Prediction Method, System, Device and Medium for Cloud Physical Parameters

Abstract

The present invention relates to the field of icing prediction, and particularly to a prediction method, system, device and medium for cloud physical parameters, which includes: obtaining a first meteorological sample set and inputting it into a convolutional neural network for iterative calculation. The input end of the convolutional neural network is T , RH , P , Q cloud , Q rain , IWC , W , Q ice , N ice and N rain ; the output end of the convolutional neural network is the probability value of MVD; meteorological sample combinations with the probability value of the median volume diameter belonging to a set probability interval and the temperature being greater than a set temperature threshold are combined to form a second meteorological sample set; the second meteorological sample set is input into the convolutional neural network again for updated iterative calculation, and finally a corresponding prediction model is obtained. The output of the prediction model is: LWC , MVD . The present invention proposes a method that can accurately identify large and small droplets and use their characteristics for LWC , MVD accurate prediction.

Description

Technical Field

[0001] The present invention relates to the technical field of icing prediction, and particularly to a prediction method, system, device and medium for cloud physical parameters. Background Art

[0002] When an aircraft passes through clouds, icing is likely to occur when it encounters supercooled water, which will significantly reduce the aerodynamic performance of the aircraft and may lead to serious aviation safety accidents. Improving the accuracy of aircraft icing prediction by accurately predicting cloud physical parameters in the clouds through which the aircraft passes is one of the effective ways to ensure aviation safety. In recent years, a large number of scholars have studied aircraft icing prediction. Bernstein et al. proposed a Current Icing Potential (CIP) algorithm, and through fusing the latest observation data and the output of the NOAA Rapid Refresh (RAP) model, hourly icing potential and Supercooled Large Droplets (SLD) forecasts were made for the airspace of the United States. Subsequently, a Numerical Model-based Icing Potential Prediction (FIP) algorithm was developed on the basis of the CIP algorithm, which laid the foundation for establishing an icing prediction system for an aviation meteorological center. However, with the development of aircraft anti-icing and de-icing technologies, the impact of trace icing on aircraft is relatively small, and the technical difficulty of icing prediction has also increased significantly.

[0003] In recent years, researchers have used advanced data assimilation technologies and improved microphysical schemes to improve the prediction accuracy. Through the cloud microphysical scheme proposed by Thompson et al., the FAA Aviation Weather Research Program has strengthened the icing prediction for aircraft and the ground as well as quantitative precipitation prediction, so as to be able to effectively predict the cloud droplet number concentration. Davis et al. conducted a simulation study on a wind power icing event in Sweden, tested different combinations of 3 microphysical schemes and 3 indefinite boundary layer schemes, and found that the combination of the Thompson microphysical scheme and the Edelhammer scheme had relatively good effects, but the false alarm rate was still relatively high.

[0004] In addition, Patent Application CN111738481A discloses a prediction method for aircraft icing meteorological parameter MVD based on BP neural network. According to the relationship among liquid water content LWC, mean effective droplet diameter MVD, and ambient air temperature T listed in Appendix C of "China Airworthiness Standards for Transport Category Aircraft [CCAR-25-R4]", the icing conditions under different flight conditions are calculated, and an icing thickness database with the change of ice thickness over time is established; the BP neural network model is trained using the icing thickness database, and the prediction of the icing meteorological parameter MVD is carried out with the mapping relationship of flight conditions (flight angle of attack, flight speed), temperature, icing thickness, and icing time provided in real time as the input. Patent Application CN114880947A discloses a prediction method for aircraft icing meteorological parameters MVD and LWC, including obtaining the icing thickness and icing rate at the measurement point position according to the ice shape calculated by icing, and establishing a database of effective icing thickness and icing rate; using flight speed, ambient temperature, wing angle of attack, effective icing thickness, and icing rate as input parameters, and training an Elman neural network with optimized initial weights and thresholds using a genetic algorithm to predict and output the meteorological parameters mean effective droplet diameter MVD and liquid water content LWC.

[0005] However, the above prediction methods still have defects in prediction reliability and practicability. Summary of the Invention

[0006] The object of the present invention is to provide a prediction method for cloud physical parameters, which partially solves or alleviates the above deficiencies in the prior art and can improve the prediction accuracy of LWC and MVD in typical application scenarios such as stratiform clouds. To solve the above-mentioned technical problems, the present invention specifically adopts the following technical solutions: A prediction method for cloud physical parameters, including the steps: Training a model, which includes the steps:

[0007] S100. Obtain a first meteorological sample set, where the first meteorological sample set includes: meteorological sample groups at multiple moments, and a meteorological sample group includes: temperature, air pressure, relative humidity, cloud water mass mixing ratio, rain water mass mixing ratio, vertical velocity, ice water mass mixing ratio, ice water content, raindrop number concentration, ice number concentration, and a first liquid water content; S101. Input the first meteorological sample set into a convolutional neural network for iterative calculation. At this time, the input end of the convolutional neural network is: temperature, air pressure, relative humidity, cloud water mass mixing ratio, rain water mass mixing ratio, vertical velocity, ice water mass mixing ratio, ice water content, raindrop number concentration, ice number concentration, and a first liquid water content; the output end of the convolutional neural network is: the probability value that the median volume diameter belongs to large droplets; wherein, the activation function uses the normalized exponential function, and the normalized exponential function is used to calculate the probability value under the corresponding meteorological sample group, and the loss function uses the binary cross-entropy function, and the binary cross-entropy function is used to calculate the convergence state of the current probability value; S102. Combine the corresponding meteorological sample groups whose probability values belong to a set probability interval and the temperature is greater than a set temperature threshold to form a second meteorological sample set; S103. Input the second meteorological sample set into the convolutional neural network again for updated iterative calculation, and finally obtain a corresponding first prediction model. The output of the first prediction model is: a second liquid water content and a median volume diameter; during this process, the activation function uses the rectified linear unit function, and the rectified linear unit function is used to calculate the second liquid water content and the median volume diameter, and the loss function uses the mean square error function, and the mean square error function is used to calculate the convergence state of the second liquid water content and the median volume diameter.

[0008] In some embodiments, before S101, it further includes the steps of: obtaining a grid model of the area to be predicted, where the grid model has multiple sample grid points; performing vertical linear interpolation on the first meteorological sample set to match corresponding meteorological sample values for the multiple sample grid points, and the multiple meteorological sample values form a meteorological sample group corresponding to one sample grid point. In some embodiments, the first meteorological sample set is derived from aircraft detection data and the WRF model.

[0009] In some embodiments, before S102, the method further includes the steps of: setting a probability interval; setting the probability interval includes the steps of: selecting a plurality of discrimination thresholds between [a, b], where the discrimination thresholds are used to determine large droplets and small droplets, and a and b are the end values of the set probability interval; setting a first verification index, the first verification index includes: a first ratio R1 of the number of samples correctly classified as large droplets to the number of actual large droplet samples, and a second ratio R2 of the number of samples classified as large droplets that are actually small droplets to the number of actual small droplet samples; wherein, when the probability value is greater than or equal to the discrimination threshold, it is considered a large droplet, and when the probability value is less than the discrimination threshold, it is considered a small droplet; calculating a plurality of receiver operating characteristic curves corresponding to the plurality of discrimination thresholds according to the first ratio R1 and the second ratio R2; calculating the area size of the receiver operating characteristic curve; and selecting at least one discrimination threshold corresponding to the largest area size as the discrimination threshold during model training.

[0010] In some embodiments, before S102, the method further includes: calculating a true skill statistic value for each discrimination threshold according to the first ratio R1 and the second ratio R2; when the true skill statistic value is greater than or equal to a set minimum statistic value, the corresponding discrimination threshold is considered to meet the requirements of model training.

[0011] In some embodiments, the method further includes: model application discrimination, which includes the steps of: S200, obtaining the vertical movement and relative humidity in the flight area of the aircraft; S201, using a first determination mechanism to determine whether the first prediction model is applicable to the flight area; the first determination mechanism requires that the vertical movement is lower than a preset vertical movement threshold and the relative humidity is higher than a preset relative humidity threshold; if so, it is recommended to use the first prediction model.

[0012] In some embodiments, before S102, the method further includes: S203, obtaining the altitude in the flight area; S204, using a second determination mechanism to determine whether the first prediction model is applicable to the flight area, and if so, it is recommended to use the first prediction model; the second determination mechanism requires that the altitude is less than a set altitude threshold.

[0013] The present invention also provides a prediction system for cloud physical parameters and a model training subsystem, which includes: a sample acquisition module for acquiring a first meteorological sample set, the first meteorological sample set including: meteorological sample groups at multiple moments, and the meteorological sample group including: temperature, air pressure, relative humidity, cloud water mass mixing ratio, rain water mass mixing ratio, vertical velocity, ice water mass mixing ratio, ice water content, raindrop number concentration, ice number concentration, and a first liquid water content; a first training module for inputting the first meteorological sample set into a convolutional neural network for iterative calculation. At this time, the input end of the convolutional neural network is: temperature, air pressure, relative humidity, cloud water mass mixing ratio, rain water mass mixing ratio, vertical velocity, ice water mass mixing ratio, ice water content, raindrop number concentration, ice number concentration, and a first liquid water content; the output end of the convolutional neural network is: the probability value that the median volume diameter belongs to large droplets; wherein, the activation function adopts the normalized exponential function, and the normalized exponential function is used to calculate the probability value of the corresponding meteorological sample group, and the loss function adopts the binary cross-entropy function, and the binary cross-entropy function is used to calculate the convergence state of the current probability value; a sample screening module for combining the corresponding meteorological sample groups whose probability values belong to a set probability interval and the temperature is greater than a set temperature threshold to form a second meteorological sample set; a second training module for inputting the second meteorological sample set into the convolutional neural network again for updated iterative calculation, and finally obtaining a corresponding first prediction model, and the output of the first prediction model is: a second liquid water content and a median volume diameter; at this time, the activation function adopts the rectified linear unit function, and the rectified linear unit function is used to calculate the second liquid water content and the median volume diameter, and the loss function adopts the mean square error function, and the mean square error function is used to calculate the convergence state of the second liquid water content and the median volume diameter.

[0014] The present invention also provides an electronic device, including: a memory for storing a computer program; a processor for implementing the steps of the prediction method for cloud physical parameters according to any one of the embodiments when executing the computer program. The present invention also provides a computer-readable storage medium, in which a computer program is stored, and the computer program, when executed by a processor, implements the steps of the prediction method for cloud physical parameters according to any one of the embodiments. The present invention also provides a computer program product, including a computer program / instructions, and the computer program / instructions, when executed by a processor, implement the steps of the prediction method for cloud physical parameters according to any one of the embodiments.

[0015] Beneficial technical effects: It should be noted that traditional LWC or MVD prediction algorithms are usually extremely dependent on aircraft flight data, such as flight speed, wing angle of attack, icing time, ice thickness and other measured requirements (or non-meteorological data), and the applicant has found that this will make the application of icing prediction algorithms very limited. On the contrary, the present invention proposes a model training and application method for predicting LWC and MVD based on meteorological parameters. This scheme of directly predicting based on pure meteorological parameters can expand the application scenarios of prediction algorithms. For example, before an aircraft performs a flight mission, such as in the design of the aircraft's related de-icing system or the flight mission planning stage, the prediction of the degree of icing is also crucial. The prediction algorithm proposed in the present invention can use the meteorological parameters obtained by meteorological forecasts to predict the degree of icing in the future long-term process (such as the next day, the next month, etc.), reducing dependence on other non-meteorological parameters.

[0016] Moreover, the first type of prediction method (or the first prediction model) proposed in the present invention has significant prediction advantages in typical flight areas such as layered clouds. Layered clouds refer to uniform (referring to uniform thickness, grayscale and light transmittance) clouds covering the entire sky or part of the sky. For example, the macroscopic structure of large-scale precipitation layered clouds associated with low-value weather systems is layered, sometimes in two layers and sometimes in three layers. There is a cloudless area between the two layers of clouds. The high-level clouds produce ice crystal particles and fall to the bottom-level clouds, which are catalytic clouds; the ice crystal particles entering the low-level clouds continue to grow, and the moisture and environment required for the growth of ice crystals in the bottom-level clouds are supply clouds.

[0017] During the flight, flying through layered clouds is one of the most typical flight scenes. However, due to the complexity of the cloud structure and meteorological characteristics inside the layered clouds, the difficulty of predicting the relevant parameters of the layered cloud area has also increased sharply. In this regard, the secondary training method based on the probability interval classification of large and small droplets proposed in the present invention can effectively improve the prediction model's recognition sensitivity for the droplet size inside the cloud layer, thereby significantly improving the model's prediction accuracy for LWC and MVD.

[0018] Furthermore, experiments have shown that the prediction model proposed in the present invention has shown extremely high reliability in predicting the parameters of small droplets. This also enables the present invention to significantly reduce the false negative rate when the freezing potential is relatively low (such as when there are more small droplets). In other words, the prediction method proposed in the present invention can alleviate the problem of low prediction accuracy for small droplets in the prior art.

[0019] In addition, the secondary training method based on the classification of large and small droplet probability intervals adopted by the present invention can also reduce the dependence on the number of training samples, and it also shows extremely high reliability in the case of small sample training. Further, since aircraft icing prediction is crucial for flight safety, the present invention also provides two training methods for users to choose from. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts do not necessarily draw according to the actual scale. Obviously, the following described drawings are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained according to these drawings.

[0021] Figure 1 Schematic diagram of the method flow in an exemplary embodiment of the present invention;

[0022] Figure 2 Schematic diagram of the WRF model domain configuration in an exemplary embodiment;

[0023] Figure 3 Schematic diagram of the NAC0012 airfoil and the change curve of the ice shape at the leading edge of the wing with the exposure time; among them, Figure (a) is a schematic diagram of the NAC0012 airfoil, and Figure (b) is the change curve of the ice shape at the leading edge of the wing with the exposure time under a given state;

[0024] Figure 4 For the large droplet classification prediction based on the CNN-Attention 0 ROC curve graph;

[0025] Figure 5 For the large droplet classification prediction based on the CNN-Attention 0 TSS curve graph;

[0026] Figure 6 For the CNN-Attention 0 Model prediction result graph of microphysical meteorological parameters; among them, Figure (a) shows the relationship between the model prediction accuracy and vertical motion and relative humidity, Figure (b) shows the relationship between the model prediction accuracy and pressure and liquid water content, Figure (c) shows the relationship between the model prediction accuracy and ice water content and temperature, and Figure (d) shows the relationship between the model prediction accuracy and raindrop number concentration and ice number concentration;

[0027] Figure 7 For the based on CNN-Attention1 Model prediction - actual observation density scatter plot of LWC;

[0028] Figure 8 Based on CNN - Attention 0 Model prediction - actual observation density scatter plot of LWC;

[0029] Figure 9 Based on CNN - Attention 1 Model prediction - actual observation density scatter plot of MVD;

[0030] Figure 10 Based on CNN - Attention 0 Model prediction - actual observation density scatter plot of MVD. Detailed implementation manners

[0031] To make the objectives, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention. In this document, suffixes such as "module", "component" or "unit" used to represent elements are only for the convenience of describing the present invention and have no specific meaning in themselves. Therefore, "module", "component" or "unit" can be used interchangeably. In addition, the terms "first" and "second" are only used for descriptive purposes and cannot be construed as indicating or implying relative importance. In this document, unless otherwise clearly specified and defined, terms such as "connection" shall be understood in a broad sense. For example, "connection" can be a direct connection or an indirect connection through an intermediate medium, and can be the internal connection of two elements. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to specific situations. In this document, "and / or" includes any and all combinations of one or more of the listed related items. In this document, "a plurality" means two or more, that is, it includes two, three, four, five, etc. In this specification, the term "about" typically represents + / - 5% of the stated value, more typically + / - 4% of the stated value, more typically + / - 3% of the stated value, more typically + / - 2% of the stated value, even more typically + / - 1% of the stated value, and even more typically + / - 0.5% of the stated value. In this specification, some embodiments may be disclosed in a format within a certain range. It should be understood that this description of "within a certain range" is only for convenience and brevity and should not be construed as a rigid limitation on the disclosed range. Therefore, the description of the range should be considered to have specifically disclosed all possible sub-ranges and the individual numerical values within this range. For example, the description of the range 1-6 should be considered to have specifically disclosed sub-ranges such as from 1 to 3, from 1 to 4, from 1 to 5, from 2 to 4, from 2 to 6, from 3 to 6, etc., and the individual numbers within this range, such as 1, 2, 3, 4, 5, and 6. The above rules apply regardless of the breadth of the range.

[0032] Embodiment 1: Refer to Figure 1 As shown, the present invention provides a method for predicting cloud physical parameters, including the steps of: training a model (equivalent to providing a first training method), which includes the steps of:

[0033] S100, obtaining a first meteorological sample set, the first meteorological sample set including: meteorological sample groups at multiple moments, the meteorological sample groups including: temperature ( T ), air pressure ( P), relative humidity ( RH ), cloud water mass mixing ratio ( Q cloud ), rain water mass mixing ratio ( Q rain ), vertical velocity ( W ), ice water mass mixing ratio ( Q ice ), ice water content ( IWC ), raindrop number concentration ( N rain ), ice number concentration ( N ice ), the first liquid water content ( LWC );

[0034] S101, input the first meteorological sample set into a convolutional neural network for iterative calculation. At this time, the input end of the convolutional neural network is: temperature, air pressure, relative humidity, cloud water mass mixing ratio, rain water mass mixing ratio, vertical velocity, ice water mass mixing ratio, ice water content, raindrop number concentration, ice number concentration, and the first liquid water content; the output end of the convolutional neural network includes: the probability value that the median volume diameter (MVD) belongs to large droplets;

[0035] Preferably, the activation function of the convolutional neural network adopts the normalized exponential function, which is used to calculate the probability value corresponding to the meteorological sample group, and the loss function of the convolutional neural network adopts the binary cross-entropy function, which is used to calculate the convergence state of the current probability value;

[0036] That is to say, the process of S101 is equivalent to a classification step. Specifically, in order to improve the accuracy of large droplet classification, the present invention adopts the normalized exponential function to convert the unnormalized MVD value into a probability value to improve the iterative convergence speed of the model and the accuracy of model training; and adopts the binary cross-entropy function to calculate the convergence situation during the iteration process. Subsequently, the classification of different sample data is completed by means of the probability value; or rather, the rapid classification of samples is completed through the probability value.

[0037] Preferably, in the classification training process of S101 in this embodiment, an initial model can also be obtained for predicting the actual liquid water content (equivalent to the second liquid water content) and the median volume diameter according to the temperature, air pressure, relative humidity, cloud water mass mixing ratio, rain water mass mixing ratio, vertical velocity, ice water mass mixing ratio, ice water content, raindrop number concentration, ice number concentration, and predicted liquid water content (equivalent to the first liquid water content) output by the WRF model; specifically, in this process, the output end of the convolutional neural network includes: the second liquid water content and the median volume diameter.

[0038] Specifically, the probability value in this embodiment can be used to represent the probability that a droplet belongs to a large droplet.

[0039] In some embodiments, the input layer in the convolutional neural network inputs a meteorological sample group and an MVD value (such as the average MVD (μm)); the activation function can calculate the probability values corresponding to different meteorological sample groups through the association between the meteorological sample group and the MVD value (such as the average MVD (μm)).

[0040] In some embodiments, the loss function is used to calculate the difference between the probability value in the current iteration process and the probability value in the previous iteration process. When the difference is small, it is considered that the two converge, and the iteration ends; otherwise, the iteration continues.

[0041] S102, combine the corresponding meteorological sample combinations for which the probability value belongs to a set probability interval and the temperature is greater than a set temperature threshold to form a second meteorological sample set; preferably, when the probability value belongs to the set probability interval and the temperature is greater than the set temperature threshold, it is considered that the corresponding sample droplet has a high probability of belonging to a small droplet. That is to say, in this embodiment, the second meteorological sample set is a sample set with a high probability of belonging to a small droplet.

[0042] S103, input the second meteorological sample set into the convolutional neural network again for updated iterative calculation, and finally obtain the corresponding first prediction model (also referred to as CNN - Attention in this article) 0 ), and the output of the first prediction model is: the second liquid water content and the median volume diameter;

[0043] In this embodiment, the first liquid water content can be a sample value output according to the WRF model; the second liquid water content refers to a predicted value output according to a prediction model (such as the first prediction model). Preferably, the activation function adopts a rectified linear unit function, and the rectified linear unit function is used to calculate the second liquid water content and the median volume diameter. The loss function adopts a mean square error function, and the mean square error function is used to calculate the convergence state of the second liquid water content and the median volume diameter. That is to say, this process can update the initial model obtained in S101.

[0044] It should be noted that a model training method based on sample interval classification for secondary training is proposed in this embodiment. Specifically, the present invention roughly screens out sample data belonging to small droplets through the probability value of the droplets, and inputs the screened small droplet sample data into the convolutional neural network again for enhanced training, so as to enhance the recognition ability of the first prediction model for droplets of different particle sizes and improve the prediction accuracy of LWC and MVD.

[0045] For example, in this embodiment, first, the T , RH , P , LWC , Q cloud , Q rain , IWC , W , Q ice , N ice and N rain and other parameters are transmitted to the convolutional layer through the input layer. The convolutional layer includes 4 one-dimensional convolutions. The first convolution reads the input sequence and projects the result onto the feature map. The second convolution performs the same operation on the feature map created by the first layer, and performs max pooling after each convolution to amplify its significant features. In the first training stage, for the large droplet classification algorithm, the activation function is the normalized exponential function, i.e., the Softmax function, which is used to calculate the probability value that the corresponding droplet belongs to a large droplet, and the loss function is the binary cross-entropy function. Subsequently, it enters the max pooling layer, and the maximum value in the sample is used as the sampled sample value. The max pooling layer is used to simplify the feature map, and then the feature map is flattened into a long vector for decoding. Finally, it enters the fully connected layer, and the spatio-temporally matched LWC and MVD values are used as the model output. Droplets with an MVD less than 100 μm are identified as small droplets, and the sample data identified as small droplets is input into the convolutional neural network again. At this time, the activation function uses the RELU function, and the loss function uses the mean squared error function.

[0046] The loss function is calculated after each iteration cycle of the training set and the test set. After 500 iteration training cycles, the model is terminated, and the corresponding first prediction model is obtained.

[0047] In some embodiments, before S101, it further includes the steps of: obtaining a grid model of the area to be predicted, the grid model having a plurality of sample grid points; performing vertical linear interpolation on the first meteorological sample set to match corresponding meteorological sample values for the plurality of sample grid points, and a plurality of meteorological sample values form a meteorological sample group corresponding to one sample grid point. In some embodiments, the first meteorological sample set is derived from aircraft detection data and the WRF model.

[0048] In some embodiments, before S102, the method further includes the steps of: setting a probability interval; the step of setting the probability interval includes: selecting a plurality of discrimination thresholds between [a, b], where the discrimination thresholds are used to determine large droplets and small droplets, and a and b are set probability end values (set by the user); setting a first verification index, where the first verification index includes: a first ratio R1 of the number of samples correctly classified as large droplets to the number of actual large-droplet samples, and a second ratio R2 of the number of samples classified as large droplets that are actually small droplets to the number of actual small-droplet samples; where when the probability value is greater than or equal to the discrimination threshold, it is considered a large droplet, and when the probability value is less than the discrimination threshold, it is considered a small droplet; calculating a plurality of Receiver Operating Characteristic (ROC) curves corresponding to the plurality of discrimination thresholds according to the first ratio R1 and the second ratio R2; calculating the size of the area under the Receiver Operating Characteristic curve (Area Under Curve, AUC); selecting at least one discrimination threshold corresponding to the largest area size as the discrimination threshold c during model training, and generating a corresponding probability interval based on the discrimination threshold, such as the probability interval can be [a, c]. The Receiver Operating Characteristic curve refers to a curve plotted with the false alarm probability (second ratio R2) as the abscissa and the hit probability (first ratio R1) as the ordinate, and the area under the curve can characterize the discrimination ability of the model. For example, when the AUC value is 0.5, it means there is no discrimination ability; the closer the AUC value is to 1, the better the prediction performance.

[0049] After the experimental verification below, the present invention, in cooperation with the above model training and probability setting scheme, can make the AUC value of the entire sample set reach 0.979, indicating that the model training method has high reliability.

[0050] In some embodiments, the first verification index may further include: a third ratio R3 of the number of samples correctly classified as small droplets to the number of actual small-droplet samples; a fourth ratio R4 of the number of samples classified as small droplets that are actually large droplets to the number of actual large-droplet samples; accuracy, that is, the ratio of the number of correctly classified droplet samples (including large droplets and small droplets) to the total number of samples.

[0051] Further, in some embodiments, in order to improve the accuracy of model training, the method may further include the steps of: calculating the true skill statistic (TSS) under each discrimination threshold according to the first ratio R1 and the second ratio R2; when the true skill statistic is greater than or equal to the set minimum statistic value, the corresponding discrimination threshold is considered to meet the requirements of model training.

[0052] For example, in this embodiment, the minimum statistical value is set at approximately 0.85. Refer to Figure 4 and Figure 5 As shown, when the TSS reaches 0.87, the first ratio is 0.98 and the second ratio is 0.11; when the TSS is 0.94, the training set exhibits better performance, the first ratio is 0.99 and the second ratio is 0.05. The reliability of the training method adopted by the present invention can be further verified through the test results.

[0053] In some embodiments, the method further includes: model application discrimination, which includes the steps of:

[0054] S200, obtaining the vertical speed and relative humidity in the flight area of the aircraft; S201, using a first determination mechanism to determine whether the first prediction model is applicable to the flight area; the first determination mechanism requires that the vertical speed is lower than a preset vertical motion threshold and the relative humidity is higher than a preset relative humidity threshold; for example, in some embodiments, the relative humidity threshold can be approximately 70%; if so, it is recommended to use the first prediction model.

[0055] It should be noted that the applicant has found that the first prediction model obtained by secondary training based on droplet classification is particularly applicable to flight areas such as stratiform clouds. Therefore, in order to improve the reliability of model prediction in the actual application process, the present invention also proposes a corresponding judgment mechanism.

[0056] In some embodiments, before S102, it further includes: S203, obtaining the altitude in the flight area; S204, using a second determination mechanism to determine whether the first prediction model is applicable to the flight area; if so, it is recommended to use the first prediction model; the second determination mechanism requires that the altitude is lower than a set altitude threshold. For example, in some embodiments, when the altitude is lower than approximately 800 hPa, it is recommended to use the first prediction model.

[0057] In some embodiments, the present invention also provides a second prediction model directly trained based on a regression algorithm; the corresponding model training steps (i.e., the second training method) include:

[0058] Obtaining a first meteorological sample set, the first meteorological sample set including: meteorological sample groups at multiple moments, the meteorological sample groups including: temperature ( T ), air pressure ( P ), relative humidity ( RH ), cloud water mass mixing ratio ( Q cloud ), rain water mass mixing ratio ( Q rain ), vertical speed ( W ), ice water mass mixing ratio (Q ice ), ice - water content ( IWC ), raindrop number concentration ( N rain ), ice number concentration ( N ice ), liquid water content ( LWC );

[0059] Input the first meteorological sample set into a convolutional neural network for iterative calculation. At this time, the input end of the convolutional neural network is: temperature, air pressure, relative humidity, cloud - water mass mixing ratio, rain - water mass mixing ratio, vertical velocity, ice - water mass mixing ratio, ice - water content, raindrop number concentration, ice number concentration, and the first liquid water content; the output end of the convolutional neural network includes: the second liquid water content and the median volume diameter; preferably, the activation function uses the rectified linear unit function, and the rectified linear unit function is used to calculate the second liquid water content and the median volume diameter. The loss function uses the mean - square error function, and the mean - square error function is used to calculate the convergence state of the second liquid water content and the median volume diameter.

[0060] For example, in some embodiments, the T , RH , P , Q cloud , Q rain , IWC , W , Q ice , N ice , N rain 、 LWC output by WRF is set as the input of the network, and the spatiotemporally - matched LWC and MVD values are used as the output of the network. It can be understood that the samples used in the model training process in this article can be the measurement results collected by a meteorological monitoring system (such as a meteorological satellite).

[0061] After the iterative process converges, obtain the corresponding second prediction model (also referred to as: CNN - Attention 1 model) in this article;

[0062] Different from the above - mentioned scheme, the model training in this embodiment directly performs regression training on the overall samples without introducing the step of distinguishing large and small droplet samples.

[0063] In some embodiments, the above - mentioned two model training methods can be used cross - adaptively.

[0064] For example, in some embodiments, different training schemes can be selected according to different training scenarios, such as the number of samples and the distribution of MVD values of the samples.

[0065] For example, in this embodiment, before model training, the steps further include: obtaining a meteorological sample set, which includes: T , RH , P , LWC , Q cloud , Q rain , IWC , W , Q ice , N ice and N rain , and MVD; calculating the difference between at least one maximum MVD value and at least one minimum MVD value in the meteorological sample set; when the difference is less than the set difference interval, it is recommended to use the first training method for sample training; otherwise, it can be recommended to use the second training method for sample training;

[0066] Again, for example, in some embodiments, before model training, the steps further include: obtaining a meteorological sample set; calculating the number of samples; when the number of samples is less than the set sample size, it is recommended to use the first training method for sample training; otherwise, it can be recommended to use the second training method for sample training.

[0067] It should be noted that the first training method provided by the present invention, which adopts the mode of secondary training based on droplet classification, has significant prediction advantages in scenarios where the number of samples is limited and the sample data distribution is relatively concentrated. Or in some embodiments, a small number of samples can be used simultaneously with two training methods to obtain a preliminary prediction model; the recommended training method is selected by comparing the prediction accuracies of the two prediction models. Or in some embodiments, two prediction models can be selected synchronously for prediction to cross-validate the prediction results.

[0068] Embodiment 2: The present invention also provides a method for verifying model training, including the steps:

[0069] S300, obtaining a first meteorological sample set and the corresponding median volume diameter, where the first meteorological sample set includes: meteorological sample groups at multiple moments, and the meteorological sample groups include: temperature, air pressure, relative humidity, cloud water mass mixing ratio, rain water mass mixing ratio, vertical velocity, ice water mass mixing ratio, ice water content, raindrop number concentration, ice number concentration, and liquid water content;

[0070] S301. Input the first meteorological sample set into a convolutional neural network for iterative calculation. At this time, the input end of the convolutional neural network is: temperature, air pressure, relative humidity, cloud water mass mixing ratio, rain water mass mixing ratio, vertical velocity, ice water mass mixing ratio, ice water content, raindrop number concentration, ice number concentration, and the first liquid water content. The activation function of the convolutional neural network is used to calculate the probability value that the median volume diameter of the sample group belongs to large droplets. The probability value represents the probability that the corresponding sample belongs to large droplets or small droplets. Correspondingly, the output end of the convolutional neural network includes: probability value.

[0071] S302. Combine the meteorological sample groups whose probability values of the median volume diameter belong to a set probability interval and whose temperature is greater than a set temperature threshold to form a second meteorological sample set.

[0072] S303. Input the second meteorological sample set into the convolutional neural network again for updated iterative calculation, and finally obtain the corresponding first prediction model. The output of the first prediction model includes: the second liquid water content and the median volume diameter.

[0073] S304. Calculate the first verification index. The first verification index includes: the first ratio R1 of the number of samples correctly classified as large droplets to the number of actual large droplet samples, and / or the second ratio R2 of the number of samples classified as large droplets that are actually small droplets to the number of actual small droplet samples.

[0074] S305. When it is judged whether the first verification index meets the set first verification condition, if not, adjust the probability interval and execute S302 again according to the new probability interval.

[0075] For application scenarios with a relatively small amount of trainable samples, the present invention provides a secondary training method based on the classification of large and small droplet probability intervals, and uses the classification accuracies of large droplets and small droplets (corresponding to R1 and R2) as key indicators to guide the adjustment direction of secondary training, and then improves the prediction accuracy of small droplets based on this key indicator. Finally, the accuracy of icing prediction is effectively improved in scenarios with a small amount of training samples or typical application scenarios such as stratiform clouds.

[0076] For example, in some embodiments, when both the first ratio and the second ratio are greater than the corresponding set values, it is considered that the first verification index meets the set first verification condition.

[0077] For another example, in some embodiments, the probability interval is set by a discrimination threshold. When the probability value is greater than or equal to the discrimination threshold, it is considered a large droplet. When the probability value is less than the discrimination threshold, correspondingly, S305 includes the steps of: calculating a plurality of receiver operating characteristic curves corresponding to the discrimination threshold according to the first ratio R1 and the second ratio R2; calculating the area size of the receiver operating characteristic curves; when the area size is greater than a set area threshold, it is considered that the first verification index determines whether it meets the set verification conditions. In some embodiments, it further includes the step of: when the area size is less than or equal to the area threshold, adjusting the discrimination threshold and re-executing S302 and S303 according to the newly formed probability interval.

[0078] In some embodiments, the convolutional neural network includes at least four convolutional layers.

[0079] Embodiment 3: The present invention also provides a method for applying a cloud physical parameter prediction model, including the steps of:

[0080] S401, obtaining a first meteorological sample set containing a combination of multiple meteorological elements and the corresponding median volume diameter; S402, selecting a first test sample set from the first meteorological sample set, and respectively training through the first test sample set using a first training method and a second training method to obtain a first prediction model and a second prediction model;

[0081] Among them, the step of training the first prediction model using the first training method includes:

[0082] Inputting the first test sample set into a convolutional neural network, at this time, the output of the convolutional neural network includes: the probability value of the median volume diameter, and the corresponding probability value represents the probability that the corresponding sample belongs to a large droplet or a small droplet;

[0083] Inputting the samples predicted as small droplets into the convolutional neural network again according to the probability value to update the convolutional neural network, and outputting the first prediction model after the training converges; at this time, the output of the convolutional neural network includes: liquid water content and median volume diameter;

[0084] The step of training the second prediction model using the second training method includes: inputting the first test sample set into a convolutional neural network, the output of the convolutional neural network is the second liquid water content and the median volume diameter, and outputting the second prediction model after the training converges;

[0085] S403. Select a second test sample set from the second meteorological sample set, where the second test sample set does not completely overlap with the first test sample set; in some embodiments, the second meteorological sample set may be a part of the first meteorological sample set. Alternatively, the second meteorological sample set may also be the meteorological prediction results obtained for a future period for the current icing prediction task (or flight task).

[0086] S404. Calculate the prediction accuracies of the first prediction model and the second prediction model respectively using the second test sample set;

[0087] S405. Recommend and select the corresponding prediction model according to the prediction accuracy.

[0088] It should be noted that due to the very complex aircraft icing mechanism, and there may be great differences in aircraft icing under different flight scenarios, climate conditions and aircraft types. Moreover, targeted modeling for icing prediction in different scenarios will incur extremely high creation costs.

[0089] In this regard, the applicant chose to start from the perspective of droplet size prediction accuracy and provided two model training methods respectively. One is to improve the droplet classification accuracy (and improve the effective learning depth for small droplets) in special scenarios (such as when the sample data volume is relatively limited, or when flying into the cumulus cloud icing meteorological area) through the secondary classification training method (i.e., the first training method); the other is to directly use neural network for regression training (corresponding to the second training method) to quickly model for another type of flight scenario (such as when the sample data volume is relatively large, or the stratus cloud meteorological environment).

[0090] Thus, when predicting icing for a new flight task, the user can use a small number of samples to complete preliminary test training through the first and second training methods in advance, and then select the recommended training method through the preliminary prediction results of the two prediction models to complete the final model training using the recommended training method.

[0091] For another example, in some embodiments, the two prediction models can also be used for synchronous prediction at the same time for the user to refer to.

[0092] In some embodiments, S403 includes the steps of: selecting from the second meteorological sample set the combination of meteorological elements belonging to the set vertical motion interval and the set relative humidity interval, and forming the second test sample set according to the combination of meteorological elements.

[0093] Furthermore, the applicant noticed that the first prediction model may show extremely high sensitivity in areas with weak vertical motion and high relative humidity. Therefore, during model verification, it is preferably to observe its prediction accuracy in specific vertical motion and relative humidity intervals.

[0094] In some embodiments, S403 includes the steps of: selecting, from the second meteorological sample set, the combination of meteorological elements belonging to a set altitude range, and forming the second test sample set according to the combination of meteorological elements.

[0095] Furthermore, the applicant notes that the first prediction model may exhibit extremely high sensitivity in low altitude regions. Therefore, when validating the model, it is preferably to observe its prediction accuracy in a specific altitude range.

[0096] In some embodiments, the meteorological sample group includes: temperature, air pressure, relative humidity, cloud water mass mixing ratio, rain water mass mixing ratio, vertical velocity, ice water mass mixing ratio, ice water content, raindrop number concentration, ice number concentration, and liquid water content.

[0097] In order to illustrate the technical solutions adopted in the present invention and the technical effects, the following will be explained through a specific model training and application example:

[0098] I. Data sources for model training and testing: This embodiment uses the flight data published by the ICICLE project, which focuses on enhancing the understanding of aircraft icing by studying atmospheric microphysical processes. During the implementation of this project, the National Center for Atmospheric Research (NCAR) conducted multiple flight tests in the Great Lakes region of North America (between 37.588 - 45.666° north latitude) from January 28 to March 8, 2019. By using the detection instruments equipped on the aircraft, cloud physics parameter data were collected to analyze the thermodynamics, cloud dynamics, and specific cloud and aerosol characteristics of cold clouds and precipitation.

[0099] This embodiment also uses the data of the airborne instruments on the Convair 580 aircraft of the National Research Council of Canada (NRC) participating in this project. Among them, the flight altitude, longitude, and latitude data are provided by the KVH 1750 IMU sensor installed in the cabin; the microphysical data of liquid and ice particles are obtained by the Forward Scattering Spectrometer Probe (FSSP), Two - Dimensional Stereo (2DS) probe, and High - Volume Precipitation Spectrometer (HVPS). The particle sizes measured by the three instruments are 3 - 35μm, 40 - 670μm, and 750 - 38400μm respectively. Taking 100μm as the threshold, the droplets are divided into small droplets (<100μm) and large droplets (>100μm) for further analysis.

[0100] A total of 30 flight detections were carried out on this aircraft. Since the first flight was a test flight and no detections were made during the second and thirtieth flights, the flight detection data from the third to the twenty-ninth flights were mainly used. It should be noted that the tenth flight was a transfer flight, and the probe instrument was activated for a very short time, and no relevant cloud physical parameters were detected. The average root mean square error of the detection data was 30 s, which roughly corresponded to a path length of 3 km. The LWC data points, large droplet data points, average LWC, average MVD, and large droplet percentage within 30 s for each flight were listed in Tables 1.1 - 1.2.

[0101] Table 1.1 Aircraft Statistical Data Table

[0102]

[0103] Table 1.2 Aircraft Statistical Data Table

[0104]

[0105] It can be understood that in some embodiments, the sample values of spatiotemporally matched LWC and MVD can be obtained through aircraft detection.

[0106] II. Data Processing Method - WRF-Based Meteorological Parameter Forecasting Method

[0107] The Weather Research and Forecasting (WRF) model is a fully compressible, non-hydrostatic numerical model. In this embodiment, the initial and boundary conditions of the fifth-generation atmospheric reanalysis dataset (ERA5) of the European Centre for Medium-Range Weather Forecasts (ECMWF) are used, with a horizontal grid spacing of 0.25°, a time resolution of 6 h. A two-way nesting strategy is adopted to create two nested domains, with a horizontal resolution of 27 km and 9 km for each domain, a vertical resolution of 44 altitude levels, and a time resolution of 1 h, as Figure 2 shown. The black line is the aircraft flight path. Each time, the flight time simulated by the WRF model is at least 6 h to ensure reliable simulation accuracy.

[0108] The WRF parameterization scheme selected in this embodiment is as follows: The Thompson microphysics scheme is used, which can improve the explicit prediction of aircraft icing and plays an important role in icing prediction; the Dudhia shortwave radiation scheme is adopted, the rapid radiative transfer model is used for the longwave radiation scheme, and the Noah land surface model is used for the Eta surface layer. These configurations are determined based on the research results under winter conditions; the Kain-Fritsch cumulus scheme is used for the cumulus scheme.

[0109] In this embodiment, the following meteorological variables are selected to describe the aircraft icing environment, including altitude ( z ), temperature ( T ), pressure (P ), relative humidity ( RH ), cloud water mass mixing ratio ( Q cloud ), rain water mass mixing ratio ( Q rain ), vertical velocity ( W ), ice water mass mixing ratio ( Q ice ), ice water content ( IWC ), raindrop number concentration ( N rain ), ice number concentration ( N ice ), liquid water content ( LWC ), as shown in Table 2. Among them, temperature, altitude, air pressure, relative humidity, cloud water mass mixing ratio, rain water mass mixing ratio, vertical velocity, and ice water mass mixing ratio are provided by WRF output parameters. Ice water content, raindrop number concentration, ice number concentration, and liquid water content are obtained by multiplying the output parameters by air density.

[0110] Table 2 List of Variables and Definitions

[0111]

[0112] III. Training Tools

[0113] In this embodiment, a feed-forward neural network model based on convolution (also known as a convolutional neural network model, or CNN network model) is preferably adopted. The CNN model consists of a convolutional layer, a pooling layer, and a fully connected layer. Among them, the convolutional layer is composed of multiple convolutional units, and the parameters of each convolutional unit are optimized through the backpropagation algorithm. In this embodiment, the goal of the convolutional operation is to extract various features of the input. The input data obtains a feature map through the convolutional layer, and the feature mapping represents the features in the neural space through the convolution of the neural network on the input data. The pooling layer is used to sample the feature map, mainly by reducing the number of network parameters to reduce the computational complexity, so as to reduce the network scale and obtain the invariant features of the input data. The fully connected layer mainly combines all local features, multiplies the local features by their corresponding weights, and then sums through the convolutional operation to form global features for calculating the final scores of each category.

[0114] In this embodiment, the adoption of the CNN model can reduce the complexity of meteorological sample data and extract data features; by introducing the attention mechanism in the CNN model, the data dimension of meteorological sample data can be reduced and the training efficiency can be improved.

[0115] Among them, the calculation formula for extracting data features using the CNN-attention mechanism model is as follows:

[0116] (1);

[0117] (2);

[0118] Among them x is the input data, y is the output data, σ is the activation function, W k and b k are the weight coefficient and the bias function respectively, k is the number of convolution kernels (representing discrete convolution operation), α and β are the size parameters of the convolution kernels, W m,n represents the weight coefficient of the weight matrix convolution kernel, m and n represent the row index and column index of the eigenvalue in the convolution kernel respectively, i, j represents the row index and column index of the input data.

[0119] IV. Sampling Strategy for Model Observation Comparison

[0120] In terms of time, in order to ensure the complete matching of the model output result with the flight detection data, in this embodiment, the aircraft detection results 15 minutes before and after the model output time are matched with the model output. In each 30-minute sampling interval, the output value of a model grid point is selected as the matching value for average observation. In terms of space, in this embodiment, the method of vertical linear interpolation is adopted to model the aircraft sampling height output to achieve the optimal matching of the flight height. The aircraft detection data and WRF prediction data after space-time matching are divided into two parts: a training set (70%) and a test set (30%). Among them, the training set is used to train the CNN-attention model, and the test set is used to evaluate the accuracy of the model.

[0121] Specifically, the model evaluation and statistical method adopted in this embodiment is:

[0122] In this embodiment, the first ratio (also known as: True Positive Rate (TPR)), the second ratio (also known as: False Positive Rate (FPR)), the third ratio (also known as: True Negative Rate (TNR)), the fourth ratio (also known as: False Negative Rate (FNR)), and the accuracy rate are used to evaluate the recognition ability of the CNN-Attention model for water droplets. Large droplets are considered positive and small droplets are considered negative. TPR is the ratio of the number of samples correctly classified as large droplets to the number of actual large-droplet samples. FPR is the ratio of the number of samples that classify small droplets as large droplets to the number of actual small-droplet samples. TNR is the ratio of the number of samples correctly classified as small droplets to the number of actual small-droplet samples. FNR is the ratio of the number of samples that classify large droplets as small droplets to the number of actual large-droplet samples. The accuracy is the ratio of the number of correctly classified droplet samples (including large and small droplets) to the total number of samples.

[0123] The Root Mean Square Error (RMSE) and Pearson Correlation Coefficient (Pearson Corr) are used to evaluate the prediction ability of the model for LWC and MVD in the case of small droplets.

[0124] The calculation formula for the Root Mean Square Error is:

[0125] (3);

[0126] The calculation formula for the Pearson Correlation Coefficient is:

[0127] (4);

[0128] Where RMSE is the Root Mean Square Error, MSE is the Mean Square Error, Corr(output, output) is the Pearson Correlation Coefficient, n is the number of samples, N is the sample size, output is the predicted value, and observe is the observed value.

[0129] V. Ice accretion numerical simulation: In this embodiment, the LEWICE software is used to perform numerical simulation of aircraft ice accretion. By inputting the wing model, aircraft state parameters, cloud microphysical parameters, and exposure time, the ice accretion mode of the aircraft and the ice accretion thickness of each region of the wing can be obtained, and the maximum ice accretion rate can be calculated.

[0130] By adjusting different parameters, a lookup table (LUT) of the maximum icing rate of the aircraft is established to determine the severity of aircraft icing. The selection of the parameter range is crucial for accurate aircraft icing prediction. In this embodiment, based on the previous results and referring to Appendix C of FAR Part 25, by statistically analyzing tens of thousands of icing events, the meteorological elements covering 95% of the icing events are extracted as the input parameter range of the LEWICE software, as shown in Table 3. Among them, the temperature is set to -28 - 4°C, the relative humidity is set to 70% - 100%, the flight speed is set to 60 - 120 m / s, the LWC is set to 0 - 1 g / m3, and the MVD is set to 5 - 100 μm.

[0131] The method for establishing the lookup table (LUT) of the maximum icing rate in this embodiment can refer to the aircraft icing degree prediction method based on icing numerical simulation disclosed in Patent Application CN117493738A.

[0132] Table 3 LEWICE Parameter Setting Table

[0133]

[0134] The wing model adopted in this embodiment is the NACA0012 airfoil, as Figure 3 shown. Figure 3 In (a) of [], it is a schematic diagram of the NAC0012 airfoil; in (b), it is the curve of the ice shape at the leading edge of the wing changing with the exposure time under a given state; in the lower right of (b), it is a schematic diagram of the ice shape change at the leading edge of the wing from 1 to 20 minutes under given conditions.

[0135] This airfoil is the standard numerical model of the Lewis wind tunnel and is widely used in icing simulation research. To examine the correlation between the time exposed to icing conditions and the resulting ice accumulation, a series of experiments were conducted under fixed environmental parameters (temperature -8°C, flight speed 80 m / s, relative humidity 100%, MVD 20 µm, LWC 0.8 g / cm 3 , exposure time 1 - 20 minutes, test time interval 1 minute). The test results show that the maximum icing thickness increases uniformly with the increase of the exposure time, as Figure 3 shown. This indicates that the growth rate of the icing thickness is not affected by the exposure time and can be used to evaluate the severity of aircraft icing.

[0136] VI. Model Verification

[0137] The size of droplets is classified using a CNN-attention model. Multiple thresholds are systematically established within the interval [0, 1], and each threshold is a discriminant criterion: if the predicted value exceeds the threshold, the diagnosis is classified as large droplets, while values below the threshold are considered small droplets. Subsequently, the true positive rate and false positive rate are calculated for these different thresholds to evaluate the diagnostic ability of the CNN-attention model. According to the receiver operating characteristic curve (ROC), the relationship between the true positive rate and false positive rate at different threshold levels is shown, and the area under the curve (AUC) is a quantitative indicator of the model's discrimination ability. When the AUC value is 0.5, it indicates no discrimination ability; the closer the AUC value is to 1, the better the prediction performance. Figure 4 For the large droplet classification ROC curve based on the CNN-Attention 0 model, the results show that the model has good training ability, and the AUC value of the entire sample set reaches 0.979.

[0138] To determine the most suitable threshold for large liquid classification, the true skill statistic (TSS) is also calculated at different thresholds in this embodiment. TSS integrates the true positive rate and false positive rate, providing an overall measure of the model's classification performance. Figure 4 It shows that as the threshold increases from 0 to 1.0, the TSS first rises and then gradually decreases, reaching the maximum value at a threshold of 0.12. Therefore, the decisive standard threshold for distinguishing large and small droplets in this embodiment is set to 0.12. Considering that the missed reporting of large droplets in the cloud may lead to serious icing consequences, while false reporting may lead to unnecessary anti-icing system activation measures, through comprehensive consideration, 0.85 is used as the minimum TSS value in this embodiment to achieve a balance between false reporting and missed reporting. When the TSS reaches 0.87, the true positive rate is 0.98 and the false positive rate is 0.11; when the TSS is 0.94, the training set shows better performance, with a true positive rate of 0.99 and a false positive rate of 0.05.

[0139] CNN-Attention 0 model and CNN-Attention 1 The true positive rate, false negative rate, and accuracy rate of the CNN-Attention 0 model for large and small droplet classification evaluation of all sample data are listed in Table 4. The results show that during the process of large and small droplet classification by the CNN-Attention

[0140] Table 4 Prediction accuracy table of the model

[0141]

[0142] Further utilize the CNN-Attention 0 model to predict microphysical meteorological parameters, including vertical velocity - relative humidity, liquid water content - air pressure, temperature - ice water content, raindrop number concentration - ice number concentration, as Figure 6 shown. The applicant also found that the CNN-Attention 0 model has significant advantages when predicting flight regions such as stratiform clouds. Especially in areas with weak vertical motion and high relative humidity (as Figure 6 shown in (a)). The applicant predicts that this is because the CNN-Attention 0 model can accurately identify droplets through droplet classification and secondary training and optimize the model based on the accurate identification results. The two enable it to have high accuracy when applied to scenarios such as stratiform clouds with complex meteorological characteristics such as reduced airflow and increased moisture. For example, when the RH level is greater than approximately 70%, the prediction reliability is significantly enhanced.

[0143] From Figure 6 it can also be seen from (b) that the CNN-Attention 0 also shows significant prediction advantages in areas such as low altitude and low liquid water content. For example, when the altitude is less than or equal to 800 hPa or the LWC is less than or equal to 0.3 g / m 3 ³, the applicability of this model is significantly enhanced. Figure 6 It is shown in (c) that the samples where the CNN-Attention 0 model correctly predicts large droplets are mainly found in environments with low ice water content. This low ice water content setting is conducive to the growth of large droplets because excessive ice water content will trigger freezing, thus hindering the formation and maintenance of large droplets. In addition, from Figure 6 it can also be seen from (d) that the CNN-Attention 0 model is particularly suitable for predicting regions with high N rain .

[0144] Prediction ability of the CNN-Attention model for small droplets MVD and LWC: Evaluate the performance of two CNN-Attention models in predicting LWC and MVD in the case of small droplets (MVD < 100 μm), with RMSE and Corr as evaluation indicators. The results are listed in Table 5. It can be seen that the two models show basically similar abilities in predicting LWC, and the performance of the CNN-Attention 0 model is slightly better than that of the CNN-Attention 1 model. In terms of predicting MVD, the CNN-Attention 0The model performs excellently, with the RMSE reduced to 25.5 and the correlation coefficient reaching 0.75, showing significant effects compared with the CNNAttention 1 model. This may be because the CNN-Attention 0 model can accurately distinguish large and small droplets, and this targeted focusing enables the model to converge more effectively, thus achieving high-precision prediction of MVD.

[0145] Table 5 Prediction evaluation of LWC and MVD by two CNN-Attention models

[0146]

[0147] To more precisely understand the proficiency of the two models in predicting LWC and MVD under small droplet conditions, density scatter plots of model predictions - actual observations are generated for the entire dataset. The tightness of the scatter points to the diagonal represents the prediction accuracy, as Figures 7 - 10 shown. The LWC scatter plots of the two models ( Figures 7 - 10 ) show a distinct concentration of points along the diagonal, indicating that both models demonstrate extraordinary capabilities in predicting the LWC of small droplets, and the prediction results are very close to the actual observations. However, there are differences in the performance of the two models in MVD prediction ( Figures 7 - 8 ). The scatter plot of the CNN-Attention 1 model has more dispersed point distributions, while the scatter plot of the CNN-Attention 0 model shows a tighter clustering around the diagonal, indicating a more accurate estimation of the MVD value. Therefore, the CNN-Attention 0 model has lower systematic bias and overall error in predicting cloud microphysical parameters, especially for small water droplets. Generally speaking, the performance evaluation on the full sample dataset shows that both of these CNN-Attention models perform excellently in distinguishing large and small droplets. In the current sample dataset, the CNN-Attention 0 model becomes the preferred option due to its higher true positive rate and accuracy.

[0148] VII. Forecast of aircraft icing severity

[0149] 7.1 Icing severity forecasting algorithm: It includes two components. The first part obtains the meteorological parameters (T, RH, LWC, and MVD) of the target area through meteorological parameter forecasting and distinguishes large and small droplets. In this embodiment, WRF and CNN-Attention 0A hybrid model combined with the model. The second part is the icing severity prediction. Using meteorological parameters and aircraft fuselage parameters as inputs, linear interpolation is performed on the maximum icing rate to achieve rapid evaluation and prediction of the maximum icing rate in the target area. The icing severity classification based on the icing rate is as follows: trace icing < 0.6 mm / min, light icing is 0.6 - 1.0 mm / min, moderate icing is 1.1 - 2.0 mm / min, and severe icing > 2.0 mm / min.

[0150] 7.2 Evaluation of the icing severity prediction algorithm for typical icing events:

[0151] 1) Single flight

[0152] The 17th flight of the ICICLE project provides a valuable case for evaluating the accuracy of the icing severity prediction algorithm. According to the flight crew report, the aircraft took off from Terre Haute Airport at 12:04 pm and flew south. There was no icing between 12:04 pm and 12:50 pm; at 13:13 pm, the pilot detected slight icing above the aircraft using radar; between 13:26 pm and 13:39 pm, the aircraft entered the cloud and encountered SLD; between 13:39 pm and 15:23 pm, SLD was encountered on the round-trip route between Bloomington Normal and Springfield; subsequently, the aircraft returned to Terre Haute Airport. Based on the icing severity prediction algorithm, the icing conditions in the altitude range of 750 to 900 hPa from 13:00 to 15:00 on February 17, 2019 were simulated. The simulation results showed that for the time periods of 14:00 and 15:00, the aircraft might encounter moderate icing or SLD in the pressure layer of 750 - 800 hPa, and this simulation result was consistent with the pilot's flight report.

[0153] 2) All flights

[0154] To comprehensively evaluate the accuracy of the icing severity prediction algorithm proposed in this paper, all data points in the sampling area were calculated using the LEWICE software, and the TPR, FNR, and accuracy of the aircraft icing severity algorithm were calculated. The results are shown in Table 6 below:

[0155] Table 6 TPR, FNR, and accuracy of icing severity prediction

[0156]

[0157] It can be found that the true positive rates of the icing severity prediction algorithm for slight icing, light icing, moderate icing, and severe icing reach 90%, 76%, 72%, and 50% respectively, and at the same time have a low false negative rate, being able to accurately predict the aircraft icing severity. To sum up, the present invention uses the flight data published by the US ICICLE project as input, establishes a CNN model based on the attention mechanism, provides a new method for distinguishing large and small droplets of liquid water in clouds, and can accurately predict cloud physical parameters such as LWC and MVD in the case of small droplets. First, based on the Thomson scheme, the cloud physical parameters output by the WRF model are obtained, and it is found that the WRF model underestimates the LWC result and further correction is needed. Subsequently, two CNN-Attention models are established to train the output parameters, and it is found that the CNN-Attention 0 model has a significant advantage in the prediction accuracy of droplet classification, reaching 90.1%, and at the same time has extremely high prediction ability for the small droplet MVD and LWC. Finally, based on the WRF model, the CNN-Attention 0 model and the icing numerical simulation calculation, an aircraft icing severity prediction algorithm is established to realize the prediction of the aircraft icing severity. Generally speaking, this embodiment adopts the method of classification simulation, excludes large droplets, and obtains accurate predictions of the icing microphysical parameters (MVD and LWC) under small droplet conditions at low cost. The CNN-Attention0 model constructed in the research shows good capabilities in both droplet classification and the prediction of MVD and LWC under small droplet conditions, and combined with the icing numerical simulation calculation, significantly improves the classification prediction accuracy of the aircraft icing severity. The map data shown in this article is sourced from: World Standard Map (Map Review Number: GS(2021)5444).

[0158] It can be understood that the present invention also provides corresponding system products for the above-mentioned various methods. For example, the present invention also provides a prediction system for cloud physical parameters. The model training subsystem includes: a sample acquisition module for acquiring a first meteorological sample set, where the first meteorological sample set includes: meteorological sample groups at multiple moments, and the meteorological sample groups include: temperature, air pressure, relative humidity, cloud water mass mixing ratio, rain water mass mixing ratio, vertical velocity, ice water mass mixing ratio, ice water content, raindrop number concentration, ice number concentration, and a first liquid water content; a first training module for inputting the first meteorological sample set into a convolutional neural network for iterative calculation. At this time, the input end of the convolutional neural network is: temperature, air pressure, relative humidity, cloud water mass mixing ratio, rain water mass mixing ratio, vertical velocity, ice water mass mixing ratio, ice water content, raindrop number concentration, ice number concentration, and a first liquid water content; the output end of the convolutional neural network is: the probability value that the median volume diameter belongs to large droplets; where the activation function uses the normalized exponential function, and the normalized exponential function is used to calculate the probability value corresponding to the meteorological sample group, and the loss function uses the binary cross-entropy function, and the binary cross-entropy function is used to calculate the convergence state of the current probability value; a sample screening module for combining the meteorological sample groups whose probability values belong to a set probability interval and whose temperature is greater than a set temperature threshold to form a second meteorological sample set; a second training module for inputting the second meteorological sample set into the convolutional neural network again for updated iterative calculation, and finally obtaining a corresponding first prediction model. The output of the first prediction model is: a second liquid water content and a median volume diameter; at this time, the activation function uses the rectified linear unit function, and the rectified linear unit function is used to calculate the second liquid water content and the median volume diameter, and the loss function uses the mean square error function, and the mean square error function is used to calculate the convergence state of the second liquid water content and the median volume diameter.

[0159] In some embodiments, it further includes: a probability setting system for: selecting multiple discrimination thresholds between [a, b], where the discrimination thresholds are used for determining large droplets and small droplets, and a and b are set probability interval values; setting a first verification index, where the first verification index includes: a first ratio R1 of the number of samples correctly classified as large droplets to the number of actual large droplet samples, and a second ratio R2 of the number of samples misclassified as large droplets to the number of actual small droplet samples; where when the probability value is greater than or equal to the discrimination threshold, it is considered a large droplet, and when the probability value is less than the discrimination threshold, it is considered a small droplet; calculating multiple receiver operating characteristic curves corresponding to the multiple discrimination thresholds according to the first ratio R1 and the second ratio R2; calculating the area size of the receiver operating characteristic curves; and selecting at least one discrimination threshold corresponding to the largest area size as the discrimination threshold during the model training process.

[0160] The present invention also correspondingly provides a verification system for model training, including: a sample acquisition module, configured to acquire a first meteorological sample set and the corresponding median volume diameter. The first meteorological sample set includes: meteorological sample groups at multiple moments, and the meteorological sample group includes: temperature, air pressure, relative humidity, cloud water mass mixing ratio, rain water mass mixing ratio, vertical velocity, ice water mass mixing ratio, ice water content, raindrop number concentration, ice number concentration, and first liquid water content; a first training module, configured to input the first meteorological sample set into a convolutional neural network for iterative calculation. At this time, the input end of the convolutional neural network is: temperature, air pressure, relative humidity, cloud water mass mixing ratio, rain water mass mixing ratio, vertical velocity, ice water mass mixing ratio, ice water content, raindrop number concentration, ice number concentration, and first liquid water content; the activation function of the convolutional neural network is configured to calculate the probability value of the median volume diameter of the sample group, and the probability value represents the probability that the corresponding sample belongs to a large droplet or a small droplet; correspondingly, the output end of the convolutional neural network is: probability value; a sample screening module, configured to form a second meteorological sample set from the meteorological sample combinations where the probability value belongs to a set probability interval and the temperature is greater than a set temperature threshold; a second training module, configured to input the second meteorological sample set into the convolutional neural network again for updated iterative calculation, and finally obtain a corresponding first prediction model. The output of the first prediction model is: second liquid water content and median volume diameter; a first verification module, configured to calculate a first verification index, and the first verification index includes: a first ratio R1 of the number of samples correctly classified as large droplets to the number of actual large droplet samples, and / or, a second ratio R2 of the number of samples misclassified as large droplets from small droplets to the number of actual small droplet samples; a second verification module, configured to determine whether the first verification index meets a set first verification condition. If not, adjust the probability interval and enter the second training module again.

[0161] Further, the probability interval is set by a discrimination threshold. When the probability value is greater than or equal to the discrimination threshold, it is considered a large droplet. When the probability value is less than the discrimination threshold, correspondingly, the second verification module is further configured to perform the following steps: calculate multiple receiver operating characteristic curves corresponding to the discrimination threshold according to the first ratio R1 and the second ratio R2; calculate the area size of the receiver operating characteristic curve; when the area size is greater than a set area threshold, it is considered that the first verification index meets the set verification condition.

[0162] In some embodiments, when the area size is less than or equal to the area threshold, the discrimination threshold is adjusted, and the sample screening module is entered again according to the newly formed probability interval. In some embodiments, the second sample acquisition module is configured to select, from the second meteorological sample set, the combination of meteorological elements belonging to a set vertical motion interval and a set relative humidity interval, and form the second test sample set according to the combination of meteorological elements. In some embodiments, the second sample acquisition module is configured to select, from the second meteorological sample set, the combination of meteorological elements belonging to a set altitude interval, and form the second test sample set according to the combination of meteorological elements.

[0163] The present invention also provides an electronic device, including: a memory for storing a computer program; a processor for implementing the steps of the cloud physical parameter prediction method according to any one of the embodiments when executing the computer program. The present invention also provides a computer-readable storage medium, in which a computer program is stored, and the computer program, when executed by a processor, implements the steps of the cloud physical parameter prediction method according to any one of the embodiments. The present invention also provides a computer program product, including a computer program / instructions, and the computer program / instructions, when executed by a processor, implement the steps of the cloud physical parameter prediction method according to any one of the embodiments. It should be noted that the activation function and loss function required by the present invention are the preferred solutions for the small sample training scenario. When the present invention is applied to other types (such as the large sample training scenario), other activation functions or loss functions can also be selected.

[0164] It should be noted that in this text, the term "including", "comprising" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed, or further includes elements inherent to such process, method, article or device. Without further limitation, an element defined by the statement "including one..." does not exclude the existence of additional identical elements in the process, method, article or device including that element. Through the description of the above embodiments, those skilled in the art can clearly understand that the above-described embodiment methods can be implemented by means of software plus a necessary general hardware platform. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present invention, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions for causing a computer terminal (which can be a mobile phone, computer, server, or network device, etc.) to execute the methods described in various embodiments of the present invention. The above has described the embodiments of the present invention in conjunction with the accompanying drawings. However, the present invention is not limited to the above specific embodiments. The above specific embodiments are merely illustrative and not restrictive. Under the inspiration of the present invention, those of ordinary skill in the art can also make many forms without departing from the purpose of the present invention and the scope protected by the claims. These all fall within the protection scope of the present invention.

Claims

1. A method for predicting cloud physical parameters, characterized in that: Includes steps: Training the model includes the following steps: S100, obtaining a first meteorological sample set, the first meteorological sample set comprising: a meteorological sample group at multiple moments, the meteorological sample group comprising: temperature, air pressure, relative humidity, cloud water mass mixing ratio, rain water mass mixing ratio, vertical speed, ice water mass mixing ratio, ice water content, raindrop number concentration, ice number concentration and first liquid water content; S101, inputting the first meteorological sample set into a convolutional neural network for iterative calculation, wherein the input end of the convolutional neural network is: temperature, air pressure, relative humidity, cloud water mass mixing ratio, rain water mass mixing ratio, vertical speed, ice water mass mixing ratio, ice water content, raindrop number concentration, ice number concentration and first liquid water content; the output end of the convolutional neural network includes: a probability value that the median volume diameter belongs to a large droplet; wherein the activation function in the convolutional neural network adopts a normalized exponential function, and the normalized exponential function is used to calculate the probability value corresponding to the meteorological sample group, and the loss function in the convolutional neural network adopts a binary cross point function, and the binary cross point function is used to calculate the convergence state of the current probability value; S102, combining the meteorological samples corresponding to the probability values ​​belonging to the set probability interval and the temperatures greater than the set temperature threshold to form a second meteorological sample set; S103, inputting the second meteorological sample set into the convolutional neural network again for updating and iterative calculation, and finally obtaining the corresponding first prediction model, the output of the first prediction model includes: the second liquid water content and the median volume diameter; wherein, the activation function in the convolutional neural network adopts a linear rectification function, and the linear rectification function is used to calculate the second liquid water content and the median volume diameter; the loss function in the convolutional neural network adopts a mean square error function, and the mean square error function is used to calculate the convergence state of the second liquid water content and the median volume diameter.

2. A cloud physical parameter prediction method according to claim 1, characterized in that: Before S101, the following steps are also included: Acquire a grid model of the area to be predicted, wherein the grid model has a plurality of sample grid points; The first meteorological sample set is vertically linearly interpolated to match corresponding meteorological sample values ​​for a plurality of the sample grid points, and the plurality of meteorological sample values ​​constitute a meteorological sample group corresponding to one of the sample grid points.

3. The method for predicting cloud physical parameters according to claim 1, characterized in that: The first meteorological sample set is derived from aircraft detection data and the WRF model.

4. The method for predicting cloud physical parameters according to claim 1, characterized in that: Before S102, the step further includes: Setting the probability interval; the setting of the probability interval comprises the steps of: Select multiple discrimination thresholds between [a, b], where the discrimination thresholds are used to determine large droplets and small droplets, and a and b are set probability end values; A first verification index is set, wherein the first verification index includes: a first ratio R1 of the number of samples correctly classified as large droplets to the actual number of large droplet samples, and a second ratio R2 of the number of samples classified as small droplets to the actual number of small droplet samples; wherein when the probability value is greater than or equal to the discrimination threshold, it is considered to be a large droplet, and when the probability value is less than the discrimination threshold, it is considered to be a small droplet; Calculating a plurality of receiver operating characteristic curves corresponding to the plurality of discrimination thresholds according to the first ratio R1 and the second ratio R2; Calculating the area of ​​the receiver operating characteristic curve; A discrimination threshold corresponding to at least one largest area size is selected as the discrimination threshold in the model training process, so as to form a corresponding probability interval according to the discrimination threshold.

5. A cloud physical parameter prediction method according to claim 4, characterized in that: Prior to S102, it also included: Calculate the true skill statistics under each discrimination threshold according to the first ratio R1 and the second ratio R2; When the true skill statistical value is greater than or equal to the set minimum statistical value, it is considered that the corresponding discrimination threshold meets the requirements of model training.

6. A cloud physical parameter prediction method according to claim 2, characterized in that: Also includes: Model application discrimination, which includes the steps of: S200, obtaining the vertical motion and relative humidity of the aircraft's flight area; S201, using a first determination mechanism to determine whether the first prediction model is applicable to the flight area; the first determination mechanism requires that the vertical motion is lower than a preset vertical motion threshold, and the relative humidity is higher than a preset relative humidity threshold; If yes, it is recommended to adopt the first prediction model.

7. A cloud physical parameter prediction method according to claim 2, characterized in that: Prior to S102, it also included: S203, obtaining the altitude of the flight area; S204, using a second determination mechanism to determine whether the first prediction model is applicable to the flight area, and if so, recommending the use of the first prediction model; the second determination mechanism requires that the altitude is less than a set poster altitude threshold.

8. A cloud physical parameter prediction system, characterized in that: Model training subsystem, which includes: A sample acquisition module is used to acquire a first meteorological sample set, wherein the first meteorological sample set includes: a meteorological sample group at multiple moments, wherein the meteorological sample group includes: temperature, air pressure, relative humidity, cloud water mass mixing ratio, rain water mass mixing ratio, vertical speed, ice water mass mixing ratio, ice water content, raindrop number concentration, ice number concentration and first liquid water content; A first training module is used to input the first meteorological sample set into a convolutional neural network for iterative calculation. At this time, the input end of the convolutional neural network is: temperature, air pressure, relative humidity, cloud water mass mixing ratio, rain water mass mixing ratio, vertical speed, ice water mass mixing ratio, ice water content, raindrop number concentration, ice number concentration and first liquid water content; the output end of the convolutional neural network is: the probability value of the median volume diameter belonging to large droplets; wherein the activation function in the convolutional neural network adopts a normalized exponential function, and the normalized exponential function is used to calculate the probability value corresponding to the meteorological sample group, and the loss function in the convolutional neural network adopts a binary cross point function, and the binary cross point function is used to calculate the convergence state of the current probability value; A sample screening module, used for combining the meteorological samples corresponding to the probability values ​​belonging to the set probability interval and the temperatures greater than the set temperature threshold to form a second meteorological sample set; The second training module is used to input the second meteorological sample set into the convolutional neural network again for updating and iterative calculation, and finally obtain the corresponding first prediction model, and the output of the first prediction model is: the second liquid water content and the median volume diameter; wherein, the activation function in the convolutional neural network adopts a linear rectification function, and the linear rectification function is used to calculate the second liquid water content and the median volume diameter; the loss function in the convolutional neural network adopts a mean square error function, and the mean square error function is used to calculate the convergence state of the second liquid water content and the median volume diameter.

9. An electronic device, characterized in that: include: Memory for storing computer programs; A processor, configured to implement the steps of the method for predicting cloud physical parameters according to any one of claims 1 to 6 when executing the computer program.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the method for predicting cloud physical parameters according to any one of claims 1 to 6 are implemented.

Images