Verification method, system, device, medium and product of icing prediction model

Abstract

The application relates to the field of icing prediction, in particular to a verification method, system, equipment, medium and product of an icing prediction model, which comprises the following steps: acquiring a first meteorological sample set and inputting the first meteorological sample set into a neural network for iterative calculation, wherein the input end is T 、 RH 、 P 、 Q cloud 、 Q rain 、 IWC 、 LWC 、 W 、 Q ice 、 N ice and N rain ; the output end is a probability value of MVD; meteorological samples with the probability value belonging to a set probability interval are combined to form a second meteorological sample set, and the second meteorological sample set is input into a convolutional neural network for updated iterative calculation; a first verification index is calculated, which comprises a first ratio of the number of samples correctly classified as large droplets to the actual number of large droplet samples and a second ratio of the number of samples classified as small droplets as large droplets to the actual number of small droplet samples; and whether it is necessary to continue updating is judged according to the first verification index. The application improves the icing prediction accuracy under the condition of small samples.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of icing prediction, and in particular to a method, system, device, medium and product for verifying an icing prediction model. BACKGROUND

[0002] When an aircraft passes through a cloud layer, icing phenomenon is likely to occur when encountering supercooled water, which can significantly reduce the aerodynamic performance of the aircraft and may lead to serious aviation safety accidents. Precise prediction of cloud physical parameters when the aircraft passes through the cloud layer and improvement of the accuracy of aircraft icing prediction are 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, which integrates the latest observation data and the output of the NOAA rapid refresh (RAP) model to predict the hourly icing potential and supercooled large droplet (SLD) in the airspace of the United States. Subsequently, a numerical model-based icing potential prediction (FIP) algorithm was developed based on the CIP algorithm, which laid the foundation for the establishment of the icing prediction system of the aviation meteorological center. However, with the development of aircraft deicing technology, the impact of micro-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 techniques and improved microphysical schemes to improve prediction accuracy. Through the cloud microphysical scheme proposed by Thompson et al., the FAA aviation weather research program has strengthened the prediction of aircraft and ground icing and quantitative precipitation prediction, which can effectively predict cloud droplet 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 unbounded boundary layer schemes, and found that the combination of Thompson microphysical scheme and Edelhammer scheme had relatively good effect, but the false positive rate was still high.

[0004] In addition, patent application CN111738481A discloses an aircraft icing weather parameter MVD prediction method based on a BP neural network. According to the relationship among liquid water content LWC, average effective water droplet diameter MVD and ambient air temperature T listed in Appendix C of China Transport Aircraft Airworthiness Standard [CCAR-25-R4], the icing conditions under different flight conditions are calculated, and an icing thickness database of icing thickness changing with time is established. The BP neural network model is trained using the icing thickness database, and the flight conditions (flight attack angle, flight speed), temperature, and icing thickness and icing time mapping relationship provided in real time are used as inputs to predict the icing weather parameter MVD. Patent application CN114880947A discloses a method for predicting aircraft icing weather parameters MVD and LWC, including obtaining the icing thickness and icing rate of the measurement point position according to the icing calculation icing shape, establishing a database of effective icing thickness and icing rate; taking flight speed, environmental temperature, wing attack angle, effective icing thickness and icing rate as input parameters, and using genetic algorithm to optimize the initial weight and threshold value of Elman neural network for training to predict the output meteorological parameters average effective water droplet diameter MVD and liquid water content LWC.

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

[0006] The purpose of the present application is to provide a verification method, system, device, medium and product for icing prediction model, which partially solves or alleviates the above-mentioned deficiencies in the prior art, and can improve the prediction accuracy of LWC and MVD in typical application scenarios such as stratiform clouds. In order to solve the above-mentioned technical problems, the present application specifically adopts the following technical solutions: a verification method for an icing prediction model, comprising the steps of:

[0007] S300, obtaining a first meteorological sample set and a corresponding median volume diameter, the first meteorological sample set comprising: a plurality of meteorological sample groups at different times, the meteorological sample group comprising: temperature, air pressure, relative humidity, cloud water mass mixing ratio, rainwater mass mixing ratio, vertical velocity, ice water mass mixing ratio, ice water content, raindrop number concentration, ice number concentration and first liquid water content;

[0008] S301, input the first weather sample set into a convolutional neural network for iterative calculation, wherein, the input 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 used to calculate the probability value of the median volume diameter of the sample group, which represents the probability that the corresponding sample belongs to large droplets or small droplets; correspondingly, the output of the convolutional neural network is: the probability value;

[0009] S302, combine the weather sample group corresponding to the probability value belonging to the set probability interval and the temperature greater than the set temperature threshold to form a second weather sample set;

[0010] S303, input the second weather sample set into the convolutional neural network again for updated iterative calculation, and finally output a corresponding first prediction model, wherein, the output of the first prediction model is: liquid water content and median volume diameter;

[0011] S304, calculate a first verification index, which includes: a first ratio R1 of the number of samples correctly classified as large droplets to the actual number of large droplet samples, and / or a second ratio R2 of the number of samples classified as large droplets to the actual number of small droplet samples;

[0012] S305, when the first verification index is determined to meet the set first verification condition, if not, adjust the probability interval, and execute S302 again according to the new probability interval.

[0013] 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 to be a large droplet, and when the probability value is less than the discrimination threshold, correspondingly, S305 includes the steps of:

[0014] According to the first ratio R1 and the second ratio R2, calculate a plurality of receiver operating characteristic curves corresponding to the discrimination threshold;

[0015] Calculate the area size of the receiver operating characteristic curve;

[0016] When the area size is greater than a set area threshold, it is considered that the first verification index meets the set verification condition.

[0017] In some embodiments, it further includes the steps of: when the area size is less than or equal to the area threshold, adjust the discrimination threshold, and execute S302 again according to the new probability interval formed correspondingly.

[0018] In some embodiments, the convolutional neural network comprises at least four convolutional layers.

[0019] The application also provides a model training verification system, comprising:

[0020] a sample obtaining module, configured to obtain a first meteorological sample set and a corresponding median volume diameter, the first meteorological sample set comprising: a plurality of meteorological sample groups at different time points, each meteorological sample group comprising: 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;

[0021] a first training module, configured to input the first meteorological sample set into a convolutional neural network for iterative calculation, wherein the input end of the convolutional neural network is configured to receive: 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, the probability value representing the probability that the corresponding sample belongs to large droplets or small droplets; correspondingly, the output end of the convolutional neural network is configured to output the probability value;

[0022] a sample screening module, configured to combine the meteorological sample groups whose probability values of the median volume diameter belong to a set probability interval and whose temperatures are greater than a set temperature threshold to form a second meteorological sample set;

[0023] a second training module, configured to input the second meteorological sample set into the convolutional neural network again for updated iterative calculation, and finally output a corresponding first prediction model, wherein the output of the first prediction model is: second liquid water content and median volume diameter;

[0024] a first verification module, configured to calculate a first verification index, wherein the first verification index comprises: a first ratio R1 of the number of samples correctly classified as large droplets to the actual number of large droplet samples, and / or a second ratio R2 of the number of samples classified as large droplets to the actual number of small droplet samples;

[0025] a second verification module, configured to determine whether the first verification index meets a set first verification condition, and if not, adjust the probability interval and re-enter the second training module.

