CN117993307A - Earth system simulation result consistency assessment method based on deep learning - Google Patents
Earth system simulation result consistency assessment method based on deep learning Download PDFInfo
- Publication number
- CN117993307A CN117993307A CN202410404783.3A CN202410404783A CN117993307A CN 117993307 A CN117993307 A CN 117993307A CN 202410404783 A CN202410404783 A CN 202410404783A CN 117993307 A CN117993307 A CN 117993307A
- Authority
- CN
- China
- Prior art keywords
- bgru
- training
- earth system
- reconstruction error
- deep learning
- 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
- 238000004088 simulation Methods 0.000 title claims abstract description 71
- 238000000034 method Methods 0.000 title claims abstract description 68
- 238000013135 deep learning Methods 0.000 title claims abstract description 15
- 238000004364 calculation method Methods 0.000 claims abstract description 38
- 238000011156 evaluation Methods 0.000 claims abstract description 16
- 230000035945 sensitivity Effects 0.000 claims abstract description 5
- 238000012549 training Methods 0.000 claims description 79
- 238000013528 artificial neural network Methods 0.000 claims description 39
- 230000008569 process Effects 0.000 claims description 39
- 238000013136 deep learning model Methods 0.000 claims description 28
- 238000012360 testing method Methods 0.000 claims description 28
- 230000006870 function Effects 0.000 claims description 24
- 238000010168 coupling process Methods 0.000 claims description 22
- 230000008878 coupling Effects 0.000 claims description 19
- 238000005859 coupling reaction Methods 0.000 claims description 19
- 230000008859 change Effects 0.000 claims description 15
- 238000010200 validation analysis Methods 0.000 claims description 12
- 238000012795 verification Methods 0.000 claims description 12
- 230000004913 activation Effects 0.000 claims description 10
- 230000006835 compression Effects 0.000 claims description 7
- 238000007906 compression Methods 0.000 claims description 7
- 238000002474 experimental method Methods 0.000 claims description 6
- 210000002569 neuron Anatomy 0.000 claims description 6
- 238000004458 analytical method Methods 0.000 claims description 5
- 238000012545 processing Methods 0.000 claims description 4
- NAWXUBYGYWOOIX-SFHVURJKSA-N (2s)-2-[[4-[2-(2,4-diaminoquinazolin-6-yl)ethyl]benzoyl]amino]-4-methylidenepentanedioic acid Chemical compound C1=CC2=NC(N)=NC(N)=C2C=C1CCC1=CC=C(C(=O)N[C@@H](CC(=C)C(O)=O)C(O)=O)C=C1 NAWXUBYGYWOOIX-SFHVURJKSA-N 0.000 claims description 3
- 238000005215 recombination Methods 0.000 claims description 3
- 230000006798 recombination Effects 0.000 claims description 3
- 238000012935 Averaging Methods 0.000 claims description 2
- 238000007405 data analysis Methods 0.000 abstract description 2
- 238000011160 research Methods 0.000 abstract description 2
- 238000012986 modification Methods 0.000 description 10
- 230000004048 modification Effects 0.000 description 10
- 238000005457 optimization Methods 0.000 description 9
- 238000013461 design Methods 0.000 description 6
- 125000004122 cyclic group Chemical group 0.000 description 4
- 230000009466 transformation Effects 0.000 description 4
- 238000013459 approach Methods 0.000 description 2
- 238000000605 extraction Methods 0.000 description 2
- 230000010354 integration Effects 0.000 description 2
- 239000000463 material Substances 0.000 description 2
- 230000000644 propagated effect Effects 0.000 description 2
- 230000000306 recurrent effect Effects 0.000 description 2
- 238000011426 transformation method Methods 0.000 description 2
- 230000002159 abnormal effect Effects 0.000 description 1
- 238000009825 accumulation Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000002457 bidirectional effect Effects 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000013210 evaluation model Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 238000003058 natural language processing Methods 0.000 description 1
- 238000003062 neural network model Methods 0.000 description 1
- 238000010606 normalization 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
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- 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
- G06F2119/02—Reliability analysis or reliability optimisation; Failure analysis, e.g. worst case scenario performance, failure mode and effects analysis [FMEA]
-
- 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
- Y02A90/00—Technologies having an indirect contribution to adaptation to climate change
- Y02A90/10—Information and communication technologies [ICT] supporting adaptation to climate change, e.g. for weather forecasting or climate simulation
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Evolutionary Computation (AREA)
- Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Medical Informatics (AREA)
- Software Systems (AREA)
- Computer Hardware Design (AREA)
- Geometry (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention provides a simulation result consistency evaluation method of an earth system based on deep learning, and the method is smoothly applied to a many-core heterogeneous super computing platform. By the technical scheme, the short-time-step earth system simulation result consistency evaluation method is small in calculated amount, time cost is saved, calculation cost is saved, and the like. According to the invention, the nonlinear characteristics of the multi-module component data are analyzed, and the consistency of the simulation result is evaluated more accurately. The invention is oriented to a many-core heterogeneous supercomputer platform, and can detect the existence of software or human errors when the disturbance related to hardware is considered. The invention can provide technical support and data support for scientific research work and business departments, scientific basis for transplanting, researching and optimizing the earth system mode under the many-core heterogeneous super computing platform, and data analysis method and theoretical support for discussing the sensitivity of the earth system simulation result to the many-core heterogeneous architecture.
