Method and device for replacing sea-gas coupling parameterization in atmospheric waveguide forecast model based on physical constraint deep learning, and storage medium

Abstract

The invention discloses a method and equipment for replacing sea-gas coupling parameterization in an atmospheric waveguide forecast model based on physical constraint deep learning, and a storage medium. The method comprises the following steps: step 1, preparing training data; step 2, increasing physical constraint to improve generalization ability; step 3, constructing a network for calculation; according to the method disclosed by the invention, the occupancy rate of computing resources can be reduced, the computing speed and the model forecasting performance are improved at the same time, a new idea is provided for application of artificial intelligence in the meteorological field, and a new technical development direction is provided for innovation in the meteorological field.

Description

Technical Field

[0001] The present invention belongs to the field of meteorology and artificial intelligence, and in particular to a method, device and storage medium in this field for replacing the sea-air coupling parameterization in an atmospheric waveguide forecast model based on physical constraint deep learning. Background Art

[0002] Uncertainty in the parameterization of air-sea coupled mixing in the atmospheric duct forecast model is the main source of deviation in the simulation of meteorological parameters of the atmospheric duct. The latest progress in deep learning methods and the availability of long-term turbulence measurements provide opportunities to explore data-driven parameterization methods for atmospheric duct air-sea coupled ocean vertical mixing processes. How to improve the simulation of air-sea coupled atmospheric duct by integrating data-driven parameterization schemes into the atmospheric duct model is a prerequisite for solving the accuracy of model forecasts in the atmospheric duct environment.

[0003] Physically driven atmospheric duct parameterization schemes have been used in many ocean and climate models. However, there are large uncertainties in the turbulent properties estimated based on the mesoscale turbulence parameterization method, which leads to large errors in the prediction of atmospheric duct environmental parameters evaluated by model parameters. For example, the K-profile parameterization significantly overestimates the downward turbulent heat flux in the Pacific cold tongue, which makes it impossible to accurately predict the atmospheric duct with non-uniform characteristics in space. However, this parameterization is widely used in many ocean and climate models; more than one-third of the international climate models adopt this parameterization scheme, which increases the uncertainty of climate simulation and projection. The lack of process understanding is the main reason for the poor performance of traditional physically driven parameterization methods. For example, the vertical eddy diffusion coefficient (KT) estimated by the parameterization method based on mesoscale turbulence increases monotonically with the decrease of Ri. However, at the same Ri value, different flow states can exist, so the parameterized diffusion coefficient is significantly different from the observed value. In order to reduce the uncertainty in the vertical mixing parameterization, new methods are needed to explore the functional relationship between the vertical eddy diffusion coefficient and the resolvable ocean variables. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide a method, device and storage medium for replacing the sea-air coupling parameterization in the atmospheric waveguide prediction model based on physical constraint deep learning.

[0005] The present invention adopts the following technical solution:

[0006] A method for replacing the air-sea coupling parameterization in an atmospheric duct forecast model based on physical constraint deep learning, the improvement of which is that it includes the following steps:

[0007] Step 1, training data preparation:

[0008] The neural network is trained based on the existing ocean mooring datasets TAO and PIRATA, which include: Ocean stratification index N 2 、Ocean shear S 2 , density ρ and velocity U, the output is the ocean turbulence parameterization value K in the atmospheric model T ;

[0009] Step 2: Add physical constraints to improve generalization ability:

[0010] The turbulence index data generated by traditional physical methods is calculated as follows:

[0011]

[0012] In the above formula, Ri = N 2 / S 2 ;

[0013] Step 3: Build a network for calculation;

[0014] First, candidate variables are constructed, including temperature T, salinity SA, and ocean current U profile;

[0015] Secondly, the input vector contains four characteristic temperature T, salinity SA, ocean current U and Ri, and 850 training samples based on physical driving parameterization are added to the training dataset;

[0016] Finally, by introducing physical constraints into the neural network, a deep learning model based on physical information was built to construct the mapping relationship between the ocean model vertical eddy diffusion coefficient KT, vertical eddy diffusion viscosity Kv and the complex information hybrid parameterization scheme;

[0017] Step 4: Verify the model:

[0018] The MOM model is used to model the global ocean, and relevant parameters are obtained and substituted into the model for verification. The horizontal resolution of MOM5 is 1° in longitude, 1 / 3° at the equator, and 1° at high latitudes; the vertical resolution is 10m at depths shallower than 220m.

