CN117313551A - Radionuclide diffusion prediction method and system based on GAT-LSTM - Google Patents
Radionuclide diffusion prediction method and system based on GAT-LSTM Download PDFInfo
- Publication number
- CN117313551A CN117313551A CN202311595153.0A CN202311595153A CN117313551A CN 117313551 A CN117313551 A CN 117313551A CN 202311595153 A CN202311595153 A CN 202311595153A CN 117313551 A CN117313551 A CN 117313551A
- Authority
- CN
- China
- Prior art keywords
- radionuclide
- gat
- data
- graph
- time
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000009792 diffusion process Methods 0.000 title claims abstract description 99
- 238000000034 method Methods 0.000 title claims abstract description 58
- 239000013598 vector Substances 0.000 claims abstract description 8
- 238000012544 monitoring process Methods 0.000 claims description 41
- 238000012549 training Methods 0.000 claims description 26
- 230000008569 process Effects 0.000 claims description 20
- 238000012360 testing method Methods 0.000 claims description 13
- 238000000605 extraction Methods 0.000 claims description 10
- 238000011156 evaluation Methods 0.000 claims description 8
- 238000007781 pre-processing Methods 0.000 claims description 4
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 claims description 3
- 238000004364 calculation method Methods 0.000 claims description 3
- 230000003993 interaction Effects 0.000 claims description 2
- 230000005855 radiation Effects 0.000 abstract description 3
- 239000011159 matrix material Substances 0.000 description 14
- 230000008859 change Effects 0.000 description 9
- 238000004088 simulation Methods 0.000 description 9
- 238000000151 deposition Methods 0.000 description 8
- 230000008021 deposition Effects 0.000 description 8
- 238000013527 convolutional neural network Methods 0.000 description 6
- 230000000694 effects Effects 0.000 description 6
- 238000011160 research Methods 0.000 description 6
- 238000010586 diagram Methods 0.000 description 5
- 239000003344 environmental pollutant Substances 0.000 description 5
- 230000000007 visual effect Effects 0.000 description 5
- 238000013528 artificial neural network Methods 0.000 description 4
- 238000002474 experimental method Methods 0.000 description 4
- 239000012530 fluid Substances 0.000 description 4
- 230000006870 function Effects 0.000 description 4
- 231100000719 pollutant Toxicity 0.000 description 4
- 238000013135 deep learning Methods 0.000 description 3
- 230000002285 radioactive effect Effects 0.000 description 3
- 101001095088 Homo sapiens Melanoma antigen preferentially expressed in tumors Proteins 0.000 description 2
- 102100037020 Melanoma antigen preferentially expressed in tumors Human genes 0.000 description 2
- 230000008901 benefit Effects 0.000 description 2
- 230000021615 conjugation Effects 0.000 description 2
- 238000013136 deep learning model Methods 0.000 description 2
- 230000005484 gravity Effects 0.000 description 2
- 230000007246 mechanism Effects 0.000 description 2
- 230000015654 memory Effects 0.000 description 2
- 238000003058 natural language processing Methods 0.000 description 2
- 239000002245 particle Substances 0.000 description 2
- 238000012545 processing Methods 0.000 description 2
- 230000000306 recurrent effect Effects 0.000 description 2
- 238000005070 sampling Methods 0.000 description 2
- 238000004062 sedimentation Methods 0.000 description 2
- 239000007787 solid Substances 0.000 description 2
- 238000012800 visualization Methods 0.000 description 2
- XLYOFNOQVPJJNP-UHFFFAOYSA-N water Substances O XLYOFNOQVPJJNP-UHFFFAOYSA-N 0.000 description 2
- 230000001133 acceleration Effects 0.000 description 1
- 230000004913 activation Effects 0.000 description 1
- 239000000443 aerosol Substances 0.000 description 1
- 239000000809 air pollutant Substances 0.000 description 1
- 231100001243 air pollutant Toxicity 0.000 description 1
- 238000003915 air pollution Methods 0.000 description 1
- 210000004027 cell Anatomy 0.000 description 1
- 239000003245 coal Substances 0.000 description 1
- 239000000356 contaminant Substances 0.000 description 1
- 238000011109 contamination Methods 0.000 description 1
- 239000013078 crystal Substances 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000008034 disappearance Effects 0.000 description 1
- 238000000469 dry deposition Methods 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000004880 explosion Methods 0.000 description 1
- 239000007789 gas Substances 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 230000002427 irreversible effect Effects 0.000 description 1
- 230000007787 long-term memory Effects 0.000 description 1
- 230000007774 longterm Effects 0.000 description 1
- 230000035800 maturation Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 210000002569 neuron Anatomy 0.000 description 1
- 239000003758 nuclear fuel Substances 0.000 description 1
- 230000005658 nuclear physics Effects 0.000 description 1
- 239000001301 oxygen Substances 0.000 description 1
- 229910052760 oxygen Inorganic materials 0.000 description 1
- 230000037361 pathway Effects 0.000 description 1
- 230000000704 physical effect Effects 0.000 description 1
- 238000011176 pooling Methods 0.000 description 1
- 238000001556 precipitation Methods 0.000 description 1
- 230000000630 rising effect Effects 0.000 description 1
- 238000001179 sorption measurement Methods 0.000 description 1
- 239000002915 spent fuel radioactive waste Substances 0.000 description 1
- 239000000126 substance Substances 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
- 238000013519 translation Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
- G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/10—Pre-processing; Data cleansing
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
- G06F30/28—Design optimisation, verification or simulation using fluid dynamics, e.g. using Navier-Stokes equations or computational fluid dynamics [CFD]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/044—Recurrent networks, e.g. Hopfield networks
- G06N3/0442—Recurrent networks, e.g. Hopfield networks characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2113/00—Details relating to the application field
- G06F2113/08—Fluids
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2119/00—Details relating to the type or aim of the analysis or the optimisation
- G06F2119/14—Force analysis or force optimisation, e.g. static or dynamic forces
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2123/00—Data types
- G06F2123/02—Data types in the time domain, e.g. time-series data
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Software Systems (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Computational Linguistics (AREA)
- Biomedical Technology (AREA)
- Health & Medical Sciences (AREA)
- Biophysics (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Geometry (AREA)
- Computer Hardware Design (AREA)
- Pure & Applied Mathematics (AREA)
- Mathematical Optimization (AREA)
- Mathematical Analysis (AREA)
- Fluid Mechanics (AREA)
- Algebra (AREA)
- Medical Informatics (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a radionuclide diffusion prediction method and a radionuclide diffusion prediction system based on GAT-LSTM, which relate to the technical field of nuclear radiation protection, and comprise the following steps: acquiring radionuclide diffusion data at each time step, converting the radionuclide diffusion data into graph data, and constructing graph time sequence data; respectively extracting spatial features of the graph data under each time step in the graph time sequence data through a GAT network module; taking the extracted spatial features of the graph data under each time step as time sequence embedded vectors, and inputting the time sequence embedded vectors into an LSTM network module for extracting the time features; and inputting the space-time characteristics of the extracted time sequence data of the graph into an FC network module to obtain the predicted value of the radionuclide concentration.