Description
Technical Field
The invention relates to the technical field of numerical simulation of an earth system, in particular to a consistency evaluation method of simulation results of the earth system based on deep learning.
Background
Consistency assessment during optimization and development of the earth system model is critical to ensuring reliability of the earth system model. For consistency assessment, there are two approaches. First, the climate scientist analyzes model simulation data of hundreds of years on a new machine and compares it with the same simulated data on a trusted machine. And secondly, processing an annual average earth system simulation result in a linear transformation mode, and comparing a newly operated simulation result with a collection result from a trusted machine.
The existing method is mainly applied to the multi-core isomorphic supercomputer system. In the face of unavoidable computational disturbances caused by many-core heterogeneous hardware designs and mixed disturbance scenarios composed of software or human errors, there is an urgent need to evaluate the scientific consistency of the simulation results of the earth system on many-core heterogeneous supercomputer systems, so that the consistency evaluation model can accept the influence of many-core heterogeneous computational disturbances to further detect software or human errors generated when optimizing and developing the earth system model.
As the resolution and complexity of the earth system modes increase, stringent requirements are placed on rapid consistency assessment. At present, a consistency evaluation method for analyzing the simulation results of the earth system with short time steps and multiple component modules is lacking. In addition, the existing linear transformation method cannot accurately analyze the nonlinear characteristics of data. In the face of the nonlinear relation generated by the combination of multi-component module data in the earth system, the nonlinear transformation is urgently needed to realize the characteristic extraction and analysis of the data.
Disclosure of Invention
In order to make up for the defects of the prior art, the invention provides a simulation result consistency evaluation method of an earth system based on deep learning. The present invention aims to solve two problems: 1. developing an earth system simulation result consistency assessment method based on deep learning; 2. the method is smoothly applied to a many-core heterogeneous super computing platform.
Aiming at the first problem, the existing linear transformation method cannot accurately analyze the nonlinear characteristics of data. In the face of the nonlinear relation generated by the combination of multi-component module data in the earth system, nonlinear transformation such as a deep learning method is urgently needed to realize the feature extraction and analysis of the data. The invention provides a deep learning-based earth system simulation result consistency assessment method, which is based on a bi-directional gating cyclic neural network, analyzes the characteristics of a trusted earth system mode simulation result variable, obtains the passing rate of the mode simulation result under different scenes through coarse-grained test, designs scenes such as compiling option change, mode parameter change and the like, and assesses the consistency of the earth system simulation result under a mixed disturbance scene.
Aiming at the second problem, with the requirement of high-resolution and high-complexity earth system simulation, the general application of the many-core heterogeneous architecture high-performance supercomputer to the earth system mode will become a future trend. The earth system mode is transplanted from the multi-core isomorphic super computing platform to the many-core heterogeneous architecture platform, and consistency of an earth system mode simulation result under the many-core heterogeneous architecture needs to be evaluated. However, the common consistency assessment method is mainly oriented to the multi-core isomorphic architecture platform. For the simulation result of the earth system after being disturbed by the many-core heterogeneous architecture, an objective and efficient consistency assessment method is needed to assess the scientific consistency of the simulation result of the earth system mode under the many-core heterogeneous architecture. Therefore, the consistency assessment method based on deep learning is applied to a many-core heterogeneous supercomputer system.
The invention is realized by the following technical scheme: the earth system simulation result consistency assessment method based on deep learning specifically comprises the following steps:
S1, realizing a two-way gating cycle-self-encoder neural network: the neuron GRU neural network of the gate control loop neural network GRNN firstly needs to calculate the values of an update gate and a reset gate according to a hidden layer h t-1 at the time t-1 and an input x t at the current time t; the calculation of update gate z t is shown in equation (1), the calculation of reset gate r t is shown in equation (2),
zt=(Wz* [ht-1,xt] ), (1)
rt=(Wr* [ht-1,xt] ), (2)
Wherein W z、Wr is a training parameter,To activate the function. After obtaining the signal of the reset gate r t, the GRU neural network calculates the reset result/>This process is similar to the memorization phase of LSTM; /(I)The calculation process of (1) is shown in the formula (3),
= tanh (W* [rt*ht-1,xt] ), (3)
Wherein W is a training parameter and tanh is an activation function. GRU neural network according to reset resultAnd updating the gate z t to calculate the result of the hidden layer h t; the calculation process of h t is shown in formula (4),
ht= (1-zt) *ht-1+zt*, (4)
The input and output of the automatic encoder AE are consistent, with the goal of reconstructing itself using sparse higher-order feature recombination; AE calculates a loss function through a predicted value obtained by training and an original true value to train AE neural network parameters; for the input data x, the calculation process of the intermediate hidden layer compression vector C of AE is shown in formula (5), the calculation process is shown in formula (6) of the reconstruction data of which the output result O of AE is C,
C= f (W(1)*x+b(1)), (5)
O = f (W(2)*C+b(2)), (6)