[0019] A device for replacing the air-sea coupling parameterization in an atmospheric waveguide forecast model based on physical constraint deep learning, the improvement of which is that it comprises: a processor; a memory in which executable instructions of the processor are stored; wherein the processor is configured to execute the steps of the above method by executing the executable instructions.

[0020] A computer-readable storage medium is used to store a program, wherein the improvement lies in that the steps of the above method are implemented when the program is executed.

[0021] The beneficial effects of the present invention are:

[0022] The method disclosed in the present invention can reduce the occupancy rate of computing resources while improving the operation speed and model forecasting performance, providing new ideas for the application of artificial intelligence in the field of meteorology and a new technical development direction for innovation in the field of meteorology. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 It is a schematic flow diagram of the method of the present invention;

[0024] Figure 2 It is an algorithm model diagram. DETAILED DESCRIPTION

[0025] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0026] Embodiment 1, this embodiment discloses a method for replacing the sea-air coupling parameterization in the atmospheric waveguide prediction model based on physical constraint deep learning, such as Figure 1 , as shown in 2, comprising the following steps:

[0027] Step 1, training data preparation:

[0028] The neural network is trained based on the existing ocean mooring datasets TAO and PIRATA, which include: Ocean stratification index N 2 、Ocean shear S 2 , density ρ and velocity U, the output is the ocean turbulence parameterization value K in the atmospheric model T ;

[0029] Step 2: Add physical constraints to improve generalization ability:

[0030] Turbulence data sources are few and sparsely distributed. In order to further improve the generalization ability of the model, the present invention also uses turbulence index data generated by traditional physical methods. By fitting this part of data, the model's understanding of the existing turbulence mechanism can be significantly increased. The specific formula is as follows:

[0031]

[0032] In the above formula, Ri = N 2 / S 2 ;

[0033] Step 3: Build a network for calculation;

[0034] Based on the universal approximation theorem, it is feasible to parameterize shear-driven mixing using deep learning methods. The sparse coverage of observations will reduce the generalization ability of deep learning parameterization.

[0035] Physics-informed deep learning is a promising approach to address this challenge by introducing physical constraints into neural networks. Knowledge is directly acquired from turbulence observations. To improve the generalization of parameterizations, traditional physics-driven parameterizations are used as physical constraints for deep learning applications. Physically constrained deep learning models are constructed to replace the air-sea coupled parameterization scheme in the atmospheric waveguide forecast model.

[0036] Firstly, candidate variables with strong correlation with the vertical eddy diffusion coefficient KT of the ocean model are constructed, including temperature T(z), salinity SA(z) and ocean current U(z) profiles;

[0037] Secondly, in order to better understand the impact of the ocean-air coupling process on the vertical eddy diffusion coefficient KT of the ocean model, an additional parameter Ri is included in the neural network-based parameterization, so the input vector contains four characteristic temperatures T(z), salinity SA(z), ocean current U(z) and Ri, and 850 training samples based on physical-driven parameterization are added to the training dataset;

[0038] Finally, by introducing physical constraints into the neural network, a deep learning model based on physical information is built, and the mapping relationship between the ocean model vertical eddy diffusion coefficient KT, vertical eddy diffusion viscosity Kv and the complex information hybrid parameterization scheme is constructed, thereby improving the generalization of parameterization;

[0039] Parameterization scheme for turbulence calculations:

[0040] The turbulent Prandtl number Pr = Kv / KT, which defines the ratio of the vertical viscosity coefficient to the vertical diffusion coefficient. For stable stratified shear flow, Pr is close to 1 at low Ri and increases with the increase of Ri. Therefore, different empirical parameterization schemes are proposed based on Ri. For example, Pr = 0.8 + 3Ri based on large eddy simulation; based on microstructure observations in the tropical eastern Pacific, Pr = 1 + 5Ri. In the present invention, Pr = min (1 + 5Ri, 10), not only because it comes from observations of the ocean in an atmospheric waveguide environment, but also because it has been adopted by many ocean circulation models.

[0041] Step 4: Verify the model:

[0042] In order to test the performance of the parameterization improvement based on the neural network, a pure ocean simulation based on MOM5 was carried out in the ocean simulation. The MOM model was used to model the global ocean, and the relevant parameters were obtained and substituted into the model for verification. The horizontal resolution of MOM5 is 1° in longitude, 1 / 3° at the equator, and 1° at high latitudes; the vertical resolution is 10m at a depth of 220m, and there are 50 vertical levels from the sea surface to a depth of 220m, with a total of 50 layers.

[0043] In the control experiment (KPP experiment), KPP was used to parameterize boundary layer mixing, driven mixing, and double diffusion mixing. In the sensitivity run (NN run), the driven mixing component of KPP was replaced by a deep learning-based parameterization in the tropical Pacific (10°S~10°N, 120°E~80°W). Both runs were initialized with observed January temperature and salinity and integrated using data from 1959 to 2017. Model outputs from the past 13 years (2005-2017) were used for the analysis. In the climate simulation, a coupled simulation based on CM2.1 was performed. The ocean component was MOM5 with exactly the same configuration as the pure ocean simulation. The horizontal resolution of the atmospheric component AM2.1 was 2.0° latitude × 2.5° longitude with 24 vertical levels, and the horizontal resolution of the land component LM2.1 was the same as AM2.1. The ocean, atmosphere, and land components exchange fluxes every 2 hours.

[0044] This embodiment also discloses a device for replacing the sea-air coupling parameterization in the atmospheric waveguide forecast model based on physical constraint deep learning, comprising: a processor; a memory, in which executable instructions of the processor are stored; wherein the processor is configured to execute the steps of the above method by executing the executable instructions.

[0045] This embodiment also discloses a computer-readable storage medium for storing a program, which implements the steps of the above method when executed.

Claims

1. A method for replacing the air-sea coupling parameterization in the atmospheric waveguide forecast model based on physical constraint deep learning, characterized in that: The steps include: Step 1, training data preparation: The neural network is trained based on the existing ocean mooring datasets TAO and PIRATA, which include: Ocean stratification index N 2 、Ocean shear S 2 , density ρ and velocity U, the output is the ocean turbulence parameterization value K in the atmospheric model T ; Step 2: Add physical constraints to improve generalization ability: The turbulence index data generated by traditional physical methods is calculated as follows: In the above formula, Ri = N 2 / S 2 ; Step 3: Build a network for calculation; First, candidate variables are constructed, including temperature T, salinity SA, and ocean current U profile; Secondly, the input vector contains four characteristic temperature T, salinity SA, ocean current U and Ri, and 850 training samples based on physical driving parameterization are added to the training dataset; Finally, by introducing physical constraints into the neural network, a deep learning model based on physical information was built to construct the mapping relationship between the ocean model vertical eddy diffusion coefficient KT, vertical eddy diffusion viscosity Kv and the complex information hybrid parameterization scheme; Step 4: Verify the model: The MOM model is used to model the global ocean, and relevant parameters are obtained and substituted into the model for verification. The horizontal resolution of MOM5 is 1° in longitude, 1 / 3° at the equator, and 1° at high latitudes; the vertical resolution is 10m at depths shallower than 220m.

2. A device for replacing the air-sea coupling parameterization in the atmospheric waveguide forecast model based on physical constraint deep learning, characterized in that: include: A processor; a memory in which executable instructions of the processor are stored; wherein the processor is configured to perform the steps of the method of claim 1 by executing the executable instructions.

3. A computer-readable storage medium for storing a program, characterized in that: When the program is executed, the steps of the method according to claim 1 are implemented.

Images