Description
Technical Field
The invention relates to the technical field of nuclear radiation protection, in particular to a radionuclide diffusion prediction method and a radionuclide diffusion prediction system based on GAT-LSTM.
Background
Nuclear safety concerns have been raised and radioactive contamination has caused irreversible damage to humans and the natural environment. Therefore, the nuclide diffusion simulation research in nuclear accident outcome evaluation is very important, and is helpful for scientifically guiding nuclear emergency decisions. Nuclide diffusion simulation research is gradually rising, the research relates to a plurality of subjects such as nuclear physics, atmospheric environment science, computer science and the like, mature theoretical basis and application modes have been developed, and the research is gradually changed from real scene monitoring experiments to simulation by means of computer. With the continuous development and maturation of deep learning technology, deep learning is also increasingly widely used in the field of nuclide diffusion simulation.
Convolutional Neural Networks (CNNs) are a deep learning model that implement feature extraction and classification through structures such as convolutional layers, pooling layers, fully connected layers, and the like. It is widely used in the fields of image recognition, voice recognition, natural language processing, etc. Convolutional Neural Networks (CNNs) also have good effects in time series prediction, such as Liu Xulin, etc., which proposes a PM2.5 concentration prediction model CNN-Seq2Seq. Sahin proposes a method for modeling air pollution monitoring problems using a CNN model that can predict the concentration of missing air pollutants.
A Recurrent Neural Network (RNN) is a neural network structure with memory capabilities. Its basic structure comprises a hidden layer and an output layer, wherein the hidden layer can receive and store information from an input sequence. RNNs are used in many fields for processing and predicting sequence data, such as natural language processing, speech recognition, machine translation, etc. Because RNN has certain memory capacity, the method can be used for predicting pollutant concentration, such as Li Shugang, and the like, and the RNN network model is used for predicting the gas concentration of a coal mine working face; huang Jie et al predict the PM2.5 concentration using the RNN-CNN hybrid model.
As a recurrent neural network, long-term memory networks (LSTM) have a cell structure that makes LSTM more advantageous for long-term sequential training than traditional RNNs. Xu et al propose a HighAir model for LSTM based encoder-decoder architecture, predicting air quality for a long triangle being a city; ma et al propose a novel Geo-LSTM model based on LSTM, predicting pollutants in Washington. In addition, a variety of hybrid models have been proposed: yang Jianan proposes the use of CNN-LSTM to predict negative oxygen ion concentration; liu Meng et al propose an LSTM-FC model of LSTM mixed with a fully connected network to predict six contaminant concentrations of PM 2.5.
In summary, various deep learning methods have been well demonstrated to be able to predict atmospheric and environmental pollutants. Although the diffusion of the radionuclide in the atmosphere has a certain similarity with the diffusion of common pollutants, the traditional diffusion prediction model cannot completely adapt to the diffusion prediction requirement of the radionuclide, and compared with the diffusion prediction model of common atmospheric pollutants, the diffusion prediction model of the radionuclide is more complex, and the influence of factors such as decay of the radionuclide, dry and wet deposition and the like on the diffusion of the radionuclide is required to be considered, wherein the dry and wet deposition mainly refers to the sedimentation process of the radionuclide in the atmosphere; dry deposition refers to the direct deposition of radionuclides from the air to the ground or water surface in the form of solid particles which can be suspended in the air for a period of time and then deposited by gravity or other substances attached to the particles; wet deposition refers to the process of settling the radionuclide carried to the ground or the surface of a body of water by precipitation (such as rain, snow, dew), in which the radionuclide is usually present in dissolved or suspended form, falling to the ground in the form of raindrops or snow crystals; dry and wet deposition is the primary sedimentation pathway for radionuclides in the atmosphere, depending on factors such as the physical properties of the nuclides, air movement, and meteorological conditions. In addition, radionuclide diffusion is also comprehensively influenced by meteorological factors such as temperature, relative pressure, wind speed, wind direction and the like, and is also influenced by different types of underlying factors. For example, in urban undersides, the distribution, height, shape and street valley structure of urban buildings can have an influence on the wind direction and wind speed of incoming wind, thereby influencing the diffusion of airborne radionuclides; in the vegetation subgasket, both single-type vegetation and multi-type vegetation more or less affect the radionuclide diffusion due to the vegetation's ability to adsorb and change the flow field. For the research on the influence of different types of under-pad factors on the diffusion of nuclides, the influence of each single under-pad on the diffusion of nuclides is analyzed and researched in a concentrated manner in the past, and the influence of different under-pads on the diffusion of nuclides is compared transversely. However, there is little concern about the diffusion of species under complex underlying surfaces combined from different underlying surfaces in the same diffusion scenario.
Disclosure of Invention
In order to overcome the defects in the prior art, the invention provides the radionuclide diffusion prediction method based on GAT-LSTM, which comprehensively considers the influence of various factors on the radionuclide diffusion and has higher nuclide diffusion prediction precision.
In order to achieve the above purpose, the present invention adopts the following technical scheme, including:
the radionuclide diffusion prediction method based on GAT-LSTM comprises the following steps of: a GAT network module, an LSTM network module, and an FC network module;
the radionuclide diffusion prediction method comprises the following steps:
s1, acquiring radionuclide diffusion data in each time step, respectively converting the radionuclide diffusion data in each time step into graph data, and constructing graph time sequence data;
s2, respectively extracting spatial features of the graph data under each time step in the graph time sequence data through a GAT network module to obtain the spatial features of the graph data under each time step;
s3, taking the extracted spatial features of the graph data in each time step as time sequence embedding vectors, inputting the time sequence embedding vectors into an LSTM network module for time feature extraction, and obtaining the time-space features of the graph time sequence data;
s4, inputting the space-time characteristics of the extracted time sequence data of the graph into an FC network module to obtain a predicted value of the radionuclide concentration.