Wherein W (1)、W(2)、b(1)、b(2) is a training parameter obtained according to loss function feedback, and f is an activation function; the BGRU-AE deep learning model is that an encoder and a decoder are replaced by BGRU neural network on the basis of AE, and a layer of fully-connected neural network is finally connected with the decoder; the encoder is composed of BGRU, and the hidden layer output H e of BGRU is taken as an output compression vector C of the encoder; the calculation process of H e is shown in the formula (7),
He=Hf t Hb 1=C, (7)
Wherein H f t and H b 1 are the last hidden layer states generated by the GRU forward and backward processes, respectively, and the decoder is also composed of BGRU, the compressed vector C is input into BGRU, the calculation process of the hidden layer output H d,Hd of BGRU is shown in formula (8),
Hd=H'f 1 H’b t, (8)
Wherein H 'f 1 and H' b t are the last hidden layer states generated by the GRU forward and backward processes respectively, and finally, H d is input into the fully connected layer FC for controlling and obtaining the vector O with the same dimension as the initial input x, the computing process of FC is shown in a formula (9),
O= f (Wfc*Hd), (9)
Wherein W fc is a training parameter and f is an activation function;
s2, selecting an earth system coupling model comprising an ocean module, an atmosphere module and corresponding coupling processes or couplers;
S3, setting the coupling frequency of the atmosphere module and the ocean module and the time step number of the earth system simulation, and verifying whether the short-time-step earth system simulation data after the ocean mode coupling frequency is changed can contain enough set variability; setting the ocean mode coupling frequency to be 8 times in 1 day, adding disturbance of O (10 -14) magnitude to the initial atmospheric temperature of the mode, comparing the disturbance with the undisturbed atmospheric and ocean result variable, proving the sensitivity dependence of the atmospheric and ocean simulation result in the mode on the initial condition, and providing enough aggregate variability for the atmospheric and ocean variable after the ocean mode coupling frequency is changed;
S4, carrying out an earth system simulation set experiment, namely realizing earth system simulation on a trusted multi-core isomorphic super computing platform, carrying out area weight average and minimum maximum standardization processing on atmosphere and ocean variables of simulation results, and constructing a training set and a verification set for the deep learning model. The area weight average calculation process is shown in formula (10), and the maximum and minimum normalization calculation process is shown in formula (11).
(10),
Where x i is the value of the variable x in the ith grid and w i is the area of the ith grid.
(11),
Where x min is the minimum value of variable x and x max is the maximum value of variable x.
S5, BGRU-AE deep learning model training: performing BGRU-AE deep learning model training by adopting a training set and a verification set on a multi-core isomorphic platform; at each time step of training BGRU-AE models, the validation dataset is input into the BGRU-AE model and the reconstruction error is calculated using the mean square error MSE (Mean Squared Error) function and the reconstruction error of the validation set is output. Adjusting BGRU-AE model parameters, repeating BGRU-AE deep learning model training until the reconstruction error of the verification data set is minimum, and storing a final BGRU-AE model; the reconstruction error calculation process is shown in formula (12).
, (12)
Where y i is the original value of variable i and y i' is the value of y i after BGRU-AE model analysis.
S6, calculating a reconstruction error threshold of the training data set, wherein the reconstruction error threshold of the training data set is used as an index to obtain a pass or fail result of consistency evaluation; re-inputting the training data set into the saved BGRU-AE deep learning model, and re-obtaining the maximum value of the reconstruction error; the reconstruction error threshold value is the maximum value of the reconstruction error.
S7, simulating an earth system to realize many-core heterogeneous programming: designing a scene comprising compiling option change and mode parameter change, realizing earth system simulation on a many-core heterogeneous super computing platform, and constructing a test data set used by BGRU-AE deep learning model;
s8, evaluating consistency of simulation results of the earth system under the mixed disturbance scene: the test dataset was input into a saved BGRU-AE deep learning model and the pass rate was calculated using the reconstruction error.
Preferably, in step S6, if the reconstruction error of a member of the set of test data sets is greater than the reconstruction loss threshold of the training data set, the consistency assessment method returns the member as "failed".
Preferably, in step S8, if the reconstruction error of the test data set is greater than the reconstruction error threshold of the training set, the training set is determined to be "failed", otherwise, the training set is passed.
The invention adopts the technical proposal, and compared with the prior art, the invention has the following beneficial effects:
1. compared with the earth system simulation result consistency evaluation method based on the annual average data, the short-time-step earth system simulation result consistency evaluation method provided by the invention has the advantages of small calculated amount, time cost saving, calculation cost and the like.
2. Compared with the method for evaluating the consistency of the simulation result of the earth system based on linear transformation, the method provided by the invention analyzes the nonlinear characteristics of the multi-module component data and evaluates the consistency of the simulation result more accurately.
3. Compared with the existing method for evaluating the consistency of the simulation results of the earth system, the invention is oriented to a many-core heterogeneous supercomputer platform, and can detect the existence of software or human errors when the disturbance related to hardware is considered.
4. The invention can provide technical support and data support for scientific research work and business departments, scientific basis for transplanting, researching and optimizing the earth system mode under the many-core heterogeneous super computing platform, and data analysis method and theoretical support for discussing the sensitivity of the earth system simulation result to the many-core heterogeneous architecture.
Additional aspects and advantages of the invention will be set forth in part in the description which follows, or may be learned by practice of the invention.
