Atmospheric density prediction method, device and equipment based on multi-model fusion

Abstract

The present application discloses an atmospheric density prediction method, device and equipment based on multi-model fusion. The method comprises: obtaining target training data; training the first metamodel using the target training data, so that the first metamodel assigns corresponding M first weight values ​​to M prediction values ​​of each prediction period based on the historical environmental parameters corresponding to each prediction period; the training of the first metamodel is completed when the first loss function value of the first metamodel satisfies the conditions, the first loss function value is the mean square error between the first prediction value corresponding to each prediction period in multiple prediction periods and the measured value, the first prediction value is obtained by weighted summing the M prediction values ​​corresponding to each prediction period using the M first weight values ​​assigned by the first metamodel in each prediction period; assigning M first weight values ​​to M target prediction values ​​through the first metamodel; weighted summing the M target prediction values ​​using the M first weight values ​​to determine the atmospheric density of the target prediction period.

Description

Technical Field

[0001] The present application relates to the field of space environment detection technology, in particular to the field of atmospheric density detection technology, and in particular to an atmospheric density prediction method, device and equipment based on multi-model fusion. Background Art

[0002] The accurate calculation of the density of the upper atmosphere near the Earth is crucial for the accurate calculation of the orbits of low-orbit and ultra-low-orbit near-Earth satellites. With the increase in the launch missions of low-orbit satellites, the demand for accurate prediction of the density of the upper atmosphere near the Earth is growing to ensure the accurate orbit insertion and orbit maintenance of satellites. This is crucial for the design, launch and on-orbit operation of satellites.

[0003] At present, when predicting atmospheric density, the (Jacchia-Bowman, JB) series of models, (Mass Spectrometer Incoherent Scatter, MSIS) series of models, and (Drag Temperature Model, DTM) series of models are usually used. However, the performance of the above-mentioned empirical models in predicting atmospheric density in different space environments varies, and there will be obvious prediction errors under certain environmental conditions. For example, the latest model of the MSIS series, MSIS20, has an error of 15% during the solar calm period and can reach 25%-48% during the solar flare period. It can be seen that under different environmental conditions, the accuracy of the atmospheric density predicted by relying solely on a single empirical model is insufficient and cannot meet the precise calculation requirements of the near-Earth ultra-low orbit satellite orbit. Summary of the invention

[0004] In view of this, the embodiments of the present application provide an atmospheric density prediction method, device, equipment, medium and product based on multi-model fusion, which can combine the prediction results of multiple basic prediction models to improve the overall prediction performance of the first element model, make up for the shortcomings of using a single basic prediction model for atmospheric density prediction, and reduce the overall prediction error when predicting atmospheric density when facing different environmental conditions.

[0005] Accurately predict the atmospheric density under various space environment conditions, reduce the atmospheric density error, and take into account a variety of space environments.

[0006] In a first aspect, an embodiment of the present application provides an atmospheric density prediction method based on multi-model fusion, the method comprising: obtaining target training data, wherein the target training data comprises historical environmental parameters, measured values ​​and M predicted values ​​corresponding to each prediction period in multiple prediction periods, the measured values ​​are the atmospheric density of each prediction period actually detected, and the M predicted values ​​are obtained by predicting the atmospheric density of each prediction period by M basic prediction models based on the historical environmental parameters of each prediction period; training a first meta-model using the target training data so that the first meta-model assigns corresponding M first weight values ​​to the M predicted values ​​of each prediction period based on the historical environmental parameters corresponding to each prediction period; when a first loss function value of the first meta-model satisfies a first preset Under the following conditions, it is determined that the training of the first metamodel is completed, wherein the first loss function value is the mean square error between the first prediction value corresponding to each prediction period in multiple prediction periods and the measured value, and the first prediction value is obtained by weighted summing the M prediction values ​​corresponding to each prediction period using the M first weight values ​​assigned by the first metamodel in each prediction period; the historical environmental parameters corresponding to the target prediction period and the M target prediction values ​​are input into the first metamodel, and M first weight values ​​are assigned to the M target prediction values ​​through the first metamodel, wherein the M target prediction values ​​are output by M basic prediction models based on the historical environmental parameters corresponding to the target prediction period; the atmospheric density of the target prediction period is determined by weighted summing the M target prediction values ​​using the M first weight values.

[0007] In a second aspect, an embodiment of the present application provides an atmospheric density prediction device based on multi-model fusion, the device comprising: an acquisition module, used to acquire target training data, wherein the target training data includes historical environmental parameters, measured values ​​and M predicted values ​​corresponding to each prediction period in multiple prediction periods, the measured values ​​are the atmospheric density of each prediction period actually detected, and the M predicted values ​​are obtained by predicting the atmospheric density of each prediction period by M basic prediction models based on the historical environmental parameters of each prediction period; a model training module, used to train a first metamodel using the target training data, so that the first metamodel assigns corresponding M first weight values ​​to the M predicted values ​​of each prediction period based on the historical environmental parameters corresponding to each prediction period; a determination module, used to determine whether the first loss function value of the first metamodel satisfies Under the first preset condition, it is determined that the training of the first metamodel is completed, wherein the first loss function value is the mean square error between the first prediction value corresponding to each prediction period in multiple prediction periods and the measured value, and the first prediction value is obtained by weighted summing the M prediction values ​​corresponding to each prediction period using the M first weight values ​​assigned by the first metamodel in each prediction period; a weight allocation module is used to input historical environmental parameters and M target prediction values ​​corresponding to the target prediction period into the first metamodel, and assign M first weight values ​​to the M target prediction values ​​through the first metamodel, wherein the M target prediction values ​​are output by M basic prediction models based on the historical environmental parameters corresponding to the target prediction period; a determination module is also used to weighted sum the M target prediction values ​​using the M first weight values ​​to determine the atmospheric density of the target prediction period.

[0008] In a third aspect, an embodiment of the present application provides an electronic device, comprising: a processor and a memory storing computer program instructions; when the processor executes the computer program instructions, the steps of the atmospheric density prediction method based on multi-model fusion as in the first aspect are implemented.

[0009] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium having computer program instructions stored thereon. When the computer program instructions are executed by a processor, the steps of the atmospheric density prediction method based on multi-model fusion as in the first aspect are implemented.

[0010] In a fifth aspect, an embodiment of the present application provides a computer program product, which is stored in a non-volatile storage medium. When the computer program product is executed by a processor, it implements the steps of the atmospheric density prediction method based on multi-model fusion as in the first aspect.

[0011] In a sixth aspect, an embodiment of the present application provides a chip, which includes a processor and a communication interface, wherein the communication interface and the processor are coupled, and the processor is used to run programs or instructions to implement the steps of the atmospheric density prediction method based on multi-model fusion as in the first aspect.

[0012] The present application provides an atmospheric density prediction method, device, equipment, medium and product based on multi-model fusion. Considering that different basic prediction models have different prediction accuracy performance capabilities for atmospheric density under different environmental conditions, the present application uses historical environmental parameters, measured values, and M prediction values ​​output by M basic prediction models for training under different prediction periods. Different prediction periods correspond to different spatial environmental conditions, so training is based on target training data, considering enough spatial environmental features, and adopting a method of hybrid fusion of spatial environmental features and basic prediction models, so that the first meta-model can better learn the weights of different basic prediction models in various environmental features, which is equivalent to dynamically weighted summing each basic prediction model, combining the prediction results of multiple basic prediction models to improve the overall prediction performance of the first meta-model, so that the first meta-model has better generalization ability, can strengthen the atmospheric density prediction advantages of different basic prediction models under different environmental conditions, make up for the shortcomings of using a single basic prediction model for atmospheric density prediction, improve the accuracy and robustness of the first meta-model, and reduce the overall prediction error when predicting atmospheric density when facing different environmental conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] In order to more clearly illustrate the technical application of the embodiments of the present application, the drawings in the embodiments of the present application are briefly introduced below.

[0014] Figure 1 It is a flowchart of an atmospheric density prediction method based on multi-model fusion provided in one embodiment of the present application;

[0015] Figure 2 is a flow chart of an atmospheric density prediction method based on multi-model fusion provided by another embodiment of the present application;

[0016] Figure 3 It is a structural schematic diagram of an atmospheric density prediction device based on multi-model fusion provided in an embodiment of the present application;

[0017] Figure 4 It is a schematic diagram of the hardware structure of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0018] The principles and spirit of the present application will be described below with reference to several exemplary embodiments. It should be understood that the purpose of providing these embodiments is to make the principles and spirit of the present application clearer and more thorough, so that those skilled in the art can better understand and implement the principles and spirit of the present application. The exemplary embodiments provided herein are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments herein, all other embodiments obtained by ordinary technicians of the art without creative work are within the scope of protection of this application.

[0019] In this document, terms such as first, second, third, etc. are only used to distinguish one entity (or operation) from another entity (or operation), but not to require or imply any order or relationship between these entities (or operations).

[0020] The following is a brief description of the concepts and technical terms that may be involved in the embodiments of the present application.

[0021] Model fusion technology: Model fusion technology is an integrated learning method that improves the overall prediction performance by combining the prediction results of multiple models. This method can make up for the shortcomings of a single model and improve the accuracy and robustness of the model.

[0022] Stacking: Stacking is an integration method that combines multiple prediction models through a two-layer or multi-layer model structure. In Stacking, the first layer contains the prediction results of multiple base models, and the second layer learner learns how to adjust the weights of each model based on these prediction results to generate the final prediction.

[0023] Base prediction models: Each base prediction model is trained and predicted independently of other models, and generates its own prediction results based on different learning algorithms or different data subsets. Different base prediction models have different learning biases and capabilities, and they capture different patterns or features in the data.

[0024] Weight model: The weight model is the second layer model in ensemble learning, which is responsible for weight calculation based on environmental characteristics. The main task of the weight model is to assign appropriate weights to these models based on the performance of each basic prediction model under different environmental characteristics.

[0025] Metamodel: A model used to integrate the prediction results of basic prediction models. In this application, the weight model is part of the metamodel.

[0026] Solar radiation parameters: Solar Indices, including F10.7, S10.7, M10.7, and Y10.7. F10.7 data is daily frequency, and its historical value can be obtained from NOAA (or the Atmospheric Density Forecast Center). The forecast value for the next 40 days can be obtained from NOAA, or it can be predicted using the F10.7 forecast model; S10.7, Y10.7, and M10.7 are daily frequency, and the forecast value for the next 40 days can be obtained using the corresponding forecast model; M10.7 is data with an update frequency of 3h, and its historical value can be obtained using an algorithm, and the forecast value can be obtained using the corresponding forecast model.

[0027] F10.7: This is the 10.7 cm radio flux index, which is a commonly used indicator of solar activity. The F10.7 index reflects the intensity of the sun's radio radiation at a wavelength of 10.7 cm and is usually used to represent the intensity of the sun's ultraviolet (UV) radiation.

[0028] S10.7 (EUV): This is the solar radiation index in the extreme ultraviolet band, representing the solar radiation in the 26-34 nanometer band. This radiation mainly comes from solar active regions.

[0029] M10.7 (FUV): This is the solar radiation index in the far ultraviolet band, representing the solar radiation around 280 nanometers. This part of the radiation includes the photospheric continuous spectrum and chromospheric line radiation.

[0030] Y10.7: This is a mixed solar index representing mainly solar X-rays and Lyman-alpha radiation. X-rays come mainly from the solar canopy, while Lyman-alpha radiation is produced in the upper chromosphere and transition region of the sun.

[0031] Geomagnetic parameters: geomagnetic index, including ap, Ap, Ap* and Dst index. The ap index is predicted by the deep learning model with 24 steps (each step is 3 days); Ap is the daily geomagnetic index, which is obtained by averaging the 8 ap values ​​of the day; Ap* is the sliding geomagnetic index, which is obtained by averaging the ap values ​​of the first 8 prediction periods of the target prediction period; Dst is an hourly frequency geomagnetic index.

[0032] ap index: also known as the equivalent planetary amplitude of geomagnetic disturbance, the unit is 2nT, a new ap value is generated every 3 hours, and 8 ap indexes are generated every day in chronological order. According to relevant standards, an ap value greater than 39 corresponds to a magnetic storm period. The maximum value of the ap index is 400.

[0033] Ap index: It is an index of the global daily geomagnetic disturbance intensity, also known as the planetary equivalent daily amplitude. It is an indicator to measure the level of geomagnetic activity. A new Ap value is generated every day, which is the average of 8 AP values ​​every day. Therefore, the Ap index can only reflect the overall situation of geomagnetic disturbance throughout the day.

[0034] Ap* index: It is an index of the intensity of geomagnetic disturbances in a predicted period. It is a real-time indicator for measuring the level of geomagnetic activity. Compared with the Ap index, it is more real-time and can therefore more accurately evaluate the overall situation of geomagnetic disturbances in a certain predicted period.

[0035] Dst Index: This is the Disturbance Storm Time (Dst) index, which is used to indicate the strength of the ring current during a magnetospheric storm. The Dst index is determined from hourly measurements of the magnetic field at four points around the Earth's equator and is used to simulate global density changes during geomagnetic storms.

[0036] The time and space parameters include altitude, longitude and latitude, universal time, DOY, month, and local time (or mean solar time, which cancels the whole hour restriction).

[0037] Atmospheric density prediction models can be divided into physical models and empirical models. Currently, the commonly used JB series models (Jacchia-Bowman), MSIS series models (Mass Spectrometer Incoherent ScatterMSIS), and DTM series models (Drag Temperature Model) are all empirical models.

[0038] JB series models (Jacchia-Bowman): This is a series of empirical atmospheric density models developed by Jacchia in the 1960s and further developed and improved by Bowman et al. These models use statistical and empirical methods to describe the changes in atmospheric density with factors such as altitude, season, and solar activity. The JB series models include versions such as JB2006 and JB2008, which use the driven solar index from on-orbit sensor data when calculating the impact of solar radiation on atmospheric density.

[0039] MSIS series models (Mass Spectrometer Incoherent Scatter): This series of models is also an empirical atmospheric model developed by the U.S. Naval Research Laboratory (NRL). The MSIS series of models describe atmospheric physical properties such as neutral atmospheric density and temperature from the ground to the thermosphere. They include MSISE-90, NRLMSISE-00, MSISE-20, etc.

[0040] DTM series models (Drag Temperature Model): The DTM series models are also empirical atmospheric models. They use the analytical solution of the diffusion balance equation to estimate the model coefficients in combination with satellite drag data, atmospheric composition and temperature data. The DTM series includes versions such as DTM94, DTM2000, and DTM2020. The orbit prediction performance of these models under different solar and geomagnetic conditions has been fully evaluated.

[0041] In the related technology, the above-mentioned empirical models have different performance capabilities in predicting atmospheric density under different space environments, and there will be obvious errors under certain environmental conditions. For example, the error of MSIS20, the latest model in the MSISE series, is 15% during the solar quiet period, and can reach 25%-48% during the solar eruption period; the model version before JB2008 does not have the model semi-annual density changes and the complete effect of solar ultraviolet radiation on the heating of the thermosphere, resulting in the inability to accurately predict; the granularity of the input parameters of some models is relatively coarse, resulting in large errors in the short-term changes in orbital atmospheric density in the model prediction results, which limits the application scenarios that require high-precision short-term forecasts; some models (MSISE, DTM) do not take into account the impact of changes in different bands of solar radiation on atmospheric density at different altitudes, resulting in insufficient prediction accuracy of atmospheric density at certain altitudes for these models. It can be seen that under different environmental conditions, the accuracy of atmospheric density predicted by relying solely on a single empirical model is insufficient and cannot meet the precise calculation requirements of near-Earth ultra-low orbit satellite orbits.

[0042] Based on this, this application uses historical environmental parameters, measured values, and M predicted values ​​output by M basic prediction models under different prediction periods for training. Different prediction periods correspond to different spatial environmental conditions, so training is based on target training data, taking into account enough spatial environmental features, and adopting a method of hybrid fusion of spatial environmental features and basic prediction models, so that the first meta-model can better learn the weights of different basic prediction models in various environmental features, which is equivalent to dynamically weighted summing each basic prediction model, combining the prediction results of multiple basic prediction models to improve the overall prediction performance of the first meta-model, so that the first meta-model has better generalization ability, can strengthen the atmospheric density prediction advantages of different basic prediction models under different environmental conditions, make up for the shortcomings of using a single basic prediction model for atmospheric density prediction, improve the accuracy and robustness of the first meta-model, and reduce the overall prediction error when predicting atmospheric density when facing different environmental conditions, so as to solve at least one of the above technical problems.

[0043] The following, in conjunction with the accompanying drawings, describes in detail the atmospheric density prediction method based on multi-model fusion provided by the embodiment of the present application through specific embodiments and their application scenarios.

[0044] Figure 1 It is a flow chart of an atmospheric density prediction method based on multi-model fusion provided in one embodiment of the present application. The executor of the atmospheric density prediction method based on multi-model fusion may be an atmospheric density prediction device.

[0045] The following takes the atmospheric density prediction method based on multi-model fusion as an example, and describes the atmospheric density prediction method based on multi-model fusion of the present application. It should be noted that the above-mentioned execution subject and application scenario do not constitute a limitation on the present application.

[0046] like Figure 1 As shown, the atmospheric density prediction method based on multi-model fusion provided in the embodiment of the present application may include steps 110 to 150.

[0047] Step 110, obtaining target training data, the target training data including historical environmental parameters, measured values ​​and M predicted values ​​corresponding to each prediction period in multiple prediction periods;

[0048] Step 120, training the first meta-model using the target training data, so that the first meta-model assigns corresponding M first weight values ​​to the M prediction values ​​of each prediction period based on the historical environment parameters corresponding to each prediction period;

[0049] Step 130, when the first loss function value of the first metamodel satisfies the first preset condition, determining that the training of the first metamodel is completed;

[0050] Step 140, inputting historical environment parameters corresponding to the target prediction period and M target prediction values ​​into the first metamodel, and assigning M first weight values ​​to the M target prediction values ​​through the first metamodel, where the M target prediction values ​​are output by the M basic prediction models based on the historical environment parameters corresponding to the target prediction period;

[0051] Step 150, performing weighted summation on the M target prediction values ​​using the M first weight values ​​to determine the atmospheric density of the target prediction period.

[0052] The atmospheric density prediction method based on multi-model fusion provided in the embodiment of the present application takes into account that different basic prediction models have different prediction accuracy performance capabilities for atmospheric density under different environmental conditions. The present application uses historical environmental parameters, measured values, and M prediction values ​​output by M basic prediction models for training under different prediction periods. Different prediction periods correspond to different spatial environmental conditions, so training is based on target training data, considering enough spatial environmental features, and adopting a method of hybrid fusion of spatial environmental features and basic prediction models, so that the first meta-model can better learn the weights of different basic prediction models in various environmental features, which is equivalent to dynamically weighted summing each basic prediction model, combining the prediction results of multiple basic prediction models to improve the overall prediction performance of the first meta-model, so that the first meta-model has better generalization ability, can strengthen the atmospheric density prediction advantages of different basic prediction models under different environmental conditions, make up for the shortcomings of using a single basic prediction model for atmospheric density prediction, improve the accuracy and robustness of the first meta-model, and reduce the overall prediction error when predicting atmospheric density when facing different environmental conditions.

[0053] The specific implementation of the above steps will be described in detail below in conjunction with specific embodiments.

[0054] Involving step 110, target training data is obtained, and the target training data includes historical environmental parameters, measured values, and M predicted values ​​corresponding to each prediction period in multiple prediction periods.

[0055] In step 110, M is a positive integer, and the multiple prediction periods can be a plurality of consecutive historical prediction periods, or a plurality of historical prediction periods with large differences in environmental conditions selected from all historical prediction periods, and the period length of the prediction period can be 3 hours. The historical environmental parameters include geomagnetic parameters, solar radiation parameters, and spatiotemporal parameters. The historical environmental parameters corresponding to the prediction period include input parameters corresponding to M basic prediction models, among which the geomagnetic parameters and solar radiation parameters are time series features, including ap, Ap*, F10.7, M10.7, S10.7, Y10.7, etc., and the spatiotemporal parameters are non-time series features. , including longitude, latitude, altitude, DOY, month, and mean solar time.

[0056] The measured value is the actual detected atmospheric density of each prediction period. For example, the actual satellite telemetry data at a large number of historical moments can be combined, and the detection value of the atmospheric density detector carried by the low-orbit satellite can be used as the measured value. The M predicted values ​​are obtained by predicting the atmospheric density of each prediction period based on the historical environmental parameters of each prediction period by M basic prediction models.

[0057] Exemplarily, the historical prediction periods include period 1, period 2, and period 3. Taking period 1 as an example, the target training data includes historical environmental parameters, measured values, and M predicted values ​​corresponding to period 1. The measured values ​​are the actually detected atmospheric density of period 1, and the M predicted values ​​are obtained by predicting the atmospheric density of period 1 based on the historical environmental parameters of period 1 by M basic prediction models.

[0058] The basic prediction model is the basic atmospheric forecast model, including but not limited to the JB series model, MSIS series model, and DTM series model.

[0059] In some optional embodiments, the M basic prediction models may include at least two of the MSIS20 model, the MSIS00 model, the DTM2020 model, and the JB2008 model.

[0060] In some optional embodiments, the M basic prediction models may include an MSIS20 model, an MSIS00 model, a DTM2020 model and a JB2008 model. This application uses MSIS20, MSIS00, DTM2020 and JB2008 as basic prediction models to make a preliminary prediction of atmospheric density.

[0061] It should be noted that the basic prediction model can be set according to specific needs, and its number can be increased or decreased. This application only points out 4 basic atmospheric density prediction models. In fact, more basic prediction models can be added to improve the generalization ability of the overall model.

[0062] Involving step 120, the first metamodel is trained using the target training data so that the first metamodel assigns corresponding M first weight values ​​to the M prediction values ​​of each prediction period based on the historical environment parameters corresponding to each prediction period.

[0063] In step 120, feature extraction is performed on the geomagnetic parameters and solar radiation parameters of N consecutive steps before each prediction period to obtain time series features, and feature extraction is performed on the spatiotemporal parameters of the prediction period to obtain non-time series features. The time series features, non-time series features, and M predicted values ​​corresponding to each prediction period are used as input features, and M measured values ​​are used as labels to train the first meta-model. N is a positive integer and can be set according to specific needs, for example, it is set to 40.

[0064] The first meta-model obtains M prediction values ​​for each prediction period, and based on all the prediction values ​​under multiple prediction periods, learns how to adjust the weight values ​​of each basic prediction model under different environmental conditions, and assigns the first weight value to the M prediction values, that is, assigns the first weight value to the M basic prediction models.

[0065] For example, the M basic prediction models include the MSIS20 model, the MSIS00 model, the DTM2020 model, and the JB2008 model. The four basic prediction models can predict the atmospheric density of each prediction period based on the historical environmental parameters of the prediction period, and obtain the four prediction values ​​corresponding to each prediction period. , , , , where the JB2008 prediction values ​​for three consecutive hours are averaged and used as the prediction value with a step length of 3 hours The first meta-model may assign four first weight values ​​to the fourth prediction value based on the historical environmental parameters corresponding to the Nth prediction period, perform weighted summation on the four prediction values ​​based on the four first weight values, and output the fused atmospheric density prediction value of the Nth period.

[0066] In some embodiments of the present application, the first meta-model includes a weight model and a fusion layer. The network structure of the weight model is transformer encoder or Lstm. The weight model is used to assign corresponding M first weight values ​​to M prediction values ​​of different prediction periods. The fusion layer is used to perform weighted summation of the M prediction values ​​using the M first weight values. Therefore, the fusion layer can also be a weighted summation layer.

[0067] Specifically, when the network structure of the weight model is transformer encoder, transformer encoder is used as a learner to train the weight model. When the network structure of the weight model is Lstm, Lstm is used as a learner to train the weight model. The weights of each basic prediction model are dynamically calculated according to the environmental characteristics corresponding to the historical environmental parameters. Finally, the prediction values ​​of each basic prediction model are weighted and summed to obtain the prediction result of the first meta-model.

[0068] The input features of the weight model include the features corresponding to the geomagnetic parameters, solar radiation parameters, and spatiotemporal parameters; the output of the weight model is the first weight value of each basic prediction model (the number of output weights is the same as the number of models). In addition to the input of the weight model, the input features of the first meta-model also include the prediction results of the basic prediction models, that is, M prediction values; the output of the first meta-model is the final result of this fusion.

[0069] In an example, for example, there are 4 basic prediction models, and the prediction values ​​of the basic prediction models of the prediction period are [pred1, pred2, pred3, pred4] respectively. At the same time, the weight vector W=[w1, w2, w3, w4, b] output by the weight model, where w1, w2, w3, w4 are the 4 first weight values, and b is the bias term, then the prediction result of the fusion layer of the first meta-model at this moment is As shown in formula (1):

[0070] (1)

[0071] In an embodiment of the present application, various atmospheric density forecast models are used as basic prediction models to make preliminary predictions on atmospheric density, and the weight model and the final weighted sum layer are connected together to form a first meta-model that is fused together. The input of the first meta-model includes the input features of the weight model and the prediction results of each basic prediction model. The first meta-model is trained using a mature machine learning algorithm transformer encoder or LSTM, so that the first meta-model can analyze the performance of each basic prediction model under different spatial environmental conditions, and combine the prediction accuracy of each basic prediction model for atmospheric density under different spatial environmental conditions. Different first weight values ​​are assigned to each basic prediction model under different spatial environmental conditions, so that the prediction results of each basic prediction model after fusion are more accurate, and more accurate atmospheric density prediction results are output.

[0072] In some embodiments of the present application, when the network structure of the weight model is Lstm, the above step 120 uses the target training data to train the first meta-model, which may specifically include the following steps:

[0073] The feature extraction of geomagnetic parameters and solar radiation parameters of N consecutive steps before each prediction period is performed to obtain the time series features. , and extract the features of the spatiotemporal parameters of each prediction period to obtain non-time series features ;

[0074] For time series characteristics After normalization, the high-order features H of the last hidden layer of the weight model in round N are obtained. ;

[0075] Merge H with non-temporal features, input into the fully connected layer (FC), and then connect to a softmax layer to output the weight vector W, which contains M first weight values;

[0076] Performing weighted summation on the prediction values ​​of the M basic prediction models based on the weight vector to obtain a first prediction value;

[0077] The measured values ​​of atmospheric density are used as labels, the mean square error between the measured values ​​and the first predicted values ​​is used as the loss function, and the back propagation algorithm is used to train the first meta-model.

[0078] Specifically, geomagnetic parameters and solar radiation parameters are time series characteristics , including ap, Ap*, F10.7, M10.7, S10.7, and Y10.7. The update frequency of ap, Ap*, and M10.7 is 3 hours, and the update frequency of F10.7, S10.7, and Y10.7 is 1 day. The time series features take 3 hours as the step length. Time series features Before inputting into Lstm, normalization operation is required. Spatiotemporal parameters are non-sequential features. , including longitude, latitude, altitude, DOY, month, and mean solar time. The Lstm sequence length can be 40 or other values, the hidden layer depth can be 1, 2 or other values, and the hidden layer dimension can be 32, 64 or other values, which are not specifically limited in this application.

[0079] The input dimension of the fully connected layer is H and The sum of the dimensions of , the output dimension is 4, which is the number of basic prediction models, and the weight vector W is shown in formula (2):

[0080] W=soft max ( FC(concat[H, X ns ])) (2)

[0081] It should be noted that N can be set according to specific needs, for example, setting N to 40 or other values, and this application does not make any specific limitations on this.

[0082] In some embodiments of the present application, when the network structure of the weight model is a transformer encoder, the above step 120 uses the target training data to train the first meta-model, which may specifically include the following steps:

[0083] The feature extraction of geomagnetic parameters and solar radiation parameters of N consecutive steps before each prediction period is performed to obtain the time series features. , and extract the features of the spatiotemporal parameters of the prediction period to obtain non-time series features ;

[0084] For time series characteristics After normalization, N token original vectors are obtained;

[0085] Input N token original vectors into the weight model so that the token original vectors are converted into embedding vectors of dimension d through a fully connected network to obtain N embedding vectors;

[0086] The embedded vector is processed by a multi-layer encoder to output N output vectors of dimension d;

[0087] The dimension of N output vectors is converted to 1 through the fully connected layer to obtain N output features;

[0088] Merge the N output features with the non-temporal features, input a fully connected layer (FC), and then a softmax layer to output the weight vector W;

[0089] Performing weighted summation on the prediction values ​​of the M basic prediction models based on the weight vector to obtain a first prediction value;

[0090] The measured values ​​of atmospheric density are used as labels, the mean square error between the measured values ​​and the first predicted values ​​is used as the loss function, and the back propagation algorithm is used to train the first meta-model.

[0091] Wherein, d can be 32, 64 or other values, and the number of encoder layers can be 1, 2 or other values, which is not specifically limited in this application.

[0092] In step 130, when the first loss function value of the first metamodel satisfies a first preset condition, it is determined that the training of the first metamodel is completed.

[0093] In step 130, the first loss function value is the mean square error between the first predicted value and the measured value corresponding to each prediction period in multiple prediction periods, and the first predicted value is obtained by weighted summing the M predicted values ​​corresponding to each prediction period using the M first weight values ​​assigned by the first metamodel in each prediction period. The present application uses the stacking method to fuse the models.

[0094] The first meta-model is a fusion model for fusing the prediction results of M basic prediction models, and the first prediction value is the fusion result obtained by weighted summing the prediction results of the M basic prediction models by the first meta-model. Therefore, the mean square error between the first prediction value and the measured value of multiple prediction cycles is calculated. When the mean square error is less than the first preset error threshold, it is determined that the first loss function value of the first meta-model meets the first preset condition, and the training of the first meta-model is completed.

[0095] In some embodiments of the present application, the first metamodel can be trained using a cross-validation method. In this case, the first metamodel includes multiple weight models, and the first weight value output by the first metamodel is the average of multiple weight values ​​output by the multiple weight models.

[0096] Exemplarily, the target training data can be divided into a training set and a test set with a ratio of 5:2 or 5:3. The training set is divided into five parts of training data and validation data with a ratio of 4:1 according to a 5-fold cross validation, so that the first meta-model can include five weight models.

[0097] Involving step 140, historical environmental parameters of the target prediction period and M target prediction values ​​are input into the first metamodel, and M first weight values ​​are assigned to the M target prediction values ​​through the first metamodel.

[0098] In step 140, the target prediction period is a future period, and M target prediction values ​​are obtained by respectively inputting historical environmental parameters corresponding to the target prediction period into M basic prediction models, and the M target prediction values ​​are output by the M basic prediction models based on the historical environmental parameters corresponding to the target prediction period. Through the weight model of the first meta-model, the historical environmental parameters of the target prediction period can be analyzed, and M first weight values ​​can be assigned to the M target prediction values, so that the weight assignment can meet the spatial environmental conditions corresponding to the target prediction period.

[0099] In related technologies, the granularity of input parameters of some models is relatively coarse, resulting in large errors in the short-term changes in orbital atmospheric density in the model prediction results, which limits the application scenarios that require high-precision short-term forecasts. Some models (such as MSISE and DTM) do not take into account the impact of changes in different bands of solar radiation on atmospheric density at different altitudes, which is caused by the fact that the input parameters related to solar radiation are too single (only F10.7).

[0100] In some embodiments of the present application, the M basic prediction models may include at least two of the MSIS20 model, the MSIS00 model, the DTM2020 model, and the JB2008 model, and the historical environmental parameters may include geomagnetic parameters, solar radiation parameters, and space-time parameters.

[0101] In the above embodiment, in order to optimize the input parameters and make the atmospheric density calculation more suitable for the atmospheric density calculation in the ultra-low orbit scenario, before the above step 140 inputs the historical environmental parameters of the target prediction period and the M target prediction values ​​into the first meta-model, the method may further include the following steps:

[0102] If the atmospheric density of the target forecast period is below 500 km, the historical environmental parameters input to the MSIS20 model, MSIS00 model and DTM2020 model are updated;

[0103] The geomagnetic index Ap in the geomagnetic parameters is replaced by the sliding geomagnetic index Ap*, the 81-day central average value of F10.7 in the solar radiation parameters is replaced by the 81-day central average value of M10.7, and the F10.7 value of the previous day in the solar radiation parameters is replaced by the M10.7 mean value of the previous day;

[0104] Among them, the sliding geomagnetic index Ap* is obtained by averaging the geomagnetic indices ap of the target prediction period and the seven prediction periods before the target prediction period. F10.7 is updated daily, and M10.7 is updated every 3 hours.

[0105] Specifically, the model prediction frequency of the MSIS20 model is every 3 hours, and its original input parameters are as follows: space-time parameters: DOY, UTSEC (universal time seconds), geoid height, geodetic latitude and geodetic accuracy; solar radiation parameters: 81-day central average value of F10.7, F10.7 value of the previous day; geomagnetic parameters: Ap value of the day, Ap value of multiple time periods before the current moment.

[0106] For this application, the optimizations made on the input parameters are: ① Change the 81-day central average value of F10.7 to the 81-day central average value of M10.7; ② Change the F10.7 value of the previous day (daily frequency) to the M10.7 value of the previous period (3h frequency); ③ Change the Ap value of the day to the sliding geomagnetic index Ap* corresponding to the target prediction period. Among them, ①② are only performed when calculating the atmospheric density below 500km, and ③ can be used for calculating all atmospheric densities, or it can be performed only when calculating the atmospheric density below 500km. The model prediction frequency of the MSIS00 model is every 3 hours, and its input parameters and optimization methods are basically the same as those of MSIS20.

[0107] The DTM2020 model is suitable for atmospheric density prediction above 120km, and the model prediction frequency is every 3 hours. The original input parameters of the model are as follows: Spatiotemporal parameters: the same as the MSIS series content, but in a slightly different form; Solar radiation parameters: 81-day central average of F10.7, F10.7 value of the previous day; Geomagnetic index: average kp value of the previous 24 hours (3-hour frequency), kp value 3 hours ago (3-hour frequency). The kp value here can be accurately converted from the ap value. The input optimization methods of this model are ① and ②.

[0108] JB2008 model, the model prediction frequency is 1 hour each time, and its input parameters include: Spatiotemporal parameters: the same as the MSIS series content, but in a slightly different form; Solar radiation parameters: daily values ​​and central 81-day averages of the F10.7 (daily frequency), S10.7 (daily frequency), M10.7 (3h frequency), and Y10.7 (daily frequency) indices. For the F10.7 and S10.7 indices, the best lag time is 1 day. This means that if today is the day when atmospheric density prediction is required, yesterday's F10.7 and S10.7 index values ​​can be used. For the M10.7 index, the best lag time is 2 days, and for the Y10.7 index, the best lag time is 5 days. Geomagnetic index: the previous Dst index (1h frequency) and the previous ap index (3h frequency).

[0109] In an embodiment of the present application, the geomagnetic index Ap in the geomagnetic parameters is replaced with the sliding geomagnetic index Ap*, the 81-day central average of F10.7 in the solar radiation parameters is replaced with the 81-day central average of M10.7, and the F10.7 value of the previous day in the solar radiation parameters is replaced with the M10.7 mean of the previous day. In this way, the time granularity of the sliding geomagnetic index Ap* and the high-frequency M10.7 index is finer, making the model more sensitive to changes in solar activity and more accurately reflecting the real-time state of solar activity. Therefore, by using the M10.7 and Ap* indices, the sensitivity of the overall model to short-term changes in atmospheric density can be enhanced, thereby improving the accuracy of the estimated atmospheric density of the near-ground high-level at the target time. At the same time, the update frequency of the input features is designed to be 3 hours, so that the prediction of changes in atmospheric density is more timely.

[0110] In some embodiments of the present application, the historical environmental parameters include ap, F10.7, M10.7, S10.7, and Y10.7. Thus, through the short-term forecast values ​​of these parameters, it is possible to predict the short-term atmospheric density, so that JB2008 can participate in model fusion, and the input parameters of the first meta-model include ap, F10.7, M10.7, S10.7, and Y10.7. Therefore, the first meta-model can combine the influence of the changes in different bands of solar radiation on the atmospheric density at different altitudes to predict the atmospheric density, so that the prediction results of the atmospheric density include the influence of radiation in different bands, thereby improving its prediction accuracy.

[0111] Involving step 150, M first weight values ​​are used to perform weighted summation on M target prediction values ​​to determine the atmospheric density of the target prediction period.

[0112] In step 150, M target prediction values ​​are weighted and summed using M first weight values ​​to obtain the atmospheric density prediction value of the first meta-model for the target prediction period. Optionally, the present application may directly use the atmospheric density prediction value as the final atmospheric density of the target prediction period, or may further calculate based on the atmospheric density prediction value to obtain the final atmospheric density of the target prediction period.

[0113] In some embodiments of the present application, in order to further improve the accuracy of atmospheric density prediction, after obtaining the target training data in step 110, the method may further include: Figure 2 As shown in step 210 and step 220, the above step 150 may specifically include step 230 to step 250.

[0114] Step 210, training the second meta-model using the target training data, so that the second meta-model predicts the atmospheric density of each prediction period based on the historical environmental parameters and the M prediction values ​​corresponding to each prediction period, and obtains a second prediction value of each prediction period;

[0115] Step 220, when a second loss function value of the second meta-model satisfies a second preset condition, determining that the training of the second meta-model is completed, wherein the second loss function value is a mean square error between a second predicted value and a measured value corresponding to each prediction period in a plurality of prediction periods;

[0116] Step 230, using M first weight values ​​to perform weighted summation on the M target prediction values ​​to obtain the atmospheric density prediction value of the first element model for the target prediction period;

[0117] Step 240, inputting historical environmental parameters and M target prediction values ​​into the second meta-model to obtain the atmospheric density prediction value of the second meta-model for the target prediction period;

[0118] Step 250 , combining the atmospheric density prediction values ​​of the target prediction period with the first metamodel and the second metamodel to determine the atmospheric density of the target prediction period.

[0119] Specifically, feature extraction is performed on the geomagnetic parameters and solar radiation parameters of N consecutive steps before each prediction period to obtain time series features, and feature extraction is performed on the spatiotemporal parameters of the prediction period to obtain non-time series features. The time series features, non-time series features, and M predicted values ​​corresponding to each prediction period are used as input features, and M measured values ​​are used as labels to train the second metamodel. The second metamodel is a fusion model for fusing the prediction results of M basic prediction models. The second prediction value is the fusion result obtained by fitting the prediction results of the M basic prediction models by the second metamodel. Therefore, the mean square error between the second predicted values ​​and the measured values ​​of multiple prediction periods is calculated. When the mean square error is less than the second preset error threshold, it is determined that the second loss function value of the second metamodel meets the second preset condition, and the second metamodel training is completed. During training, the loss function uses the mean square error, and is evaluated through multi-fold cross validation (such as 5 folds) to ensure the robustness of the model performance.

[0120] In an embodiment of the present application, various atmospheric density forecast models are first used as basic prediction models to make a preliminary prediction of atmospheric density, and then different machine learning algorithms are used to train to obtain the first meta-model and the second meta-model. The first meta-model and the second meta-model both take into account enough spatial environment and spatiotemporal parameter characteristics, and adopt a method of mixed fusion of environmental characteristics and basic prediction models, so that the fusion model can better learn the weights of different basic prediction models in various environmental characteristics, which is equivalent to dynamically weighted summing each basic prediction model. Based on this, the first meta-model and the second meta-model respectively output atmospheric density prediction values, and each atmospheric density prediction value integrates the prediction results of each basic prediction model. Therefore, by combining the two fusion results, a more accurate prediction result after the two machine learning algorithms can be obtained.

[0121] In some embodiments of the present application, the second meta-model may be any one of lightGBM, random forest algorithm, vector regression SVR algorithm, Xgboost regression model, KNN regression algorithm, and time series prediction algorithm informer.

[0122] As a specific example, the time series feature takes a step length of 3 hours, and takes 40 consecutive steps of data as model input. The M basic prediction models include MSIS20, MSIS00, DTM2020, and JB2008. The specific training process of different second meta-models is as follows:

[0123] lightGBM as the second meta-model: In this fusion scheme, lightGBM is used as a learner to directly train the meta-model to predict atmospheric density. The input features include 40 time series features (ap, Ap*, F10.7, M10.7, S10.7, Y10.7) with a frequency of 3 hours, a total of 240; 6 spatiotemporal parameter features longitude, latitude, altitude, DOY, month, and mean universal time (whole hour); 6 prediction results of 4 basic prediction models , , , , , The input features of the model do not need to be standardized, and the output of the model is the predicted value of this model. The model's hyperparameters: objective is set to regression, and boosting_type is set to gbdt. Other hyperparameters can be searched through grid search to obtain the best combination. The commonly used hyperparameters are recommended to be n_estimators 100, max_depth 5, feature_fraction 1, min_data_in_leaf 10, and bagging_fraction 0.9 to 1.

[0124] Random forest as the second meta-model: Use the random forest algorithm as a learner to directly train the meta-model to predict the atmospheric density. The input of the second meta-model is the same as the above lightGBM, and the output is The hyperparameters usually involved in tuning are n_estimators, max_depth, min_sample_split, min_sample_leaf, max_leaf_nodes, etc.

[0125] SVM as the second element model: Use the support vector regression (SVR) algorithm in the support vector machine. SVM is suitable for processing data in high-dimensional space, especially when there are a large number of features like in this application. SVR can fit the data well by finding a complex function and has good generalization ability. The input features include 40 time series features with a frequency of 3 hours (ap, Ap*, F10.7, M10.7, S10.7, Y10.7), a total of 240; 6 spatiotemporal parameter features: longitude, latitude, altitude, DOY, month, and mean universal time (whole hour); 6 prediction results of 4 basic prediction models , , , , , Since SVM is sensitive to the dimension of features, the features need to be standardized or normalized. For example, StandardScaler can be used to standardize the features to a mean of 0 and a standard deviation of 1. The output of this model is the predicted value of atmospheric density. .

[0126] SVM includes the following training hyperparameters:

[0127] Kernel is the most important hyperparameter in SVM, which determines how to map data from the original space to a higher dimensional space. In this fusion solution, the kernel function selects the rbf kernel.

[0128] C (regularization parameter): controls the penalty of the error term. The larger the C value, the more the model will tend to fit the training data (which may lead to overfitting); a smaller C value will improve generalization ability, but may lead to underfitting. In this application, you can try different values ​​from 1 to 100, and perform grid search or cross-validation to optimize.

[0129] Epsilon (ε-insensitive loss function parameter): controls the error range between the model prediction and the true value, that is, within what error range the model will not be penalized for deviating from the target value. A smaller epsilon will be sensitive to all errors, while a larger epsilon allows the model to ignore some errors. The commonly used value range is 0.1 to 1.0, which needs to be adjusted according to the data distribution.

[0130] Gamma (kernel function coefficient, applicable to nonlinear kernel functions such as rbf and poly): controls the influence of a single training sample on the decision boundary. A larger gamma value will cause the model to be more dependent on local samples (possibly overfitting), while a smaller gamma value will be smoother and have better generalization ability. This application adjusts from 0.001 to 1.

[0131] In some embodiments of the present application, the above step 250 may include: determining the average of the atmospheric density prediction values ​​of the first metamodel and the second metamodel for the target prediction period as the atmospheric density of the target prediction period.

[0132] In other embodiments of the present application, in order to further improve the prediction accuracy of the atmospheric density in the target prediction period, the above step 250 may specifically include the following steps:

[0133] The first predicted value and the second predicted value corresponding to each prediction period are used as input features, and the measured value corresponding to each prediction period is used as output feature, the linear regression model is trained to obtain the target fusion model, and the model parameters of the target fusion model are updated using the gradient descent method;

[0134] The atmospheric density prediction values ​​of the target prediction period from the first meta-model and the second meta-model are input into the target fusion model, and the output fusion result is used as the atmospheric density of the target prediction period.

[0135] Specifically, in this stage, a linear regression algorithm is used to further fuse the prediction results of the first meta-model and the second meta-model in the first stage to obtain the final prediction result. The target fusion model can be specifically shown as formula (3):

[0136] (3)

[0137] Among them, w and d are the parameters of the linear regression model. is the prediction result of the meta-model and its corresponding weight.

[0138] In the embodiment of the present application, a two-stage model fusion technology is adopted. In the first stage, a machine learning algorithm is used to fuse various basic atmospheric forecast models (basic prediction models), build and train the fusion model, and then use a variety of different machine learning machines to train multiple fusion models to obtain a first meta-model and a second meta-model; in the second stage, a simple linear regression model is used to further fuse the prediction results of the first meta-model and the second meta-model, and the two are fitted. Combining the fusion operation of the atmospheric density prediction values ​​of the two stages, the prediction accuracy of the atmospheric density can be effectively improved.

[0139] It should be noted that the above linear regression model can also be replaced by other more complex models to improve the fitting ability of the target fusion model.

[0140] In other embodiments of the present application, the first predicted value and the second predicted value corresponding to each prediction period are used as input features, and the measured value corresponding to each prediction period is used as an output feature, and the linear regression model is trained to obtain a target fusion model.

[0141] In some embodiments of the present application, the above step 250 combines the first metamodel and the second metamodel for the atmospheric density prediction value of the target prediction period to determine the atmospheric density of the target prediction period, which may specifically include the following steps:

[0142] Inputting historical environmental parameters of the target prediction period and M target prediction values ​​into the third meta-model, assigning M second weight values ​​to the M target prediction values ​​through the weight model in the third meta-model, and performing weighted summation of the M target prediction values ​​by the fusion layer in the third meta-model using the M second weight values ​​to obtain the atmospheric density prediction value of the third meta-model for the target prediction period;

[0143] Inputting historical environmental parameters of a target prediction period and M target prediction values ​​into the fourth element model, obtaining the atmospheric density prediction value of the fourth element model for the target prediction period;

[0144] Inputting historical environmental parameters of the target prediction period and M target prediction values ​​into the fifth element model, obtaining the atmospheric density prediction value of the fifth element model for the target prediction period;

[0145] Input the atmospheric density prediction values ​​of the five meta-models for the target prediction period into the pre-trained target fusion model, and obtain the atmospheric density of the target prediction period after fusion;

[0146] Specifically, the weight model in the first metamodel is transformer encoder, the second metamodel is lightGBM, the weight model in the third metamodel is Lstm, the fourth metamodel is random forest algorithm, and the fifth metamodel is vector regression SVR algorithm. The third metamodel, the fourth metamodel and the fifth metamodel are all trained based on the target training data. The second weight value is the weight value assigned by the third metamodel.

[0147] In an embodiment of the present application, in order to make a comprehensive and integrated prediction of the atmospheric density of the target prediction period, the present application continues to introduce Lstm, random forest algorithm, and SVR algorithm on the basis of transformer encoder and lightGBM. Among them, in combination with transformer encoder and Lstm, it is possible to assign appropriate weights, i.e., a first weight value and a second weight value, to these basic prediction models and their prediction results based on the performance of different basic prediction models under the spatial environmental conditions corresponding to the target prediction period, and combine the two weight values ​​to output a more accurate atmospheric density prediction result. On this basis, combined with the advantages of the integrated learning method of lightGBM, random forest algorithm, and SVR algorithm, the prediction results of the five meta-models are further integrated, and the prediction results of the five meta-models are combined to improve the overall prediction performance of atmospheric density, which can make up for the shortcomings of a single model and reduce the overall prediction error when predicting atmospheric density when facing different environmental conditions.

[0148] In some embodiments of the present application, before inputting the historical environment parameters of the target prediction period and the M target prediction values ​​into the third meta-model, the method may further include the following steps:

[0149] The third meta-model is trained using the target training data, so that the third meta-model assigns corresponding M second weight values ​​to the M prediction values ​​of each prediction period based on the historical environment parameters corresponding to each prediction period;

[0150] When the third loss function value of the third meta-model satisfies the third preset condition, determining that the training of the third meta-model is completed, wherein the third loss function value is the mean square error between the third predicted value corresponding to each prediction period in the multiple prediction periods and the measured value, and the third predicted value is obtained by weighted summing the M predicted values ​​corresponding to each prediction period using the M second weight values ​​assigned by the third meta-model in each prediction period;

[0151] The fourth meta-model and the fifth meta-model are trained by using the target training data, so that the fourth meta-model and the fifth meta-model predict the atmospheric density of each prediction period based on the historical environmental parameters and the M prediction values ​​corresponding to each prediction period, and obtain the fourth prediction value and the fifth prediction value of each prediction period;

[0152] When a fourth loss function value of the fourth meta-model satisfies a fourth preset condition, determining that the training of the fourth meta-model is completed, wherein the fourth loss function value is a mean square error between a fourth predicted value and a measured value corresponding to each prediction period in a plurality of prediction periods;

[0153] When the fifth loss function value of the fifth meta-model satisfies the fifth preset condition, it is determined that the training of the fifth meta-model is completed, wherein the fifth loss function value is the mean square error between the fifth predicted value and the measured value corresponding to each prediction period in multiple prediction periods.

[0154] Specifically, the five meta-models are all trained based on the target training data. The third preset condition is that the third loss function value is less than the third preset error threshold, the fourth preset condition is that the fourth loss function value is less than the fourth preset error threshold, and the fifth preset condition is that the fifth loss function value is less than the fifth preset error threshold. The five preset error thresholds corresponding to the five meta-models can be set according to specific needs, and can be set to the same value or to different values. This application does not make any specific limitations on this.

[0155] When the atmospheric density of the target prediction period is determined based on five meta-models, it also includes: taking the first prediction value, the second prediction value, the third prediction value, the fourth prediction value and the fifth prediction value corresponding to each prediction period as input features, and the measured value corresponding to each prediction period as output features, training the linear regression model to obtain the target fusion model, and using the gradient descent method to update the model parameters of the target fusion model.

[0156] In the embodiment of the present application, the third meta-model, the fourth meta-model and the fifth meta-model can all be trained using the same batch of target training data, which can effectively utilize the limited sampling data and improve the utilization rate of the sampling data. Moreover, based on the advantages of different machine learning algorithms, different meta-models can be trained separately, and the prediction results of the five meta-models can be further integrated. The prediction results of the five meta-models can be combined to improve the overall prediction performance of the atmospheric density, which can make up for the shortcomings of a single model and reduce the overall prediction error when predicting the atmospheric density when facing different environmental conditions.

[0157] As a specific example, combining a large number of satellite telemetry data at historical moments, the corresponding solar index, space environment data and spatiotemporal parameters, the predicted values ​​calculated using five meta-models are , , , , .Will , , , , As input features, the atmospheric density measured by satellite As a label, a linear regression model is trained. The loss function of the regression model uses the mean square error (MSE) and the gradient descent method is used to update the model parameters.

[0158] Corresponding to the method embodiment of the present application, the present application also provides an atmospheric density prediction device based on multi-model fusion.

[0159] Figure 3 Schematic diagram of the structure of an atmospheric density prediction device based on multi-model fusion provided in an embodiment of the present application. Figure 3 As shown, the atmospheric density prediction device 300 based on multi-model fusion may include: an acquisition module 310, a model training module 320, a determination module 330, and a weight allocation module 340.

[0160] Among them, the acquisition module 310 is used to acquire target training data, wherein the target training data includes historical environmental parameters, measured values ​​and M predicted values ​​corresponding to each prediction period in multiple prediction periods, the measured value is the atmospheric density of each prediction period actually detected, and the M predicted values ​​are obtained by predicting the atmospheric density of each prediction period by M basic prediction models based on the historical environmental parameters of each prediction period; the model training module 320 is used to train the first metamodel using the target training data, so that the first metamodel assigns corresponding M first weight values ​​to the M predicted values ​​of each prediction period based on the historical environmental parameters corresponding to each prediction period; the determination module 330 is used to determine the first metamodel when the first loss function value of the first metamodel meets the first preset condition. The model training is completed, wherein the first loss function value is the mean square error between the first prediction value corresponding to each prediction period in multiple prediction periods and the measured value, and the first prediction value is obtained by weighted summing the M prediction values ​​corresponding to each prediction period using the M first weight values ​​assigned by the first meta-model in each prediction period; the weight allocation module 340 is used to input the historical environmental parameters and M target prediction values ​​corresponding to the target prediction period into the first meta-model, and assign M first weight values ​​to the M target prediction values ​​through the first meta-model, wherein the M target prediction values ​​are output by M basic prediction models based on the historical environmental parameters corresponding to the target prediction period; the determination module 330 is also used to perform weighted summing of the M target prediction values ​​using the M first weight values ​​to determine the atmospheric density of the target prediction period.

[0161] The atmospheric density prediction device based on multi-model fusion provided by the embodiment of the present application takes into account that different basic prediction models have different prediction accuracy performance capabilities for atmospheric density under different environmental conditions. The present application uses historical environmental parameters, measured values, and M prediction values ​​output by M basic prediction models for training under different prediction periods. Different prediction periods correspond to different spatial environmental conditions, so training is based on target training data, considering enough spatial environmental features, and adopting a method of hybrid fusion of spatial environmental features and basic prediction models, so that the first meta-model can better learn the weights of different basic prediction models in various environmental features, which is equivalent to dynamically weighted summing each basic prediction model, combining the prediction results of multiple basic prediction models to improve the overall prediction performance of the first meta-model, so that the first meta-model has better generalization ability, can strengthen the atmospheric density prediction advantages of different basic prediction models under different environmental conditions, make up for the shortcomings of using a single basic prediction model for atmospheric density prediction, improve the accuracy and robustness of the first meta-model, and reduce the overall prediction error when predicting atmospheric density when facing different environmental conditions.

[0162] In some embodiments of the present application, the first meta-model includes a weight model and a fusion layer. The network structure of the weight model is transformer encoder or Lstm. The weight model is used to assign corresponding M first weight values ​​to M prediction values ​​of different prediction periods. The fusion layer is used to perform weighted summation of the M prediction values ​​using the M first weight values.

[0163] In some embodiments of the present application, the model training module is also used to train the second metamodel using the target training data after obtaining the target training data, so that the second metamodel predicts the atmospheric density of each prediction period based on the historical environmental parameters and M prediction values ​​corresponding to each prediction period, and obtains the second prediction value of each prediction period; the determination module is also used to determine that the training of the second metamodel is completed when the second loss function value of the second metamodel satisfies the second preset condition, wherein the second loss function value is the mean square error between the second prediction value corresponding to each prediction period in multiple prediction periods and the measured value; the determination module is specifically used to: use M first weight values ​​to weighted sum the M target prediction values ​​to obtain the atmospheric density prediction value of the first metamodel for the target prediction period; input the historical environmental parameters and M target prediction values ​​into the second metamodel to obtain the atmospheric density prediction value of the second metamodel for the target prediction period; combine the atmospheric density prediction values ​​of the first metamodel and the second metamodel for the target prediction period to determine the atmospheric density of the target prediction period.

[0164] In some embodiments of the present application, the second meta-model is any one of lightGBM, random forest algorithm, vector regression SVR algorithm, Xgboost regression model, KNN regression algorithm, and time series prediction algorithm informer.

[0165] In some embodiments of the present application, the determination module is specifically used to: use the first prediction value and the second prediction value corresponding to each prediction period as input features, and the measured value corresponding to each prediction period as output features, train the linear regression model to obtain a target fusion model, and use the gradient descent method to update the model parameters of the target fusion model; input the atmospheric density prediction values ​​of the first metamodel and the second metamodel for the target prediction period into the target fusion model, and use the output fusion result as the atmospheric density of the target prediction period.

[0166] In some embodiments of the present application, the determination module is specifically used to: input historical environmental parameters and M target prediction values ​​of the target prediction period into the third metamodel, assign M second weight values ​​to the M target prediction values ​​through the weight model in the third metamodel, and the fusion layer in the third metamodel uses the M second weight values ​​to weighted sum the M target prediction values ​​to obtain the atmospheric density prediction value of the third metamodel for the target prediction period; input historical environmental parameters and M target prediction values ​​of the target prediction period into the fourth metamodel to obtain the atmospheric density prediction value of the fourth metamodel for the target prediction period; input historical environmental parameters and M target prediction values ​​of the target prediction period into the fifth metamodel to obtain the atmospheric density prediction value of the fifth metamodel for the target prediction period; input atmospheric density prediction values ​​of the five metamodels for the target prediction period into the pre-trained target fusion model, and obtain the atmospheric density of the target prediction period after fusion; wherein the weight model in the first metamodel is transformer encoder, the second metamodel is lightGBM, the weight model in the third metamodel is Lstm, the fourth metamodel is the random forest algorithm, and the fifth metamodel is the vector regression SVR algorithm. The third metamodel, the fourth metamodel and the fifth metamodel are all trained based on the target training data.

[0167] In some embodiments of the present application, the model training module is also used to: before inputting the historical environment parameters of the target prediction period and M target prediction values ​​into the third metamodel, train the third metamodel using the target training data so that the third metamodel assigns corresponding M second weight values ​​to the M prediction values ​​of each prediction period based on the historical environment parameters corresponding to each prediction period; determine that the training of the third metamodel is completed when the third loss function value of the third metamodel satisfies the third preset condition, wherein the third loss function value is the mean square error between the third prediction value corresponding to each prediction period in multiple prediction periods and the measured value, and the third prediction value is obtained by weighted summing the M prediction values ​​corresponding to each prediction period using the M second weight values ​​assigned by the third metamodel in each prediction period; train the fourth metamodel and the fifth metamodel using the target training data, So that the fourth meta-model and the fifth meta-model predict the atmospheric density of each prediction period based on the historical environmental parameters and M prediction values ​​corresponding to each prediction period, and obtain the fourth prediction value and the fifth prediction value of each prediction period; when the fourth loss function value of the fourth meta-model satisfies the fourth preset condition, it is determined that the training of the fourth meta-model is completed, wherein the fourth loss function value is the mean square error between the fourth prediction value corresponding to each prediction period in multiple prediction periods and the measured value; when the fifth loss function value of the fifth meta-model satisfies the fifth preset condition, it is determined that the training of the fifth meta-model is completed, wherein the fifth loss function value is the mean square error between the fifth prediction value corresponding to each prediction period in multiple prediction periods and the measured value.

[0168] In some embodiments of the present application, the M basic prediction models include at least two of the MSIS20 model, the MSIS00 model, the DTM2020 model, and the JB2008 model, the historical environmental parameters include geomagnetic parameters, solar radiation parameters, and spatiotemporal parameters, and also include a parameter adjustment module for inputting the historical environmental parameters of the target prediction period and the M target prediction values ​​into the first meta-model, if the atmospheric density of the target prediction period is below 500 km atmospheric density, then adjusting the atmospheric density of the target prediction period to the MSIS20 model, the MSIS00 model, and the DTM2020 model. The historical environmental parameters of the 020 model are updated; the geomagnetic index Ap in the geomagnetic parameters is replaced by the sliding geomagnetic index Ap*, the 81-day central average of F10.7 in the solar radiation parameters is replaced by the 81-day central average of M10.7, and the F10.7 value of the previous day in the solar radiation parameters is replaced by the M10.7 mean of the previous day; among them, the sliding geomagnetic index Ap* is obtained by averaging the geomagnetic index ap of the target prediction period and the seven prediction periods before the target prediction period, F10.7 is updated daily, and M10.7 is updated every 3 hours.

[0169] The atmospheric density prediction device based on multi-model fusion provided in the embodiment of the present application can achieve Figure 1-2 The various processes implemented by the service platform in the method embodiment can achieve the same technical effect. To avoid repetition, they will not be described here.

[0170] Figure 4 It is a schematic diagram of the hardware structure of an electronic device provided in an embodiment of the present application.

[0171] like Figure 4 As shown, the electronic device 400 includes a memory 401 , a processor 402 , and a computer program stored in the memory 401 and executable on the processor 402 .

[0172] In an example, the processor 402 may include a central processing unit (CPU), or an application specific integrated circuit (ASIC), or may be configured to implement one or more integrated circuits of the embodiments of the present application.

[0173] The memory 401 may include a read-only memory (ROM), a random access memory (RAM), a disk storage medium device, an optical storage medium device, a flash memory device, an electrical, optical or other physical / tangible memory storage device. Therefore, generally, the memory includes one or more tangible (non-transitory) computer-readable storage media (e.g., a memory device) encoded with software including computer executable instructions, and when the software is executed (e.g., by one or more processors), it can be operated to perform the operations described with reference to the card opening method in the embodiment according to the first aspect of the present application.

[0174] The processor 402 runs a computer program corresponding to the executable program code by reading the executable program code stored in the memory 401, so as to implement the card activation method in the embodiment of the first aspect.

[0175] In some examples, the electronic device 400 may further include a communication interface 403 and a bus 410. Figure 4 As shown, the memory 401, the processor 402, and the communication interface 403 are connected via a bus 410 and communicate with each other.

[0176] The communication interface 403 is mainly used to realize the communication between the modules, devices, units and / or equipment in the embodiment of the present application. The input device and / or output device can also be accessed through the communication interface 403.

[0177] The bus 410 includes hardware, software, or both, coupling the components of the electronic device 400 to each other. By way of example and not limitation, the bus 410 may include an Accelerated Graphics Port (AGP) or other graphics bus, an Enhanced Industry Standard Architecture (EISA) bus, a Front Side Bus (FSB), a Hyper Transport (HT) interconnect, an Industry Standard Architecture (ISA) bus, an InfiniBand interconnect, a Low Pin Count (LPC) bus, a Memory Bus, a Micro Channel Architecture (MCA) bus, a Peripheral Component Interconnect (PCI) bus, a PCI-Express (PCI-E) bus, a Serial Advanced Technology Attachment (SATA) bus, a Video Electronics Standards Association Local Bus (VLB) bus, or other suitable buses or a combination of two or more of these. Where appropriate, the bus 410 may include one or more buses. Although embodiments of the present application describe and illustrate a particular bus, the present application contemplates any suitable bus or interconnect.

[0178] The electronic device provided in the embodiment of the present application can realize Figure 1-2 The various processes implemented by the electronic device in the method embodiment can achieve the same technical effect, and to avoid repetition, they will not be described here.

[0179] In combination with the atmospheric density prediction method based on multi-model fusion in the above embodiment, the embodiment of the present application can provide a computer storage medium for implementation. The computer storage medium stores computer program instructions; when the computer program instructions are executed by the processor, the steps of any one of the atmospheric density prediction methods based on multi-model fusion in the above embodiment are implemented.

[0180] In combination with the atmospheric density prediction method based on multi-model fusion in the above embodiment, the embodiment of the present application can provide a computer program product for implementation. The (computer) program product is stored in a non-volatile storage medium, and when the program product is executed by at least one processor, it implements the steps of any one of the atmospheric density prediction methods based on multi-model fusion in the above embodiment.

[0181] An embodiment of the present application further provides a chip, which includes a processor and a communication interface, wherein the communication interface and the processor are coupled, and the processor is used to run programs or instructions to implement the various processes of the above-mentioned atmospheric density prediction method embodiment based on multi-model fusion, and can achieve the same technical effect. To avoid repetition, it will not be repeated here.

[0182] It should be understood that the chip mentioned in the embodiments of the present application can also be called a system-level chip, a system chip, a chip system or a system-on-chip chip, etc.

[0183] It should be clear that the present application is not limited to the specific configuration and processing described above and shown in the figures. For the sake of simplicity, a detailed description of the known method is omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of the present application is not limited to the specific steps described and shown, and those skilled in the art can make various changes, modifications and additions, or change the order between the steps after understanding the spirit of the present application.

[0184] The functional blocks shown in the structural block diagram described above can be implemented as hardware, software, firmware or a combination thereof. When implemented in hardware, it can be, for example, an electronic circuit, an application-specific integrated circuit (ASIC), appropriate firmware, a plug-in, a function card, etc. When implemented in software, the elements of the present application are programs or code segments used to perform the required tasks. The program or code segment can be stored in a machine-readable medium, or transmitted on a transmission medium or a communication link via a data signal carried in a carrier. "Machine-readable medium" may include any medium capable of storing or transmitting information. Examples of machine-readable media include electronic circuits, semiconductor memory devices, ROMs, flash memories, erasable ROMs (EROMs), floppy disks, CD-ROMs, optical disks, hard disks, optical fiber media, radio frequency (RF) links, etc. The code segment can be downloaded via a computer network such as the Internet, an intranet, etc.

[0185] It should also be noted that the exemplary embodiments mentioned in this application describe some methods or systems based on a series of steps or devices. However, this application is not limited to the order of the above steps, that is, the steps can be performed in the order mentioned in the embodiment, or in a different order from the embodiment, or several steps can be performed simultaneously.

[0186] Aspects of the present disclosure are described above with reference to the flowcharts and / or block diagrams of the methods, devices (systems) and computer program products according to the embodiments of the present disclosure. It should be understood that each box in the flowchart and / or block diagram and the combination of each box in the flowchart and / or block diagram can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device to produce a machine so that these instructions executed by the processor of the computer or other programmable data processing device enable the implementation of the functions / actions specified in one or more boxes of the flowchart and / or block diagram. Such a processor can be, but is not limited to, a general-purpose processor, a special-purpose processor, a special application processor, or a field programmable logic circuit. It can also be understood that each box in the block diagram and / or flowchart and the combination of boxes in the block diagram and / or flowchart can also be implemented by dedicated hardware that performs the specified function or action, or can be implemented by a combination of dedicated hardware and computer instructions.

[0187] The above is only a specific implementation of the present application. Those skilled in the art can clearly understand that for the convenience and simplicity of description, the specific working processes of the systems, modules and units described above can refer to the corresponding processes in the aforementioned method embodiments, and will not be repeated here. It should be understood that the protection scope of the present application is not limited to this. Any technician familiar with the technical field can easily think of various equivalent modifications or replacements within the technical scope disclosed in this application, and these modifications or replacements should be included in the protection scope of this application.

Claims

1. A method for predicting atmospheric density based on multi-model fusion, characterized in that: include: Obtain target training data, wherein the target training data includes historical environmental parameters, measured values, and M predicted values ​​corresponding to each prediction period in multiple prediction periods, wherein the measured values ​​are actually detected atmospheric density of each prediction period, and the M predicted values ​​are obtained by predicting the atmospheric density of each prediction period by M basic prediction models based on the historical environmental parameters of each prediction period; Using the target training data to train the first meta-model, so that the first meta-model assigns corresponding M first weight values ​​to the M prediction values ​​of each prediction period based on the historical environment parameters corresponding to each prediction period; When a first loss function value of the first metamodel satisfies a first preset condition, determining that the training of the first metamodel is completed, wherein the first loss function value is a mean square error between a first predicted value corresponding to each prediction period in a plurality of prediction periods and an actual value, and the first predicted value is obtained by weighted summing the M predicted values ​​corresponding to each prediction period using the M first weight values ​​assigned by the first metamodel in each prediction period; Inputting historical environment parameters corresponding to a target prediction period and M target prediction values ​​into the first meta-model, and assigning M first weight values ​​to the M target prediction values ​​through the first meta-model, wherein the M target prediction values ​​are output by the M basic prediction models based on the historical environment parameters corresponding to the target prediction period; The M target prediction values ​​are weightedly summed using the M first weight values ​​to determine the atmospheric density of the target prediction period.

2. The method according to claim 1, characterized in that The first meta-model includes a weight model and a fusion layer. The network structure of the weight model is transformer encoder or Lstm. The weight model is used to assign corresponding M first weight values ​​to M prediction values ​​of different prediction periods. The fusion layer is used to perform weighted summation of the M prediction values ​​using the M first weight values.

3. The method according to claim 1, characterized in that After acquiring the target training data, the method further includes: The target training data is used to train the second meta-model, so that the second meta-model predicts the atmospheric density of each prediction period based on the historical environmental parameters and the M prediction values ​​corresponding to each prediction period, and obtains the second prediction value of each prediction period; When the second loss function value of the second meta-model satisfies the second preset condition, determining that the training of the second meta-model is completed, wherein the second loss function value is the mean square error between the second predicted value and the measured value corresponding to each prediction period in multiple prediction periods; The method of weighting and summing the M target prediction values ​​by using the M first weight values ​​to determine the atmospheric density of the target prediction period includes: Using the M first weight values ​​to perform weighted summation on the M target prediction values, to obtain an atmospheric density prediction value of the first element model for the target prediction period; Inputting the historical environmental parameters and the M target prediction values ​​into the second meta-model to obtain the atmospheric density prediction value of the second meta-model for the target prediction period; The atmospheric density of the target prediction period is determined by combining the atmospheric density prediction values ​​of the target prediction period with the atmospheric density of the target prediction period by the first meta-model and the second meta-model.

4. The method according to claim 3, characterized in that The second meta-model is any one of lightGBM, random forest algorithm, vector regression SVR algorithm, Xgboost regression model, KNN regression algorithm, and time series prediction algorithm informer.

5. The method according to claim 3, characterized in that: The step of combining the first meta-model and the second meta-model with respect to the atmospheric density prediction value of the target prediction period to determine the atmospheric density of the target prediction period includes: The first predicted value and the second predicted value corresponding to each prediction period are used as input features, and the measured value corresponding to each prediction period is used as an output feature, a linear regression model is trained to obtain a target fusion model, and a model parameter of the target fusion model is updated using a gradient descent method; The atmospheric density prediction values ​​of the target prediction period by the first meta-model and the second meta-model are input into the target fusion model, and the output fusion result is used as the atmospheric density of the target prediction period.

6. The method according to claim 3, characterized in that The step of combining the first meta-model and the second meta-model with respect to the atmospheric density prediction value of the target prediction period to determine the atmospheric density of the target prediction period includes: Inputting historical environmental parameters of the target prediction period and M target prediction values ​​into the third metamodel, assigning M second weight values ​​to the M target prediction values ​​through the weight model in the third metamodel, and performing weighted summation on the M target prediction values ​​by the fusion layer in the third metamodel using the M second weight values ​​to obtain the atmospheric density prediction value of the third metamodel for the target prediction period; Inputting the historical environmental parameters of the target prediction period and the M target prediction values ​​into the fourth element model to obtain the atmospheric density prediction value of the fourth element model for the target prediction period; Inputting the historical environmental parameters of the target prediction period and the M target prediction values ​​into the fifth element model to obtain the atmospheric density prediction value of the fifth element model for the target prediction period; Inputting the atmospheric density prediction values ​​of the five meta-models for the target prediction period into the pre-trained target fusion model, and obtaining the atmospheric density of the target prediction period after fusion; Among them, the weight model in the first metamodel is transformer encoder, the second metamodel is lightGBM, the weight model in the third metamodel is Lstm, the fourth metamodel is random forest algorithm, and the fifth metamodel is vector regression SVR algorithm. The third metamodel, the fourth metamodel and the fifth metamodel are all trained based on the target training data.

7. The method according to claim 6, characterized in that Before inputting the historical environment parameters of the target prediction period and the M target prediction values ​​into the third meta-model, the method further includes: The third meta-model is trained using the target training data, so that the third meta-model assigns corresponding M second weight values ​​to the M prediction values ​​of each prediction period based on the historical environment parameters corresponding to each prediction period; In the case where the third loss function value of the third meta-model satisfies the third preset condition, determining that the training of the third meta-model is completed, wherein the third loss function value is the mean square error between the third predicted value corresponding to each prediction period in multiple prediction periods and the measured value, and the third predicted value is obtained by weighted summing the M predicted values ​​corresponding to each prediction period using the M second weight values ​​assigned by the third meta-model in each prediction period; The fourth meta-model and the fifth meta-model are trained by using the target training data, so that the fourth meta-model and the fifth meta-model predict the atmospheric density of each prediction period based on the historical environmental parameters and the M prediction values ​​corresponding to each prediction period, and obtain the fourth prediction value and the fifth prediction value of each prediction period; When a fourth loss function value of the fourth meta-model satisfies a fourth preset condition, determining that the training of the fourth meta-model is completed, wherein the fourth loss function value is a mean square error between a fourth predicted value and a measured value corresponding to each prediction period in a plurality of prediction periods; When the fifth loss function value of the fifth meta-model satisfies the fifth preset condition, it is determined that the training of the fifth meta-model is completed, wherein the fifth loss function value is the mean square error between the fifth predicted value and the measured value corresponding to each prediction period in multiple prediction periods.

8. The method according to claim 1, characterized in that The M basic prediction models include at least two of the MSIS20 model, the MSIS00 model, the DTM2020 model, and the JB2008 model; the historical environmental parameters include geomagnetic parameters, solar radiation parameters, and spatiotemporal parameters; and before inputting the historical environmental parameters of the target prediction period and the M target prediction values ​​into the first meta-model, the following is further included: If the atmospheric density of the target prediction period is below 500 km, the historical environmental parameters input to the MSIS20 model, the MSIS00 model and the DTM2020 model are updated; The geomagnetic index Ap in the geomagnetic parameters is replaced by the sliding geomagnetic index Ap*, the 81-day central average value of F10.7 in the solar radiation parameters is replaced by the 81-day central average value of M10.7, and the F10.7 value of the previous day in the solar radiation parameters is replaced by the M10.7 mean value of the previous day; The sliding geomagnetic index Ap* is obtained by averaging the geomagnetic indices ap of the target prediction period and the seven prediction periods before the target prediction period. F10.7 is updated daily, and M10.7 is updated every 3 hours.

9. An atmospheric density prediction device based on multi-model fusion, characterized in that: include: An acquisition module is used to acquire target training data, wherein the target training data includes historical environmental parameters, measured values ​​and M predicted values ​​corresponding to each prediction period in multiple prediction periods, wherein the measured values ​​are actually detected atmospheric density of each prediction period, and the M predicted values ​​are obtained by predicting the atmospheric density of each prediction period by M basic prediction models based on the historical environmental parameters of each prediction period; A model training module, configured to train the first meta-model using the target training data, so that the first meta-model assigns corresponding M first weight values ​​to M prediction values ​​of each prediction period based on the historical environment parameters corresponding to each prediction period; A determination module, configured to determine that the training of the first metamodel is completed when a first loss function value of the first metamodel satisfies a first preset condition, wherein the first loss function value is a mean square error between a first predicted value corresponding to each prediction period in a plurality of prediction periods and an actual value, and the first predicted value is obtained by weighted summing the M predicted values ​​corresponding to each prediction period using the M first weight values ​​assigned by the first metamodel in each prediction period; A weight assignment module, configured to input historical environment parameters corresponding to a target prediction period and M target prediction values ​​into the first meta-model, and assign M first weight values ​​to the M target prediction values ​​through the first meta-model, wherein the M target prediction values ​​are output by the M basic prediction models based on the historical environment parameters corresponding to the target prediction period; The determination module is further used to perform weighted summation on the M target prediction values ​​using the M first weight values ​​to determine the atmospheric density of the target prediction period.

10. An electronic device, characterized in that: The electronic device comprises: a processor and a memory storing computer program instructions; when the electronic device executes the computer program instructions, the method according to any one of claims 1 to 8 is implemented.

Images