In step S1, in the graph data, each monitoring area is taken as a node, radionuclide concentration, wind speed, wind direction, relative pressure and underlying surface information of the monitoring area are taken as node characteristics, and mutual influence among the monitoring areas is taken as an edge to establish a connection relationship among the nodes.
Preferably, for a complex underlying scene composed of multiple underlying surfaces, the monitoring areas are respectively provided in different underlying surfaces, and the plurality of monitoring areas are provided in the same underlying surface.
Preferably, the connection relation between the nodes is established by using the k neighbor node and the pearson correlation coefficient of each node: each node is connected with the nearest k nodes by default, on the basis, the relevance between the two nodes is calculated through the pearson relevance coefficient, a relevance threshold is set to determine the additional connection relation between the nodes, if the pearson relevance coefficient between the two nodes is larger than the relevance threshold, the connection between the two nodes is established, and otherwise, the connection relation between the two nodes is not established.
Preferably, the pearson correlation coefficient is calculated as follows:
in the method, in the process of the invention,rrepresenting the pearson correlation coefficient between two nodes,x i andy i the first two nodes respectivelyiThe characteristics of the individual nodes are such that,nrepresenting the total number of node features;rthe larger the value, the greater the correlation between the two nodes; when (when)rEqual to 1, this indicates a complete correlation between the two nodes; when (when)rEqual to 0, this means that the two nodes are completely independent.
Preferably, the global training process of the GAT-LSTM network model is as follows:
s11, setting the number of time steps input by the GAT-LSTM network model as M and the number of time steps output by the GAT-LSTM network model as P according to a prediction target;
s12, acquiring radionuclide concentration diffusion time sequence data, converting the radionuclide concentration diffusion time sequence data into graph time sequence data, and creating sample data; the sample data includes: time series data of graphG={G t+1 ,G t+2 ,…,G t+M Real value of radionuclide concentration to be predictedY t ={Y t M++1 ,Y t M++2 ,…,Y t M+P+ -a }; wherein,G t M+ represent the firstt+MThe graph data at the time step is a graph,Y t M P++ representing the first to be predictedt+M+PGraph data at individual time steps;
s13, dividing sample data into a training set and a testing set;
s14, training the GAT-LSTM network by using a training set to obtain a trained GAT-LSTM network model;
s15, predicting graph time sequence data in the test set by using the trained GAT-LSTM network model to obtain a radionuclide concentration predicted valuePredicted value of radionuclide concentration +.>And true valueY t And comparing, and testing the prediction accuracy of the trained GAT-LSTM network model.
Preferably, in the global training process of the GAT-LSTM network model, an Adam optimizer is adopted to update parameters, and the mean square error is selectedMSESelecting average absolute percentage error as a function of lossMAPEAs an evaluation index; wherein,
in the method, in the process of the invention,mrepresenting the total number of samples,represents a predicted value of the concentration of the radionuclide,Y t representing the true value of radionuclide concentration.
Preferably, in step S12, a complex underlying surface geometric model composed of a plurality of underlying surfaces is built, and radionuclide diffusion time series data is simulated and generated by adopting CFD software assuming a radionuclide release source.
Preferably, different pressure loss coefficients are set to represent different underlying surfaces, i.e., underlying surface information is the pressure loss coefficient.
The invention also provides a radionuclide diffusion prediction system based on GAT-LSTM, which is suitable for the radionuclide diffusion prediction method based on GAT-LSTM, and comprises the following steps: the data preprocessing module is used for a GAT-LSTM network model;
the data preprocessing module is used for converting radionuclide diffusion data into graph data;
the GAT-LSTM network model takes picture time sequence data as input and radionuclide concentration predicted values as output, and is used for predicting radionuclide concentrations according to the picture time sequence data;
the GAT-LSTM network model comprises: a GAT network module, an LSTM network module, and an FC network module; the GAT network module is used for extracting spatial features of the graph data in each time step to obtain spatial features of the graph data in each time step; the LSTM network module is used for extracting time characteristics of the spatial characteristics of the graph data in each time step to obtain the time-space characteristics of the graph time sequence data; the FC network module is used for predicting the radionuclide concentration according to the time-space characteristics of the time sequence data of the graph.
The invention has the advantages that:
(1) According to the method, spatial features can be extracted from a plurality of monitoring areas by means of different deep learning models, time sequence features are extracted from historical change data under long time, attention mechanisms, namely correlation coefficients, are introduced to capture the mutual influence among different monitoring areas, and the method can be used for predicting the change of the radionuclide concentration of the target monitoring area well.
(2) The invention presents the diffusion condition of the radionuclide under different underlying scenes through the visualization of the attention weight, namely the correlation coefficient, and presents the influence of different underlying surfaces on the diffusion of the radionuclide to a certain extent.
(3) The GAT network used by the invention can learn the association relation among any nodes, automatically learn the association degree among different nodes through an attention mechanism, and fully mine the spatial characteristics among the nodes.
(4) The LSTM network used by the invention can learn the dependency relationship of a long sequence and effectively process time sequence data; compared with a simple RNN, the method can relieve the problems of gradient disappearance and gradient explosion, and enables the network to be more easy to train.
(5) The GAT-LSTM network model of the invention starts from radionuclide concentration meteorological data and underlying surface characteristic data, and predicts diffusion under the complex underlying surface scene of the radionuclide by utilizing the space characteristic extraction capability of the GAT network module and the time characteristic extraction capability of the LSTM network module, and has higher prediction precision.
(6) Based on the complex underlying surface model, the interaction effect on nuclide diffusion among different underlying surface models is researched, the method has important significance on radiation protection work, and more reference information can be provided for nuclear emergency decision under the condition of radioactive leakage accidents.
Drawings
FIG. 1 is a flow chart of the radionuclide diffusion prediction method based on GAT-LSTM of the invention.
FIG. 2 is a complex underlying geometric model of three different vegetation.
FIG. 3 is a schematic diagram of the positional relationship between each monitoring region and the Source of radionuclide release.
Fig. 4 is a graph of radionuclide concentration variation for each monitoring region.
FIG. 5 is a diagram of the data of the graph; wherein 5a is the visual display of the graph data, 5b is the node feature matrix, and 5c is the node adjacency matrix.
Fig. 6 is a schematic diagram of a process for extracting spatiotemporal features of the time series data of the graph.
Fig. 7 is a block diagram of a GAT-LSTM network.
FIG. 8 is a graph showing the result of predicting the concentration of the nuclide in the monitoring region A1.
Fig. 9 is a view showing the attention weighting value of the monitoring area A1 by other nodes when the diffusion is performed for 40 minutes.
FIG. 10 is a graph showing the result of predicting the concentration of the nuclide in the monitoring region B2.
Fig. 11 is a view showing the attention weighting value of the monitoring area B2 by other nodes when the diffusion is performed for 40 minutes.
FIG. 12 is a graph showing the result of predicting the concentration of the nuclide in the monitoring region C1.
Fig. 13 is a view showing the weight of attention of other nodes to the monitoring area C1 at 60 minutes of diffusion.
Detailed Description
The following description of the embodiments of the present invention will be made clearly and completely with reference to the accompanying drawings, in which it is apparent that the embodiments described are only some embodiments of the present invention, but not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.
The diffusion of radionuclides in the atmosphere is comprehensively influenced by meteorological factors such as relative pressure, wind speed, wind direction and the like, and is also influenced by different types of underlying factors. The invention provides a mixed nuclide diffusion prediction model based on GAT and LSTM, namely a GAT-LSTM network model for predicting the concentration in radionuclide diffusion. The GAT-LSTM network model takes graph time sequence data as input and radionuclide concentration predicted values as output; the GAT-LSTM network model includes: GAT network module, LSTM network module, FC network module.
As shown in fig. 1, the radionuclide diffusion prediction method of the present invention includes the steps of:
s1, radionuclide diffusion data under each time step are acquired, and are respectively converted into graph data, so that graph time sequence data are constructed.
In the graph data, each monitoring area is taken as a node, the radionuclide concentration, the wind speed, the wind direction, the relative pressure and the underlying surface information of the monitoring area are taken as node characteristics, and the mutual influence among the monitoring areas is taken as an edge so as to establish a connection relation among the nodes.
For a complex underlying surface scene composed of multiple underlying surfaces, monitoring areas are respectively arranged in different underlying surfaces, and a plurality of monitoring areas are arranged in the same underlying surface.
The k neighbor nodes and the pearson correlation coefficient of each node are used for carrying out common evaluation to establish the connection relation between the nodes: each node is connected with the nearest k nodes by default, on the basis, the relevance between different nodes is calculated through the pearson relevance coefficient, a relevance threshold is set to determine the additional connection relation between the nodes, if the pearson relevance coefficient between two different nodes is larger than the relevance threshold, the connection between the two different nodes is established, and otherwise, the connection relation between the two different nodes is not established.
S2, respectively carrying out spatial feature extraction on the graph data under each time step in the graph time sequence data through the GAT network module to obtain the spatial features of the graph data under each time step.
And S3, taking the extracted spatial features of the graph data in each time step as time sequence Embedding vectors (time sequence Embedding), and inputting the time sequence Embedding vectors into an LSTM network module for time feature extraction to obtain the space-time features of the graph time sequence data.
S4, inputting the space-time characteristics of the extracted time sequence data of the graph into an FC network module to obtain a predicted value of the radionuclide concentration.
The present embodiment uses radionuclide diffusion data obtained by actual monitoring because it is difficult to obtain and training of neural networks (GAT-LSTM networks) requires a large amount of reliable dataOpenFOAM simulation of CFD software generates radionuclide diffusion timing data. In the embodiment, a complex underlying surface geometric model formed by three different vegetation (underlying surfaces) is established by taking leakage accidents of spent fuel (irradiated nuclear fuel) in the transportation process as a research scene, and the model is shown in fig. 2. Wherein, calculation region of model 1000X 1000m 3 The left plane of the cube is an air inlet, the right plane of the cube is an air outlet, the airflow direction is set to be along the positive direction of the x axis, and the wind speed is set to be 4m/s. The solid dots in the figure represent the source of radionuclide I-131 released, coordinates (400 m,500m,1 m), and this example assumes a continuous release of the source for 1 hour, i.e., 3600 seconds, with a release rate of 3.3 (KBq/m) 3 ) And/s, sampling frequency is 1 second once, and sampling is continuously carried out for 1 hour. The half-life of radionuclide I-131 is 8.02d. The complex underlying surface geometric model comprises three different vegetation groups A, B, C shown in fig. 2, wherein different vegetation has different adsorption and flow field changing capacities, which is called a vegetation effect, the essence of the vegetation effect is that the vegetation is taken as a porous medium, the pressure loss is generated on the diffusion fluid passing through the surface of the vegetation, and the pressure loss generated by different vegetation types and different density degrees is different. Therefore, the present embodiment sets different pressure loss coefficients to represent different vegetation undersides, wherein the pressure loss coefficient of the vegetation group A is set to 0.5m -1 The space size is 120m multiplied by 60m multiplied by 15m; the pressure loss coefficient of the vegetation group B is set to 2m -1 The space size is 60m×140m×15m; the pressure loss coefficient of the vegetation group C is 8m -1 The spatial dimensions were 60m×140m×15m.
In order to accurately simulate the distribution of leaked nuclides, the standard k-epsilon turbulence model is introduced in the embodiment, and the vegetation effect is provided in the underlying surface of the embodiment, so that the decay and deposition effects of nuclides are considered, and the loss of a momentum source is considered. In this embodiment, an improved model based on computational fluid dynamics method is also built to simulate the diffusion process of nuclides under accident conditions.
The simplified mass conservation equation and momentum conservation equation are as follows:
in the method, in the process of the invention,ρrepresenting the density of the radioactive aerosol,tthe time is represented by the time period of the day,uthe speed is indicated by the velocity of the light,u i 、u j all representing the velocity component of the fluid,x i 、x j all representing components of the spatial coordinates,μthe dynamic viscosity is indicated as the dynamic viscosity,gindicating the acceleration of gravity and,λthe coefficient of pressure loss is indicated as such,indicating fluid density->Representing loss of momentum source->Indicating that the velocity component is inx i Conjugation in the direction, ++>Indicating that the velocity component is inx j Conjugation in the direction.
The corrected radionuclide concentration taking into account the decay and deposition of the nuclideCThe equation can be written as:
in the method, in the process of the invention,λ d representing the decay factor of the nuclide,v d representing the nuclide deposition factor, T 1/2 Indicating the half-life of the nuclide,Cindicating the concentration of the radionuclide(s),representing the turbulence diffusion coefficient.