Drawings
The foregoing and/or additional aspects and advantages of the invention will become apparent and may be better understood from the following description of embodiments taken in conjunction with the accompanying drawings in which:
FIG. 1 is a GRU neuron structure.
FIG. 2 is a BGRU-AE neural network structure.
Fig. 3 is a time series of reconstruction errors for a training set and a validation set. Blue lines are training sets and red lines are test sets.
Fig. 4 is a probability density function of the reconstruction error of the training set. The red line is the threshold for reconstruction error.
Fig. 5 compiles test set reconstruction errors in an optimization options modification scenario. The test set reconstruction errors-O0, -O1, -O2 do not exceed the training set reconstruction error threshold. The abscissa represents the-O0, -O1, -O2 compilation option experiments, and the ordinate represents the reconstruction error.
FIG. 6 is a graph of test set reconstruction errors in a pattern parameter change and initial disturbance addition scenario. The test set reconstruction errors in the c0_lnd, c0_ocn and initial disturbance addition scenarios all exceed the training set reconstruction error threshold. The abscissa indicates the modification of the c0_ lnd and c0_ocn parameters and the addition of O (10 -6) initial perturbation experiment, and the ordinate indicates the reconstruction error.
Detailed Description
In order that the above-recited objects, features and advantages of the present application will be more clearly understood, a more particular description of the application will be rendered by reference to the appended drawings and appended detailed description. It should be noted that, without conflict, the embodiments of the present application and features in the embodiments may be combined with each other.
In the following description, numerous specific details are set forth in order to provide a thorough understanding of the present invention, but the present invention may be practiced otherwise than as described herein, and therefore the scope of the present invention is not limited to the specific embodiments disclosed below.
The following describes a method for evaluating consistency of simulation results of an earth system based on deep learning according to an embodiment of the present invention in detail with reference to fig. 1 to 6.
The invention provides a simulation result consistency assessment method of an earth system based on deep learning. Secondly, the coupling frequency of the atmosphere module and the ocean module and the time step number of the earth system simulation are set, and whether the short time step earth system simulation data after the ocean mode coupling frequency is changed can contain enough set variability is verified. And (3) carrying out an earth system simulation set experiment, and constructing a training set and a verification set for the deep learning model. And carrying out BGRU-AE deep learning model training by adopting a training set and a verification set on the multi-core isomorphic platform. At each time step of training BGRU-AE models, the validation dataset is input into the BGRU-AE model and the reconstruction error is calculated using a mean square error function. Parameters of BGRU-AE model were adjusted until the reconstruction error of the validation dataset was minimal and the final BGRU-AE model was saved. The reconstruction error threshold of the training dataset is calculated as an indicator to obtain a "pass" or "fail" result of the consistency assessment. And (3) simulating the earth system to realize many-core heterogeneous programming, designing scenes such as compiling option change, mode parameter change and the like, realizing the earth system simulation on a many-core heterogeneous super computing platform, and evaluating the consistency of the earth system simulation result under the mixed disturbance scene. The method specifically comprises the following steps:
S1, realizing a two-way gating cycle-self-encoder neural network: the gated recurrent neural network GRNN (Gate Recurrent Neural Network) is a neural network model widely applied in the fields of natural language processing, speech recognition, machine translation and the like. Compared with the traditional cyclic neural network, the neuron GRU (GateRecurrent Unit) of the gated cyclic neural network has simpler structure, only two gates of the update gate and the reset gate are used for controlling the flow of information, network parameters are fewer, and the convergence is faster. The GRU neuron structure is shown in FIG. 1. The neuron GRU neural network of the gate control loop neural network GRNN firstly needs to calculate the values of an update gate and a reset gate according to a hidden layer h t-1 at the time t-1 and an input x t at the current time t; the calculation of update gate z t is shown in equation (1), the calculation of reset gate r t is shown in equation (2),
zt=(Wz* [ht-1,xt] ), (1)
rt=(Wr* [ht-1,xt] ), (2)
Wherein W z、Wr is a training parameter,To activate the function. After obtaining the signal of the reset gate r t, the GRU neural network calculates the reset result/>This process is similar to the memorization phase of LSTM; /(I)The calculation process of (1) is shown in the formula (3),
= tanh (W* [rt*ht-1,xt] ), (3)
Wherein W is a training parameter and tanh is an activation function. GRU neural network according to reset resultAnd updating the gate z t to calculate the result of the hidden layer h t; the calculation process of h t is shown in formula (4),
ht= (1-zt) *ht-1+zt*, (4)
The unidirectional neural network can only acquire memory from past time points, so that unidirectional propagation information is realized. And the bidirectional neural network can be propagated bidirectionally, and can acquire information from a future time point. The basic idea of the two-way neural network is to input the same input sequence into a forward and backward propagated cyclic neural network respectively, and the two neural networks are connected with the same output layer, so that the output of the neural network can acquire the context information of the past and future moments. The automatic encoder AE (Auto-Encoder) is a neural network of unsupervised data dimension compression and data feature expression. The input and output of the automatic encoder AE are consistent, with the goal of reconstructing itself using sparse higher-order feature recombination; AE calculates a loss function through a predicted value obtained by training and an original true value to train AE neural network parameters; for the input data x, the calculation process of the intermediate hidden layer compression vector C of AE is shown in formula (5), the calculation process is shown in formula (6) of the reconstruction data of which the output result O of AE is C,
