CN114896826A - Planet boundary layer parameterization method based on physics and residual error attention network - Google Patents

Abstract

The invention relates to a planet boundary layer parameterization method based on physics combined with a residual attention network, which designs a residual attention network simultaneously combined with physics and deep learning, and comprises eight space-time combined modules which are sequentially connected, wherein each space-time combined module comprises a channel attention module CAB and a space attention module PNSAB, important characteristics are extracted and a specific physical process is simulated through a convolution module CAB module and the PNSAB module, the CAB module is used for simulating energy exchange in an atmospheric turbulence process, the network can continuously learn atmospheric physical common knowledge, the PNSAB module selects meteorological elements which are important for a prediction result, parameterization in a traditional mode can be effectively replaced, and a plurality of designed modules cooperatively simulate the physical process in a boundary layer. The model of the invention can better simulate the velocity, temperature, wind speed and vertical distribution of water vapor in the boundary layer, and simultaneously uses lower calculation cost and shorter time.

Description

Planet boundary layer parameterization method based on physics and residual error attention network

Technical Field

The invention relates to the field of meteorological forecast numerical calculation, in particular to a planet boundary layer parameterization method based on physics and a residual error attention network.

Background

Solar radiation is the main power source of atmospheric motion, and after the lower cushion surface absorbs the solar radiation, the heat exchange between the lower cushion surface and the free atmosphere is mainly completed through turbulent heat transmission of the boundary layer. The transport of matter and water vapor within the boundary layer plays a very important role in the formation of clouds and precipitation, the diffusion of pollutants, and atmospheric chemical changes. Therefore, a reasonable description of the physical processes within the boundary layer is essential in the numerical model. However, due to the vertical and horizontal resolution limitations of the modes themselves and the complexity of turbulent motion within the boundary layer, some physical processes within the boundary layer cannot be described by the dynamic framework in the modes, and therefore, need to be parameterized.

In the numerical mode, boundary layer parameterization mainly solves the problem of closing vertical turbulence flux terms. It represents these physical processes in the form of a system of equations. There are also different parameterization schemes to represent these physical processes. Therefore, boundary layer structural features obtained by different scheme simulations of different parametric scheme simulations are greatly different. Experiments prove that the temperature and the ground wind speed can be better simulated by the non-local BLK scheme, the boundary layer mixing and the entrainment in the daytime are strong and are also strong, which are simulated by the non-local YSU and ACM2 schemes, so that higher temperature and lower humidity are simulated, and the temperature and the humidity which are simulated are lower and higher due to weaker mixing and entrainment in the local MYJ scheme. The mixing effect of YSU is stronger than that of ACM2 and MYJ schemes at night, and the simulated temperature is higher and the humidity is lower. Overall, the non-local YSU, ACM2 scheme is better than the local MYJ scheme. The non-local BLK scheme is used in the MM5 mode to simulate the climate conditions of a region. And YSU, ACM2 and MYJ are used in WRF mode for simulations of various regions. Therefore, the current common method is to use the WRF mode to simulate the situation of each region, but the simulation performance of different boundary layer parameterizations in different regions is very different. Furthermore, the WRF model is a constantly evolving and newer mesoscale model with new boundary layer parameterizations for almost every version.

At present, for strong convection weather forecast, deep learning and various meteorological big data are effectively combined to form an effective tool. For example, based on factors such as cloud top height and cloud top temperature detected by a MSG satellite rotation enhanced visible light and infrared imager (SEVIRI), a scholars adopts random forests or adopts two machine learning algorithms of a neural network and an SVM to carry out quantitative precipitation estimation. Based on the research, many people also use physical knowledge to guide the design of the parameterized network structure of the substitute boundary layer, which also lays the research foundation of the invention.

The prior art scheme has the following defects:

1. computing resource constraints

For the current global and regional numerical prediction mode business application, due to the limitation of computing resources, a large vortex simulation method using a resolution of a hundred meter scale in business in a short period is difficult to realize. The time-limited boundary layer parameterization behaves in the form of a set of equations in the model, so that a considerable amount of time is spent in the simulation, which may miss significant climate changes.

2. Instability of the gas

Since the pattern prediction has different sensitivities to different physical parameterization schemes, there is no unified parameterization scheme, so that it is clearly unreasonable to select different parameterization schemes according to the real-time situation.

3. Resolution limit

With the increasing demand for climate prediction, the resolution is also increased accordingly. However, the current numerical model cannot meet the requirements of weather forecasting, so a new method is needed for realizing the climate prediction.

4. Problems of physical processes

Although there are methods that use machine learning instead of specific parameterization, they do not fully consider the physical processes present in the boundary layer and therefore do not achieve sufficient accuracy.

Disclosure of Invention

Aiming at the defects of the prior art, the invention provides a planet boundary layer parameterization method based on physics combined with a residual attention network, which designs the residual attention network combined with physics and deep learning, learns the characteristics of input parameter data through a channel attention module and a space attention module in eight space-time combined modules which are sequentially connected, and specifically comprises the following steps:

step 1: acquiring a boundary layer parameterized data set, installing a mesoscale weather forecast mode, namely a WRF mode, on a server, and acquiring the boundary layer parameterized data set through the WRF mode, wherein the boundary layer parameterized data set comprises a first data set and a second data set, the first data of the first data set comprises 16 meteorological element parameter data including ground heat flux and height of a planetary boundary layer of each layer of planetary boundary layer, and the second data of the second data set is meteorological element attribute parameters of each layer of boundary layer output after the first data is input into the WRF mode;

each layer of the 17 atmospheric boundary layers comprises a first data set and a second data set corresponding to the first data set;

step 2: dividing the boundary layer parameterized data set into a training set, a verification set and a test set according to an appointed proportion;

and step 3: preprocessing the first data set to enable the first data to obtain dimensions of 17 x 16 and be more suitable for the structure of a parameterized network;

and 4, step 4: constructing a parameterized network based on a residual attention network, and sending the first data set processed in the step 3 into the constructed parameterized network for training, wherein the parameterized network comprises eight space-time combination modules TSI connected in sequence, each space-time combination module comprises a channel attention module CAB and a space attention module PNSAB, the channel attention module CAB is used for simulating energy exchange between vertical sections and learning the internal relation of the vertical sections, and the space attention module PNSAB is used for extracting important input meteorological variables, and specifically comprises the following steps:

step 41: obtaining a first dimension reduction characteristic by passing first data of a first data set in a training set through a convolution layer;

step 42: feeding the first dimensionality reduction feature into a first time-space bonding module (TSI) ₁ The first time-space bonding module TSI ₁ The first two convolution layers process the first dimension reduction characteristic to obtain a second dimension reduction characteristic;

step 43: feeding the second dimension-reduced feature to the channel attention module CAB and the spatial attention module PNSAB, respectively, at the same time;

the channel attention module CAB comprises two branches: the second dimension reduction feature obtains a profile global feature through the global feature extractor AAP, obtains a profile local feature through the local feature extractor AMP, adds the profile global feature and the profile local feature, and obtains a first profile attention matrix through an activation function;

the spatial attention module PNSAB comprises two branches: the second dimension reduction feature obtains important meteorological element features through the key feature extractor, interfering meteorological element features are removed through the feature suppressor, the output result of the key feature extractor and the output result of the feature suppressor are spliced through a channel, and then a first meteorological element attention matrix is obtained through a convolution layer;

finally, multiplying the first section attention matrix and the first meteorological element attention moment matrix to obtain a first global attention distribution matrix;

step 44: obtaining a reconstructed global attention distribution matrix by passing the first global attention distribution matrix through a convolutional layer and an active layer, and then multiplying the reconstructed global attention distribution matrix by the second dimension reduction feature obtained in the step 42 to obtain a global feature;

step 45: adding the global feature and the first dimension reduction feature obtained in the step 41 to obtain a first reconstructed global feature;

step 46: the first reconstructed global feature of the step 45 is sequentially sent to the remaining seven space-time combination modules, the structure of each space-time combination module is the same, the operations from the step 42 to the step 45 are repeated for seven times, finally, the eighth space-time combination module outputs the final eighth reconstructed global feature, and the eighth reconstructed global feature passes through a full connection layer to obtain the final predicted second data;

step 47: continuously updating the training network by taking data with the batch size equal to 256 as an iterative training network;

and 48: if a batch of complete data is trained, namely all data are sent to the model for training, verifying the model and storing the model;

step 49: judging whether the total times of training iteration is reached, wherein the training index is the root mean square error RMSE, if the RMSE values which are continuously updated for three times are not updated, ending the training, and storing the training model which updates the RMSE values for the last time, otherwise, returning to the step 41;

and 5: and inputting the trained network storage parameters into a first data set of the test set for testing to obtain second test data.

According to a preferred embodiment, the pretreatment step of step 3 specifically comprises:

step 31: sampling the few classes by adopting a synthetic minority class oversampling technology SMOTE to expand the few classes and enhance data, balancing class distribution and avoiding the overfitting problem in training;

step 32: standardizing the first data set processed in the step 31 to enable the first data set to be in accordance with standard normal distribution;

step 33: the first data is processed into 17 x 16 dimensions, which is convenient for convolution to extract features.

The invention has the beneficial effects that:

1. although traditional parameterization schemes simplify physical processes, the computation of these physical processes still requires significant computational resources and computation time. The deep learning model has stronger data processing capability, can generate relatively accurate results in reasonable time, saves a large amount of resources and greatly improves the calculation efficiency.

2. The numerical pattern is a pattern that uses gridding to compute the result. Thus limiting the resolution to some extent. The invention provides a parameterized model of the residual attention network to obtain corresponding output resolution according to the input data resolution, so that if finer resolution is input, the finer resolution is also obtained, and the parameterized model can break through the barrier of resolution and realize finer resolution output.

3. Conventional numerical models require different parameterization schemes to be chosen depending on the situation (e.g., location). The model of the invention can fit the prediction result according to a large amount of data sets, does not need to select various parameterization schemes, and can adapt to the situation of the current place according to the distribution of the data.

4. Compared with the previous machine learning method, the method of the invention designs the channel attention module CAB and the space attention module PNSAB to simulate the specific physical process and physical reality in the boundary layer, the CAB module is used for simulating the energy exchange in the turbulent flow process, so that the network can continuously learn the physical common knowledge, and the PNSAB module selects meteorological elements which are important for the prediction result, so that the predicted second data is more accurate, and the prediction reliability is greatly improved.

5. The parameterized model of the invention is trained, and the traditional parameterization needs to be calculated on line, so the parameterized model of the invention has almost real-time prediction results, is particularly suitable for medium-short term and sudden prediction in daily life, and has the advantages of low cost and quick response.

Drawings

Fig. 1 is a schematic structural diagram of a PBLNet network according to the present invention;

FIG. 2 is a schematic diagram of the structure of channel attention module CBA;

fig. 3 is a schematic structural diagram of the spatial attention module PNSAB;

FIG. 4 (a), (b) and (c) are graphs showing the variation of the RMSE and R indices of U, V, W in 17 vertical cross sections, respectively;

fig. 5 (a) and (b) are graphs showing the variation of RMSE and R indices of tK and qvapro in 17 vertical cross sections, respectively.

Detailed Description

In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention will be described in further detail with reference to the accompanying drawings in conjunction with the following detailed description. It should be understood that the description is intended to be exemplary only, and is not intended to limit the scope of the present invention. Moreover, in the following description, descriptions of well-known structures and techniques are omitted so as to not unnecessarily obscure the concepts of the present invention.

The following detailed description is made with reference to the accompanying drawings.

In the present invention, U, V, W, tK, QVAPOR represent the latitudinal wind, the longitudinal wind, the vertical speed,

Temperature, water-vapor mixing ratio.

Aiming at the defects of the prior art, the invention provides a planet boundary layer parameterization method based on physics combined with a residual attention network, and FIG. 1 is a schematic structural diagram of the PBLNet network. The method specifically comprises the following steps:

step 1: the method comprises the steps of obtaining a boundary layer parameterized data set, installing a mesoscale weather forecast mode, namely a WRF mode, on a server, obtaining the boundary layer parameterized data set through the WRF mode, wherein the boundary layer parameterized data set comprises a first data set and a second data set, the first data of the first data set comprises 16 meteorological element parameter data including ground heat flux and height of a planetary boundary layer of each layer of planetary boundary layer, and the second data of the second data set is meteorological element attribute parameters of each layer of boundary layer output after the WRF mode is input into the first data set.

The first data of the first data set specifically includes: a specific humidity Q2 at a height of 2m above ground, a temperature T2 at a height of 2m above ground, a latitudinal component U10 of a wind field of 10m above ground, a longitudinal component V10 of a wind field of 10m above ground, a ground heat flux GRDFLX, a short wave flux SWDOWN from ground, a long wave radiation flux GLW down at ground, a latent heat flux LH at ground surface, a heat flux (sensible heat flux) HFX up at ground surface, a planetary boundary layer height PBLH, a surface friction velocity UST, a skin sea surface temperature TSK, a soil temperature TSLB of 2 meters underground, a soil humidity SMIOS of 0-0.3 cm underground, a wind-turning latitudinal component Ug at 700hPa, a wind-turning longitudinal component Vg at 700 hPa.

The second data of the second data set comprises: warp wind U, weft wind V, vertical velocity W, temperature tK, water-vapor mixture ratio QVAPOR.

Each layer of the 17 boundary layers includes a first data set and a second data set corresponding thereto.

Step 2: dividing the boundary layer parameterized data set into a training set, a verification set and a test set according to an appointed proportion;

and step 3: preprocessing the first data set, specifically including:

step 31: sampling the few classes by adopting a synthetic minority class oversampling technology SMOTE to expand the few classes and enhance data, balancing class distribution and avoiding the overfitting problem in training;

step 32: carrying out standardization processing on the data set processed in the step 31 to enable the data set to be in accordance with standard normal distribution;

step 33: the first data set is processed into dimensions of 17 x 16, which facilitates convolution to extract features. This is due to the boundary layer having 17 layers, each layer having 16 parameters of the first data.

And 4, step 4: constructing a parameterized network, and sending the first data set preprocessed in the step 3 into the constructed parameterized network for training, wherein the parameterized network comprises eight space-time combination modules TSI connected in sequence, each space-time combination module comprises a channel attention module CAB and a space attention module PNSAB, and important features are extracted and a specific physical process is simulated through a convolution module CAB module and the PNSAB module, and the method specifically comprises the following steps:

step 41: obtaining a first dimension reduction characteristic by passing first data of a first data set in a training set through a convolution layer;

step 42: feeding the first dimensionality reduction feature into a first time-space bonding module (TSI) ₁ The first time-space bonding module TSI ₁ The first two convolution layers process the first dimension reduction feature to obtain a second dimension reduction feature.

In the planetary boundary layer, energy exchange exists among the sections due to the turbulent flow process, but the energy exchanged by the adjacent vertical sections is more frequent, so that the adjacent sections are more important for predicting the result. Of course, different meteorological elements are also very important for the prediction result in the same section. Based on the common sense of physics, the CAB module and the PNSAB module are designed in the invention. The CAB module is used to simulate the energy exchange during such turbulence, allowing the network to constantly learn this physical knowledge. The PNSAB module is used for selecting meteorological elements which are important for the prediction result. And the CAB module and the PNSAB module make the network pay more attention to the adjacent section and the important meteorological elements, the invention is realized by an attention mechanism,

step 43: feeding the second dimension-reduced feature to the channel attention module CAB and the spatial attention module PNSAB, respectively, at the same time:

the channel attention module CAB comprises two branches: a global feature extractor AAP (right side of fig. 2) and a local feature extractor AMP (left side of fig. 2). In the planetary boundary layer, there is an exchange of energy between the profiles due to the turbulent flow process, but the energy exchanged by the adjacent vertical profiles is more and more frequent. These two branches work together to learn the turbulence process and energy exchange between the vertical sections. And obtaining the global feature of the profile through a global feature extractor by the second dimension reduction feature, wherein the global feature extractor explicitly considers the global structure of the vertical profile to obtain the global compression feature of the vertical profile. The local feature extractor is used for learning nearest neighbor relation between the vertical sections, the section global feature and the section local feature are added, then a larger weight is generated for the closer vertical section, a smaller weight is generated for the farther vertical section, and a first section attention matrix is obtained through an activation function. Moreover, a high turbulent mixing rate exists between the adjacent vertical sections, so that the energy transfer efficiency can be improved. Thus, the modules of CAB are multifunctional in the network architecture of the present invention.

The spatial attention module PNSAB is also divided into two parts: the key feature extractor (top of fig. 3) and the feature suppressor (bottom of fig. 3), which together give a large weight to the features of the vital meteorological elements, which are the first dataset, i.e. the 16 meteorological elements. And obtaining important meteorological element characteristics by the second dimension reduction characteristics through a key characteristic extractor, removing the interfering meteorological element characteristics through a characteristic suppressor, splicing the output result of the key characteristic extractor and the output result of the characteristic suppressor through a channel, and then obtaining a first meteorological element attention matrix through a convolution layer. And multiplying the first section attention characteristic and the first meteorological element attention characteristic to obtain a first global attention distribution matrix.

In addition to the effects between vertical profiles, the predictor variables of the present invention are also closely related to meteorological variables such as latent heat flux, specific humidity at 2m height above ground, and therefore the present invention designs the PNSAB module to give more attention to important variables and to give higher weight to important meteorological variable features. Meanwhile, in order to highlight more important elements, the method also removes the interfering meteorological variables through the negative value characteristic. The PNSAB and CAB adopt the same structure and are also divided into two parts, a key feature extractor (upper part of fig. 3) and a feature suppressor (lower part of fig. 3). The key feature extractor is used for extracting important meteorological elements and simultaneously considering the elements with smaller influence factors. And the feature suppressor removes the interfering features by inverting the features. These two branches work together to highlight important meteorological elements, emphasizing those features that should be given large weight and which should be removed, respectively, and finally get the final output through the fully connected layer.

Step 44: and obtaining a reconstructed global attention distribution matrix by passing the first global attention distribution matrix through a convolutional layer and an active layer, and then multiplying the reconstructed global attention distribution matrix by the second dimension-reducing feature obtained in the step 42 to obtain a global feature. Namely, a residual structure is adopted, and the residual structure is used in the parameterized network to prevent the network from degrading due to the fact that the parameterized network has a deeper layer number.

Step 45: and adding the global feature and the first dimension reduction feature obtained in the step 41 to obtain a first reconstructed global feature.

Step 46: and (4) sequentially sending the first reconstructed global features of the step (45) to the rest seven space-time combination modules, wherein the structure of each space-time combination module is the same, repeating the operations from the step (42) to the step (45) for seven times, finally outputting the final eighth reconstructed global features by an eighth space-time combination module, and obtaining the final predicted second data by the eighth reconstructed global features through a full connection layer.

Step 47: continuously updating the training network by taking data with the batch size equal to 256 as an iterative training network;

and 48: if a batch of complete data is trained, namely all data are sent to the model for training, verifying the model and storing the model;

step 49: judging whether the total times of training iteration is reached, wherein the training index is Root Mean Square Error (RMSE), if the RMSE values of three consecutive times are not updated, namely the RMSE values of the last three consecutive times are all larger than the RMSE value of the fourth last time, and the RMSE values are not updated, ending the training, and storing the training model of which the RMSE values are updated for the last time, namely the training model of the fourth last time, otherwise, returning to the step 41;

and 5: and inputting the trained network storage parameters into a first data set of the test set for testing to obtain second test data.

In order to more fully prove the effectiveness of the method, the method is compared with the prior art, and the adopted quantitative evaluation indexes comprise: root mean square error RMSE, mean absolute error MAE, R2, PCC. R2 refers to the correlation between the real value and the predicted value, and the numerical range is 0-1, and the higher the score is, the more accurate the prediction is. RMSE is used to measure the error, with smaller values representing better results. PCC means: PCC refers to the correlation coefficient of two vectors of predicted values and true values. The closer the correlation coefficient is to 1, the stronger the correlation, the closer the correlation coefficient is to 0, and the weaker the correlation.

Fig. 4 is a graph showing the variation of the quantitative index of U, V, W in 17 vertical cross sections, and fig. 5 is a graph showing the variation of the quantitative indices of tK and qvaper in 17 vertical cross sections. The PBLNet of the invention has 16 input variables in total, but not all of the input variables have an effect on vertical profile variable prediction, and the vertical profile input variables of different heights have different degrees of importance. CAB assigns different weights to each channel to make each vertical profile more closely related. Meanwhile, PNSAB strengthens important input variables to improve the prediction effect. Experimental results prove that the method provided by the invention has obvious effect.

As can be seen from fig. 4 and 5: as PBL height increases, RMSE of U, V, W, tK and qvapro becomes higher and R2 becomes lower, indicating that the near-surface variation has a greater effect on lower vertical profiles. And it is readily seen that the RMSE of M1 and M2 increases sharply with increasing height. M3 maintained the same trend on fig. 4 and 5 as PBLNet, indicating that the CAB module is more important than PNSAB. As can be seen from fig. 4 and 5, the CAB module provided in the present invention can capture the neighborhood relationship, so as to better simulate the energy exchange between vertical sections and effectively reflect the physical process.

M1 is the basic model, ResNet, no CAB and PNSAB module are added, M2 is the basic model with PNSAB module added. M3 is a parameterized network provided by the invention, wherein CAB module is added on the basic model, and PBLNet is a parameterized network provided by the invention, wherein PNSAB module and CAB module are added on the basic model.

To validate the effectiveness of PBLNet, the present invention compares it to several recent methods, FFN, HPC, HAC, DeepPE, quantitatively on the same test dataset. Table 1 shows a quantitative comparison of the different methods in U, V, W. FFN, HPC, HAC, DeepPE are the results of other parameterization schemes; PBLNet is the result of the proposed method of the invention. To verify the effectiveness of PBLNet, it was compared in experiments with several recent methods, FFN, HPC, HAC, deep pe. And a quantitative comparison was performed on the same test data set. Tables 1 and 2 show the experimental results of PBLNet and the most advanced method. It can be seen that the method of the present invention achieves the lowest RMSE and MAE, and higher R2 and PCC for all output variables, including U, V, W, tK, qvapro. This means that the model of the invention has a higher accuracy and is closer to the observed values. Specifically, PBLNet improved in RMSE of U, V, W, tK and qvaper by 23.7%, 25.7%, 20.2%, 36.4%, and 42.7%, respectively, compared to the better network DeepPE. The greatest improvement was 42.7% over qvapro, indicating that the CAB of the present invention well simulates turbulent motion of vertical heat and water vapor exchange. It is also worth noting that the model of the method of the present invention improves U, V, W, tK, QVAPOR more significantly than DeepPE. These show that PBLNet has more significant advantages in energy delivery and selection of important features, better reflecting the actual motion of the satellite boundary layer in real life.

TABLE 1

Table 2 shows the quantitative comparison of the different methods in tK, qvapro.

TABLE 2

Tables 1 and 2 show the average results of the different model experiments, and the last row shows that PBLNet achieves the best performance among all the evaluation indices of the five predictor variables. All these statements indicate that PBLNet has better vertical blend fit capability.

Table 3 ablation experiments at U, V, W, and tables 3 and 4 were performed to verify the effect of CAB and PNSAB in fig. 4, and it can be seen that M1, M2 and M3, although better results than FFN, HPC, HAC, deep pe were also achieved. As can be seen from tables 3 and 4, M1, M2, M3 still did not perform as well as PBLNets. This further shows that the CAB and PNSAB proposed by the present invention make a very large contribution in the parameterization process. Specifically, when CAB and PNSAB are removed, it means that interdependencies between vertical sections, physical processes, near-surface variables and meteorological variables are not taken into account. In addition, M3 gave better results than M2, which also indicates that CAB plays a greater role than PNSAB. It can be concluded that the intrinsic link between turbulent transport and vertical profile is more critical than near surface variables. Ablation experimental results verify that the CAB and PNSAB modules of the invention can promote the model to analyze the physical process of planetary boundary layer parameterization.

TABLE 3

As can be seen from Table 3, as the planet boundary layer height increases, the RMSE of U, V, W, tK and QVAPOR becomes higher and R2 becomes lower, indicating that the near-ground variables have a greater effect on the lower layers. Meanwhile, in order to verify the effectiveness of the CAB and PNSAB modules and further illustrate the influence relationship between different vertical sections, experiments resulted in the variation of RMSE and R2 of 5 output variables of each section over 17 vertical sections. It is readily seen that the RMSE of M1 and M2 increases sharply with increasing height. M3 maintained the same trend on the graph as PBLNet, indicating that CAB was more important than PNSAB. As can be seen from tables 3 and 4 and fig. 4 and 5, the CAB module can capture the neighborhood relationship, so that the energy exchange between vertical sections can be better simulated, and the physical reality can be effectively reflected. In summary, the foregoing ablation experiments demonstrate that the PBLNet scheme, CAB and PNSAB modules designed by the present invention are very effective for planetary boundary layer parameterization.

Table 4 ablation experiments in tK, qvapro.

TABLE 4

It should be noted that the above-mentioned embodiments are exemplary, and that those skilled in the art, having benefit of the present disclosure, may devise various arrangements that are within the scope of the present disclosure and that fall within the scope of the invention. It should be understood by those skilled in the art that the present specification and figures are illustrative only and are not limiting upon the claims. The scope of the invention is defined by the claims and their equivalents.