The OpenFOAM simulation of CFD software generates diffusion data of the radionuclide I-131, wherein the time step of collecting the diffusion data is 10 seconds, the total simulated diffusion time is 2 hours, namely 7200 seconds, and finally 720 concentration distribution data are obtained. The characteristics of radionuclide diffusion data include time, radionuclide concentration, relative pressure, X-axis, Y-axis, Z-axis wind speed, and three-dimensional coordinates. Some examples of data generated by CFD software are shown in table 1 below.
TABLE 1 radionuclide diffusion data
In this embodiment, three monitoring areas are selected from three different vegetation groups A, B, C, and a total of nine monitoring areas, namely A1, A2, A3, B1, B2, B3, C1, C2, and C3, are obtained, and fig. 3 shows the positional relationship between each monitoring area and the radionuclide releasing Source. The coordinates of each monitoring area are shown in table 2 below.
Table 2 monitoring area coordinates
As shown in FIG. 4, the variation of the radionuclide Concentration in each monitoring region is represented by time on the abscissa, s, and the radionuclide Concentration on the ordinate, kBq/m 3 . In the early stages of radionuclide diffusion, A1, B1 and B2 first monitor the nuclide concentration due to geographical location and wind direction factors. With the continuous release of the release source and the progress of the nuclide diffusion, each monitoring area, i.e., each node, successively monitored the nuclide concentration around 30 minutes after the release of the release source. A1, B1, B2 and B3 are less affected by the factors of the underlying surface because they are positioned at the left part of the vegetation group, and the nuclide concentrations of A1, B2 and B3 tend to be stable after the release source stops releasing, i.e. after 1 hour. Other nodes are comprehensively affected by the underlying surface and the meteorological phenomena, and the nuclide concentration still changes after the release source stops releasing.
And converting radionuclide diffusion data of each monitoring area into graph data, wherein in the graph data, each monitoring area is taken as a node, radionuclide concentration, wind speed, wind direction, relative pressure and underlying surface information, namely pressure loss coefficients of the monitoring areas are taken as node characteristics, and the mutual influence among the monitoring areas is taken as edges so as to establish a connection relation among the nodes. The graph data structure is shown in fig. 5, wherein 5a is a visual display of the graph data; 5b is a node characteristic matrix, wherein rows represent each node, and columns represent node characteristics, including radionuclide concentration, wind speed, wind direction, relative pressure and pressure loss coefficient of the node; and 5c is a node adjacency matrix, wherein rows represent each node, and columns represent nodes with connection relations.
In this embodiment, the connection relationship between the nodes is established by using the pearson correlation coefficient and the 3 neighbor node of each node: each node is connected with the nearest 3 nodes by default, on the basis, the relevance among different nodes is calculated through the pearson relevance coefficient, and a relevance threshold value is set to determine the additional connection relation among the nodes.
The calculation formula of the pearson correlation coefficient is as follows:
in the method, in the process of the invention,rrepresenting the pearson correlation coefficient between two nodes,x i andy i respectively the first two nodesiThe characteristics of the individual nodes are such that,nrepresenting the total number of node features; when (when)rEqual to 1, this indicates a complete correlation between the two nodes; when (when)rWhen equal to 0, the two nodes are completely independent;rthe larger the value, the greater the correlation between the two nodes; in this embodiment, the correlation threshold is set to 0.9, if the pearson correlation coefficient between two nodesrAnd if the number is more than 0.9, the connection is established between the two nodes, otherwise, the connection is not established between the two nodes.
For example, pearson correlation coefficients between nodes at the diffusion time t=60 min are shown in table 3 below.
Table 3 pearson correlation coefficients between nodes at diffusion time t=60 min
The node adjacency matrix at which the diffusion time t=60 min is obtained is shown in table 4 below.
Table 4 node adjacency matrix at diffusion time t=60 min
In the node adjacency matrix, an element value of 1 indicates that a connection relationship exists between two nodes, and an element value of 0 indicates that no connection relationship exists between the two nodes.
The degree matrix is a diagonal matrix, and only the diagonal position records that the node has several nodes connected with it. The degree matrix at diffusion time t=60 min is shown in table 5 below.
Table 5 degree matrix at diffusion time t=60 min
After diffusion data of the radionuclide I-131 are obtained through OpenFOAM simulation of CFD software and are converted into graph time sequence data, sample data can be created and used for training and generating a GAT-LSTM network model.
Fig. 6 illustrates a spatio-temporal feature extraction process of graph time series data based on the GAT-LSTM network model.
In this embodiment, the global training process of the GAT-LSTM network model is specifically as follows:
s11, setting the number of time steps in the past as M according to a prediction target, and outputting the number of time steps in the future as P.
S12, creating sample data according to graph time sequence data converted from diffusion time sequence data of the radionuclide I-131 obtained through simulation of CFD software. The sample data includes: time series data of graphG={G t+1 ,G t+2 ,…,G t+M Real value of radionuclide concentration to be predictedY t ={Y t M++1 ,Y t M++2 ,…,Y t M+P+ }. Wherein,G t+1 represent the firsttPlot data for +1 time step (past time step),Y t M++1 representing the first to be predictedt+MPlot data for +1 time steps (future time steps).
In the graph time sequence data, the graph data structure at each time step comprises a node characteristic matrix, a node adjacent matrix and a degree matrix.
S13, dividing the sample data into a training set and a testing set according to the proportion of 9:1.
And S14, training the GAT-LSTM network by using the training set to obtain a trained GAT-LSTM network model.
S15, predicting graph time sequence data in the test set by using the trained GAT-LSTM network model to obtain a radionuclide concentration predicted valuePredicted value of radionuclide concentration +.>And true valueY t And comparing, evaluating by using the evaluation index, and analyzing the prediction accuracy of the trained GAT-LSTM network model.
In the global training process of the GAT-LSTM network model, an Adam optimizer is adopted to update parameters, and mean square error is selectedMSESelecting average absolute percentage error as a function of lossMAPEAs an evaluation index. Wherein,
in the method, in the process of the invention,mrepresenting the total number of samples,the predicted value is represented by a value of the prediction,Y t representing the true value.
In the embodiment, based on radionuclide diffusion data generated by CFD software simulation, graph time sequence data of radionuclide diffusion is created through a data processing process. And inputting the training set into a GAT-LSTM network for training, finally obtaining a trained GAT-LSTM network model, namely a prediction model, testing on test set data by using the prediction model, and calculating the error between a predicted value and a true value, namely a simulation value by using an evaluation index, thereby evaluating the prediction precision of the prediction model. And the attention weight, namely the data characteristic, at different moments in the training process is extracted, and the diffusion trend of the radionuclide under the complex underlying surface is presented through the visual attention weight.
The GAT-LSTM network model is complex, the GAT network module is required for node space feature extraction in graph data, the LSTM network module is required for time feature extraction of graph time sequence data, and finally the predicted value of the radionuclide concentration is required to be output through the FC network module. Each network module realizes functions by setting the respective layer number, unit number and activation function. The configuration parameters, evaluation index selection, and optimizer settings of each network module of the GAT-LSTM network model in this embodiment are shown in table 6 below.
TABLE 6 parameter settings for GAT-LSTM network model
The GAT-LSTM network of this embodiment has the structure shown in FIG. 7. Wherein, N_GAT represents the number of layers of the GAT layer and is set to 2; N_LSTM represents the number of LSTM layers, and is set to 4; n_fc represents the number of layers of the FC layer, and is set to 2; N_Unites represents the number of neuron units of the hidden layer, and is set to 16.
In this embodiment, for a target node, i.e., a monitoring area to be predicted, map data of P time steps in the future are predicted from map data of M time steps in the past, i.e., radionuclide concentrations of P time steps in the future are predicted, and attention weights, i.e., correlation coefficients, in a certain time step are intercepted to present a radionuclide diffusion trend. In this embodiment, m=15 and p=3 are set in the nuclide concentration prediction, that is, the radionuclide concentration of 30s in the future is predicted from the radionuclide concentration variation data and the meteorological data in the first 150s at this time.
In the embodiment, the radionuclide concentration is predicted at each selected node in the three vegetation underlying surfaces, and then the attention weight of the radionuclide is intercepted at a certain time step to present the diffusion trend of the radionuclide. Specifically, a node A1 is selected from the vegetation group a, a node B2 is selected from the vegetation group B, a node C1 is selected from the vegetation group C, and the prediction results of the radionuclide concentrations of the three nodes are sequentially shown below.
In the radionuclide concentration prediction experiment on the node A1, 600s to 3310s data are used as model training sets, and 3320s to 3620s data are used as model test sets. The overall radionuclide diffusion prediction result of node A1 is shown in fig. 8 and 9. FIG. 8 is a graph showing the result of predicting the concentration of a nuclide at A1, and the mean square errorMSEMean absolute percentage error =1.41MAPE=0.33; fig. 9 is a visual diagram of the attention weights of other nodes to A1, namely the correlation coefficients, when the nodes diffuse for 40 minutes, and the attention weights are marked on the corresponding edges.
In the radionuclide concentration prediction experiment on node B2, 600s to 3310s data were used as model training sets and 3320s to 3620s data were used as model test sets. The overall radionuclide diffusion prediction results for node B2 are shown in fig. 10 and 11. Wherein FIG. 10 shows the result of predicting the concentration of the nuclide at B2,MSE=1.37,MAPE=0.43; fig. 11 is a visual illustration of the attention weights of other nodes to B2 at 40 minutes of diffusion, with the attention weights noted on the respective edges.
In the radionuclide concentration prediction experiment on the node C1, the data from 1200s to 5350s are used as model training sets, and the data from 5360s to 5820s are used as model test sets. The overall radionuclide diffusion prediction result of the node C1 is shown in fig. 12 and 13. Wherein FIG. 12 shows the result of predicting the concentration of the nuclide at C1,MSE=2.96,MAPE=0.62; FIG. 13 is a view of expansionAnd (3) the attention weight visualization graph of other nodes on C1 when the current node looses for 60 minutes, wherein the attention weight is marked on the corresponding side.
By combining the prediction results of the three nodes, the GAT-LSTM network model provided by the invention can be used for fitting the integral change of the radionuclide concentration, has better performance on the radionuclide concentration prediction of the three nodes, and has the advantages ofMSEAndMAPEgood results are obtained on indexes.
In fig. 9, 11 and 13, the present invention visualizes the attention weight of each connection node to the target node in time steps, and is used to represent the influence degree of each connection node nuclide concentration change, meteorological information and underlying information on the target node nuclide concentration change. In fig. 9, the nodes connected to the node A1 are the node A2, the node A3, the node B1, and the node B2, respectively; the weight of attention from high to low is respectively: node A1 self-attention weight, node B2 attention weight to node A1, node B1 attention weight to node A1, node A3 attention weight to node A1, and node A2 attention weight to node A1. In fig. 11, the nodes connected to the node B2 are the node A1, the node A2, the node B1, the node B3, the node C1, the node C2, and the node C3, respectively; the weight of attention from high to low is respectively: node B2 self-attention weighting, node A1 attention weighting to node B2, node B1 attention weighting to node B2, node C1 attention weighting to node B2, node B3 attention weighting to node B2, node C2 attention weighting to node B2, and node A2 attention weighting to node B2. In fig. 13, the nodes connected to the node C1 are the node A3, the node B2, the node C2, and the node C3, respectively; the weight of attention from high to low is respectively: node C1 self-attention weight, node B2 attention weight to node C1, node C3 attention weight to node C1, node C2 attention weight to node C1, and node A3 attention weight to node C1.
In the prediction tasks of the three nodes, the node which has the greatest influence on the nuclide concentration change of the target node is the target node, and the node which is closest to the target node or is close to the air inlet side is the next node. This conforms to the law of radionuclide diffusion to some extent. However, it should be noted that these attention weights are not necessarily required, and in the whole time sequence data set of the graph, the attention weights only exist in a certain time step, but in the training process of the network model, the weights change with time and are also affected by the node characteristic data, and as the model training progresses, the attention weights gradually tend to be stable and also conform to the physical rule of the actual diffusion scene. Therefore, the GAT-LSTM network model provided by the invention can better predict the change of the target node nuclide concentration, can show the trend of radionuclide diffusion to a certain extent, and can show the influence of different underlayers on nuclide diffusion to a certain extent.
The above embodiments are merely preferred embodiments of the present invention and are not intended to limit the present invention, and any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the scope of the present invention.
Claims (10)
1. The radionuclide diffusion prediction method based on GAT-LSTM is characterized in that the GAT-LSTM network model comprises the following steps: a GAT network module, an LSTM network module, and an FC network module;
the radionuclide diffusion prediction method comprises the following steps:
s1, acquiring radionuclide diffusion data in each time step, respectively converting the radionuclide diffusion data in each time step into graph data, and constructing graph time sequence data;
s2, respectively extracting spatial features of the graph data under each time step in the graph time sequence data through a GAT network module to obtain the spatial features of the graph data under each time step;
s3, taking the extracted spatial features of the graph data in each time step as time sequence embedding vectors, inputting the time sequence embedding vectors into an LSTM network module for time feature extraction, and obtaining the time-space features of the graph time sequence data;
s4, inputting the space-time characteristics of the extracted time sequence data of the graph into an FC network module to obtain a predicted value of the radionuclide concentration.
2. The method according to claim 1, wherein in step S1, each monitoring region is defined as a node, the radionuclide concentration, wind speed, wind direction, relative pressure, and underlying information of the monitoring region are defined as node characteristics, and the interaction between each monitoring region is defined as an edge to establish a connection relationship between nodes.
3. The method according to claim 2, wherein for a complex under-pad scene composed of a plurality of under-pads, monitoring areas are provided in different under-pads, respectively, and a plurality of monitoring areas are provided in the same under-pad.
4. The method for predicting radionuclide diffusion based on GAT-LSTM according to claim 2, wherein the connection relationship between nodes is established by using k neighbor nodes and pearson correlation coefficients of each node: each node is connected with the nearest k nodes by default, on the basis, the relevance between the two nodes is calculated through the pearson relevance coefficient, a relevance threshold is set to determine the additional connection relation between the nodes, if the pearson relevance coefficient between the two nodes is larger than the relevance threshold, the connection between the two nodes is established, and otherwise, the connection relation between the two nodes is not established.
5. The method for predicting radionuclide diffusion based on GAT-LSTM according to claim 3, characterized in that the calculation formula of pearson correlation coefficient is as follows:
in the method, in the process of the invention,rrepresenting the pearson correlation coefficient between two nodes,x i andy i the first two nodes respectivelyiThe characteristics of the individual nodes are such that,nrepresenting the total number of node features;rthe larger the value, the greater the correlation between the two nodes; when (when)rEqual to 1, this indicates a complete correlation between the two nodes; when (when)rEqual to 0, this means that the two nodes are completely independent.
6. The method of radionuclide diffusion prediction based on GAT-LSTM according to any of claims 1 to 5, characterized by the global training procedure of the GAT-LSTM network model, specifically as follows:
s11, setting the number of time steps input by the GAT-LSTM network model as M and the number of time steps output by the GAT-LSTM network model as P according to a prediction target;
s12, acquiring radionuclide concentration diffusion time sequence data, converting the radionuclide concentration diffusion time sequence data into graph time sequence data, and creating sample data; the sample data includes: time series data of graphG={G t+1 ,G t+2 ,…,G t+M Real value of radionuclide concentration to be predictedY t ={Y t M++1 ,Y t M++2 ,…,Y t M+P+ -a }; wherein,G t M+ represent the firstt+MThe graph data at the time step is a graph,Y t M P++ representing the first to be predictedt+M+PGraph data at individual time steps;
s13, dividing sample data into a training set and a testing set;
s14, training the GAT-LSTM network by using a training set to obtain a trained GAT-LSTM network model;
s15, predicting graph time sequence data in the test set by using the trained GAT-LSTM network model to obtain a radionuclide concentration predicted valuePredicted value of radionuclide concentration +.>And true valueY t And comparing, and testing the prediction accuracy of the trained GAT-LSTM network model.
7. The method of claim 6, wherein the GAT-LSTM network model uses an Adam optimizer to update parameters during global training to select mean square errorMSESelecting average absolute percentage error as a function of lossMAPEAs an evaluation index; wherein,
in the method, in the process of the invention,mrepresenting the total number of samples->Represents a predicted value of the concentration of the radionuclide,Y t representing the true value of radionuclide concentration.
8. The method according to claim 6, wherein in step S12, a complex underlying surface geometric model composed of a plurality of underlying surfaces is created, and radionuclide diffusion time series data is simulated and generated by using CFD software assuming a radionuclide release source.
9. The method of claim 8, wherein the different pressure loss coefficients are set to represent different underlying surfaces, i.e., underlying surface information is a pressure loss coefficient.
10. A GAT-LSTM based radionuclide diffusion prediction system, characterized in that the system is adapted for a GAT-LSTM based radionuclide diffusion prediction method according to any of claims 1 to 5;
the system comprises: the data preprocessing module is used for a GAT-LSTM network model;
the data preprocessing module is used for converting radionuclide diffusion data into graph data;
the GAT-LSTM network model takes picture time sequence data as input and radionuclide concentration predicted values as output, and is used for predicting radionuclide concentrations according to the picture time sequence data;
the GAT-LSTM network model comprises: a GAT network module, an LSTM network module, and an FC network module; the GAT network module is used for extracting spatial features of the graph data in each time step to obtain spatial features of the graph data in each time step; the LSTM network module is used for extracting time characteristics of the spatial characteristics of the graph data in each time step to obtain the time-space characteristics of the graph time sequence data; the FC network module is used for predicting the radionuclide concentration according to the time-space characteristics of the time sequence data of the graph.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202311595153.0A CN117313551A (en) | 2023-11-28 | 2023-11-28 | Radionuclide diffusion prediction method and system based on GAT-LSTM |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202311595153.0A CN117313551A (en) | 2023-11-28 | 2023-11-28 | Radionuclide diffusion prediction method and system based on GAT-LSTM |
Publications (1)
Publication Number | Publication Date |
---|---|
CN117313551A true CN117313551A (en) | 2023-12-29 |
Family
ID=89273964
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202311595153.0A Pending CN117313551A (en) | 2023-11-28 | 2023-11-28 | Radionuclide diffusion prediction method and system based on GAT-LSTM |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN117313551A (en) |
Citations (10)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111639748A (en) * | 2020-05-15 | 2020-09-08 | 武汉大学 | Watershed pollutant flux prediction method based on LSTM-BP space-time combination model |
CN111832814A (en) * | 2020-07-01 | 2020-10-27 | 北京工商大学 | Air pollutant concentration prediction method based on graph attention machine mechanism |
CN112884045A (en) * | 2021-02-25 | 2021-06-01 | 河北工业大学 | Classification method of random edge deletion embedded model based on multiple visual angles |
CN113516304A (en) * | 2021-06-29 | 2021-10-19 | 上海师范大学 | Space-time joint prediction method and device for regional pollutants based on space-time graph network |
CN113811009A (en) * | 2021-09-24 | 2021-12-17 | 之江实验室 | Multi-base-station cooperative wireless network resource allocation method based on space-time feature extraction reinforcement learning |
CN114154702A (en) * | 2021-11-25 | 2022-03-08 | 深圳中兴网信科技有限公司 | Pollutant concentration prediction method and device based on multi-granularity graph space-time neural network |
US20220214322A1 (en) * | 2021-01-07 | 2022-07-07 | Tsinghua University | Air pollutants concentration forecasting method and apparatus and storage medium |
CN115438582A (en) * | 2022-08-30 | 2022-12-06 | 河北工业大学 | PM2.5 concentration prediction method combining multiple elements and attention |
CN115795344A (en) * | 2022-11-30 | 2023-03-14 | 湖北工业大学 | Graph convolution network document classification method and system based on mixed diffusion |
CN115859861A (en) * | 2022-12-07 | 2023-03-28 | 国网福建省电力有限公司电力科学研究院 | Substation fire atmospheric pollution diffusion analysis method based on urban environment influence |
-
2023
- 2023-11-28 CN CN202311595153.0A patent/CN117313551A/en active Pending
Patent Citations (10)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111639748A (en) * | 2020-05-15 | 2020-09-08 | 武汉大学 | Watershed pollutant flux prediction method based on LSTM-BP space-time combination model |
CN111832814A (en) * | 2020-07-01 | 2020-10-27 | 北京工商大学 | Air pollutant concentration prediction method based on graph attention machine mechanism |
US20220214322A1 (en) * | 2021-01-07 | 2022-07-07 | Tsinghua University | Air pollutants concentration forecasting method and apparatus and storage medium |
CN112884045A (en) * | 2021-02-25 | 2021-06-01 | 河北工业大学 | Classification method of random edge deletion embedded model based on multiple visual angles |
CN113516304A (en) * | 2021-06-29 | 2021-10-19 | 上海师范大学 | Space-time joint prediction method and device for regional pollutants based on space-time graph network |
CN113811009A (en) * | 2021-09-24 | 2021-12-17 | 之江实验室 | Multi-base-station cooperative wireless network resource allocation method based on space-time feature extraction reinforcement learning |
CN114154702A (en) * | 2021-11-25 | 2022-03-08 | 深圳中兴网信科技有限公司 | Pollutant concentration prediction method and device based on multi-granularity graph space-time neural network |
CN115438582A (en) * | 2022-08-30 | 2022-12-06 | 河北工业大学 | PM2.5 concentration prediction method combining multiple elements and attention |
CN115795344A (en) * | 2022-11-30 | 2023-03-14 | 湖北工业大学 | Graph convolution network document classification method and system based on mixed diffusion |
CN115859861A (en) * | 2022-12-07 | 2023-03-28 | 国网福建省电力有限公司电力科学研究院 | Substation fire atmospheric pollution diffusion analysis method based on urban environment influence |
Non-Patent Citations (8)
Title |
---|
YUANZHAO ZHAI: "DIVERSIFYING MESSAGE AGGREGATION IN MULTI-AGENT COMMUNICATION VIA NORMALIZED TENSOR NUCLEAR NORM REGULARIZATION", IEEE, pages 1 - 5 * |
刘全根: "地球科学新学科新概念集成", 31 August 1995, 《北京:地震出版社》, pages: 41 - 42 * |
匿名: "《核科学与工程》", 31 December 2020, 匿名, pages: 236 * |
孙小新: "基于图神经网络的PM2.5浓度预测算法研究", 中国博士学位论文电子期刊网, pages 027 - 22 * |
王奎良: "马达加斯加主要矿产资源潜力分析", 31 December 2021, 中国地质大学出版社, pages: 19 * |
阮灵盼: "基于长短期记忆网络的涉核运输事故后果预测", 辐射研究与辐射工艺学报, pages 1 - 13 * |
陈自艳: "密集型监测场景下PM<sub>2.5</sub>浓度时序预测研究", 中国优秀硕士论文电子期刊网, pages 027 - 2164 * |
魏东;董法军;董希琳;: "核事故中放射性核素扩散浓度的理论预测", 中国安全科学学报, no. 03, pages 111 - 117 * |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN111860205B (en) | Forest fire evaluation method based on multisource remote sensing images and grids and storage medium | |
CN110346517B (en) | Smart city industrial atmosphere pollution visual early warning method and system | |
Boris | The threat of chemical and biological terrorism: preparing a response | |
KR100982447B1 (en) | Landslide occurrence prediction system and predicting method using the same | |
CN114266200A (en) | Nitrogen dioxide concentration prediction method and system | |
CN113516304B (en) | Regional pollutant space-time joint prediction method and device based on space-time diagram network | |
CN115688636A (en) | Atmospheric environmental pollution remediation simulation method and system based on computational fluid dynamics | |
CN106845080A (en) | Scene Tourist meteorological disaster intelligent Forecasting based on difference amendment | |
CN115293545A (en) | Typhoon disaster risk assessment method and system based on power grid district division and storage medium | |
CN115471679A (en) | Method and intelligent system for synchronously assimilating water level and flow of natural river | |
CN114255250A (en) | River protection engineering slope collapse detection method based on deep learning | |
CN116384565A (en) | Hierarchical atmospheric ozone concentration prediction method based on missing data filling | |
CN114694767A (en) | PM2.5 concentration prediction method based on space-time diagram ordinary differential equation network | |
Abdelhady et al. | Realizing hurricane resilient communities through distributed computing | |
Patnaik et al. | Large scale urban simulations with the MILES approach | |
CN113742983A (en) | Long-span structural damage identification method based on depth self-encoder neural network | |
Yan et al. | An Efficient Method for Optimizing Sensors’ Layout for Accurate Measurement of Underground Ventilation Networks | |
CN117313551A (en) | Radionuclide diffusion prediction method and system based on GAT-LSTM | |
Patnaik et al. | Large scale urban simulations with FCT | |
Brandt et al. | Using a combination of two models in tracer simulations | |
Shigenaka et al. | Estimating pedestrian flow in crowded situations with data assimilation | |
Green et al. | Modeling wildfire using evolutionary cellular automata | |
Viner | Coming and going: Transport and tracking | |
Wong et al. | Integrating computational fluid dynamics (CFD) models with GIS: An evaluation on data conversion formats | |
Haiqin et al. | Data driven based PM2. 5 concentration forecasting |
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 |