C= f (W(1)*x+b(1)), (5)
O = f (W(2)*C+b(2)), (6)
Wherein W (1)、W(2)、b(1)、b(2) is a training parameter obtained according to loss function feedback, and f is an activation function; the BGRU-AE deep learning model provided by the invention is that an encoder and a decoder are replaced by BGRU neural network on the basis of AE, and a layer of fully-connected neural network is finally connected with the decoder; the BGRU-AE neural network structure is shown in FIG. 2. Wherein the blue dashed box represents an encoder and the red dashed box represents a decoder. The encoder is composed of BGRU, and the hidden layer output H e of BGRU is taken as an output compression vector C of the encoder; the calculation process of H e is shown in the formula (7),
He=Hf t Hb 1=C, (7)
Wherein H f t and H b 1 are the last hidden layer states generated by the GRU forward and backward processes, respectively, and the decoder is also composed of BGRU, the compressed vector C is input into BGRU, the calculation process of the hidden layer output H d,Hd of BGRU is shown in formula (8),
Hd=H'f 1 H’b t, (8)
Wherein H 'f 1 and H' b t are the last hidden layer states generated by the GRU forward and backward processes respectively, and finally, H d is input into the fully connected layer FC for controlling and obtaining the vector O with the same dimension as the initial input x, the computing process of FC is shown in a formula (9),
O= f (Wfc*Hd), (9)
Wherein W fc is a training parameter and f is an activation function;
s2, selecting an earth system coupling model comprising an ocean module, an atmosphere module and corresponding coupling processes or couplers so as to meet the requirement of consistency evaluation of the earth system multi-component modules;
S3, setting the coupling frequency of the atmospheric module and the ocean module and the time step number of the earth system simulation so as to meet the requirement of short-time-step earth system simulation consistency evaluation, and verifying whether the short-time-step earth system simulation data after the ocean mode coupling frequency is changed can contain enough set variability; in the existing earth system mode, the ocean low frequency coupling causes calculation errors to be spread out by a large number of integration steps. The ocean mode coupling frequency is set to be 8 times in 1 day, so that errors can be rapidly diffused, and uncertainty of a calculation result is caused. On this basis, the invention verifies whether the data after changing the ocean mode coupling frequency can contain enough set variability. According to the invention, the disturbance of O (10 -14) magnitude is added to the initial atmospheric temperature of the mode, the difference between the initial atmospheric temperature and the undisturbed atmospheric and marine result variable is compared, the sensitivity dependence of the atmospheric and marine simulation result in the mode on the initial condition is proved, and the atmospheric and marine variable after the coupling frequency of the marine mode is changed can provide enough aggregate variability;
S4, carrying out an earth system simulation set experiment, selecting variable result data of an atmosphere and ocean module, realizing earth system simulation on a trusted multi-core isomorphic super computing platform, carrying out area weight averaging and standardization processing on the simulation result data, and constructing a training set and a verification set for a deep learning model;
The area weight average calculation process is shown in a formula (10), the maximum and minimum standardization calculation process is shown in a formula (11),
(10),
Where x i is the value of the variable x in the ith grid, w i is the area of the ith grid,
(11),
Wherein x min is the minimum value of variable x and x max is the maximum value of variable x;
S5, BGRU-AE deep learning model training: performing BGRU-AE deep learning model training by adopting a training set and a verification set on a multi-core isomorphic platform; at each time step of training BGRU-AE models, inputting the validation dataset into the BGRU-AE model, and calculating a reconstruction error using a mean square error MSE (Mean Squared Error) function, and outputting the reconstruction error of the validation set; adjusting BGRU-AE model parameters, repeating BGRU-AE deep learning model training until the reconstruction error of the verification data set is minimum, and storing a final BGRU-AE model; the reconstruction error calculation process is shown in formula (12).
, (12)
Where y i is the original value of variable i and y i' is the value of y i after BGRU-AE model analysis.
Fig. 3 shows a reconstruction error time sequence for a training set and a validation set. The present patent uses training and validation datasets to train BGRU-AE models. Wherein the training set is used to train the model and the validation set is used to fit the minimum loss and select the model parameters. The invention selects the model training state of the minimum loss of the validation set to determine the neural network parameters, and the result of the figure 3 shows that the BGRU _AE network under training has no over-fitting phenomenon.
S6, calculating a reconstruction error threshold of the training data set, wherein the reconstruction error threshold of the training data set is used as an index to obtain a pass or fail result of consistency evaluation; re-inputting the training data set into the saved BGRU-AE deep learning model, and re-obtaining the maximum value of the reconstruction error; it is generally assumed that reconstruction errors for normal inputs may be lower because they approach the training data set, while reconstruction errors for abnormal inputs may become higher. Thus, if the reconstruction error of a member of a set of test data sets is greater than the reconstruction loss threshold for the training data set, then the consistency assessment method returns that member as "failed". The reconstruction error threshold value is the maximum value of the reconstruction error;
Fig. 4 shows a probability density function of the reconstruction error of the training set. After re-inputting the training dataset into the saved BGRU-AE model, the present invention calculates the reconstruction error. Thus, if the reconstruction error of one member of the test dataset may be higher for the member of the maximum of the training dataset, then our consistency assessment tool returns that member as "failed".
