, the normalized attention distribution matrix is combined with a value tensorCombining to get hidden tensor with time weight information>Wherein the value tensor->To hide tensor->And weight matrix->Is the product of (1), weight matrix->、/>And->Parameters to be learned.
Further, in step S2, capturing the spatiotemporal features of the input data specifically includes:
firstly, a time gating convolution technology is used for applying one-dimensional convolution along a time dimension, and time dimension characteristics are extracted to obtain a hidden tensor；
Secondly, capturing nonlinear characteristics through a linear gating unit to obtain a hidden tensor；
Finally, carrying out space diagram convolution operation by adopting a spectrogram method, extracting space dimension characteristics, and obtaining a hidden tensorThe method specifically comprises the following steps:
definition of the drawing volume symbol'"the graph convolution operation in the spatial dimension is defined as the input +. >And the picture volume kernel->Is expressed as a multiplication of:
（6）
in the formula (6) of the present invention,is an input tensor, which is the functional state of each side of the traffic network; />Is a normalized graph Laplace matrix, +.>The method comprises the steps of carrying out a first treatment on the surface of the Wherein (1)>Is an identity matrix>Is a weighted adjacency matrix, ">Is a degree matrix->The method comprises the steps of carrying out a first treatment on the surface of the Definition of Fourier basis->Is->Feature vector matrix, ">Is->Diagonal matrix of eigenvectors of (a) and thus satisfy +.>The method comprises the steps of carrying out a first treatment on the surface of the The graph convolution operation is further expressed as:
（7）
the computational complexity of equation (7) isThen adoptThe chebyshev approximation reduces the computational complexity, and the graph convolution operation is rewritten as follows:
（8）
in the formula (8), the expression "a",for vector set, ++>Is vector set->Element of (a)>，/>Is->The order chebyshev polynomial is approximately in the Laplace operator +.>The value at which the approximate Laplace matrix is +.>，/>Is->Maximum characteristic value; />Is the kernel size of the graph convolution, which determines the maximum radius of the convolution from the center node; by recursive calculation +.>Order graph convolution to reduce computational complexity to +.>。
The beneficial effects of the invention are as follows:
the invention develops an end-to-end data driving agent model based on a time-space diagram convolution network technology of time attention, and can directly predict the functional state of a traffic network from rainfall forecast.
1. The invention establishes the end-to-end data driving agent model based on the deep learning technology, and the model can rapidly and accurately predict the time-space change of the traffic network function in the extreme storm event, and greatly reduces the calculation resources and the operation time compared with the traditional physical model so as to support the requirement of real-time community emergency response planning and has the capability of assimilating real-time traffic information.
2. The invention establishes a multi-mode input mechanism, so that the data-driven proxy model can simultaneously input rainfall sequences and observation values of historical traffic network function states; in addition, a disaster feature capture module is created that provides a data driven proxy model with the ability to learn nonlinear interactions between two inputs.
3. The invention develops a time attention module based on a self-attention mechanism, so that a data-driven agent model can well capture the non-stationary characteristic of rainfall variation along with time in an extreme storm event, and realizes the multi-step traffic network function prediction stable under the non-stationary rainfall condition.
Drawings
Fig. 1 is a schematic diagram of a model constructed by a space-time prediction method of traffic network function states in an extreme heavy rain event according to an embodiment of the present invention.
Fig. 2 is a network structure diagram of a space-time prediction model of traffic network function states in an extreme storm event according to an embodiment of the invention.
Fig. 3 is a schematic diagram of an input data processing module according to an embodiment of the present invention.
Fig. 4 is a schematic diagram of a disaster feature capturing module according to an embodiment of the present invention.
Fig. 5 is a schematic diagram of a time attention module according to an embodiment of the present invention.
FIG. 6 is a schematic diagram of a spatio-temporal feature capture module according to an embodiment of the present invention.
Fig. 7 is an iterative prediction schematic diagram adopting a sliding window prediction method according to an embodiment of the present invention.
Fig. 8 is a regression diagram of model 1, model 4, model 6, model 7 according to an embodiment of the present invention.
FIG. 9 shows the average WAPE of each time step in a future 12-hour predicted time for model 1-model 5 according to an embodiment of the present invention.
Fig. 10 shows average WAPE for each time step in a future 12-hour predicted time for model 4, model 6, and model 7 according to an embodiment of the present invention.
FIG. 11 is a graph showing the error distribution of model 4 over a test dataset according to an embodiment of the present invention.
Detailed Description
The invention will now be described in further detail with reference to the drawings and the specific examples, which are given by way of illustration only and are not intended to limit the scope of the invention, in order to facilitate a better understanding of the invention to those skilled in the art.
