MIGHTI mid- and upper-atmosphere temperature correction method based on genetic algorithm optimized BP neural network

Abstract

The present invention discloses a method for correcting the temperature of the middle and upper atmosphere of MIGHTI based on a BP neural network optimized by a genetic algorithm. The present invention optimizes the BP neural network inversion model through a genetic algorithm, establishes an optimal mapping relationship with the calibration atmospheric temperature, and optimizes the BP neural network inversion model through a genetic algorithm. Compared with the existing "peeling onion" algorithm for obtaining the temperature of the middle and upper atmosphere of MIGHTI, the use of the BP neural network inversion method solves the problems of large amount of calculation and long time required for inversion while ensuring temperature accuracy. The BP neural network of the present invention makes full use of the line of sight intensity value data related to temperature in the MIGHTI data set for feature learning, reduces the amount of calculation and shortens the time consumption while ensuring inversion accuracy, and has the characteristics of a simple and fast model and high inversion accuracy.

Description

Technical Field

[0001] The present invention relates to the field of atmospheric science research, and in particular to a MIGHTI middle and upper atmospheric temperature correction method based on a genetic algorithm optimized BP neural network. Background Art

[0002] The mid- and upper-level atmosphere is a crucial component of the "five-in-one" battlefield of "sea, land, air, space, and electronic warfare," and a vital component of national aerospace security. Accurately detecting the temperature field in the mid- and upper-level atmosphere improves the accuracy of medium- and long-term upper-atmospheric weather forecasts, enhances the safety of launch and re-entry of various spacecraft, and provides assurance for aerospace and national defense security. The MIGHTI payload on the ICON satellite boasts global-scale observations and high temporal and spatial resolution, enabling high-precision temperature calculations from observed data. The current MIGHTI temperature inversion method uses a "peeling onion" algorithm: MIGHTI's direct observation data is the signal intensity along the limb observation path of the oxygen A band. The "peeling onion" algorithm is used to derive relative intensity values ​​for target layers from these observations. The relative intensity values ​​for target layers in channels with different temperature responses are then compared and, based on the partition function, the atmospheric temperature profile is inverted. While the algorithm is stable and highly accurate, it requires consideration of the real-time center wavelength shift of the optical filter for each inversion, resulting in high computational complexity and long inversion time.

[0003] MIGHTI uses the "onion peeling" algorithm to extract relative intensity data from the standard intensity data for the target layer using the line-of-sight intensity data. The relative intensity data for the target layers of channels with different temperature responses are then compared to obtain the relative intensity ratio. Based on the partition function relationship between the relative intensity ratio and temperature, the atmospheric temperature profile is inverted. To achieve more accurate atmospheric temperature profiles, the temperature calculation must take into account the real-time changes in the filter center wavelength. As the filter temperature changes, the center wavelength drifts, causing the temperature sensitivity of each channel to change. This ultimately manifests as a shift in the partition function relationship between the relative intensity ratio of the target layer and atmospheric temperature. Therefore, when using relative intensity ratios to invert the atmospheric temperature profile, appropriate inversion curves must be selected for different altitudes. This significantly increases the computational complexity of the temperature inversion, resulting in longer MIGHTI temperature inversion times and preventing real-time inversion of upper atmospheric temperatures. Furthermore, because MIGHTI's sampling channels detect oxygen A-band intensity data, which ranges from 90 to 140 km during the day, the "peeling onion" algorithm can accurately obtain temperature information. However, at night, when the oxygen A-band intensity decreases, the signal strength within the MIGHTI sampling channel decreases, reducing the signal-to-noise ratio within the channel. Therefore, when using the "peeling onion" algorithm to invert temperature, temperature errors accumulate at each target layer, resulting in larger errors in the final nighttime temperature inversion.

[0004] It should be noted that the information disclosed in the background technology section is only used to enhance the understanding of the background of the present invention, and therefore may include information that does not constitute prior art known to ordinary technicians in the field. Summary of the Invention

[0005] In order to solve the above problems in the prior art, the present invention provides a MIGHTI middle and upper atmospheric temperature correction method based on genetic algorithm optimized BP neural network.

[0006] By establishing relationships between multiple MIGHTI observations and temperature, this method fully utilizes physical quantities related to temperature inversion, thereby overcoming the computational complexity and time-consuming nature of the "onion peeling" algorithm. Furthermore, given that the initial weights and thresholds of the BP neural network may not be optimal, resulting in the trained model weights and thresholds being prone to local optima, a genetic algorithm is introduced to iteratively approximate the optimal weights and thresholds to meet the accuracy requirements of the global inversion correction.

[0007] To achieve the above-mentioned object of the invention, the present invention provides a MIGHTI mid- and upper-level atmospheric temperature correction method based on a genetic algorithm optimized BP neural network, comprising the following steps:

[0008] Step S1: In the inversion model, data is crucial. The MIGHTI observation data is obtained from the ICON official website.

[0009] Step S2: Select the data set of line-of-sight intensity values ​​of the channel at 92-140 km within θ consecutive days from the MIGHTI observation data.

[0010]

[0011] Among them, I A ,I B ,I C ,I D ,I E These are the data sets of line-of-sight intensity values ​​at 92-140 km for channels A, B, C, D, and E over consecutive θ days;

[0012] Then, remove the sight intensity value dataset After the star point outliers in the sampling channel, the sight intensity value dataset I B , I C , I D The sight intensity value dataset I with the background channel subtracted A , I E Eliminate errors and obtain a standard intensity value data set without background signals

[0013]

[0014] Among them, E B 、E C With E D These are the standard intensity value datasets at 92-140 km within consecutive θ days for sampling channels B, C, and D, respectively;

[0015] Step S3, use the "onion peeling" algorithm to analyze the standard intensity value data set Perform inversion and calculate the relative intensity value data set of the target layer height h within the range of 92-140km within θ consecutive days and Where h is the height of the target layer, 92≤h≤140km;

[0016] Step S4: The relative intensity value dataset of the target layer with height h Respectively By making a ratio, we can obtain a data set of relative intensity ratios of the target layer with a height of h within the range of 92-140 km on consecutive θ days. and

[0017] Extracting the temperature partition function R from the high-resolution transmission database HITRAN2016Partition , the relative intensity ratio data set and Substitute the corresponding temperature distribution function R Partition , the calibration temperature data set within the range of 92-140km for continuous θ days is calculated

[0018] Step S5: Establish a standard intensity value data set based on the principle of correspondence between the target layer height and the observation time node With the calibration temperature data set One-to-one correspondence;

[0019] Step S6: standard intensity value data set With the calibration temperature data set Normalize to obtain dimensionless standard intensity value data set With dimensionless scaled temperature dataset

[0020] Step S7: Constructing a BP neural network inversion model

[0021] 7.1) Define the topology of the BP neural network, which includes an input layer, a hidden layer, and an output layer. Adjacent layers are connected using a fully connected method. The transfer function from the input layer to the hidden layer is a tan-sigmoid function, and the transfer function from the hidden layer to the output layer is a Purelin linear function. The training function is the traingdm function.

[0022] 7.2) Determine the number of neurons in the input layer, hidden layer, and output layer of the BP neural network inversion model, set the number of training times O of the BP neural network, and the training target minimum error E min ,

[0023] 7.3) The dimensionless standard intensity value data set processed in step S5 With dimensionless scaled temperature dataset Divide the training set (E, T) Train and test set (E,T) Test ,

[0024] Among them, the training set (E,T) Train is the dimensionless standard intensity value data set for consecutive θ-x days The corresponding dimensionless scaled temperature data set A set of , x is an integer between 1-θ;

[0025] Test set (E, T) Test is the dimensionless standard intensity value data set for x consecutive days The corresponding dimensionless scaled temperature data set A collection of

[0026] 7.4) The training set (E, T) Train Substitute it into the BP neural network inversion model for training, establish the mapping relationship between the dimensionless standard intensity value data and the dimensionless calibration temperature data, and obtain the training set (E, T) Train Output temperature dataset Training set (E, T) up to θ-x days Train Output temperature dataset and dimensionless scaled temperature datasets The error value between them is not greater than the minimum error E of the training target min Or when the number of training times reaches O, the trained BP neural network inversion model is obtained;

[0027] Step S8, test set (E, T) Train Substitute the trained BP neural network inversion model into the test set (E, T) Train According to the data forward propagation and error back propagation method, the test set (E, T) of the trained BP neural network inversion model is obtained. Train Output temperature data And the initial weights W and initial thresholds J of the x groups of trained BP neural network inversion models;

[0028] Step S9: constructing a BP neural network inversion model based on genetic algorithm GA-BP

[0029] 9.1) After several cycles of genetics, selection, crossover, and mutation of the fitness value F, the optimal values ​​of the initial weight W and initial threshold J are obtained.

[0030] in,

[0031]

[0032] e x is the test error of the test set;

[0033] 9.2) Input the optimal values ​​of the initial weight W and initial threshold J into the trained BP neural network inversion model to build a standard intensity value dataset With the calibration temperature data set The optimal mapping relationship between them is obtained, that is, the BP neural network inversion model GA-BP based on genetic algorithm;

[0034] Step S10: Select the continuous ρ-day standard intensity value data set Input the BP neural network inversion model GA-BP based on genetic algorithm, and the output temperature is obtained Where ρ≥1, then the output temperature After denormalization, we can get the atmospheric temperature at 92-140 km for ρ consecutive days.

[0035] Preferably, in step S2, a method for processing outliers using MATLAB software is used to remove star point outliers in the sight intensity dataset.

[0036] Preferably, in step S5, the standard intensity value data set is analyzed by an EXCEL table. With the calibration temperature data set Perform data processing and establish a standard intensity value data set With the calibration temperature data set One-to-one correspondence.

[0037] Beneficial effects

[0038] First, compared with the "onion peeling algorithm", the present invention utilizes the strong nonlinear mapping capability of the BP neural network to establish the relationship between multiple MIGHTI observation data and temperature, making full use of the physical quantities related to temperature inversion. While ensuring the accuracy of temperature inversion, it overcomes the problems of the "onion peeling algorithm" with large computational complexity and long operation cycle, greatly reducing the time cost of calculation.

[0039] Secondly, the present invention has extremely high global optimization capability, optimizes the initial weights and thresholds of the initial neural network, solves the problem of inversion falling into local optimality through the optimal values ​​of the initial weights and thresholds, and improves the inversion accuracy and speed.

[0040] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] The accompanying drawings are incorporated into and constitute a part of the specification to illustrate embodiments consistent with the present invention and, together with the description, serve to explain the principles of the present invention. Obviously, the drawings described below are only some embodiments of the present invention, and it is clear that those skilled in the art can derive other drawings based on these drawings without inventive effort.

[0042] Figure 1 It is a schematic diagram of the MIGHTI middle and upper atmospheric temperature inversion method based on the genetic algorithm optimized BP neural network of the present invention.

[0043] Figure 2 This is a curve showing the relationship between the relative intensities of the three channels B, C, and D and the temperature at the target layer of 116 km after considering the drift of the center wavelength of the filter in the embodiment of the present invention;

[0044] Figure 3 is a curve showing the relationship between the relative intensity ratio of the B&C and D&C channels and the temperature at the target layer of 116 km after considering the drift of the filter center wavelength in an embodiment of the present invention;

[0045] Figure 4 It is the structural diagram of BP neural network inversion model;

[0046] Figure 5 This is a temperature comparison curve output from June 1 to 5 by the BP neural network inversion model and the GA-BP model according to an embodiment of the present invention;

[0047] Figure 6 This is a verification diagram of the calibrated atmospheric temperature obtained by the "onion peeling algorithm" in the embodiment of the present invention and the atmospheric temperature on June 3rd inverted by the present invention. DETAILED DESCRIPTION

[0048] Example embodiments will now be described more fully with reference to the accompanying drawings. However, example embodiments can be implemented in various forms and should not be construed as limited to the examples set forth herein; rather, these embodiments are provided so that the present invention will be more comprehensive and complete and the concepts of the example embodiments will be fully conveyed to those skilled in the art. The described features, structures, or characteristics may be combined in any suitable manner in one or more embodiments. In the following description, many specific details are provided to provide a full understanding of the embodiments of the present invention. However, those skilled in the art will appreciate that the technical solutions of the present invention may be practiced while omitting one or more of the specific details.

[0049] The present invention provides a MIGHTI middle and upper atmospheric temperature correction method based on genetic algorithm optimization BP neural network. The genetic algorithm optimizes the BP neural network inversion model to obtain the optimal values ​​of the initial weight W and initial threshold J of the BP neural network inversion model, so that the model can continuously capture the standard intensity data set. Characteristics and establishment and calibration of atmospheric temperature The optimal mapping relationship between them is obtained, and the BP neural network inversion model GA-BP based on genetic algorithm optimization is obtained. Then the continuous ρ-day standard intensity value dataset to be tested is The BP neural network inversion model GA-BP based on genetic algorithm optimization is input to obtain the temperature output, and then the temperature output is denormalized to obtain the atmospheric temperature.

[0050] The following will further explain and illustrate the MIGHTI middle and upper atmospheric temperature inversion method based on genetic algorithm optimized BP neural network of the present invention with reference to the accompanying drawings.

[0051] like Figure 1As shown in FIG, the MIGHTI mid- and upper-atmosphere temperature inversion method based on the genetic algorithm optimized BP neural network includes the following steps:

[0052] Step S1: Obtain MIGHTI observation data from the ICON official website (ftp: / / icon-science.ssl.berkeley.edu / pub). Log in to the website and download the LEVEL.1 / and LEVEL.2 / data in the MIGHTI database for the first five months of 2021.

[0053] Step S2: Obtain the MIGHTI oxygen A-band limb observation line of sight intensity value dataset for five channels related to the inversion of middle and upper atmospheric temperature for 151 consecutive days in the first five months of 2021.

[0054]

[0055] Among them, I A ,I B ,I C ,I D ,I E These are the line-of-sight intensity values ​​at 92-140 km for channels A, B, C, D, and E in the first five months of 2021;

[0056] Using MATLAB software to deal with outliers to remove sight intensity data sets The star point outliers in the sampling channel are as follows: B , I C , I D The sight intensity value dataset I with the background channel subtracted A , I E Eliminate errors and obtain a standard intensity value data set without background signals

[0057]

[0058] Among them, E B 、E C With E D These are the standard intensity value data at 92-140km for sampling channels B, C, and D in the first five months of 2021;

[0059] Step S3: The standard intensity value dataset is retrievaled according to the “onion peeling” algorithm described in the literature (The MIGHTI wind retrieval algorithm: Description and verification, Space Science Reviews, 212(1), 585-600.). Perform inversion and calculate the relative intensity value dataset of the target layer height h within the range of 92-140km in the first five months of 2021 and Where h is the height of the target layer, 92≤h≤140km;

[0060] Step S4: The relative intensity value dataset of the target layer with height h Respectively By making a ratio, we can obtain the relative intensity ratio data set of the target layer with a height of h within the range of 92-140 km in the first five months of 2021. and Extracting the temperature partition function R from the high-resolution transmission database HITRAN2016 Partition , the relative intensity ratio data set and Substitute the corresponding temperature distribution function R Partition , calculated to obtain the calibration temperature data set within the 92-140km altitude range in the first five months of 2021

[0061] Affected by temperature, the center wavelength of the MIGHTI payload filter drifts, so the atmospheric temperature and relative intensity ratio dataset at different target layer altitudes of 92-140km is The temperature distribution function relationship is not the same. Take the 116km position on January 1, 2021 as an example. Figure 2 Shown: Relative intensity value dataset of the target layer simulated by the high-resolution transmission database HITRAN2016 Relationship with atmospheric temperature. Figure 2 Relative intensity data of the target layer at an altitude of 116 km Respectively By making a ratio, we can get the relative intensity ratio data of the target layer at an altitude of 116 km. and Relationship with atmospheric temperature, such as Figure 3 As shown, fitting it will give the atmospheric temperature and relative intensity ratio data set at 116km on January 1, 2021. and The partition function R of the relation Partition , as shown below:

[0062]

[0063] Will and Substituting into the above formula, we can get the calibration temperature at 116km on January 1, 2021

[0064] For different observation time nodes and target layer heights, select the corresponding intensity ratio data set and The partition function corresponding to the target layer height is used to obtain the calibration temperature data set

[0065] Step S5: To facilitate data processing, the standard intensity value data set is processed using an EXCEL table. With the calibration temperature data set Perform unified processing and establish a standard intensity value data set With the calibration temperature data set One-to-one correspondence;

[0066] Specifically, enter the standard intensity value datasets of the B, C, and D channels of MIGHTI in the first five months of 2021 in row 123 of the Excel spreadsheet. Data (listed as E B 、E C With E D The standard intensity value for each day), the 4th line enters the corresponding calibration temperature data set (Listed as the calibration temperature value corresponding to each day) Then, the data is read using the MATLAB command xlsread for reading EXCLE tables, making the data more convenient to use in MATLAB software.

[0067] Step S6, due to the BP neural network input data standard intensity value data set In 10 4 Order of magnitude, and the calibration temperature data set In 10 2 In order to prevent the large difference in the order of magnitude of input and output data from causing poor network performance and to improve the training speed, the present invention normalizes the data to accelerate the convergence of the BP neural network. With the calibration temperature data set Normalize to obtain dimensionless standard intensity value data set With dimensionless scaled temperature dataset

[0068] The expression of the normalization process is:

[0069]

[0070] in, and Standard intensity value datasets With the calibration temperature data set The maximum value of and Standard intensity value datasets With the calibration temperature data set The minimum value of .

[0071] Step S7, constructing a BP neural network inversion model;

[0072] 7.1) Define the topology of the BP neural network, such as Figure 4 As shown in the figure, it includes input layer, hidden layer and output layer. Adjacent layers are connected in a fully connected manner. The transfer function from the input layer to the hidden layer is a tan-sigmoid function, and the training function is a traingdm function.

[0073] 7.2) Determine the number of neurons in the input layer, hidden layer, and output layer of the BP neural network inversion model;

[0074] Since the input data of the BP neural network inversion model is a dimensionless standard intensity value dataset The number of rows in the dimension is 3, and the dimensionless calibration atmospheric temperature data The dimension is 1, so the number of neurons in the input layer is 3 and the number of neurons in the output layer is 1.

[0075] The number of nodes in the hidden layer is determined by the empirical formula get,

[0076] Where m is the number of nodes in the hidden layer, n is the number of nodes in the input layer, n=3, l is the number of nodes in the output layer, l=1, and a is a constant between 1 and 10.

[0077] In order to make the BP neural network inversion model have the necessary learning ability and information processing ability without increasing the complexity of the network structure, the number of nodes in the hidden layer in this example is m=4.

[0078] Set the training times O of the BP neural network inversion model to 1000 and the training target minimum error E min is 10.

[0079] 7.3) The dimensionless standard intensity value dataset With dimensionless scaled temperature dataset Divide the training set (E, T) Train and test set (E,T) Test ,

[0080] Among them, the training set (E,T) Train This is the dimensionless standard intensity value dataset for the first four months of 2021. The corresponding dimensionless scaled temperature data set A collection of

[0081] Test set (E, T)Test The dimensionless standard intensity value dataset for May 2021 The corresponding dimensionless scaled temperature data set A collection of

[0082] 7.4) The training set (E, T) Train Substitute into the BP neural network inversion model for training to obtain the temperature data set of the training set Through several forward propagation and back propagation, until the training set (E, T) within the first 4 months Train Temperature dataset and dimensionless scaled temperature datasets The error value between them is not greater than the minimum error E of the training target min =10 or the number of training times reaches O=1000, and the trained BP neural network inversion model is obtained;

[0083] Step S8, test set (E, T) Test Substitute the trained BP neural network inversion model for the test set (E, T) Test Perform training to obtain the test set (E, T) of the trained BP neural network inversion model Test Output temperature data and 31 sets of initial parameters (W, J), where W is the initial weight of the BP neural network inversion model, and J is the initial threshold of the BP neural network inversion model;

[0084] Step S9: constructing a BP neural network inversion model based on genetic algorithm GA-BP

[0085] 9.1) After several cycles of genetics, selection, crossover, and mutation of the fitness value F, the optimal values ​​of the initial weight W and initial threshold J are obtained.

[0086]

[0087]

[0088] e 31 is the test error of the test set;

[0089] 9.2) Input the optimal values ​​of the initial weight W and initial threshold J into the trained BP neural network inversion model training to determine the test set (E, T) Test Dimensionless standard intensity value dataset With dimensionless scaled temperature dataset The optimal mapping relationship between them is obtained, that is, the BP neural network inversion model GA-BP based on genetic algorithm;

[0090] Step S10: Select the continuous p-day standard relative intensity value data set Input the BP neural network inversion model GA-BP based on genetic algorithm, and the output temperature is obtained Then, the output temperature After denormalization, we can get the atmospheric temperature at 92-140 km for ρ consecutive days.

[0091] Case implementation and computer simulation results analysis

[0092] The observation data of MIGHTI from June 1 to 5, 2021 were retrieved and selected, and programming simulation was performed using MatlabR2021b software.

[0093] The running results are as follows Figure 5-6 The results show that the atmospheric temperature calculated by the BP neural network inversion model GA-BP of the genetic algorithm of the present invention is basically consistent with the calibration temperature obtained by "peeling the onion" in the spatial region of 92-140 km, indicating that the atmospheric temperature obtained by the GA-BP network structure is basically close to the calibration temperature of "peeling the onion". At the same time, the difference between the GA-BP method of the present invention and the calibration temperature of "peeling the onion" is calculated using the mean absolute error MAE and the mean absolute percentage error MAPE. The mean absolute error MAE is 8.80 (since the middle and upper atmospheric temperatures of 92-140 km are generally several hundred K, the MAE difference of 8.8 can be considered to be extremely small); the mean absolute percentage error MAPE is 2.58%, both indicating the accuracy of the GA-BP method. Moreover, GA-BP takes nearly 2 seconds, which shows that the GA-BP method of the present invention has excellent inversion accuracy and efficiency.

[0094] MIGHTI observation data from June 1 to 5, July 1 to 5, and August 1 to 5, 2021, were retrieved and compared with the "Peeling Onion" algorithm and the GA-BP method of the present invention. The simulation results are shown in Table 1. The GA-BP method of the present invention significantly reduces the time required by the "Peeling Onion" algorithm, indicating that the GA-BP method of the present invention significantly reduces the amount of calculation and shortens the time consumption while ensuring inversion accuracy, and has the characteristics of being fast and having high inversion accuracy.

[0095] Table 1. Comparison of computing time between GA-BP and “Onion Peeling”

[0096] Time required (s) "Peeling the Onion" algorithm GA-BP June 1-5 903 2 July 1-5 862 2 August 1-5 924 2

[0097] It should be noted that the present invention also simulates the inversion model using only BP neural network, such as Figure 5As shown in the figure, the atmospheric temperatures outputted around groups 150 and 200 have large offset peaks compared to the calibration temperatures, indicating that the trained BP neural network inversion model has a certain degree of inversion error. However, the temperature outputted by the GA-BP method optimized by the genetic algorithm is basically consistent with the calibration temperature, indicating that the GA-BP method of the present invention has high inversion accuracy. Similarly, Figure 6 As shown in the figure, the trained BP neural network inversion model also has obvious temperature offset at 100-110 km, which also proves that the trained BP neural network inversion model has inversion error. However, the GA-BP method of the present invention is basically consistent with the calibration temperature and has better inversion stability.

[0098] The mean absolute error (MAE) and mean absolute percentage error (MAPE) are used to compare the simulation results of BP and GA-BP with the calibration temperature to judge the model performance. The results are shown in Table 2.

[0099] Table 2. Errors in atmospheric temperature retrieval using BP, GA-BP, and “Onion Peeling”

[0100] Neural Networks Network Layer Activation Function Optimization Algorithm MAE MAPE BP 3-4-1 Transig,Purelin Gradient descent 11.37 4.23％ GA-BP 3-4-1 Transig,Purelin Genetic Algorithm 8.80 2.58％

[0101] The simulation results in Table 2 also show that: when the network parameters are the same, the GA-BP error of the present invention (MAE = 8.80 / MAPE = 2.58%) is better than that of the trained BP neural network inversion model (MAE = 11.37 / MAPE = 4.23%) compared with the calibration temperature obtained by BP, that is, the error value of the calibration temperature is smaller, and therefore, it has better inversion stability.

[0102] In summary, the GA-BP method of the present invention utilizes the strong nonlinear mapping capability of the BP neural network to establish the relationship between multiple MIGHTI observation data and temperature, fully utilizes the physical quantities related to temperature inversion, converts the physical quantities into mathematical quantities, and obtains the mathematical characteristics of the input data. It overcomes the problem of large amount of calculation and long operation cycle caused by the need to consider the real-time change of the center wavelength of the filter in "peeling the onion", greatly reducing the time cost of calculation. The use of the BP neural network method can solve the difficult problems of large amount of calculation and long operation cycle of traditional physical methods. At the same time, in the process of genetic selection and cross-mutation of fitness values ​​using the genetic algorithm, a global search for the initial weight threshold is performed, which solves the temperature offset error problem existing in the trained BP neural network inversion model. Therefore, it is more consistent with the temperature obtained by "peeling the onion". Under the condition of the same network parameters, the BP neural network optimized by the genetic algorithm of the present invention has a smaller error value with the calibration temperature, has better inversion stability, and can further reduce the inversion discreteness of the BP neural network in the genetic algorithm, thereby improving the inversion accuracy.

[0103] The preferred embodiments of the present invention are described in detail above. However, the present invention is not limited to the specific details in the above embodiments. Within the scope of the technical concept of the present invention, the technical solution of the present invention can be subjected to a variety of equivalent transformations, and these equivalent transformations all belong to the protection scope of the present invention. It should also be noted that the various specific technical features described in the above specific embodiments can be combined in any suitable manner unless there is any contradiction. In order to avoid unnecessary repetition, the present invention will no longer describe various possible combinations separately. In addition, the various different embodiments of the present invention can also be arbitrarily combined, and as long as they do not violate the concept of the present invention, they should also be regarded as the contents disclosed by the present invention.

Claims

1. The MIGHTI mid- and upper-atmosphere temperature correction method based on genetic algorithm optimized BP neural network is characterized by The following steps are involved: Step S1, obtain MIGHTI observation data from the ICON official website; Step S2: Select the data set of line-of-sight intensity values ​​of the channel at 92-140 km within θ consecutive days from the MIGHTI observation data. Among them, I A ,I B ,I C ,I D ,I E These are the data sets of line-of-sight intensity values ​​at 92-140 km for channels A, B, C, D, and E over consecutive θ days; Then, remove the sight intensity value dataset After the star point outliers in the sampling channel, the sight intensity value dataset I B , I C , I D The sight intensity value dataset I with the background channel subtracted A , I E Eliminate errors and obtain a standard intensity value data set Among them, E B 、E C With E D The data sets of standard intensity values ​​at 92-140 km within consecutive θ days for sampling channels B, C, and D, respectively; Step S3, use the "peeling onion" algorithm to analyze the standard intensity value data set Calculate the relative intensity value data set of the target layer height h within the range of 92-140km on consecutive θ days and 92≤h≤140km; Step S4: The relative intensity value dataset of the target layer with height h Respectively By making a ratio, we can obtain a data set of relative intensity ratios of the target layer with a height of h within the range of 92-140 km on consecutive θ days. and The temperature partition function RPartition is extracted from the high-resolution transmission database HITRAN2016, and the relative intensity ratio dataset is converted to and Substitute the corresponding temperature partition function RPartition to calculate the calibration temperature data set within the range of 92-140km on consecutive θ days Step S5: Establish a standard intensity value data set based on the principle of correspondence between the target layer height and the observation time node With the calibration temperature data set One-to-one correspondence; Step S6: standard intensity value data set With the calibration temperature data set Normalize to obtain dimensionless standard intensity value data set With dimensionless scaled temperature dataset Step S7: Constructing a BP neural network inversion model 7.1) Define the topology of the BP neural network, which includes an input layer, a hidden layer, and an output layer. Adjacent layers are connected using a fully connected method. The transfer function from the input layer to the hidden layer is a tan-sigmoid function, and the transfer function from the hidden layer to the output layer is a Purelin linear function. The training function is the traingdm function. 7.2) Determine the number of neurons in the input layer, hidden layer, and output layer of the BP neural network inversion model, set the number of training times O of the BP neural network, and the training target minimum error E min , 7.3) The dimensionless standard intensity value data set processed in step S5 With dimensionless scaled temperature dataset Divide the training set (E, T) Train and test set (E,T) Test , Among them, the training set (E,T) Train is the dimensionless standard intensity value data set for consecutive θ-x days The corresponding dimensionless scaled temperature data set A set of , x is an integer between 1-θ; Test set (E, T) Test is the dimensionless standard intensity value data set for x consecutive days The corresponding dimensionless scaled temperature data set A collection of 7.4) The training set (E, T) Train Substitute it into the BP neural network inversion model for training, establish the mapping relationship between the dimensionless standard intensity value data and the dimensionless calibration temperature data, and at the same time, obtain the training set (E, T) Train Output temperature dataset Training set (E, T) up to θ-x days Train Output temperature dataset and dimensionless scaled temperature datasets The error value between them is not greater than the minimum error E of the training target min Or when the number of training times reaches O, the trained BP neural network inversion model is obtained; Step S8, test set (E, T) Test Substitute the trained BP neural network inversion model into the test set (E, T) Test According to the data forward propagation and error back propagation method, the test set (E, T) of the trained BP neural network inversion model is obtained. Test Output temperature data And the initial weights W and initial thresholds J of the x groups of trained BP neural network inversion models; Step S9: constructing a BP neural network inversion model based on genetic algorithm GA-BP 9.1) Calculate the fitness value F, and then use the genetic algorithm to perform several cycles of inheritance, selection, crossover, and mutation on the fitness value F to obtain the optimal values ​​of the initial weight W and initial threshold J. in, e x For the test set (E,T) Test The test error; 9.2) Input the optimal values ​​of the initial weight W and initial threshold J into the trained BP neural network inversion model training to determine the test set (E, T) Test Dimensionless standard intensity value dataset With dimensionless scaled temperature dataset The optimal mapping relationship between them is obtained, that is, the BP neural network inversion model GA-BP based on genetic algorithm; Step S10: Select the continuous ρ-day standard intensity value data set Input the BP neural network inversion model GA-BP based on genetic algorithm, and the output temperature is obtained Where ρ≥1, then the output temperature After denormalization, we can get the atmospheric temperature at 92-140 km for ρ consecutive days.

2. The method according to claim 1, characterized in that In step S2, the outlier processing method of MATLAB software is used to remove the star point outliers in the sight intensity dataset.

3. The method according to claim 1, characterized in that In step S5, the standard intensity value data set is analyzed by using an EXCEL table. With the calibration temperature data set Perform data processing and establish a standard intensity value data set With the calibration temperature data set One-to-one correspondence.

Images