S7, simulating an earth system to realize many-core heterogeneous programming: designing a scene comprising compiling option change and mode parameter change, realizing earth system simulation on a many-core heterogeneous super computing platform, and constructing a test data set used by BGRU-AE deep learning model;
S8, evaluating consistency of simulation results of the earth system under the mixed disturbance scene: the test dataset was input into a saved BGRU-AE deep learning model and the pass rate was calculated using the reconstruction error. If the reconstruction error of the test data set is greater than the reconstruction error threshold of the training set, the training set is judged to be failed, otherwise, the training set passes. We focused on the impact of compilation option modifications and schema parameter modifications on simulation result consistency in heterogeneous computation. When the disturbance related to hardware is considered, the earth system simulation result consistency assessment method based on the deep learning can detect the existence of software or human errors.
Compiling simulation result consistency assessment under the optimization option changing scene: fig. 5 illustrates test set reconstruction errors in a compile optimization option modification scenario. In converting a high-level programming language into machine language code, different compiler optimization options may result in differences in assembly code, such as different code execution orders and/or different intermediate register floating point precision, ultimately resulting in different floating point results. However, modifications in the-O0, -O1, -O2 compiler optimization options may result in non-identical floating point results, but no scientific changes are expected. The test data set in the-O0, -O1, -O2 compilation optimization options may be consistent with the training data set that considers the integration distribution, although involving many-core heterogeneous hardware design perturbations. According to the invention, simulation results of heterogeneous versions on a many-core heterogeneous supercomputer system are input into a consistency evaluation deep learning model. The reconstruction errors of the test sets of-O0, -O1 and-O2 do not exceed the reconstruction error threshold of the training set, so that the consistency evaluation method provided by the invention can be proved to be capable of receiving the mixed disturbance caused by the many-core heterogeneous hardware design and the acceptable compiling optimization option change.
Simulation result consistency evaluation under the conditions of mode parameter change and initial disturbance addition: fig. 6 illustrates test set reconstruction errors in a compile optimization option modification scenario. The climate scientist provides a list of CAM input parameters that are believed to affect the climate in a non-trivial manner, such as c0_ lnd and c0_ocn. The method modifies the values of the input parameters in the cloud accumulation convection parameterization scheme, and then tests whether the method can detect inconsistency caused by the change of the model parameters in heterogeneous calculation. Meanwhile, the invention evaluates the consistency of the simulation result under the initial disturbance adding scene. In view of the ensemble distribution, the test dataset with mixed perturbations caused by climate change modification and hardware design are inconsistent with the training dataset. According to the invention, simulation results of heterogeneous versions on a many-core heterogeneous supercomputer system are input into a consistency evaluation deep learning model. The test set reconstruction errors in the c0_lnd, c0_ocn and initial disturbance addition scenarios all exceed the training set reconstruction error threshold, proving that when considering disturbances related to many-core heterogeneous hardware design, the invention can detect software changes or human errors affecting climate consistency results.
In the description of the present invention, the term "plurality" means two or more, unless explicitly defined otherwise, the orientation or positional relationship indicated by the terms "upper", "lower", etc. are based on the orientation or positional relationship shown in the drawings, merely for convenience of description of the present invention and to simplify the description, and do not indicate or imply that the apparatus or elements referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be construed as limiting the present invention; the terms "coupled," "mounted," "secured," and the like are to be construed broadly, and may be fixedly coupled, detachably coupled, or integrally connected, for example; can be directly connected or indirectly connected through an intermediate medium. The specific meaning of the above terms in the present invention can be understood by those of ordinary skill in the art according to the specific circumstances.
In the description of the present specification, the terms "one embodiment," "some embodiments," "particular embodiments," and the like, mean that a particular feature, structure, material, or characteristic described in connection with the embodiment or example is included in at least one embodiment or example of the present invention. In this specification, schematic representations of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the particular features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.
The above is only a preferred embodiment of the present invention, and is not intended to limit the present invention, but various modifications and variations can be made to the present invention by those skilled in the art. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.