[0026] 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 to be a large droplet, and when the probability value is less than the discrimination threshold, correspondingly, the second verification module is further configured to perform the following steps:

[0027] calculating a plurality of receiver operating characteristic curves corresponding to the discrimination threshold according to the first ratio R1 and the second ratio R2;

[0028] calculating an area size of the receiver operating characteristic curve;

[0029] when the area size is greater than a set area threshold, it is considered that the first verification index determines whether to meet the set verification condition.

[0030] In some embodiments, further comprising the step of: when the area size is less than or equal to the area threshold, adjusting the discrimination threshold, and entering the sample screening module again according to the corresponding new probability interval.

[0031] The application also provides an electronic device comprising a memory for storing a computer program, and a processor for executing the computer program to implement the steps of the verification method of the icing prediction model according to any one of the embodiments. The application also provides a computer readable storage medium having a computer program stored therein, wherein the computer program is executed by a processor to implement the steps of the verification method of the icing prediction model according to any one of the embodiments. The application also provides a computer program product comprising computer programs / instructions, which are executed by a processor to implement the steps of the verification method of the icing prediction model according to any one of the embodiments.

[0032] Beneficial technical effects:

[0033] It should be noted that in the traditional LWC or MVD prediction algorithm, it is usually extremely dependent on the flight data of the aircraft, such as flight speed, wing angle of attack, icing time, icing thickness, and other measured requirements (or non-meteorological data), and the applicant found that this would greatly limit the application of the icing prediction algorithm.

[0034] On the contrary, the application proposes a model training and application method for predicting LWC and MVD based on meteorological parameters. This direct prediction scheme based on pure meteorological parameters can expand the application scenarios of the prediction algorithm. For example, before the aircraft performs a flight task, such as the design of the related deicing system of the aircraft or the planning stage of the flight task, the prediction of the icing degree is also crucial, and the prediction algorithm proposed in the application can predict the icing degree in the future long-term process (such as one day, one month, etc.) by means of meteorological parameters obtained through meteorological prediction, thereby reducing the dependence on other non-meteorological parameters.

[0035] And, the first prediction method (or the first prediction model) has a significant prediction advantage in the typical flight area such as stratiform cloud. The stratiform cloud refers to a uniform (uniform in thickness, gray scale and light transmission) cloud layer covering the whole or part of the sky. For example, the macrostructure of a large range of precipitation stratiform cloud associated with a low value weather system is layered, sometimes two layers, sometimes three layers. The cloud between the two layers is a cloud-free area, the high-level cloud produces ice crystal particles and falls to the bottom cloud, which is a catalytic cloud; the ice crystal particles entering the low-level cloud continue to grow, and the water and environment required for the growth of the ice crystal in the bottom cloud are the supply cloud.

[0036] In the flight process of the flight, crossing the stratiform cloud is one of the most typical flight scenes, but due to the complexity of the cloud layer structure and meteorological characteristics inside the stratiform cloud, the difficulty of predicting the related parameters of the stratiform cloud area also increases sharply. For this purpose, the secondary training method based on the classification of large and small droplet probability intervals proposed by the present application can effectively improve the recognition sensitivity of the prediction model for the size of the droplets inside the cloud layer, and thus significantly improve the prediction accuracy of the model for LWC and MVD.

[0037] And, through experiments, it has been proved that the prediction model proposed by the present application shows extremely high reliability when predicting the parameters of small droplets. This also enables the present application to significantly reduce the false negative rate when the icing potential is relatively low (such as when there are more small droplets). In other words, the prediction method proposed by the present application can alleviate the problem of low prediction accuracy of small droplets in the prior art.

[0038] In addition, the secondary training method based on the classification of large and small droplet probability intervals used by the present application 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.

[0039] Further, since the aircraft icing prediction is crucial to flight safety, the present application also provides two training methods for users to choose from. BRIEF DESCRIPTION OF DRAWINGS

[0040] In order to more clearly illustrate the technical solutions of the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiment or prior art description. In all the drawings, similar elements or parts are generally identified by similar reference signs. In the drawings, each element or part is not necessarily drawn according to the actual proportion. Obviously, the drawings described below are some embodiments of the present application, and those skilled in the art can obtain other drawings according to these drawings without creating any inventive labor.

[0041] Figure 1 The method flowchart in an exemplary embodiment of the present application;

[0042] Figure 2 A schematic diagram of WRF model domain configuration in an exemplary embodiment;

[0043] Figure 3 A schematic diagram of NAC0012 airfoil and wing leading edge ice shape as a function of exposure time; wherein (a) is a schematic diagram of NAC0012 airfoil, (b) is a plot of wing leading edge ice shape as a function of exposure time for a given condition;

[0044] Figure 4 A ROC curve plot of large droplet classification prediction based on CNN-Attention0 model;

[0045] Figure 5 A TSS curve plot of large droplet classification prediction based on CNN-Attention0 model;

[0046] Figure 6 A plot of CNN-Attention0 model prediction results for microphysical meteorological parameters; wherein (a) shows the relationship between model prediction accuracy and vertical motion and relative humidity, (b) shows the relationship between model prediction accuracy and pressure and liquid water content, (c) shows the relationship between model prediction accuracy and ice water content, temperature, (d) shows the relationship between model prediction accuracy and raindrop number concentration and ice number concentration;

[0047] Figure 7 A model forecast-actual observation density scatter plot of LWC based on CNN-Attention1;

[0048] Figure 8 A model forecast-actual observation density scatter plot of LWC based on CNN-Attention0;

[0049] Figure 9 A model forecast-actual observation density scatter plot of MVD based on CNN-Attention1;

[0050] Figure 10 A model forecast-actual observation density scatter plot of MVD based on CNN-Attention0. DETAILED DESCRIPTION

[0051] In order to make the objects, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some but not all of the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the present application. In this document, the suffixes such as "module", "part" or "unit" used to represent elements are merely for the convenience of description of the present application, and have no specific meaning in themselves. Therefore, "module", "part" or "unit" can be used interchangeably. In addition, the terms "first", "second" are used only for the purpose of description, and cannot be understood as indicating or implying relative importance. In this document, unless otherwise clearly specified and limited, the term "connection" and the like should be understood in a broad sense, for example, "connection" can be direct connection or indirect connection through an intermediate medium, and can be internal connection of two elements. For a person of ordinary skill in the art, the specific meaning of the above-mentioned terms in the present application can be understood according to the specific circumstances. In this document, "and / or" includes any and all combinations of one or more listed related items. In this document, "multiple" means two or more, that is, it includes two, three, four, five, etc. In this specification, the term "about" typically means + / - 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, certain embodiments can be disclosed in a format that is a range. It should be understood that this "in a range" description is merely for the convenience and brevity, and should not be interpreted 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 independent digital values within the 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., as well as individual numbers within the range, such as 1, 2, 3, 4, 5 and 6. The above rules apply regardless of the breadth of the range.