The invention discloses a space-time prediction method of traffic network function states in extreme storm events, which comprises the following steps:
s1, generating a physical simulation data set:
firstly, generating a rainfall sequence aiming at a target area, then simulating an urban flood inundation scene, carrying out functional analysis on a traffic network to obtain a traffic network functional state, and respectively carrying out pretreatment on a space dimension and a time dimension on an original data set formed by the rainfall sequence and the traffic network functional state;
s2, constructing a data driving agent model and training:
constructing a data driving agent model based on the TA-STGCN, wherein the data driving agent model at least comprises an input data processing module, a disaster feature capturing module, a space-time feature capturing module and a future traffic prediction module; randomly dividing a physical simulation data set into a training set and a testing set, and training network parameters of a data driving agent model by using a K-fold cross validation method; and the future traffic prediction is completed by carrying out fusion processing on the input data and capturing disaster characteristics and space-time characteristics of the input data.
For a better understanding of the above-described technical solutions, exemplary embodiments of the present invention will be described in detail below with reference to the accompanying drawings, but are merely exemplary embodiments of the present invention, however, it should be understood that the present invention may be implemented in various forms and is not limited to the embodiments described herein. These embodiments are provided so that this disclosure will be thorough and complete.
The space-time prediction method for the traffic network function state in the extreme storm event according to the embodiment, as shown in fig. 1, comprises the following steps:
s1, generating a physical simulation data set:
for a target area, firstly, a rainfall sequence is generated, then, an urban flood inundation scene is simulated, and then, functional analysis is carried out on a traffic network to obtain a traffic network functional state, and the original data set consisting of the rainfall sequence and the traffic network functional state is respectively preprocessed in space dimension and time dimension.
In step S1 of the present embodiment, the generation of the rainfall sequence specifically includes:
the method comprises the steps of generating a historical record of a design rain type and an extreme storm event based on a target area by adopting a Monte Carlo sampling methodBar resolution 1 hour, duration +.>An hourly rainfall sequence, wherein +.>，/>The present embodiment is preferably->=10000，/>The peak rainfall per rainfall sequence is 20-80 mm, and the cumulative rainfall is 45.5-652.5 mm.
In step S1 of the present embodiment, the model simulation urban flood inundation scene specifically includes:
simulating a flood inundation scene of a city by a hydrologic hydrodynamic analysis method based on geographic and hydrologic conditions of a target area, wherein the hydrologic hydrodynamic analysis method can adopt a classical 2D flood model, such as LISFLOOD-FP or a flood model based on cellular automaton, and can also adopt a mature business model, such as Info Works ICM or MIKE flood;
After obtaining the flood inundation scene of the city, mapping the flood inundation scene of the city onto a traffic network by utilizing ArcGIS software to extract the road inundation state, and evaluating the influence of the city flood on the traffic network function, wherein the influence of the city flood on the traffic network function at least comprises the road passing capability and the maximum safe driving speed, so that the service level of all roads in the traffic network under the extreme storm scene is obtained.
In step S1 of the present embodiment, performing functional analysis on the traffic network specifically includes:
the traffic accessibility index based on IPW evaluates the functional state of the traffic network, and the traffic network is set as a directed graphWherein->Is a set of nodes representing road nodes, such as residential areas, economic hinges, major traffic road intersections, etc.; />Representing roads forming a traffic network as an edge set; />And->Are all positive integers, & gt>The value of (2) is determined by the actual edges and nodes in the traffic network;
is provided withRepresenting road node->And (2) road node->The total number of IPWs is set to->Representing road node->And (2) road node->The%>IPWs each of which is a series of ordered edges, wherein +.>And->The method comprises the steps of carrying out a first treatment on the surface of the Therefore, the road node- >At time->Is defined as the road node +.>Mean value of reliability of all IPWs +.>The method comprises the following steps:
（1）
in the formula (1),for road node->And (2) road node->Between (1)>Bar IPW at time->Is used as a weight factor of (1),for road node->And (2) road node->Between (1)>Bar IPW at time->The formula is:
（2）
（3）
formula [ (formula ]2) And in the formula (3) of the above,and->The number representing the road and the corresponding time of the road are respectively +.>Is a service level of (2); />For road->In kilometers.
After the original data set consisting of the rainfall sequence and the traffic network function state is obtained, the original data set needs to be preprocessed so as to be used in the data driving agent model, and therefore, the original data set needs to be preprocessed in a space dimension and a time dimension respectively.
In step S1 of the present embodiment, the preprocessing of the spatial dimension of the original data set composed of the rainfall sequence and the traffic network function state specifically includes:
the rainfall sequence is structured grid data obtained by a meteorological observation and Monte Carlo sampling method, and the traffic network functional state is unstructured graph data consistent with network topology, namely, the rainfall sequence is different from the spatial characteristics of the traffic network functional state. In order to facilitate the learning of a data-driven proxy model from training data to capture the non-local spatial characteristics of traffic network functions, the invention adopts a Thiessen polygon method to divide a target area, calculates the average rainfall of polygons where each road node is located, and converts a rainfall sequence from structured grid data into graphic data with the same topological structure as the traffic network function state, which is that Wherein->For the rainfall at time 1, +.>For the rainfall at time 2, +.>For the rainfall at time 24 +.>Represents a real set,/->Representing matrix->Is composed of->A matrix of row 24 columns of real numbers;
and (3) withThe corresponding traffic network function state in the time dimension is +.>Wherein, the method comprises the steps of, wherein,for the traffic network functional state at time 1, +.>For the traffic network functional state at time 2, +.>Is the traffic network function status at time 24.
In step S1 of the present embodiment, the preprocessing of the time dimension of the original data set composed of the rainfall sequence and the traffic network function state specifically includes:
usingTime step window +.>And->Processing, setting the moving step length as +.>Hour, wherein->Representing historical data time step,/->Representing the predicted time step, +.>Is a positive integer; obtainingGroup rainfall sequence data->Andgroup traffic network functional status->；
Splitting the traffic network function state to obtain the pastTime-step traffic network function statusAnd future->Time-step traffic network function statusThe method comprises the steps of carrying out a first treatment on the surface of the Thus (S)>And->Input dataset constituting data driven proxy model, < ->Is a standard label for the output variables of the data driven proxy model.
S2, constructing a data driving agent model and training:
constructing a data driving agent model based on a TA-STGCN, wherein the network structure of the data driving agent model is shown in figure 2, and the data driving agent model at least comprises an input data processing module, a disaster feature capturing module, a space-time feature capturing module and a future traffic prediction module; randomly dividing a physical simulation data set into a training set and a testing set, training network parameters of a data driving agent model by using a K-fold cross validation method, and solidifying the trained parameters in a neural network; and the future traffic prediction is completed by carrying out fusion processing on the input data and capturing disaster characteristics and space-time characteristics of the input data.
In this embodiment, the input data processing module is configured to process input data to achieve consistency of time and feature channel dimensions;
the disaster feature capturing module is used for extracting a nonlinear relation between a rainfall sequence and a traffic network functional state and capturing time information features of the rainfall sequence and the traffic network functional state;
the space-time characteristic capturing module captures space-time characteristics of the traffic network function state change in the extreme storm event by adopting a space-time diagram convolution method;
the predictive future traffic module decodes the information using the fully connected neural network and predicts a future traffic network functional status.
In step S2 of the present embodiment, the processing of the input data specifically includes:
the input to the data driven proxy model includes a rainfall sequenceAnd pass->Time-step traffic network function status->If you will go->Rainfall recording and future +.>The rainfall forecast of the time step is connected in series, and the time dimension is longer than the observed value of the traffic network function state;
thus, in the pastTime-step traffic network function status->Is increased by a time dimension of +.>A filling tensor of the time step, which enables the time dimension of the rainfall sequence and the traffic network function state to be consistent, as shown in figure 3; all elements in the fill tensor are 0, which acts as placeholders so that the two input variables have the same length in the time dimension;
then, the two input tensors of rainfall sequence and traffic network function state are operated by connection on characteristic channel dimension ""operation is concatenated into a hidden tensor->The following are provided:
（4）
in the formula (4) of the present invention,for the hidden variable corresponding to time 1, +.>For time->Corresponding hidden variable,/->For time->Corresponding hidden variable,/->For time->Corresponding hidden variables.
In step S2 of the present embodiment, capturing disaster characteristics of input data specifically includes:
hidden tensorThe system consists of two channels of rainfall and a traffic network function, wherein the influence of the rainfall on the traffic network function shows a complex nonlinear relation, and the complex nonlinear relation comprises various physical processes such as the interaction of the rainfall and flood, the influence of the flood on the traffic network and the like. In order to more effectively capture the correlation between the two characteristics of rainfall and traffic network function, the disaster characteristic capturing module adopts two steps of up-sampling and fusion. Specifically, the present invention fills in the extended channel dimension by "0" before capturing disaster features of input data, and then +_hidden tensor +.>Extraction of flux using inverse convolution kernelThe trace features, resulting in an intermediate tensor with +.>The channel dimensions, as shown in FIG. 4, have +.>Hidden tensor of individual channel dimension>The expression is as follows:
the 1 st channel dimension is expressed as:，
the 2 nd channel dimension is expressed as:，
......
first, theThe individual channel dimensions are expressed as: />，
The pooling operation is then used to fuse the features of all channels to extract key features from the channel dimensions, and the pooling operation in the present invention may be implemented by various methods, for example, a maximum pooling method or an average pooling method may be selected The regular pooling method or the random pooling method is preferably an average pooling method, each pooling operation reduces the channel size, and a hidden tensor ++ ∈1 is obtained through a plurality of rounds of pooling operations>Hidden tensor->Expressed as:
then, in order to focus the attention of the data driven proxy model on important time steps within the global viewpoint, the present invention introduces a self-attention mechanism, i.e. by the time attention module finding information of the traffic network function status affecting future time steps and assigning them different weights; the self-attention mechanism calculates the attention distribution matrix using dot product operation, as shown in fig. 5, specifically as follows:
first, three weight matrices are set、/>And->Hide tensor->Multiplying the weight matrix ∈ ->Sum weight matrixObtaining query tensors respectively>And Key tensor->By +.>And Key tensor->Dot product is performed divided by +.>Obtaining the wholeThe attention distribution matrix of the information at different moments in the time series is normalized with the softmax function, namely: />（5）
In the formula (5) of the present invention,for query tensor->And Key tensor->Dimension of->Transpose symbols representing the matrix;
Then, the normalized attention distribution matrix is combined with a value tensorCombining to get hidden tensor with time weight information>Hidden tensor->Expressed as: />Wherein the value tensor->To hide tensor->And weight matrix->Is the product of (1), weight matrix->、/>And->Parameters to be learned.
In step S2 of the present embodiment, capturing the spatio-temporal features of the input data specifically includes:
in this embodiment, a space-time diagram convolution module is used to extract the space-time characteristics of the hidden tensor, as shown in fig. 6. Firstly, a time gating convolution technology is used for applying one-dimensional convolution along a time dimension, and time dimension characteristics are extracted to obtain a hidden tensorHidden tensor->Expressed as: />Wherein->Represented as hidden tensorsMiddle time->Hidden variable of->Is the size of the time convolution kernel; secondly, capturing nonlinear characteristics through a linear gating unit to obtain hidden tensor +.>Hidden tensor->Denoted as->Wherein->Expressed as hidden tensor +.>Middle time->Hidden variable of->Is the size of the time convolution kernel; finally, carrying out space diagram convolution operation by adopting a spectrogram method, extracting space dimension characteristics, and obtaining hidden tensor ++>Hidden tensor->Expressed as: />Wherein->Expressed as hidden tensor +. >Middle time->Hidden variable of->For the size of the time convolution kernel, the space graph convolution by adopting a spectrogram method is specifically as follows: definition of the drawing volume symbol'"the graph convolution operation in the spatial dimension is defined as the input +.>And the picture volume kernel->Is expressed as a multiplication of:
（6）
in the formula (6) of the present invention,is an input tensor, which is the functional state of each side of the traffic network; />Is a normalized graph Laplace matrix, +.>The method comprises the steps of carrying out a first treatment on the surface of the Wherein (1)>Is an identity matrix>Is a weighted adjacency matrix, ">Is a degree matrix->The method comprises the steps of carrying out a first treatment on the surface of the Definition of Fourier basis->Is->Feature vector matrix, ">Is->Diagonal matrix of eigenvectors of (a) and thus satisfy +.>The method comprises the steps of carrying out a first treatment on the surface of the Picture scrollThe product operation is further expressed as:
（7）
the computational complexity of equation (7) isAnd then, reducing the calculation complexity by using a Chebyshev approximation method, and rewriting the graph convolution operation as follows:
（8）
in the formula (8), the expression "a",for vector set, ++>Is vector set->Element of (a)>，/>Is->The order chebyshev polynomial is approximately in the Laplace operator +.>The value at which the approximate Laplace matrix is +.>，/>Is->Maximum characteristic value; />Is the kernel size of the graph convolution, which determines the maximum radius of the convolution from the center node; by recursive calculation +.>Order graph convolution to reduce computational complexity to +. >。
Further, the predicted future traffic module employs a fully connected network. Further preferably, the present embodiment uses a composite prediction strategy to complete prediction of future traffic, specifically uses a "sliding window prediction method" in the composite prediction strategy to complete prediction of future traffic, as shown in fig. 7, and the method includes dividing the whole prediction range into a plurality of equal-length intervals, preferably equal-length intervals areTime step and establish a +.>The time step output prediction model is then used to implement global prediction using an iterative prediction method, and a schematic diagram of two iterative predictions is shown in fig. 7. It is noted that increasing the number of iterations also leads to prediction errors, and therefore an appropriate single prediction time step +.>And the corresponding iteration times finally realize the prediction of the future multi-step traffic network function state.
The method for predicting the functional state of the traffic network in the extreme storm event is described by a specific case, and is specifically described by taking an eastern coastal region of China as an example. The area is a coastal area on a river alluvial plain, the area is 21.72 square kilometers, the river network is dense, the underground water is rich, and due to geographical position, the area is in a rainy season from 5 months to 7 months each year, and is in a typhoon season from 9 months later, so that serious water immersion often occurs. The traffic network in this area consists of 233 road nodes and 382 roads, with a total length of about 122 km.
The data set required for training is generated by the method described in step S1 of this embodiment, which is preferably the method of "generating a physical simulation data setTherefore, it is->And. And then the data-driven proxy model is built according to the method in the step S2 of building the data-driven proxy model and training.
To verify the validity of the data-driven proxy model, the present embodiment designs 7 models and parameters, as shown in Table 1, for single step prediction durationTime attention module selection (i.e. whether time attention module is turned on) and input data (i.e. whether +.>Hour rainfall as input information), wherein "O" in table 1 indicates activation or use of the time attention module or input data and "X" indicates non-activation or non-use of the time attention module or input data. To achieve a continuous prediction for the next 12 hours, an iterative prediction strategy as shown in fig. 7 is used. Model 1, model 2, model 3, model 4, model 5, model 6, model 7 require 12 iterations, 6 iterations, 4 iterations, 3 iterations, 2 iterations, 3 iterations, as shown in the last column of table 1. For each model, the present embodiment preferably randomly selects 70% from the dataset as the training set, with the remaining 30% as the test set. In this example, all experiments were performed in a Windows operating system (CPU: intel Core i7-10700 CPU @ 2.90GHz GPU : NVIDIA GeForce RTX 3060, ti).
TABLE 1 7 models designed and parameters for each model
To evaluate the performance of the different models on the test dataset, the present embodiment evaluates 7 models using Mean Absolute Error (MAE) and Weighted Absolute Percent Error (WAPE) as evaluation indicators. As shown in table 2, the single step 1 hour prediction error MAE of model 1 was minimal in 7 models, increasing from 1 hour of model 1 to 6 hours of model 5 (other mode settings unchanged) with increasing prediction duration, the model faced greater challenges in capturing the temporal characteristics of rainfall affecting the traffic network functional state process, and MAE increased from 0.011 to 0.015. Meanwhile, among the models 4 and 6, the model 4 with the time attention module has a better prediction effect than the model 6 without the time attention module for the same single step 4 hour prediction. Model 7 performs worst in all models because it does not use input data related to rainfall, which is the primary driving force in this task to affect changes in traffic network function status.
TABLE 2 MAE and WAPE for 7 models designed
Table 3 evaluates the cumulative MAE and WAPE predicted for the 7 models for the same fixed 12 hours. It can be readily seen that the performance ranking of the model up to the 12 hour prediction length is different from that of the model of the single step prediction, as shown in table 3. Model 1 requires 12 iterations to complete 12 hours of prediction, and compared with other models, error accumulation is obvious; on the other hand, model 4, which only needs to iterate 3 times, has the smallest cumulative error. Obviously, in one step, the duration is predicted There is a trade-off between overall length predicted duration, andand may be optimized for a particular predictive task to obtain optimal performance. Determining coefficient R of each model 2 As also shown in table 3, the regression charts of model 1, model 4, model 6, and model 7 are shown in fig. 8. The smaller the traffic network function, the higher the damage level. The larger the deviation of the linear regression fit line from the diagonal (prediction=target), the larger the prediction error, and in fig. 8, the solid line represents the linear regression fit line, that is, the diagonal. This illustrates an important feature of the model: the prediction error of the model increases as the traffic network function decreases.
TABLE 3 7 models designed cumulative MAE, cumulative WAPE and determinant R for the same fixed 12 hour prediction 2
To further study the predictive stability of the data driven proxy model, FIG. 9 plots the average WAPE for each time step for model 1-model 5 over a future 12 hour prediction time, and FIG. 10 plots the average WAPE for each time step for model 4, model 6, and model 7 over a future 12 hour prediction time. FIGS. 9 and 10 show the effect of single step prediction duration on the performance of a data driven proxy model 12 hours prediction task, model 1 @, as expected =1) the prediction error at the initial time is the smallest, but as the number of iterations increases, the error gradually accumulates; over 3 hours (i.e., 3 iterations), the cumulative prediction error exceeds all models except model 7; model 2, model 3 and model 5 generally exhibit similar trends. FIG. 10 shows the effect of the rainfall-related model input and the time attention module on the model prediction accuracy, model 7 performing worst due to the lack of significant rainfall input, as shown in tables 2 and 3; model 4 (with temporal attention module) always performs better than model 6 (without temporal attention module) in the predicted time. In fig. 8, it is worth noting that the prediction errors of all models start to decrease gradually at a later stage, becauseIn order to stop rainfall in all samples, the traffic network function is gradually restored, and thus, the prediction error is also gradually reduced.
Fig. 11 further analyzes the error distribution of model 4 over the test dataset. As is apparent from fig. 11, the prediction error (WAPE) at the 9 th hour in the future is the largest among all rainfall scenes of the test dataset. However, only 25% of the rainfall scenarios have errors greater than 3.3%. Even considering a rainfall scenario in the 1.5-fold median range (IQR), only 12.5% of the samples have a prediction error exceeding 8.3%. The median error and average error curves indicate that the prediction accuracy of model 4 is relatively stable over the prediction duration, which is an important quality of the prediction model.
The foregoing has described only the basic principles and preferred embodiments of the present invention, and many variations and modifications will be apparent to those skilled in the art in light of the above description, which variations and modifications are intended to be included within the scope of the present invention.
Claims (10)
1. A space-time prediction method for traffic network function states in extreme storm events is characterized by comprising the following steps:
s1, generating a physical simulation data set:
firstly, generating a rainfall sequence aiming at a target area, then simulating an urban flood inundation scene, carrying out functional analysis on a traffic network to obtain a traffic network functional state, and respectively carrying out pretreatment on a space dimension and a time dimension on an original data set formed by the rainfall sequence and the traffic network functional state;
s2, constructing a data driving agent model and training:
constructing a data driving agent model based on the TA-STGCN, wherein the data driving agent model at least comprises an input data processing module, a disaster feature capturing module, a space-time feature capturing module and a future traffic prediction module; randomly dividing a physical simulation data set into a training set and a testing set, and training network parameters of a data driving agent model by using a K-fold cross validation method; and the future traffic prediction is completed by carrying out fusion processing on the input data and capturing disaster characteristics and space-time characteristics of the input data.
2. The method of space-time prediction of traffic network function status in extreme stormwater events as claimed in claim 1, wherein in step S1, generating a rainfall sequence specifically comprises:
the method comprises the steps of generating a historical record of a design rain type and an extreme storm event based on a target area by adopting a Monte Carlo sampling methodBar resolution 1 hour, duration +.>An hourly rainfall sequence, wherein +.>，/>And->And->The rainfall peak value of each rainfall sequence is 20-80 mm, and the accumulated rainfall is 45.5-652.5 mm.
3. The method of space-time prediction of traffic network function status in extreme stormwater events as claimed in claim 1, wherein in step S1, modeling urban flood inundation scenarios specifically comprises:
simulating a flood inundation scene of the city by a hydrological hydrodynamic analysis method based on the geographic and hydrological conditions of the target area;
after obtaining the flood inundation scene of the city, mapping the flood inundation scene of the city onto a traffic network by utilizing ArcGIS software to extract the road inundation state, and evaluating the influence of the city flood on the traffic network function, wherein the influence at least comprises the road passing capability and the maximum safe driving speed, so that the service level of all roads in the traffic network under the extreme heavy rain scene is obtained.
4. The method for spatiotemporal prediction of traffic network functional status in extreme stormwater event according to claim 2, wherein in step S1, the functional analysis of the traffic network specifically comprises:
the traffic accessibility index based on IPW evaluates the functional state of the traffic network, and the traffic network is set as a directed graphWherein->Is a node set representing road nodes; />Representing roads forming a traffic network as an edge set; />And->Are all positive integers, & gt>The value of (2) is determined by the actual edges and nodes in the traffic network;
is provided withRepresenting road node->And (2) road node->The total number of IPWs is set to->Representing road node->And (2) road node->The%>IPWs each of which is a series of ordered edges, wherein +.>And->The method comprises the steps of carrying out a first treatment on the surface of the Therefore, the road node->At time->Is defined as the road node +.>Mean value of reliability of all IPWs +.>The method comprises the following steps:（1）
in the formula (1),for road node->And (2) road node->Between (1)>Bar IPW at time->Is used as a weight factor of (1),for road node->And (2) road node->Between (1)>Bar IPW at time->The formula is:
（2）
（3）
in the formula (2) and the formula (3),and->The number representing the road and the corresponding time of the road are respectively +. >Is a service level of (2); />For road->In kilometers.
5. The method for space-time prediction of traffic network function states in extreme stormwater events as claimed in claim 4, wherein in step S1, the preprocessing of the original data set comprising the rainfall sequence and the traffic network function states in spatial dimension specifically comprises:
the rainfall sequence is structured grid data, the traffic network functional state is unstructured graph data, namely, the spatial characteristics of the two types of data, namely, the rainfall sequence and the traffic network functional state, are different, a Thiessen polygon method is adopted to divide a target area, the average rainfall capacity of polygons where nodes of each road are positioned is calculated, and the rainfall sequence is converted from the structured grid data into graph data with the same topological structure as the traffic network functional state, so that the method is thatWherein->For the rainfall at time 1, +.>For time->Is to (1) rainfall, 18)>For time->Is used for the precipitation amount of the water,，/>represents a real set,/->Representing matrix->Is composed of->Go->A matrix of column real numbers;
and (3) withThe corresponding traffic network function state in the time dimension is thatWherein->For the traffic network functional state at time 1, +.>For time->Traffic network function status of- >For time->Traffic network function status of (c).
6. The method for space-time prediction of traffic network function states in extreme stormwater events according to claim 5, wherein in step S1, the preprocessing of the original data set comprising the rainfall sequence and the traffic network function states in time dimension specifically comprises:
usingTime step window +.>And->Processing, setting the moving step length as +.>Hour, wherein->Representing historical data time step,/->Representing the predicted time step, +.>Is a positive integer; obtainingGroup rainfall sequence data->Andgroup traffic network functional status->；
Splitting the traffic network function state to obtain the pastTime-step traffic network function statusAnd future->Time-step traffic network function statusThe method comprises the steps of carrying out a first treatment on the surface of the Thus (S)>And->Input dataset constituting data driven proxy model, < ->Is a standard label for the output variables of the data driven proxy model.
7. The method for spatiotemporal prediction of traffic network function status in extreme stormwater events as claimed in claim 1, wherein in step S2,
the input data processing module is used for processing input data so as to realize consistency of time and characteristic channel dimensions;
The disaster feature capturing module is used for extracting a nonlinear relation between a rainfall sequence and a traffic network functional state and capturing time information features of the rainfall sequence and the traffic network functional state;
the space-time characteristic capturing module captures space-time characteristics of the traffic network function state change in the extreme storm event by adopting a space-time diagram convolution method;
the predictive future traffic module decodes the information using the fully connected neural network and predicts a future traffic network functional status.
8. The method of spatiotemporal prediction of traffic network function status in extreme stormwater event as claimed in claim 6, wherein in step S2, the processing of the input data comprises:
the input to the data driven proxy model includes a rainfall sequenceAnd pass->Time-step traffic network function status->；
In the pastTime-step traffic network function status->Is increased by a time dimension of +.>Filling tensors of time steps to enable time dimensions of the rainfall sequence and the traffic network functional state to be consistent; all elements in the fill tensor are 0, which acts as placeholders so that the two input variables have the same length in the time dimension;
then, the two input tensors of rainfall sequence and traffic network function state are operated by connection on characteristic channel dimension " "operation is concatenated into a hidden tensor->The following are provided:
（4）
in the formula (4) of the present invention,for the hidden variable corresponding to time 1, +.>For time->Corresponding hidden variable,/->For the moment of timeCorresponding hidden variable,/->For time->Corresponding hidden variables.
9. The method of spatiotemporal prediction of traffic network functional status in extreme stormwater events as claimed in claim 8, wherein capturing disaster characteristics of the input data in step S2 comprises:
to hidden tensorsExtracting channel features using inverse convolution kernels to obtain an intermediate tensor with +.>A channel dimension; feature fusion of all channels using pooling operations to extract key features from channel dimensionsEach time the pooling operation reduces the channel size, a hidden tensor with the channel number of 1 is obtained through a plurality of rounds of pooling operations>；
Then, a self-attention mechanism is introduced to find out the influence on the futureInformation of the traffic network function status of the time step and assigning different weights to them; the self-attention mechanism uses dot product operation to calculate the attention distribution matrix, specifically as follows:
first, three weight matrices are set、/>And->Hide tensor->Multiplying the weight matrix ∈ ->Sum weight matrix->Obtaining query tensors respectively >And Key tensor->By +.>Sum key tensor/>Dot product is performed divided by +.>Obtaining an attention distribution matrix of information at different moments in the whole time sequence and normalizing attention coefficients by using a softmax function, namely:
（5）
in the formula (5) of the present invention,for query tensor->And Key tensor->Dimension of->Transpose symbols representing the matrix;
then, the normalized attention distribution matrix is combined with a value tensorCombining to get hidden tensor with time weight information>Wherein the value tensor->To hide tensor->And weight matrix->Is the product of (1), weight matrix->、/>And->Is the parameter to be learned.
10. The method of spatiotemporal prediction of traffic network functional status in extreme stormwater events as claimed in claim 9, wherein capturing spatiotemporal features of the input data in step S2 comprises:
firstly, a time gating convolution technology is used for applying one-dimensional convolution along a time dimension, and time dimension characteristics are extracted to obtain a hidden tensor；
Secondly, capturing nonlinear characteristics through a linear gating unit to obtain a hidden tensor；
Finally, carrying out space diagram convolution operation by adopting a spectrogram method, extracting space dimension characteristics, and obtaining a hidden tensorThe method specifically comprises the following steps:
definition of the drawing volume symbol' "the graph convolution operation in the spatial dimension is defined as the input +.>And the picture volume kernel->Is expressed as a multiplication of:
（6）
in the formula (6) of the present invention,is an input tensor, which is the functional state of each side of the traffic network; />Is a normalized graph Laplace matrix, +.>The method comprises the steps of carrying out a first treatment on the surface of the Wherein (1)>Is an identity matrix>Is a weighted adjacency matrix, ">Is a degree matrix->The method comprises the steps of carrying out a first treatment on the surface of the Definition of Fourier basis->Is->Feature vector matrix, ">Is->Diagonal matrix of eigenvectors of (a) and thus satisfy +.>The method comprises the steps of carrying out a first treatment on the surface of the The graph convolution operation is further expressed as:
（7）
the computational complexity of equation (7) isAnd then, reducing the calculation complexity by using a Chebyshev approximation method, and rewriting the graph convolution operation as follows:
（8）
in the formula (8), the expression "a",for vector set, ++>Is vector set->Element of (a)>，/>Is->The order chebyshev polynomial is approximately in the Laplace operator +.>Values at which the approximate Laplace matrix is/>，/>Is->Maximum characteristic value; />Is the kernel size of the graph convolution, which determines the maximum radius of the convolution from the center node; by recursive calculation +.>Order graph convolution to reduce computational complexity to +.>。
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202410264225.1A CN117875190B (en) | 2024-03-08 | Space-time prediction method for traffic network function state in extreme storm event |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202410264225.1A CN117875190B (en) | 2024-03-08 | Space-time prediction method for traffic network function state in extreme storm event |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN117875190A true CN117875190A (en) | 2024-04-12 |
| CN117875190B CN117875190B (en) | 2024-06-04 |
Family
ID=
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2019092146A1 (en) * | 2017-11-13 | 2019-05-16 | Suez Groupe | Device and method for processing heterogeneous data to determine inflows in time and space |
| WO2023215621A1 (en) * | 2022-05-05 | 2023-11-09 | InnerPlant, Inc. | System and method for modeling crop yield based on detection of plant stressors in crops |
| US20230375745A1 (en) * | 2022-05-17 | 2023-11-23 | Nanjing University Of Information Science & Technology | High-temperature disaster forecast method based on directed graph neural network |
| CN117494586A (en) * | 2023-12-29 | 2024-02-02 | 浙江大学 | Mountain torrent space-time prediction method based on deep learning |
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2019092146A1 (en) * | 2017-11-13 | 2019-05-16 | Suez Groupe | Device and method for processing heterogeneous data to determine inflows in time and space |
| WO2023215621A1 (en) * | 2022-05-05 | 2023-11-09 | InnerPlant, Inc. | System and method for modeling crop yield based on detection of plant stressors in crops |
| US20230375745A1 (en) * | 2022-05-17 | 2023-11-23 | Nanjing University Of Information Science & Technology | High-temperature disaster forecast method based on directed graph neural network |
| CN117494586A (en) * | 2023-12-29 | 2024-02-02 | 浙江大学 | Mountain torrent space-time prediction method based on deep learning |
Non-Patent Citations (1)
| Title |
|---|
| 杜圣东;李天瑞;杨燕;王浩;谢鹏;洪西进;: "一种基于序列到序列时空注意力学习的交通流预测模型", 计算机研究与发展, no. 08, 6 August 2020 (2020-08-06) * |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Pang et al. | Solar radiation prediction using recurrent neural network and artificial neural network: A case study with comparisons | |
| Yen et al. | Application of the deep learning for the prediction of rainfall in Southern Taiwan | |
| US20230128989A1 (en) | Systems and methods of data preprocessing and augmentation for neural network climate forecasting models | |
| Afan et al. | ANN based sediment prediction model utilizing different input scenarios | |
| Kisi et al. | River suspended sediment estimation by climatic variables implication: Comparative study among soft computing techniques | |
| Hung et al. | An artificial neural network model for rainfall forecasting in Bangkok, Thailand | |
| CN108009674A (en) | Air PM2.5 concentration prediction methods based on CNN and LSTM fused neural networks | |
| JP2019094640A (en) | Water level prediction method, water level prediction program and water level prediction device | |
| CN111126704A (en) | Multi-region precipitation prediction model construction method based on multi-graph convolution and memory network | |
| CN112149887A (en) | PM2.5 concentration prediction method based on data space-time characteristics | |
| CN115578852A (en) | Traffic prediction method based on DSTGCN | |
| CN114049545A (en) | Typhoon intensity determining method, system, equipment and medium based on point cloud voxel | |
| Nguyen et al. | Water level forecasting using neuro-fuzzy models with local learning | |
| Lin et al. | Applications of cluster analysis and pattern recognition for typhoon hourly rainfall forecast | |
| Zhang et al. | A multi-strategy-mode waterlogging-prediction framework for urban flood depth | |
| Wang et al. | Rapid prediction of urban flood based on disaster-breeding environment clustering and Bayesian optimized deep learning model in the coastal city | |
| Cobos et al. | MarineTools. temporal: A Python package to simulate Earth and environmental time series | |
| Chen et al. | Physics-guided meta-learning method in baseflow prediction over large regions | |
| CN117875190B (en) | Space-time prediction method for traffic network function state in extreme storm event | |
| CN116976702A (en) | Urban digital twin platform and method based on large-scene GIS lightweight engine | |
| CN117875190A (en) | Space-time prediction method for traffic network function state in extreme storm event | |
| Bosma et al. | Estimating solar and wind power production using computer vision deep learning techniques on weather maps | |
| Li et al. | Resilience Assessment of Urban Distribution Network Under Heavy Rain: A Knowledge-Informed Data-Driven Approach (April 2023) | |
| Kwon et al. | Assessing the impacts of dam/weir operation on streamflow predictions using LSTM across South Korea | |
| Paul et al. | Forecasting of reservoir inflow by the combination of deep learning and conventional machine learning |
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 |