Claims (3)
1. The earth system simulation result consistency assessment method based on deep learning is characterized by comprising the following steps of:
S1, realizing a two-way gating cycle-self-encoder neural network: the neuron GRU neural network of the gate control loop neural network GRNN firstly needs to calculate the values of an update gate and a reset gate according to a hidden layer h t-1 at the time t-1 and an input x t at the current time t; the calculation of update gate z t is shown in equation (1), the calculation of reset gate r t is shown in equation (2),
zt =(Wz * [ht-1, xt] ) , (1)
rt = (Wr * [ht-1, xt] ) , (2)
Wherein W z、Wr is a training parameter,Is an activation function; after obtaining the signal of the reset gate r t, the GRU neural network calculates the reset result/>This process is similar to the memorization phase of LSTM; /(I)The calculation process of (1) is shown in the formula (3),
= tanh (W * [rt *ht-1, xt] ) , (3)
Wherein W is a training parameter, and tanh is an activation function; GRU neural network according to reset resultAnd updating the gate z t to calculate the result of the hidden layer h t; the calculation process of h t is shown in formula (4),
ht = (1-zt) * ht-1 + zt * , (4)
The input and output of the automatic encoder AE are consistent, with the goal of reconstructing itself using sparse higher-order feature recombination; AE calculates a loss function through a predicted value obtained by training and an original true value to train AE neural network parameters; for the input data x, the calculation process of the intermediate hidden layer compression vector C of AE is shown in formula (5), the calculation process is shown in formula (6) of the reconstruction data of which the output result O of AE is C,
C = f (W(1) * x + b(1)), (5)
O = f (W(2) * C + b(2)), (6)
Wherein W (1)、W(2)、b(1)、b(2) is a training parameter obtained according to loss function feedback, and f is an activation function; the BGRU-AE deep learning model is that an encoder and a decoder are replaced by BGRU neural network on the basis of AE, and a layer of fully-connected neural network is finally connected with the decoder; the encoder is composed of BGRU, and the hidden layer output H e of BGRU is taken as an output compression vector C of the encoder; the calculation process of H e is shown in the formula (7),
He = H f t H b 1 = C, (7)
Wherein H f t and H b 1 are the last hidden layer states generated by the GRU forward and backward processes, respectively, and the decoder is also composed of BGRU, the compressed vector C is input into BGRU, the calculation process of the hidden layer output H d,Hd of BGRU is shown in formula (8),
Hd = H'f 1 H’b t, (8)
Wherein H 'f 1 and H' b t are the last hidden layer states generated by the GRU forward and backward processes respectively, and finally, H d is input into the fully connected layer FC for controlling and obtaining the vector O with the same dimension as the initial input x, the computing process of FC is shown in a formula (9),
O = f (Wfc * Hd), (9)
Wherein W fc is a training parameter and f is an activation function;
s2, selecting an earth system coupling model comprising an ocean module, an atmosphere module and corresponding coupling processes or couplers;
S3, setting the coupling frequency of the atmosphere module and the ocean module and the time step number of the earth system simulation, and verifying whether the short-time-step earth system simulation data after the ocean mode coupling frequency is changed can contain enough set variability; setting the ocean mode coupling frequency to be 8 times in 1 day, adding disturbance of O (10 -14) magnitude to the initial atmospheric temperature of the mode, comparing the disturbance with the undisturbed atmospheric and ocean result variable, proving the sensitivity dependence of the atmospheric and ocean simulation result in the mode on the initial condition, and providing enough aggregate variability for the atmospheric and ocean variable after the ocean mode coupling frequency is changed;
S4, carrying out an earth system simulation set experiment, selecting variable result data of an atmosphere and ocean module, realizing earth system simulation on a trusted multi-core isomorphic super computing platform, carrying out area weight averaging and standardization processing on the simulation result data, and constructing a training set and a verification set for a deep learning model;
The area weight average calculation process is shown in a formula (10), the maximum and minimum standardization calculation process is shown in a formula (11),
(10),
Where x i is the value of the variable x in the ith grid, w i is the area of the ith grid,
(11),
Wherein x min is the minimum value of variable x and x max is the maximum value of variable x;
S5, BGRU-AE deep learning model training: performing BGRU-AE deep learning model training by adopting a training set and a verification set on a multi-core isomorphic platform; at each time step of training BGRU-AE models, inputting the validation dataset into the BGRU-AE model and calculating a reconstruction error using a mean square error function; and outputting a reconstruction error of the verification set; adjusting BGRU-AE model parameters, repeating BGRU-AE deep learning model training until the reconstruction error of the verification data set is minimum, and storing a final BGRU-AE model; the reconstruction error calculation process is shown in equation (12),
, (12)
Where y i is the original value of variable i, y i' is the value of y i after BGRU-AE model analysis;
S6, calculating a reconstruction error threshold of the training data set, wherein the reconstruction error threshold of the training data set is used as an index to obtain a pass or fail result of consistency evaluation; re-inputting the training data set into the saved BGRU-AE deep learning model, and re-obtaining the maximum value of the reconstruction error; the reconstruction error threshold value is the maximum value of the reconstruction error;
S7, simulating an earth system to realize many-core heterogeneous programming: designing a scene comprising compiling option change and mode parameter change, realizing earth system simulation on a many-core heterogeneous super computing platform, and constructing a test data set used by BGRU-AE deep learning model;
S8, evaluating consistency of simulation results of the earth system under the mixed disturbance scene: inputting the test data set into a saved BGRU-AE deep learning model, and calculating the passing rate by using the reconstruction error; if the reconstruction error of the test data set is greater than the reconstruction error threshold of the training set, the training set is judged to be failed, otherwise, the training set passes.
2. The method according to claim 1, wherein if the reconstruction error of a member of the set of test data sets is greater than the reconstruction loss threshold of the training data set in step S6, the method returns the member as "failed".