[0052] Embodiment one: see Figure 1 As shown, the present application provides a cloud physical parameter prediction method, comprising the steps of: training a model (equivalent to providing a first training method), which comprises the steps of:

[0053] S100, obtaining a first weather sample set, the first weather sample set comprising: a plurality of weather sample groups at different time points, the weather sample group comprising: temperature (T) T ), air pressure (P P), relative humidity ( RH ), cloud water mass mixing ratio ( Q cloud ), rainwater 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 ), first liquid water content ( LWC );

[0054] S101: Input the first meteorological sample set into a convolutional neural network for iterative calculation. In this case, the input end of the convolutional neural network 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; the output end of the convolutional neural network includes: the probability value of the median volume diameter (MVD) belonging to large droplets;

[0055] Preferably, the activation function of the convolutional neural network adopts a normalized exponential function, which is used to calculate the probability value corresponding to the meteorological sample group; the loss function of the convolutional neural network adopts a binary intersection function, which is used to calculate the convergence state of the current probability value;

[0056] In other words, step S101 is equivalent to the classification step. Specifically, to improve the accuracy of large droplet classification, the present invention uses a normalized exponential function to convert unnormalized MVD values ​​into probability values, thereby increasing the iterative convergence speed and training accuracy of the model. A binary crossover function is also used to calculate convergence during the iteration process. Subsequently, the probability values ​​are used to classify different sample data; in other words, the probability values ​​are used to quickly classify samples.

[0057] Preferably, in the classification training process of S101 of this embodiment, an initial model for predicting the actual liquid water content (equivalent to the second liquid water content) and the median volume diameter can also be obtained based on 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.

[0058] Specifically, the probability value in the embodiment can be used to represent the probability that the droplet belongs to a large droplet.

[0059] In some embodiments, the input layer in the convolutional neural network inputs a meteorological sample group and a MVD value (such as an average MVD (μm)); and the activation function can be calculated by the association between the meteorological sample group and the MVD value (such as the average MVD (μm)) to obtain the probability value corresponding to the different meteorological sample groups.

[0060] 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 last iteration process, and when the difference is small, it is considered that the two converge, the iteration is ended, otherwise the iteration is continued.

[0061] S102, the probability value belongs to a set probability interval, and the temperature is greater than a set temperature threshold value, and the corresponding meteorological sample group is combined 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 value, it is considered that the sample droplet corresponding to the sample droplet belongs to a small droplet with high probability. That is to say, in the embodiment, the second meteorological sample set is a sample set with high probability of belonging to a small droplet.

[0062] S103, the second meteorological sample set is input again to the convolutional neural network for update iteration calculation, and finally a corresponding first prediction model (also referred to as CNN-Attention0 in the present document) is obtained, and the output of the first prediction model is: a second liquid water content and a median volume diameter;

[0063] In the embodiment, the first liquid water content can be a sample value output according to the WRF model; and 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 linear rectifier function, the linear rectifier function is used to calculate the second liquid water content and the median volume diameter, and the loss function adopts a mean square error function, 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, the process can update the initial model obtained in S101.

[0064] It is worth noting that the embodiment proposes a model training method based on sample interval classification for secondary training. Specifically, the present application roughly screens sample data belonging to small droplets by the probability value of the droplet, and inputs the screened small droplet sample data again into the convolutional neural network for reinforcement 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.

[0065] For example, in the embodiment, the first prediction model is obtained by initially predicting the liquid water content (LWC) and the median volume diameter (MVD) of the sample droplet through the weather research and forecast (WRF) model.T 、 RH 、 P 、 LWC 、 Q cloud 、 Q rain 、 IWC 、 W 、 Q ice 、 N ice and N rain and the like are transmitted to the convolutional layer through the input layer. The convolutional layer includes four one-dimensional convolutions, the first convolution reads the input sequence and projects the result onto a feature map, the second convolution performs the same operation on the feature map created by the first layer, and performs maximum merging after each convolution to amplify its significant features. Under the first training stage, for the large droplet classification algorithm, the activation function is the normalized exponential function, that is, the Softmax function, which is used to calculate the probability value of the corresponding droplet belonging to the large droplet, and the loss function is the binary cross-point function. Subsequently, enter the maximum pooling layer, take the maximum value in the sample as the sample value after sampling, simplify the feature map using the maximum pooling layer, and then flatten the feature map into a long vector for decoding. Finally, enter the fully connected layer, and take the spatiotemporal matching LWC and MVD values as the model output. The droplets with MVD less than 100 μm are identified as small droplets, and the sample data identified as small droplets are input into the convolutional neural network again, at this time, the activation function adopts the RELU function, and the loss function adopts the mean square error function.

[0066] The loss function is calculated after each iteration cycle of the training set and the test set, the model is terminated after 500 iteration training cycles, and the corresponding first prediction model is obtained.

[0067] In some embodiments, before S101, further comprising the steps of: obtaining a grid model of the region 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 forming 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 WRF mode.

[0068] In some embodiments, before S102, there is further comprising a step of setting a probability interval; the step of setting a probability interval comprises steps of: selecting a plurality of discrimination thresholds between [a, b], the discrimination thresholds are used for determining large droplets and small droplets, a and b are set probability end values (set by the user); setting a first verification index, the first verification index comprises: 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 (ROC) 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 curve (AUC) of the receiver operating characteristic curve; selecting at least one discrimination threshold corresponding to the largest area size as the discrimination threshold c in the model training process, and generating a corresponding probability interval based on the discrimination threshold, such as the probability interval [a, c]. Wherein the receiver operating characteristic curve refers to a curve graph with false alarm probability (second ratio R2) as the horizontal coordinate and hit probability (first ratio R1) as the vertical coordinate, and the area under the curve can represent the model discrimination ability, such as AUC value of 0.5, which means no discrimination ability; the closer the AUC value is to 1, the better the prediction performance.

[0069] Through the following test verification, the AUC value of the entire sample set can reach 0.979 by the above model training and probability setting scheme, indicating that the model training method has high reliability.

[0070] In some embodiments, the first verification index can further comprise: a third ratio R3 of the number of samples correctly classified as small droplets to the actual number of small droplet samples; a fourth ratio R4 of the number of samples classified as small droplets to the actual number of large droplet samples; and accuracy, which is the ratio of the number of correctly classified droplet samples (including large droplets and small droplets) to the total number of samples.

[0071] Further, in some embodiments, in order to improve the accuracy of model training, there is further comprising a step of: calculating a 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 a set minimum statistical value, it is considered that the corresponding discrimination threshold meets the requirements of model training.

[0072] For example, in this embodiment, the minimum statistical value is set at about 0.85, see Figure 4 and Figure 5 When TSS reaches 0.87, the first ratio is 0.98 and the second ratio is 0.11; when TSS is 0.94, the training set shows 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 application can be further verified by the test results.

[0073] In some embodiments, the method further comprises a model application discrimination, which comprises the steps of:

[0074] S200, obtaining the vertical speed and relative humidity under the flight area of the aircraft; S201, using a first decision mechanism to determine whether the first prediction model is applicable to the flight area; the first decision mechanism requires that the vertical speed be lower than a preset vertical motion threshold and the relative humidity be higher than a preset relative humidity threshold; for example, in some embodiments, the relative humidity threshold can be about 70%; if yes, the first prediction model is recommended.

[0075] It is worth noting that the applicant has found that the first prediction model trained based on droplet classification is particularly suitable for flight areas such as stratiform clouds, and therefore, in order to improve the reliability of model prediction in actual application, the present application further proposes a corresponding judgment mechanism.

[0076] In some embodiments, before S102, further comprising: S203, obtaining the altitude under the flight area; S204, using a second decision mechanism to determine whether the first prediction model is applicable to the flight area, if yes, the first prediction model is recommended; the second decision mechanism requires that the altitude be lower than a set altitude threshold. For example, in some embodiments, when the altitude is lower than about 800 hPa, the first prediction model is recommended.

[0077] In some embodiments, the present application further provides a second prediction model trained directly based on a regression algorithm; the corresponding model training step (i.e. the second training method) comprises:

[0078] Obtaining a first meteorological sample set, the first meteorological sample set comprising: a plurality of meteorological sample groups at different time points, each meteorological sample group comprising: temperature (T) T ), air pressure (P) P ), relative humidity (RH) RH ), cloud water mass mixing ratio (Qc) Q cloud ), rainwater mass mixing ratio (Qr) Q rain ), vertical speed (W) W ), ice water mass mixing ratio (Qi)Q ice ice water content (IWC), IWC raindrop number concentration (RNC), N rain ice number concentration (INC), N ice liquid water content (LWC), LWC

[0079] The first set of weather samples is input into a convolutional neural network for iterative calculation, where the input 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 output of the convolutional neural network includes second liquid water content and median volume diameter; preferably, the activation function is a linear rectifier function used to calculate the second liquid water content and the median volume diameter, and the loss function is a mean square error function used to calculate the convergence state of the second liquid water content and the median volume diameter.

[0080] For example, in some embodiments, the output of WRF is used as the input of the network, and the spatiotemporally matched LWC and MVD values are used as the output of the network. T 、 RH 、 P 、 Q cloud 、 Q rain 、 IWC 、 W 、 Q ice 、 N ice 、 N rain 、 LWC are set as the input of the network, and the spatiotemporally matched LWC and MVD values are set as the output of the network. It can be understood that the samples used in the model training process herein can be measurement results collected by a weather monitoring system (such as a weather satellite).

[0081] After the iterative process converges, the corresponding second prediction model (also referred to herein as: CNN-Attention1 model) is obtained.

[0082] Unlike the above scheme, the model training in this embodiment directly performs regression training on the overall samples without introducing the step of distinguishing between large and small droplet samples.

[0083] In some embodiments, the above two model training methods can be used in cross.

[0084] ​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.

[0085] For example, in the present embodiment, before the model training, the following steps are further included: 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 a set difference interval, recommending to use the first training method for sample training; otherwise, the second training method can be recommended for sample training.

[0086] For example, in some embodiments, before the model training, the following steps are further included: obtaining a meteorological sample set; calculating the number of samples; when the number of samples is less than a set sample amount, recommending to use the first training method for sample training; otherwise, the second training method can be recommended for sample training.

[0087] It is worth noting that the first training method provided by the present application uses a mode of secondary training based on droplet classification, which has significant prediction advantages in scenarios where the number of samples is limited and the sample data distribution is relatively concentrated. In some embodiments, a small number of samples can be used to obtain a preliminary prediction model using two training methods; the recommended training method is selected by comparing the prediction accuracy of the two prediction models. In some embodiments, two prediction models can be used simultaneously for prediction to cross-verify the prediction results.

[0088] Embodiment two: the present application further provides a model training verification method, including the following steps:

[0089] S300, obtaining a first meteorological sample set and corresponding median volume diameter, the first meteorological sample set including: a plurality of meteorological sample groups at different times, the meteorological sample group including: temperature, air pressure, relative humidity, cloud water mass mixing ratio, rainwater mass mixing ratio, vertical velocity, ice water mass mixing ratio, ice water content, raindrop number concentration, ice number concentration, and liquid water content;

[0090] S301, input the first weather sample set to 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, rainwater 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 used to calculate the probability value of the median volume diameter belonging to large droplets of the sample group, and the probability value represents the probability of the corresponding sample belonging to large droplets or small droplets; correspondingly, the output end of the convolutional neural network includes: the probability value;

[0091] S302, the weather sample group 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 is combined to form a second weather sample set;

[0092] S303, the second weather sample set is input to the convolutional neural network again for updated iterative calculation, and finally a corresponding first prediction model is obtained, and the output of the first prediction model includes: second liquid water content and median volume diameter;

[0093] S304, a first verification index is calculated, 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 / or a second ratio R2 of the number of samples classified as large droplets to the actual number of small droplet samples;

[0094] S305, when the first verification index is judged to be inconsistent with a set first verification condition, if not, the probability interval is adjusted, and S302 is performed again according to the new probability interval.

[0095] For the application scene with relatively small trainable sample quantity, the application provides a secondary training method based on large and small droplet probability interval classification, and adopts the classification accuracy of large droplets and small droplets (corresponding to R1 and R2) as a key index to guide the adjustment direction of the secondary training, and then improves the prediction accuracy of small droplets based on the key index. Finally, in the training sample quantity is small, or the typical application scene of stratiform cloud, the accuracy of icing prediction is effectively improved.

[0096] For example, in some embodiments, when the first ratio and the second ratio are both greater than the corresponding set value, it is considered that the first verification index meets the set first verification condition.

[0097] For 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 to be a large droplet, when the probability value is less than the discrimination threshold, correspondingly, S305 comprises 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 curve; when the area size is greater than a set area threshold, it is considered that the first verification index determines whether to meet the set verification condition. In some embodiments, it further comprises the steps of: when the area size is less than or equal to the area threshold, the discrimination threshold is adjusted, and S302, S303 are executed again according to the corresponding new probability interval.

[0098] In some embodiments, the convolutional neural network comprises at least four convolutional layers.

[0099] Embodiment three: the application further provides a cloud physical parameter prediction model application method, comprising the steps of:

[0100] S401, obtaining a first meteorological sample set comprising a plurality of meteorological element combinations and corresponding median volume diameters; S402, selecting a first test sample set from the first meteorological sample set, and respectively training the first test sample set by using a first training method and a second training method to obtain a first prediction model and a second prediction model;

[0101] The step of training the first prediction model by using the first training method comprises:

[0102] The first test sample set is input into the convolutional neural network, and the output of the convolutional neural network at this time comprises: a probability value of the median volume diameter, which corresponds to the probability value indicating the probability of the corresponding sample belonging to a large droplet or a small droplet;

[0103] According to the probability value, the sample predicted as a small droplet is input into the convolutional neural network again to update the convolutional neural network, and the first prediction model is output after training convergence; at this time, the output of the convolutional neural network comprises: liquid water content and median volume diameter;

[0104] The step of training the second prediction model by using the second training method comprises: inputting the first test sample set into the convolutional neural network, and the output of the convolutional neural network is the second liquid water content and the median volume diameter, and the second prediction model is output after training convergence;

[0105] S403, selecting a second test sample set from the second weather sample set, the second test sample set not completely overlapping with the first test sample set; in some embodiments, the second weather sample set can be a part of the first weather sample set. Alternatively, the second weather sample set can also be a weather prediction result obtained in a future period of time for the current icing prediction task (or flight task).

[0106] S404, calculating the prediction accuracy of the first prediction model and the second prediction model respectively by using the second test sample set;

[0107] S405, recommending selecting a corresponding prediction model according to the prediction accuracy.

[0108] It is worth noting that the mechanism of aircraft icing is very complex, and the aircraft icing can vary greatly under different flight scenarios, climate conditions and aircraft types. Moreover, targeted modeling for icing prediction under different scenarios will consume a very high creation cost.

[0109] In this regard, the applicant chooses to cut in from the perspective of droplet size prediction accuracy, and provides two model training methods respectively. One is to improve the droplet classification accuracy (and improve the effective learning depth of small droplets) in special scenarios (for example, the sample data volume is relatively limited, or when flying to cumulus icing weather regions) through 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 relatively large sample data volume, or stratiform cloud weather environment).

[0110] Therefore, when predicting icing for a new flight task, the user can use a small amount of sample 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.

[0111] For 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.

[0112] In some embodiments, S403 includes the step of: selecting the meteorological element combination belonging to the set vertical motion interval and the set relative humidity interval from the second weather sample set, and forming the second test sample set according to the meteorological element combination.

[0113] Further, the applicant notices that the first prediction model can exhibit high sensitivity in areas with weak vertical motion and high relative humidity. Therefore, when verifying the model, it is preferred to observe its prediction accuracy under specific vertical motion and relative humidity intervals.

[0114] In some embodiments, S403 comprises the step of: selecting, from the second meteorological sample set, the combination of meteorological elements belonging to a set altitude interval, and forming the second test sample set according to the combination of meteorological elements.

[0115] Further, the applicant notices that the first prediction model can exhibit extremely high sensitivity in low-altitude areas. Therefore, when verifying the model, it is preferable to observe its prediction accuracy at a specific altitude interval.

[0116] In some embodiments, the meteorological sample set comprises: 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.

[0117] In order to describe the technical solutions adopted by the present application and the technical effects, the following will explain a specific model training and application example:

[0118] I. Data sources for model training and testing: This embodiment uses flight data published by the ICICLE project, which focuses on improving the understanding of aircraft icing by studying atmospheric microphysical processes. During the implementation of the project, the National Center for Atmospheric Research (NCAR) carried out multiple flight tests in the Great Lakes region of North America (between latitudes 37.588-45.666°) from January 28 to March 8, 2019. Cloud physics parameter data was collected using onboard instruments, and the thermodynamics, cloud dynamics, and specific cloud and aerosol characteristics of cold clouds and precipitation were analyzed.

[0119] This embodiment also uses data from the onboard instruments of the Convair 580 aircraft of the National Research Council of Canada (NRC) participating in the project. Among them, the flight altitude, longitude, and latitude data are provided by the KVH 1750 IMU sensor loaded in the cabin; the microphysical data of liquid and ice particles are obtained by the Forward Scattering Spectrometer Probe (FSSP), the Two-Dimensional Stereo (2DS) probe, and the High Volume Precipitation Spectrometer (HVPS). The particle sizes measured by the three instruments are 3-35 μm, 40-670 μm, and 750-38400 μm, respectively. With 100 μm as the threshold, the droplets are divided into small droplets (<100 μm) and large droplets (>100 μm) for further analysis.

[0120] The aircraft carried out a total of 30 flight probes. Since the first flight was a test flight, the second and the thirtieth flights did not carry out probes, so the flight data of 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 is 30s, which is roughly equivalent to a path length of 3km. The LWC data points, large droplet data points, average LWC, average MVD, and large droplet percentage within 30s of each flight of the aircraft are listed in Tables 1.1-1.2.

[0121] Table 1.1 Aircraft statistical data table

[0122]

[0123] Table 1.2 Aircraft statistical data table

[0124]

[0125] It can be understood that in some embodiments, the spatiotemporally matched sample values of LWC and MVD can be obtained by aircraft detection.

[0126] II. Data processing method - WRF-based weather parameter prediction method

[0127] The Weather Research and Forecasting (WRF) model is a fully compressible, non-hydrostatic numerical model. In this embodiment, the initial and boundary conditions are based on the fifth generation atmospheric reanalysis dataset (ERA5) of the European Centre for Medium-Range Weather Forecasts (ECMWF), with a horizontal grid spacing of 0.25° and a temporal resolution of 6h. A two-way nesting strategy is used to create two nested domains, each with a horizontal resolution of 27km and 9km and a vertical resolution of 44 elevation layers, with a temporal resolution of 1h, as shown in Figure 2 The black line represents the flight path of the aircraft. Each WRF model simulation is at least 6h long to ensure reliable simulation accuracy.

[0128] The WRF parameterization scheme selected in this embodiment is as follows: the microphysical scheme uses the Thompson scheme, which can improve the explicit prediction of aircraft icing and plays an important role in icing prediction; the shortwave radiation scheme uses the Dudhia scheme, the longwave radiation scheme uses the fast radiation transfer model, and the Eta surface layer uses the Noah land surface model. These configurations are based on the results of winter conditions research; the cumulus scheme uses the Kain-Fritsch cumulus scheme.

[0129] The following meteorological variables are selected in this embodiment to describe the aircraft icing environment, including altitude ( z ), temperature ( T ), and air pressureP ), relative humidity ( RH ), cloud water mass mixing ratio ( Q cloud ), rainwater 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 ), as shown in Table 2. The 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, and the ice water content, raindrop number concentration, ice number concentration, and liquid water content are obtained by multiplying the output parameters by the air density.

[0130] Table 2 Variables and definitions list

[0131]

[0132] 3. Training Tools

[0133] This embodiment preferably adopts a convolution-based feedforward neural network model (also known as a convolutional neural network model, or a CNN network model). The CNN model consists of a convolution layer, a pooling layer, and a fully connected layer. Among them, the convolution layer is composed of multiple convolution units, and the parameters of each convolution unit are optimized by a back-propagation algorithm. In this embodiment, the goal of the convolution operation is to extract various features of the input. The input data passes through the convolution layer to obtain a feature map, and the feature map is convolved with the input data through the neural network to represent the features in the neural space. The pooling layer is used to sample the feature map, mainly to reduce the computational complexity by reducing the number of network parameters, so as to reduce the network size 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 them through a convolution operation to form a global feature, which is used to calculate the final score of each category.

[0134] The CNN model used in this embodiment 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.

[0135] Among them, the calculation formula for data feature extraction using the CNN-attention mechanism model is as follows:

[0136] (1) ;

[0137] (2) ;

[0138] wherein x is input data, y is output data, σ is an activation function, W k and b k are weight coefficients and bias functions, respectively, k is the number of convolution kernels (representing a discrete convolution operation), α and β are size parameters of the convolution kernel, W m,n represent weight coefficients of the weight matrix convolution kernel, m and n represent row index and column index of the feature value in the convolution kernel, respectively, i, j represent row index and column index of the input data.

[0139] Four, sampling strategy of model observation comparison

[0140] In terms of time, in order to ensure that the model output result is completely matched with the flight detection data, the embodiment matches the model output with the aircraft detection result 15 minutes before and after the model output time. In each 30-minute sampling interval, the output value of a model grid point is selected as the matching value of the average observation. In terms of space, the embodiment adopts a vertical linear interpolation method to model the aircraft sampling height output, so as to realize the optimized matching of the flight height. After the space-time matching, the aircraft detection data and the WRF prediction data are divided into a training set (70%) and a test set (30%). The training set is used to train the CNN-attention model, and the test set is used to evaluate the accuracy of the model.

[0141] Specifically, the statistical method for evaluating the model adopted by the embodiment is as follows:

[0142] The first ratio (also referred to as: true positive rate (TPR)), the second ratio (also referred to as: false positive rate (FPR)), the third ratio (also referred to as: true negative rate (TNR)), the fourth ratio (also referred to 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 positive and small droplets are negative. The TPR is the ratio of the number of samples correctly classified as large droplets to the number of actual large droplet samples, the FPR is the ratio of the number of small droplets classified as large droplets to the number of actual small droplet samples, the TNR is the ratio of the number of samples correctly classified as small droplets to the number of actual small droplet samples, the FNR is the ratio of the number of large droplets classified as small droplets to the number of actual large droplet samples, and the accuracy is the ratio of the number of correctly classified droplet samples (including large and small droplets) to the total number of samples.

[0143] The root mean square error (RMSE) and the Pearson correlation coefficient (Pearson Corr) are used to evaluate the prediction ability of the model for LWC and MVD under small droplet conditions.

[0144] The formula for calculating the root mean square error is:

[0145] (3);

[0146] The formula for calculating the Pearson correlation coefficient is:

[0147] (4);

[0148] 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.

[0149] Five, numerical simulation of icing: In this embodiment, the LEWICE software is used to numerically simulate the icing of an aircraft. By inputting the wing model, aircraft state parameters, cloud microphysical parameters and exposure time, the icing pattern of the aircraft and the icing thickness of each region of the wing can be obtained, and the maximum icing rate can be calculated.

[0150] A lookup table (LUT) of the maximum icing rate of the aircraft is established by adjusting different parameters to determine the icing severity of the aircraft. The selection of the parameter range is crucial for accurate aircraft icing prediction. Based on the previous results and referring to FAR Part 25 Appendix C, the meteorological elements covering 95% of the icing events are extracted as the input parameter range of the LEWICE software by statistically analyzing tens of thousands of icing events, as shown in Table 3. Among them, the temperature is set to -28-4℃, the relative humidity is set to 70%-100%, the flight speed is set to 60-120m / s, the LWC is set to 0-1g / m 3 , and the MVD is set to 5-100μm.

[0151] The method of establishing the maximum icing rate lookup table (LUT) in this embodiment can refer to the aircraft icing severity prediction method based on icing numerical simulation disclosed in patent application CN117493738A.

[0152] Table 3 LEWICE parameter setting table

[0153]

[0154] The wing model used in this embodiment is a NACA0012 airfoil, as shown in Figure 3 . Figure 3 Figure (a) in the above figure is a schematic diagram of the NAC0012 airfoil; figure (b) is a curve showing the change of the ice type on the leading edge of the wing with exposure time under given conditions; and the lower right of figure (b) is a schematic diagram showing the change of the ice type on the leading edge of the wing from 1min to 20min under given conditions.

[0155] This airfoil is a standard numerical model of the Lewis wind tunnel and is widely used in icing simulation research. In order to test the correlation between the exposure time under icing conditions and the ice accumulation resulting therefrom, a series of experiments were conducted under fixed environmental parameters (temperature -8°C, flight speed 80m / s, relative humidity 100%, MVD 20µm, LWC 0.8g / cm 3 , exposure time 1-20min, test time interval 1min). The test results show that the maximum icing thickness increases uniformly with the increase of the exposure time, as shown in Figure 3 . 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 the aircraft icing.

[0156] Six, model verification

[0157] The CNN-Attention model is used to classify the size of the droplets. A number of thresholds are systematically established in the interval [0, 1], each of which is a discriminative criterion: if the predicted value exceeds the threshold, the diagnosis is classified as a large droplet, while a value below the threshold is considered to be a small droplet. Subsequently, the true positive rate and the false positive rate are calculated for these different thresholds, thereby evaluating the diagnostic ability of the CNN-Attention model. According to the relationship between the true positive rate and the false positive rate at different threshold levels shown by the receiver operating characteristic curve (ROC), the area under the curve (AUC) is a quantitative indicator of the discriminative ability of the model. When the AUC value is 0.5, it indicates no discriminative ability; the closer the AUC value is to 1, the better the prediction performance. Figure 4 The ROC curve for large droplet classification based on the CNN-Attention0 model shows that the model has good training ability, with an AUC value of 0.979 for the entire sample set.

[0158] To determine the most appropriate threshold for large droplet classification, the true skill statistic (TSS) at different thresholds is also calculated in this embodiment. TSS integrates the true positive rate and the false positive rate, providing a comprehensive measure of the classification performance of the model. Figure 4 It is shown that as the threshold increases from 0 to 1.0, the TSS first increases and then gradually decreases, reaching a maximum at a threshold of 0.12. Therefore, the decisive threshold for distinguishing between large and small droplets in this embodiment is set to 0.12. Considering that the false negatives of large droplets in the cloud may lead to serious icing consequences, while false positives may lead to unnecessary activation of the de-icing system, by considering the overall situation, 0.85 is used as the minimum TSS value to balance between false positives and false negatives. 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.

[0159] The true positive rate, false negative rate, and accuracy of the CNN-Attention0 model and the CNN-Attention1 model in classifying large and small droplets for all sample data are listed in Table 4. The results show that the CNN-Attention0 model has a higher true positive rate of 95% in classifying large and small droplets, indicating that it has stronger ability to accurately diagnose large droplets; at the same time, its accuracy reaches 90.1%, reflecting its overall precision in distinguishing between large and small droplets.

[0160] Table 4. Prediction accuracy table of the model

[0161]

[0162] The CNN-Attention0 model is further used 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, such as Figure 6 As shown, the applicant also found that the CNN-Attention0 model has significant advantages in predicting flight areas such as stratiform clouds. Especially in areas with weak vertical motion and high relative humidity (such as Figure 6 (a)). Applicants predict that this is due to the CNN-Attention0 model's ability to accurately identify droplets through droplet classification and secondary training, and to optimize the model based on these precise identification results. These precise characteristics enable 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.

[0163] from Figure 6 (b) shows that CNN-Attention0 also shows significant prediction advantages in areas with low altitude and low liquid water content. For example, when the altitude is less than or equal to 800hPa or the LWC is less than or equal to 0.3g / m 3 , the applicability of the model is significantly enhanced. Figure 6 (c) in Figure 2 shows that the samples where the CNN-Attention0 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, thereby hindering the formation and maintenance of large droplets. In addition, from Figure 6 (d) in the figure also shows that the CNN-Attention0 model is particularly suitable for predicting N rain Higher areas.

[0164] CNN-Attention Models' Prediction Capabilities for MVD and LWC of Small Droplets: The performance of two CNN-Attention models in predicting LWC and MVD for small droplets (MVD < 100 μm) was evaluated, using RMSE and Correlation coefficient (Correlation coefficient) as evaluation metrics. The results are listed in Table 5. As can be seen, the two models exhibited similar capabilities for LWC prediction, with the CNN-Attention0 model performing slightly better than the CNN-Attention1 model. In terms of MVD prediction, the CNN-Attention0 model performed exceptionally well, with an RMSE reduction of 25.5 and a correlation coefficient of 0.75, demonstrating significant improvement compared to the CNN-Attention1 model. This is likely due to the CNN-Attention0 model's ability to accurately distinguish between large and small droplets. This targeted focus enabled the model to converge more effectively, resulting in highly accurate MVD prediction.

[0165] Table 5 Evaluation of LWC and MVD prediction by two CNN-Attention models

[0166]

[0167] To better understand the proficiency of both models in predicting LWC and MVD under small droplet conditions, density scatter plots of model predicted versus actual observed values were generated for the entire dataset, with the closeness of the scatter points to the diagonal line representing the prediction accuracy, as shown in FIG. 6. The LWC scatter plots ( Figures 7-10 ) for both models show a clear concentration of points along the diagonal line, indicating that both models have demonstrated exceptional ability in predicting LWC for small droplets, with the predicted results closely matching the actual observed values. However, a difference in performance is observed in MVD prediction ( Figures 7-10 ). The scatter plot for the CNN-Attention1 model shows a more dispersed distribution of points, while the scatter plot for the CNN-Attention0 model shows a tighter clustering around the diagonal line, indicating more accurate estimation of MVD values. Therefore, the CNN-Attention0 model has lower systematic bias and overall error in predicting cloud microphysical parameters, especially for small water droplets. Overall, the performance evaluation on the full sample dataset indicates that both CNN-Attention models perform well in distinguishing between large and small droplets. Among the current sample dataset, the CNN-Attention0 model is the preferred option due to its higher true positive rate and accuracy. Figures 7-8 Seven, aircraft icing severity prediction

[0168] 7.1 Icing severity prediction algorithm: includes two components. The first part obtains the meteorological parameters (T, RH, LWC, and MVD) of the target area through meteorological parameter prediction and distinguishes between large and small droplets. In this embodiment, a hybrid model combining WRF and the CNN-Attention0 model is used. The second part is the icing severity prediction, which takes meteorological parameters and aircraft body parameters as input, performs linear interpolation on the maximum icing rate, and realizes the rapid evaluation and prediction of the maximum icing rate in the target area. The icing severity based on icing rate is divided as follows: trace icing <0.6 mm / min, light icing 0.6-1.0 mm / min, moderate icing 1.1-2.0 mm / min, and heavy icing >2.0 mm / min.

[0169] 7.2 Evaluation of icing severity prediction algorithm for typical icing events:

[0170] 1) Single flight

[0171]

[0172] ​The 17th flight of the ICICLE project provided a valuable case to evaluate 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 used radar to detect light icing above the aircraft; between 13:26 pm and 13:39 pm, the aircraft entered the cloud layer and encountered SLD; between 13:39 pm and 15:23 pm, the aircraft encountered SLD while traveling back and forth between Bloomington Normal and Springfield; then, the aircraft returned to Terre Haute airport. Based on the icing severity prediction algorithm, the icing conditions at 750-900 hPa from 13:00 to 15:00 on February 17, 2019 were simulated, and it was found that the aircraft might encounter moderate icing or SLD in the pressure layer of 750-800 hPa for the time periods of 14:00 and 15:00, which is consistent with the pilot's flight report.

[0173] 2) All flights

[0174] In order to comprehensively evaluate the accuracy of the icing severity prediction algorithm proposed in this paper, the TPR, FNR and accuracy of the aircraft icing severity algorithm were calculated using the LEWICE software for all data points in the sampling area. The results are shown in Table 6 as follows:

[0175] Table 6 TPR, FNR and accuracy of icing severity prediction

[0176]

[0177] It can be found that the icing severity prediction algorithm has a true positive rate of 90%, 76%, 72% and 50% for slight icing, mild icing, moderate icing and severe icing, respectively, and has a low false negative rate, which can achieve relatively accurate aircraft icing severity prediction. In summary, the present invention uses flight data published by the US ICICLE project as input and establishes a CNN model based on the attention mechanism, providing 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. It is found that the WRF model underestimates the LWC results and needs further correction. Subsequently, two CNN-Attention models are established to train the output parameters. It is found that the CNN-Attention0 model has a significant advantage in the prediction accuracy of droplet classification, which can reach 90.1%. At the same time, it has extremely high prediction capabilities for small droplet MVD and LWC. Finally, an aircraft icing severity prediction algorithm was established based on the WRF model, the CNN-Attention0 model, and numerical icing simulations, enabling aircraft icing severity prediction. Overall, this example employs a classification simulation approach to eliminate large droplets and accurately predict icing microphysical parameters (MVD and LWC) under small droplet conditions at a low cost. The constructed CNN-Attention0 model demonstrates excellent capabilities in both droplet classification and MVD and LWC prediction under small droplet conditions. Combined with numerical icing simulations, it significantly improves the classification accuracy of aircraft icing severity predictions. The map data presented in this article is sourced from the World Standard Map (Approval Number: GS(2021)5444).

[0178] It can be understood that the present application also provides corresponding system products for the above-mentioned methods. For example, the present application also provides a cloud physical parameter prediction system, a model training subsystem, which comprises: a sample acquisition module, configured to acquire a first meteorological sample set, the first meteorological sample set comprising: a plurality of meteorological sample groups at different time points, the meteorological sample group comprising: temperature, air pressure, relative humidity, cloud water mass mixing ratio, rainwater 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, wherein the input end of the convolutional neural network is: temperature, air pressure, relative humidity, cloud water mass mixing ratio, rainwater 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 output end of the convolutional neural network is: a probability value of the median volume diameter belonging to large droplets; wherein the activation function adopts a normalized exponential function, the normalized exponential function is used to calculate the probability value corresponding to the meteorological sample group, the loss function 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, configured to combine the meteorological sample groups whose probability value belongs 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, configured to input the second meteorological sample set into the convolutional neural network again for updated iterative calculation, and finally acquire a corresponding first prediction model, the output of the first prediction model being: second liquid water content and median volume diameter; at this time, the activation function adopts a linear rectifier function, the linear rectifier function is used to calculate the second liquid water content and the median volume diameter, and the loss function adopts a mean square error function, the mean square error function is used to calculate the convergence state of the second liquid water content and the median volume diameter.

[0179] In some embodiments, further comprising: a probability setting system, configured to: select a plurality of discrimination thresholds between [a, b], the discrimination thresholds being used for discrimination of large droplets and small droplets, a and b being set probability interval values; set a first verification index, the first verification index comprising: 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 large 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 size of the receiver operating characteristic curve; selecting at least one discrimination threshold corresponding to the largest area size as the discrimination threshold in the model training process.

[0180] The application also correspondingly provides a model training verification system, comprising: a sample acquisition module, configured to acquire a first meteorological sample set and a corresponding median volume diameter, the first meteorological sample set comprising: a plurality of meteorological sample groups at different time points, the meteorological sample group comprising: temperature, air pressure, relative humidity, cloud water mass mixing ratio, rainwater 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, wherein the input end of the convolutional neural network is: temperature, air pressure, relative humidity, cloud water mass mixing ratio, rainwater 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 used 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 large droplets or small droplets; correspondingly, the output end of the convolutional neural network is: the probability value; a sample screening module, configured to combine the meteorological sample groups corresponding to the probability value belonging to a set probability interval and the temperature being greater than a set temperature threshold to form a second meteorological sample set; a second training module, configured to input the second meteorological sample set into the convolutional neural network again for updated iterative calculation, and finally acquire a corresponding first prediction model, wherein 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, wherein the first verification index comprises: a first ratio R1 of the number of samples correctly classified as large droplets to the actual number of large droplet samples, and / or a second ratio R2 of the number of samples classified as large droplets to the actual number of small droplet samples; a second verification module, configured to determine whether the first verification index meets a set first verification condition, and if not, adjust the probability interval and re-enter the second training module.

[0181] Further, the probability interval is set by a discrimination threshold value, when the probability value is greater than or equal to the discrimination threshold value, it is considered to be large droplets, and when the probability value is less than the discrimination threshold value, correspondingly, the second verification module is further configured to perform the following steps: calculate a plurality of receiver operating characteristic curves corresponding to the discrimination threshold value 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 value, it is considered that the first verification index meets the set verification condition.

[0182] In some embodiments, when the area size is less than or equal to the area threshold, then the discrimination threshold is adjusted, and the sample screening module is entered again according to the corresponding newly formed probability interval. In some embodiments, the second sample obtaining module is configured to select the meteorological element combinations belonging to a set vertical motion interval and a set relative humidity interval from the second meteorological sample set, and form the second test sample set according to the meteorological element combinations. In some embodiments, the second sample obtaining module is configured to select the meteorological element combinations belonging to a set altitude interval from the second meteorological sample set, and form the second test sample set according to the meteorological element combinations.

[0183] The present application also provides an electronic device comprising a memory for storing a computer program, and 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 application also provides a computer readable storage medium having a computer program stored therein, wherein 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 application also provides a computer program product comprising computer programs / instructions, wherein the computer programs / 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 the loss function required by the present application are priority solutions for small sample training scenarios. When the present application is applied to other types (such as large sample training scenarios), other activation functions or loss functions can also be selected.

[0184] It should be noted that, in this text, the term "comprising", "including" or any other variant thereof is intended to cover a non-exclusive inclusion, so that a process, method, article or apparatus that includes a series of elements not only includes those elements, but also includes other elements not explicitly listed, or further includes elements inherent in such a process, method, article or apparatus. Without more limitations, the element defined by the statement "comprising a" does not exclude the presence of additional identical elements in the process, method, article or apparatus including the element. Through the description of the above embodiments, those skilled in the art can clearly understand that the above example method can be realized by software plus a general hardware platform, of course, it can also be realized by hardware, but in many cases the former is a better embodiment. Based on such understanding, the technical solutions of the present application can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes a plurality of instructions for making a computer terminal (which can be a mobile phone, computer, server, or network device, etc.) execute the method described in each embodiment of the present application. The embodiments of the present application are described above in conjunction with the drawings, but the present application is not limited to the above specific embodiments, and the above specific embodiments are only illustrative, not limiting, and those skilled in the art can make many forms under the inspiration of the present application without departing from the scope of the present application and the scope of protection of the claims.

Claims

1. A method of validating an icing prediction model, the method comprising: The method comprises the steps of: S300, obtaining a first weather sample set and a corresponding median volume diameter, the first weather sample set comprising a plurality of weather sample groups at different time points, and each weather sample group comprising 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; S301, inputting the first weather sample set into a convolutional neural network for iterative calculation, wherein the input end of the convolutional neural network is connected to 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 used 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 large droplets or small droplets; correspondingly, the output end of the convolutional neural network is connected to the probability value; S302, combining the weather sample groups corresponding to the probability value belonging to a set probability interval and the temperature being greater than a set temperature threshold to form a second weather sample set; S303, inputting the second weather sample set into the convolutional neural network again for updated iterative calculation, and finally outputting a corresponding first prediction model, wherein the output of the first prediction model is liquid water content and median volume diameter; S304, calculating a first verification index, wherein the first verification index comprises a first ratio R1 of the number of samples correctly classified as large droplets to the actual number of large droplet samples, and / or a second ratio R2 of the number of samples classified as large droplets from small droplets to the actual number of small droplet samples; S305, determining whether the first verification index meets a set first verification condition, and if not, adjusting the probability interval and performing S302 again according to the new probability interval.

2. The method of claim 1, wherein, The probability interval is set by a discrimination threshold value, when the probability value is greater than or equal to the discrimination threshold value, it is considered to be large droplets, and correspondingly, S305 comprises the steps of: calculating a plurality of receiver operating characteristic curves corresponding to the discrimination threshold value according to the first ratio R1 and the second ratio R2; calculating the area size of the receiver operating characteristic curve; when the area size is greater than a set area threshold value, it is considered that the first verification index meets the set verification condition.

3. The method of claim 2, wherein, The method further comprises the steps of: when the area size is less than or equal to the area threshold value, the discrimination threshold value is adjusted, and S302 is performed again according to the new probability interval formed correspondingly.

4. The method of claim 2, wherein, The convolutional neural network comprises at least four convolutional layers.

5. A model training verification system, comprising: The method comprises: a sample acquisition module configured to obtain a first weather sample set and a corresponding median volume diameter, the first weather sample set comprising a plurality of weather sample groups at different time points, and each weather sample group comprising 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 first training module is configured to input the first set of weather samples into a convolutional neural network for iterative calculation, wherein the input end of the convolutional neural network is configured to receive 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, wherein the probability value represents the probability that the corresponding sample belongs to large droplets or small droplets; and the output end of the convolutional neural network is configured to output the probability value. The sample screening module is configured to combine the sample group 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 to form a second set of weather samples. The second training module is configured to input the second set of weather samples into the convolutional neural network again for updated iterative calculation, and finally output a corresponding first prediction model, wherein the output of the first prediction model is second liquid water content and median volume diameter. The first verification module is configured to calculate a first verification index, 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 / or a second ratio R2 of the number of samples classified as large droplets to the actual number of small droplet samples. The second verification module is configured to determine whether the first verification index meets a set first verification condition, and if not, adjust the probability interval and re-enter the second training module. 6.The model training verification system of claim 5, wherein, The probability interval is set by a discrimination threshold, 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, the second verification module is further configured to perform the following steps: Calculate a plurality of 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 curves; When the area size is greater than a set area threshold, it is considered that the first verification index meets the set verification condition. 7.The model training verification system of claim 6, wherein, Further comprising the steps of: When the area size is less than or equal to the area threshold, the discrimination threshold is adjusted, and the new probability interval is formed again to re-enter the sample screening module.

8. An electronic device, comprising: The memory is configured to store a computer program, and the processor is configured to execute the computer program to realize the steps of the verification method of the icing prediction model according to any one of claims 1 to 4. The computer readable storage medium stores a computer program, and the computer program is executed by the processor to realize the steps of the verification method of the icing prediction model according to any one of claims 1 to 4. The computer program / instructions are executed by the processor to realize the steps of the verification method of the icing prediction model according to any one of claims 1 to 4.

9. A computer-readable storage medium, characterized in that, ​ 10. A computer program product comprising computer programs / instructions, characterized in that, ​

Images