Claims (2)

1. The planet boundary layer parameterization method based on physics and residual error attention network is characterized in that a residual error attention network combining physics and deep learning is designed, and features of input parameter data are learned through a channel attention module and a space attention module in eight space-time combination modules which are sequentially connected, and the method specifically comprises the following steps:

step 1: acquiring a boundary layer parameterized data set, installing a mesoscale weather forecast mode, namely a WRF mode, on a server, and acquiring the boundary layer parameterized data set through the WRF mode, wherein the boundary layer parameterized data set comprises a first data set and a second data set, the first data of the first data set comprises 16 meteorological element parameter data including ground heat flux and height of a planetary boundary layer of each layer of planetary boundary layer, and the second data of the second data set is meteorological element attribute parameters of each layer of boundary layer output after the first data is input into the WRF mode;

each layer of the 17 atmospheric boundary layers comprises a first data set and a second data set corresponding to the first data set;

step 2: dividing the boundary layer parameterized data set into a training set, a verification set and a test set according to an appointed proportion;

and step 3: preprocessing the first data set to enable the first data to obtain dimensions of 17 x 16 and be more suitable for the structure of a parameterized network;

and 4, step 4: constructing a parameterized network based on a residual attention network, and sending the first data set processed in the step 3 into the constructed parameterized network for training, wherein the parameterized network comprises eight space-time combination modules TSI connected in sequence, each space-time combination module comprises a channel attention module CAB and a space attention module PNSAB, the channel attention module CAB is used for simulating energy exchange between vertical sections and learning the internal relation of the vertical sections, and the space attention module PNSAB is used for extracting important input meteorological variables, and specifically comprises the following steps:

step 41: obtaining a first dimension reduction characteristic by passing first data of a first data set in a training set through a convolution layer;

step 42: feeding the first dimensionality reduction feature into a first time-space bonding module (TSI) ₁ The first time-space bonding module TSI ₁ The first two convolution layers process the first dimension reduction characteristic to obtain a second dimension reduction characteristic;

step 43: feeding the second dimension-reduced feature to the channel attention module CAB and the spatial attention module PNSAB, respectively, at the same time;

the channel attention module CAB comprises two branches: the second dimension reduction feature obtains a profile global feature through the global feature extractor AAP, obtains a profile local feature through the local feature extractor AMP, adds the profile global feature and the profile local feature, and obtains a first profile attention matrix through an activation function;

the spatial attention module PNSAB comprises two branches: the second dimension reduction feature obtains important meteorological element features through the key feature extractor, the interfering meteorological element features are removed through the feature suppressor, the output result of the key feature extractor and the output result of the feature suppressor are spliced through a channel, and then a first meteorological element attention matrix is obtained through a convolution layer;

finally, multiplying the first section attention matrix and the first meteorological element attention moment matrix to obtain a first global attention distribution matrix;

step 44: obtaining a reconstructed global attention distribution matrix by passing the first global attention distribution matrix through a convolutional layer and an active layer, and then multiplying the reconstructed global attention distribution matrix by the second dimension reduction feature obtained in the step 42 to obtain a global feature;

step 45: adding the global feature and the first dimension reduction feature obtained in the step 41 to obtain a first reconstructed global feature;

step 46: the first reconstructed global feature of the step 45 is sequentially sent to the remaining seven space-time combination modules, the structure of each space-time combination module is the same, the operations from the step 42 to the step 45 are repeated for seven times, finally, the eighth space-time combination module outputs the final eighth reconstructed global feature, and the eighth reconstructed global feature passes through a full connection layer to obtain the final predicted second data;

step 47: continuously updating the training network by taking data with the batch size equal to 256 as an iterative training network;

and 48: if a batch of complete data is trained, namely all data are sent to the model for training, verifying the model and storing the model;

step 49: judging whether the total times of training iteration is reached, wherein the training index is the root mean square error RMSE, if the RMSE values which are continuously updated for three times are not updated, ending the training, and storing the training model which updates the RMSE values for the last time, otherwise, returning to the step 41;

and 5: and inputting the trained network storage parameters into a first data set of the test set for testing to obtain second test data.

2. The planetary boundary layer parameterization method according to claim 1, wherein the preprocessing step of step 3 specifically comprises:

step 31: sampling the few classes by adopting a synthetic minority class oversampling technology SMOTE to expand the few classes and enhance data, balancing class distribution and avoiding the overfitting problem in training;

step 32: standardizing the first data set processed in the step 31 to enable the first data set to be in accordance with standard normal distribution;

step 33: the first data is processed into dimensions of 17 x 16, so that the feature extraction by convolution is facilitated.

Images