3. The method for evaluating consistency of simulation results of an earth system based on deep learning according to claim 1, wherein in the step S8, if the reconstruction error of the test data set is greater than the reconstruction error threshold of the training set, the training set is determined to be "failed", otherwise, the test data set passes.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202410404783.3A CN117993307B (en) | 2024-04-07 | 2024-04-07 | Earth system simulation result consistency assessment method based on deep learning |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202410404783.3A CN117993307B (en) | 2024-04-07 | 2024-04-07 | Earth system simulation result consistency assessment method based on deep learning |
Publications (2)
Publication Number | Publication Date |
---|---|
CN117993307A true CN117993307A (en) | 2024-05-07 |
CN117993307B CN117993307B (en) | 2024-06-14 |
Family
ID=90889283
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202410404783.3A Active CN117993307B (en) | 2024-04-07 | 2024-04-07 | Earth system simulation result consistency assessment method based on deep learning |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN117993307B (en) |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113556319A (en) * | 2021-06-11 | 2021-10-26 | 杭州电子科技大学 | Intrusion detection method based on long-short term memory self-coding classifier under internet of things |
WO2022023008A1 (en) * | 2020-07-27 | 2022-02-03 | Zf Friedrichshafen Ag | Computer-implemented method and computer program for machine-learning a robustness of an acoustic classifier, acoustic classification system for automatically operable driving systems, and automatically operable driving system |
CN115545169A (en) * | 2022-11-02 | 2022-12-30 | 上海技群信息科技有限公司 | GRU-AE network-based multi-view service flow anomaly detection method, system and equipment |
CN115910083A (en) * | 2022-10-27 | 2023-04-04 | 网易(杭州)网络有限公司 | Real-time voice conversion method, device, electronic equipment and medium |
-
2024
- 2024-04-07 CN CN202410404783.3A patent/CN117993307B/en active Active
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2022023008A1 (en) * | 2020-07-27 | 2022-02-03 | Zf Friedrichshafen Ag | Computer-implemented method and computer program for machine-learning a robustness of an acoustic classifier, acoustic classification system for automatically operable driving systems, and automatically operable driving system |
CN113556319A (en) * | 2021-06-11 | 2021-10-26 | 杭州电子科技大学 | Intrusion detection method based on long-short term memory self-coding classifier under internet of things |
CN115910083A (en) * | 2022-10-27 | 2023-04-04 | 网易(杭州)网络有限公司 | Real-time voice conversion method, device, electronic equipment and medium |
CN115545169A (en) * | 2022-11-02 | 2022-12-30 | 上海技群信息科技有限公司 | GRU-AE network-based multi-view service flow anomaly detection method, system and equipment |
Non-Patent Citations (4)
Title |
---|
SHAOQING ZHANG ETC.: "Optimizing high-resolution Community Earth System Model on a heterogeneous many-core supercomputing platform", GEOSCI. MODEL DEV., 31 December 2020 (2020-12-31) * |
YANGYANG YU: "Characterizing uncertainties of Earth system modeling with heterogeneous many-core architecture computing", GEOSCI. MODEL DEV., 31 December 2022 (2022-12-31) * |
周天军 等: "区域海气耦合模式FROALS的发展及其应用", 大气科学, vol. 40, no. 6, 31 January 2016 (2016-01-31) * |
路强,等: "基于自动编码器的时间序列预测混合模型", 计算机系统应用, vol. 31, no. 7, 31 December 2022 (2022-12-31) * |
Also Published As
Publication number | Publication date |
---|---|
CN117993307B (en) | 2024-06-14 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN112202736B (en) | Communication network anomaly classification method based on statistical learning and deep learning | |
Lanubile et al. | Evaluating predictive quality models derived from software measures: lessons learned | |
Asl et al. | Optimo: A BIM-based multi-objective optimization tool utilizing visual programming for high performance building design | |
Abbas et al. | Formal property verification in a conformance testing framework | |
US20140214373A1 (en) | Framework for solving trajectory optimization problems | |
CN110571792A (en) | Analysis and evaluation method and system for operation state of power grid regulation and control system | |
CN110210654A (en) | Product model designing system and method | |
CN114389265B (en) | Power grid emergency load optimization method and system based on proliferation and reduced evolution | |
CN114818549B (en) | Method, system, equipment and medium for calculating fluid mechanics parameter of object | |
CN109961172B (en) | CPS rare event probability prediction method based on statistical model test | |
CN110874616A (en) | Transformer operation prediction method based on LSTM network and Markov chain correction error | |
CN116451514A (en) | Bridge structure state evaluation method and device and electronic equipment | |
CN116227180A (en) | Data-driven-based intelligent decision-making method for unit combination | |
Blaud et al. | From multi‐physics models to neural network for predictive control synthesis | |
CN117993307B (en) | Earth system simulation result consistency assessment method based on deep learning | |
Zargar et al. | Deep learning in early-stage structural performance prediction: Assessing morphological parameters for buildings | |
Lavaei et al. | Compositional reinforcement learning for discrete-time stochastic control systems | |
US11989656B2 (en) | Search space exploration for deep learning | |
Zunin et al. | Developing methods for combinational circuit generation | |
CN117578438A (en) | Generating countermeasure network method and system for predicting new energy power generation | |
CN117477544A (en) | LSTM ultra-short-term photovoltaic power prediction method and system integrating time mode features | |
Rauzy et al. | Model-based safety assessment: Rational and trends | |
US20240103920A1 (en) | Method and system for accelerating the convergence of an iterative computation code of physical parameters of a multi-parameter system | |
Chen et al. | Deep generative models for geometric design under uncertainty | |
CN112667957A (en) | Intelligent electric energy meter failure rate prediction method based on deep neural network |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |