Construction method and system of machine learning satellite observation operator

Abstract

The invention provides a construction method and system for a machine learning satellite observation operator, and the method comprises the steps: collecting the historical observation data of a satellite and the corresponding historical reanalysis data; taking the historical reanalysis data as input data of a deep learning model, taking historical observation data of a satellite as output data of the deep learning model, and training the deep learning model to obtain a trained artificial intelligence forward observation operator; solving a tangent linear observation operator of the artificial intelligence forward observation operator by using an automatic differential algorithm; solving an adjoint observation operator of the artificial intelligence forward observation operator by using an automatic differential algorithm; and accessing the artificial intelligence forward observation operator, the tangent linear observation operator and the adjoint observation operator into an assimilation system. According to the method, the artificial intelligence tangent linear mode and the adjoint mode are calculated by using automatic differential, so that the consistency of mathematical relationships of the three modes can be ensured, and the development difficulty of the tangent linear mode and the adjoint mode can be reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of satellite observation operator construction, and particularly to a construction method and system for a machine learning satellite observation operator. Background Art

[0002] With the frequent occurrence of global climate change and extreme weather events, accurate weather forecasting and atmospheric monitoring have become increasingly important. And the atmospheric variational data assimilation (VDA) system, as a key technology to improve the accuracy of weather forecasting, its core lies in effectively integrating observation data from multiple sources to optimize the numerical weather prediction model. However, currently, the atmospheric variational data assimilation system still faces many challenges when dealing with complex and high-dimensional satellite observation data. Since satellite observation is an indirect observation, it is necessary to achieve direct assimilation through a fast radiative transfer model. This model uses a combination of physical and statistical methods to map the model forecast space to the satellite observation radiation space. Currently, the fast radiative transfer model has a good simulation effect for clear sky conditions, but for cloud areas, due to the addition of cloud scattering simulation, the calculation speed is slow. Summary of the Invention

[0003] In order to overcome the deficiencies of the prior art, the purpose of the present invention is to provide a construction method and system for a machine learning satellite observation operator.

[0004] To achieve the above purpose, the present invention provides the following solutions:

[0005] A construction method for a machine learning satellite observation operator, comprising:

[0006] Step 1: Collect historical observation data of the satellite and its corresponding historical reanalysis data;

[0007] Step 2: Use the historical reanalysis data as the input data of the deep learning model, and use the historical observation data of the satellite as the output data of the deep learning model, and train the deep learning model to obtain a trained artificial intelligence forward observation operator;

[0008] Step 3: Use the automatic differentiation algorithm to solve the tangent linear observation operator of the artificial intelligence forward observation operator;

[0009] Step 4: Use the automatic differentiation algorithm to solve the adjoint observation operator of the artificial intelligence forward observation operator;

[0010] Step 5: Connect the artificial intelligence forward observation operator, the tangent linear observation operator and the adjoint observation operator to the assimilation system to complete weather forecasting.

[0011] Preferably, the step 3: using the automatic differentiation algorithm to solve the tangent linear observation operator of the artificial intelligence forward observation operator, includes:

[0012] Use the automatic differentiation algorithm to solve the Jacobian matrix of the artificial intelligence forward observation operator;

[0013] Multiply the Jacobian matrix by the perturbed input, and apply the chain rule to obtain the gradient of the output with respect to the perturbed input, and use it as the tangent linear observation operator.

[0014] Preferably, step 4: Use the automatic differentiation algorithm to solve the adjoint observation operator of the artificial intelligence forward observation operator, including:

[0015] Input x 0 into the artificial intelligence forward observation operator to obtain the output of the artificial intelligence forward observation operator;

[0016] Use the automatic differentiation algorithm to calculate the gradient with respect to the input x 0 ;

[0017] Use the gradient of x 0 as the adjoint observation operator of the artificial intelligence forward observation operator.

[0018] Preferably, in step 2, the deep learning model is a ResMLP network model.

[0019] The present invention also provides a construction system for a machine learning satellite observation operator, including:

[0020] A data collection module for collecting historical observation data of the satellite and its corresponding historical reanalysis data;

[0021] A training module for using the historical reanalysis data as the input data of the deep learning model and the historical observation data of the satellite as the output data of the deep learning model to train the deep learning model to obtain a trained artificial intelligence forward observation operator;

[0022] A linear observation operator calculation module for using the automatic differentiation algorithm to solve the tangent linear observation operator of the artificial intelligence forward observation operator;

[0023] An adjoint observation operator calculation module for using the automatic differentiation algorithm to solve the adjoint observation operator of the artificial intelligence forward observation operator;

[0024] An application module for connecting the artificial intelligence forward observation operator, the tangent linear observation operator and the adjoint observation operator to the assimilation system.

[0025] Preferably, in the linear observation operator calculation module, it includes:

[0026] Use the automatic differentiation algorithm to solve the Jacobian matrix of the artificial intelligence forward observation operator;

[0027] Multiply the Jacobian matrix by the perturbation input, and apply the chain rule to obtain the gradient of the output with respect to the perturbation input, and use it as the tangent linear observation operator.

[0028] Preferably, in the adjoint observation operator calculation module, it includes:

[0029] Input x 0 Into the artificial intelligence forward observation operator to obtain the output of the artificial intelligence forward observation operator;

[0030] Using the automatic differentiation algorithm, calculate the gradient with respect to the input x 0 ;

[0031] Take the gradient of x 0 As the adjoint observation operator of the artificial intelligence forward observation operator.

[0032] Preferably, in the step 2, the deep learning model is a ResMLP network model.

[0033] The present invention also provides an electronic device, including a bus, a transceiver, a memory, a processor, and a computer program stored on the memory and executable on the processor. The transceiver, the memory, and the processor are connected through the bus. It is characterized in that when the computer program is executed by the processor, it implements the steps in the above-mentioned method for constructing a machine learning satellite observation operator.

[0034] The present invention also provides a computer-readable storage medium, on which a computer program is stored. It is characterized in that when the computer program is executed by a processor, it implements the steps in the above-mentioned method for constructing a machine learning satellite observation operator.

[0035] The beneficial effect of the method for constructing a machine learning satellite observation operator provided by the present invention is that: compared with the prior art, by using automatic differentiation to calculate the tangent linear mode and adjoint mode of artificial intelligence, it can not only ensure the consistency of the mathematical relationships among the three, but also reduce the development difficulty of the tangent linear mode and adjoint mode. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0037] Figure 1 It is a flowchart of a method for constructing a machine learning satellite observation operator provided by an embodiment. DETAILED DESCRIPTION

[0038] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0039] Referring to "embodiments" herein means that the specific features, structures or characteristics described in connection with the embodiments may be included in at least one embodiment of the present application. The phrase appears at various positions in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment mutually exclusive with other embodiments. Those skilled in the art explicitly and implicitly understand that the embodiments described herein may be combined with other embodiments.

[0040] The terms "first", "second", "third", "fourth", etc. in the specification and claims of the present application and the accompanying drawings are used to distinguish different objects, rather than to describe a specific order. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion. For example, including a series of steps, processes, methods, etc. is not limited to the listed steps, but optionally also includes steps not listed, or optionally also includes other step elements inherent to these processes, methods, products or devices.

[0041] To make the above objects, features and advantages of the present invention more obvious and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0042] Please refer to Figure 1 , a method for constructing a machine learning satellite observation operator, comprising:

[0043] Step 1: Collect historical observation data of the satellite and its corresponding historical reanalysis data;

[0044] Step 2: Use the historical reanalysis data as the input data of the deep learning model, and use the historical observation data of the satellite as the output data of the deep learning model, and train the deep learning model to obtain a trained artificial intelligence forward observation operator;

[0045] Step 3: Use the automatic differentiation algorithm to solve the tangent linear observation operator of the artificial intelligence forward observation operator;

[0046] In the said Step 3, it includes:

[0047] Use the automatic differentiation algorithm to solve the Jacobian matrix of the artificial intelligence forward observation operator;

[0048] Multiply the Jacobian matrix by the perturbation input and apply the chain rule to obtain the gradient of the output with respect to the perturbation input; wherein, the calculation formula for the gradient of the output with respect to the perturbation input is:

[0049]

[0050] wherein, J represents the complete gradient approximation of the deep learning model as the derivative of the model, v represents the perturbation of the input, and J·v represents the gradient of the output with respect to the perturbation input.

[0051] Step 4: Use the automatic differentiation algorithm to solve the adjoint observation operator of the artificial intelligence forward observation operator;

[0052] In the said Step 4, it includes:

[0053] Input x 0 into the artificial intelligence forward observation operator to obtain the output of the artificial intelligence forward observation operator;

[0054] Use the automatic differentiation algorithm to calculate the gradient with respect to the input x 0 ;

[0055] Take the gradient of x 0 as the adjoint observation operator of the artificial intelligence forward observation operator.

[0056] Step 5: Connect the artificial intelligence forward observation operator, the tangent linear observation operator and the adjoint observation operator to the assimilation system.

[0057] Next, the present invention further describes a method for constructing a machine learning satellite observation operator in combination with specific embodiments:

[0058] Figure 1 The flowchart shows the construction of an artificial intelligence satellite observation operator applying an assimilation system in the present invention. This method includes four stages: historical data collection, artificial intelligence model training, artificial intelligence observation operator construction, and connection to the assimilation system.

[0059] Step 001: Historical data collection is mainly for preparing the training set for constructing the artificial intelligence observation operator. Satellite observation data collection: It can include actual satellite observation data, including radiation data obtained from geostationary orbit (GEO) satellites, data obtained from sensors of the Special Sensor Microwave Imager / Sounder (SSMIS) satellites, etc.; or satellite simulation observation data can be generated using traditional satellite observation operators such as the Rapid Transfer Radiative Transfer Model (RTTOV) as the target data for learning.

[0060] Reanalysis data collection: Reanalysis data covers meteorological elements such as air temperature, humidity, air pressure, wind speed, and precipitation, and is sourced from globally recognized reanalysis datasets such as the European Centre for Medium-Range Weather Forecasts' fifth-generation climate data reanalysis (ERA5) and the National Aeronautics and Space Administration (NASA)'s Modern-Era Retrospective Analysis for Research and Applications (MERRA). The reanalysis data is time-matched with satellite observation data and interpolated to the grid points where the observations are located to serve as the input data for the artificial intelligence model, ensuring that the collected data covers the physical parameter space with sufficient resolution.

[0061] Step 002: In the field of hybrid modeling of machine learning and earth science, deep learning models have achieved good results. A deep learning model is a type of artificial neural network model that usually contains multiple hidden layers, used to identify complex pattern features, and automatically extracts the laws of data through training with a large amount of data, so as to achieve efficient modeling and prediction of data.

[0062] In a deep learning model, the multi-layer perceptron model (MLP) has strong stability and fitting ability. The MLP model is a feedforward and supervised deep neural network, consisting of an input layer, hidden layers, and an output layer, and the layers are interconnected through a fully connected manner. The input layer is responsible for receiving the original input features, while the output layer generates the final prediction results. The hidden layers are located between the input layer and the output layer, and the neurons in them receive the output of the previous layer and perform weighted sum and activation function transformation.

[0063] Based on this, the present invention introduces a residual module, which does not consider the longitude and latitude information in the horizontal direction of the vertical grid column, but only uses the data processing method with an independent vertical grid column as a unit. Thus, both the input and output of the data are one-dimensional vectors composed of multiple physical variables, which can fully capture the relationships between all elements in the input vector to implement a deeper deep learning model for learning complex non-linear information in the data. The ResMLP deep learning model uses 7 residual modules, each residual module has 2 fully connected layers with a width of 512, the total depth of the network is 14 layers, and the number of neurons determines the width of this layer. This multi-layer structure can gradually extract higher-level features, thereby improving the learning ability and accuracy. In addition, ResMLP also has higher parameter utilization efficiency and can converge faster.

[0064] Forward propagation is performed using ResMLP, and the output result is calculated through training data and weight parameters. The calculated output result is compared with the true result in the training set, and the loss gradient is calculated through the loss function. Backpropagation calculates the gradient of the loss function with respect to each parameter through the chain rule of derivatives and updates the parameters according to the gradient. After each forward propagation ends, the loss values of consecutive multiple iterations are compared. When the loss value no longer continues to decrease and tends to be stable, it represents that the machine learning training model is approximately converged. The historical meteorological data is converted into a tensor format for calculation. In an automatic mixed-precision manner, some operators use single precision and other parts use half-precision data types, enabling the operators of the neural network to be executed quickly with low precision. At the same time, techniques such as pruning, quantization, and distillation are used to reduce the number of model parameters and the amount of calculation, and the cache is reasonably utilized to reduce memory access latency and improve the inference efficiency, which can be accelerated by more than 30% in online coupling, especially in the simulation of cloud and rain areas.

[0065] Step 003: Since artificial intelligence models are generally trained using deep learning frameworks, when applied, the trained artificial intelligence model needs to be converted into an inference model through static compilation to implement the artificial intelligence observation operator model.

[0066] To ensure the mathematical relationship and correctness between the tangent linear model and the forward model, an automatic differentiation algorithm is used to implement the construction of the tangent linear model of the machine learning model. Automatic differentiation is a core technology in the fields of deep learning and scientific computing. The key lies in decomposing a complex function into a combination of a series of simple functions and then applying the chain rule for differentiation.

[0067] In the process of implementing automatic differentiation, a computational graph needs to be constructed to describe the operation process of the function. Nodes represent the basic operations of the function, and edges represent the data dependencies in the function. Through the computational graph, the derivative of the function can be automatically deduced, thus realizing automatic differentiation. The neural network parameter calculation process is different from numerical differentiation using finite difference approximation and symbolic differentiation for symbolic derivation. It calculates step by step according to a certain calculation order, can accurately calculate the derivative, and at the same time avoids the expression explosion problem of symbolic differentiation and the precision loss of numerical differentiation, as well as the possible errors in manually calculating the derivative.

[0068] After the machine learning forward mode training is completed, automatic differentiation is used to solve the derivative of the input. When there is an n-dimensional input and an m-dimensional output, The complete gradient is a matrix of the derivatives of each output with respect to each input, that is, the Jacobian matrix, which is expressed as follows:

[0069]

[0070] When there is a perturbed function input, that is The input is n-dimensional and can be represented as equivalent to AND column vector as a single-column Jacobian matrix.

[0071] Taking the first part as the deep learning model, the second part as the perturbation input, multiplying the Jacobian matrix of the first function by the second function, and applying the chain rule, the gradient of the output with respect to the perturbation input can be obtained and is expressed as follows:

[0072]

[0073] where J represents the complete gradient approximation of the deep learning model as the derivative of the model, v represents the perturbation of the input, and J·v represents the gradient of the output with respect to the perturbation input. The machine learning tangent linear mode uses the complete gradient approximation model derivative generated by automatic differentiation to replace the tangent linear observation operator.

[0074] In the weather prediction model, the calculation process of the tangent linear observation operator combines the Jacobian matrix of the deep learning model with the input perturbation to obtain the gradient of the model output with respect to the perturbation input. Specifically, the Jacobian matrix J describes the sensitivity of the model output (such as future temperature, humidity, etc.) to the input features (such as current temperature, humidity, air pressure, etc.). By perturbing the input v and applying the chain rule, we can calculate the gradient J·v of the model output with respect to the input perturbation, which reflects the response of the model to different input changes. This process helps us to more precisely adjust the model parameters during the training of the deep learning model to improve the accuracy of weather prediction. Compared with the traditional observation operator, the tangent linear observation operator can more accurately capture complex nonlinear relationships through automatic differentiation, thus providing higher-quality gradient information and optimizing the training process of the weather prediction model.

[0075] The adjoint mode of the observation operator is an important part of the assimilation system, and the development of the traditional adjoint model is very cumbersome. When the machine learning observation operator replaces the traditional observation operator, it is difficult for the machine learning model to explicitly solve the adjoint mode. Assuming that a new machine learning model is constructed to replace the adjoint mode, it will lead to the newly constructed machine learning model being unable to correctly match the machine learning observation operator, but the gradient can be calculated by means of automatic differentiation of deep learning to bypass directly obtaining the adjoint mode. In the deep learning framework, automatic differentiation can accept the gradient of the output through the vector-Jacobian product and return the gradient of the input, making the computational graph differentiable.

[0076] Constructing the adjoint model through automatic differentiation is an important way to solve the stability of the deep learning model accessing the assimilation system. When the machine learning neural network model M: As an alternative observation operator, the physical field is a one-dimensional vector, and the adjoint model of model M in the field x to be analyzed 0 can be a matrix of dimension d×d and is defined as follows:

[0077]

[0078] Assume that the model output is another set of d-dimensional vectors. First, by inputting the field x to be analyzed 0 into the machine learning observation operator model, the inferred state x 1 can be obtained. Then, a scalar z is constructed by taking the dot product between x 1 and y.

[0079] Through automatic differentiation, the gradient at x 0 can be obtained, that is On the other hand, by manually deriving the computational graph, the representative meaning of the gradient can be found, and the expression of the gradient is as follows:

[0080]

[0081] Through computational derivation, it is found that the gradient of node x 0 represents the adjoint result acting on the y model. Since both x 0 and y are arbitrary, through this method, the computational results of the adjoint model defined at any point can be obtained and act on any vector. Thus, any differentiable prediction adjoint model can be constructed through automatic differentiation. This shows that the mechanism of calculating the gradient through automatic differentiation is exactly the same as the two processes of calculating using the traditional adjoint model-based gradient method. This also shows that it is feasible to use automatic differentiation to calculate the adjoint model of the machine learning observation operator. In the test of the online coupled assimilation system, the tangent linear model and adjoint model of the artificial intelligence observation operator have a significant improvement in computational efficiency compared to the traditional observation operator.

[0082] Suppose we train a deep learning model to predict the temperature in the next few days. In this model, the input x 0 may be the current air temperature, humidity, air pressure, etc., and the output y is the temperature prediction in the next few days. By calculating the loss function (such as the error between the predicted temperature and the actual temperature), we can calculate the gradient with respect to the input. These gradients will help us optimize the model parameters and improve the accuracy of meteorological predictions. During the training process, when calculating the gradient, automatic differentiation will directly calculate the gradient of the loss function with respect to the input x 0 (that is ) In this way, we can efficiently train the meteorological prediction model. Compared with the traditional observation operator, the machine learning model has certain advantages in terms of computational efficiency and accuracy. By constructing the adjoint model through automatic differentiation, not only can the computational efficiency be improved, but also the cumbersome derivation process in the traditional adjoint model can be avoided. For the meteorological prediction model, the adjoint model calculation provides an efficient method to optimize the model and effectively apply the machine learning observation operator to the assimilation system. This method can achieve faster gradient calculation in real-time meteorological data processing and enhance the prediction accuracy.

[0083] Step 004: It is necessary to connect the artificial intelligence forward mode, tangent linear mode, and adjoint mode to the assimilation system. The coupling technology between the machine learning model and the data assimilation system not only needs to ensure the stability of the machine learning model embedded in the traditional assimilation system but also needs to achieve flexible debugging of the machine learning model to facilitate obtaining the best results through repeated experiments. The construction of the machine learning observation operator belongs to a way of hybrid modeling. Hybrid modeling research is based on the cross-integration of multiple disciplines. The radiative transfer model usually uses the statically typed language Fortran as the standard programming tool, while the deep learning model is based on the machine learning framework of the dynamically typed language Python. When running in a coupled manner, there will be compatibility issues between the deep learning algorithm and the radiative transfer model code. The Fortran community ecosystem lacks mature and easy-to-use machine learning frameworks and tools, making the transplantation cost of the deep learning model very high.

[0084] The Fortran-Python interface (FPI) needs to solve the code compatibility problem of deploying the deep learning model in the radiative transfer model. Although the main interface of Pytorch is Python, the Python interface is on top of a large number of C++ code libraries, and libtorch is the C++ interface framework of Pytorch machine learning. Libtorch provides basic data structures and functions, including tensors and automatic differentiation, which can be used to extend the C++ code libraries based on this and is also a collection of built-in general components for neural network modeling. By customizing module extensions to the libtorch interface, the forward inference, tangent linear, and adjoint model calculations in the Pytorch framework are realized.

[0085] In terms of specific implementation, C++ of libtorch is adopted as the front end to deploy the Pytorch machine learning model. It is necessary to build libtorch into a dynamic link library, write interface code using C++ and Fortran, so as to build a call link from Fortran to C++, and achieve effective coupling between Python and Fortran. Through iso_c_binding, the Fortran program is allowed to share data with C++ and call C++ functions, and the memory pointer of the data is passed to C++, enabling C++ to directly access the arrays in the Fortran process. This method can not only take advantage of Python in terms of rapid development and rich neural network libraries, but also give full play to the strengths of Fortran in the field of high-performance numerical computing, thus enhancing the flexibility and functionality of the project while ensuring the computing efficiency. In the present invention, the forward mode, tangent linear mode and adjoint mode of the artificial intelligence observation operator are implemented using the C++ front end of libtorch, and can be embedded into any assimilation system written in Fortran.

[0086] Taking the WRFDA three-dimensional variational assimilation system as an example, the WRF forecast background field and satellite observation data are input, the artificial intelligence observation operator interface is called, the simulation results of satellite observations are obtained through the artificial intelligence forward model, then perturbations are constructed using the simulation errors, the perturbation results of the input physical quantities are obtained through the artificial intelligence adjoint model, and then the error results of the simulated observations are generated from the physical quantity perturbation results until the cost function converges to the minimum value, indicating that the assimilation convergence is completed, and the assimilation analysis field is generated. By selecting multiple examples, it is verified that the cost function of the WRFDA assimilation system coupled with the machine learning observation operator model can converge under different times and different weather processes.

[0087] In the present invention, the artificial intelligence tangent linear mode and adjoint mode are calculated using automatic differentiation, which can not only ensure the consistency of the mathematical relationships among the three, but also reduce the development difficulty of the tangent linear mode and adjoint mode.

[0088] The present invention also provides a construction system for a machine learning satellite observation operator, including:

[0089] A data collection module, which is used to collect the historical observation data of the satellite and its corresponding historical reanalysis data;

[0090] A training module, which is used to use the historical reanalysis data as the input data of the deep learning model, use the historical observation data of the satellite as the output data of the deep learning model, and train the deep learning model to obtain a trained artificial intelligence forward observation operator;

[0091] A linear observation operator calculation module, which is used to solve the tangent linear observation operator of the artificial intelligence forward observation operator using the automatic differentiation algorithm;

[0092] An adjoint observation operator calculation module, configured to use an automatic differentiation algorithm to solve the adjoint observation operator of the artificial intelligence forward observation operator;

[0093] An application module, configured to access the artificial intelligence forward observation operator, the tangent linear observation operator, and the adjoint observation operator into an assimilation system.

[0094] Preferably, in the linear observation operator calculation module, it includes:

[0095] Using an automatic differentiation algorithm to solve the Jacobian matrix of the artificial intelligence forward observation operator;

[0096] Multiplying the Jacobian matrix by the perturbation input and applying the chain rule to obtain the gradient of the output with respect to the perturbation input; wherein, the calculation formula for the gradient of the output with respect to the perturbation input is:

[0097]

[0098] Wherein, J represents the complete gradient approximation of the deep learning model as the derivative of the model, v represents the perturbation of the input, and J·v represents the gradient of the output with respect to the perturbation input.

[0099] Preferably, in the adjoint observation operator calculation module, it includes:

[0100] Input x 0 Into the artificial intelligence forward observation operator to obtain the output of the artificial intelligence forward observation operator;

[0101] Using an automatic differentiation algorithm to calculate the gradient with respect to the input x 0 ;

[0102] Taking the gradient of x 0 As the adjoint observation operator of the artificial intelligence forward observation operator.

[0103] Compared with the prior art, the beneficial effects of a construction system for a machine learning satellite observation operator provided by the present invention are the same as those of a construction method for a machine learning satellite observation operator described in the above technical solution, and will not be elaborated here.

[0104] The present invention also provides an electronic device, comprising a bus, a transceiver, a memory, a processor, and a computer program stored on the memory and executable on the processor. The transceiver, the memory, and the processor are connected through the bus. It is characterized in that when the computer program is executed by the processor, the steps in the above-mentioned method for constructing a machine learning satellite observation operator are realized. Compared with the prior art, the beneficial effects of the electronic device provided by the present invention are the same as those of the above-mentioned method for constructing a machine learning satellite observation operator, and will not be elaborated herein.

[0105] The present invention also provides a computer-readable storage medium, on which a computer program is stored. It is characterized in that when the computer program is executed by a processor, the steps in the above-mentioned method for constructing a machine learning satellite observation operator are realized. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided by the present invention are the same as those of the above-mentioned method for constructing a machine learning satellite observation operator, and will not be elaborated herein.

[0106] In the present specification, the various embodiments are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. For the same or similar parts among the various embodiments, reference can be made to each other. For the methods disclosed in the embodiments, since they correspond to the devices disclosed in the embodiments, the description is relatively simple, and reference can be made to the description of the device part for the relevant parts.

[0107] In this article, specific examples are used to elaborate on the principles and implementation manners of the present invention. The descriptions of the above embodiments are only used to help understand the method and its core idea of the present invention. At the same time, for those of ordinary skill in the art, based on the idea of the present invention, there will be changes in the specific implementation manners and application scopes. In summary, the content of this specification should not be construed as a limitation to the present invention.

Claims

1. A method for constructing a machine learning satellite observation operator, characterized in that: include: Step 1: Collect historical satellite observation data and their corresponding historical reanalysis data; Step 2: Use the historical reanalysis data as input data of the deep learning model, use the historical observation data of the satellite as output data of the deep learning model, and train the deep learning model to obtain a trained artificial intelligence forward observation operator; Step 3: Use the automatic differentiation algorithm to solve the tangent linear observation operator of the artificial intelligence forward observation operator; Step 4: Use the automatic differentiation algorithm to solve the adjoint observation operator of the artificial intelligence forward observation operator; Step 5: Connect the artificial intelligence forward observation operator, tangent observation operator and adjoint observation operator to the assimilation system to complete the weather forecast.

2. The method for constructing a machine learning satellite observation operator according to claim 1, characterized in that: The step 3: using an automatic differentiation algorithm to solve the tangent linear observation operator of the artificial intelligence forward observation operator includes: Use the automatic differentiation algorithm to solve the Jacobian matrix of the artificial intelligence forward observation operator; The Jacobian matrix is ​​multiplied by the perturbation input and the chain rule is applied to obtain the gradient of the output with respect to the perturbation input and used as the tangent linear observation operator.

3. The method for constructing a machine learning satellite observation operator according to claim 2, characterized in that: The step 4: using an automatic differentiation algorithm to solve the adjoint observation operator of the artificial intelligence forward observation operator includes: Input x0 into the artificial intelligence forward observation operator to obtain the output of the artificial intelligence forward observation operator; Using the automatic differentiation algorithm, calculate the gradient relative to the input x0; The gradient of x0 is used as the adjoint observation operator of the artificial intelligence forward observation operator.

4. The method for constructing a machine learning satellite observation operator according to claim 1, characterized in that: In step 2, the deep learning model is a ResMLP network model.

5. A system for constructing a machine learning satellite observation operator, characterized in that: include: Data collection module, used to collect historical satellite observation data and its corresponding historical reanalysis data; A training module, used to use historical reanalysis data as input data of a deep learning model, use historical satellite observation data as output data of the deep learning model, and train the deep learning model to obtain a trained artificial intelligence forward observation operator; A linear observation operator calculation module is used to solve the tangent linear observation operator of the artificial intelligence forward observation operator using an automatic differentiation algorithm; An adjoint observation operator calculation module is used to solve the adjoint observation operator of the artificial intelligence forward observation operator using an automatic differentiation algorithm; The application module is used to connect the artificial intelligence forward observation operator, the tangent observation operator and the adjoint observation operator into the assimilation system.

6. A system for constructing a machine learning satellite observation operator according to claim 5, characterized in that: In the linear observation operator calculation module, it includes: Use the automatic differentiation algorithm to solve the Jacobian matrix of the artificial intelligence forward observation operator; The Jacobian matrix is ​​multiplied by the perturbation input and the chain rule is applied to obtain the gradient of the output with respect to the perturbation input and used as the tangent linear observation operator.

7. A system for constructing a machine learning satellite observation operator according to claim 6, characterized in that: In the adjoint observation operator calculation module, it includes: Input x0 into the artificial intelligence forward observation operator to obtain the output of the artificial intelligence forward observation operator; Using the automatic differentiation algorithm, calculate the gradient relative to the input x0; The gradient of x0 is used as the adjoint observation operator of the artificial intelligence forward observation operator.

8. The method for constructing a machine learning satellite observation operator according to claim 5, characterized in that: In step 2, the deep learning model is a ResMLP network model.

9. An electronic device, comprising a bus, a transceiver, a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the transceiver, the memory, and the processor are connected via the bus, wherein: When the computer program is executed by the processor, the steps in the method for constructing a machine learning satellite observation operator as described in any one of claims 1 to 4 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps in the method for constructing a machine learning satellite observation operator as described in any one of claims 1 to 4 are implemented.

Images