A method for evaluating the effect of rocket-assisted rainfall enhancement operations

Abstract

The invention discloses a method for evaluating the effect of rocket rain enhancement operation, comprising: obtaining and cutting the operation cloud data of which the rainfall enhancement amount is to be predicted, and obtaining the rainfall enhancement evaluation interval of the operation cloud; searching the similar non-operation cloud submatrix of the operation cloud according to the operation screening condition of the rainfall enhancement evaluation interval of the operation cloud; applying a pre-constructed general similarity matching retrieval model to obtain a similar cloud similarity ranking list of the operation cloud in the similar non-operation cloud submatrix; tracking the cloud cluster trajectories of the second first calibrated number in the similar cloud similarity ranking list, and constructing a special natural rainfall model for the operation cloud; tracking the operation cloud trajectory of which the rainfall enhancement amount is to be predicted, and calculating the natural rainfall of the operation cloud under non-catalytic conditions by using the constructed special natural rainfall model; and then obtaining the rainfall of the operation cloud under catalytic conditions according to the trajectory, and evaluating the effect of rocket rain enhancement operation. The invention can effectively eliminate the influence of natural precipitation variability and improve the evaluation effect of rocket rain enhancement operation.

Description

Technical Field

[0001] The present invention belongs to the technical field of artificially impacted weather operation effect prediction and evaluation of new generation information technology, and particularly relates to a method for evaluating the effect of rocket rain-enhancing operation. Background Art

[0002] Artificial rainfall enhancement by rockets is an act of human intervention in the weather system. Artificial rainfall enhancement rockets spread catalysts in the operating clouds to cause water vapor in the clouds to condense into raindrops, thereby increasing precipitation. Artificial rainfall enhancement plays an important role in alleviating regional drought, improving agricultural production, regulating water resource allocation, and reducing natural disasters.

[0003] However, accurate assessment of the effects of artificial rainfall enhancement is an indispensable link in the sustainable and healthy development of the weather modification industry. It is of great significance for measuring the social and economic benefits of rocket artificial rainfall enhancement operations, testing whether the theories and methods of artificial weather modification are correct, and improving the scientific and technological level of artificial rainfall enhancement operations.

[0004] In traditional statistical tests, the effect evaluation conclusions of randomized experiments have always been generally accepted as more scientific results, but they require a large sample size and give up some operating opportunities, and the method itself cannot solve the problem of large variability in natural precipitation; non-randomized artificial rainfall effect evaluation methods include sequence analysis, covariate FCM and other methods. These methods are based on the correlation of precipitation and estimate natural precipitation under certain assumptions, but the variability of natural precipitation is huge, which affects the final evaluation effect. Summary of the invention

[0005] In order to address the deficiencies in the prior art, the present invention provides a method for evaluating the effect of rocket rainmaking operations, which utilizes multidimensional radar data and deep learning technology to eliminate the influence of natural precipitation variability, and uses an Arcface algorithm model trained with historical rainmaking operation cloud data to find non-operational clouds (comparison clouds) that are most similar to the operating clouds for which rainfall enhancement is to be predicted, and constructs a special natural rainfall model for the operating clouds for which rainfall enhancement is to be predicted. After predicting and calculating the natural rainfall of the operating clouds under non-catalytic conditions, the rocket rainmaking operation effect is evaluated based on the rainfall of the operating clouds under rocket rainmaking catalytic conditions, thereby improving the evaluation effect.

[0006] To achieve the above objective, a rocket rain-enhancing operation effect evaluation method according to an embodiment of the present invention, which includes:

[0007] S1. Obtain the work cloud data to predict the rainfall and crop it to obtain the work cloud rainfall evaluation range;

[0008] S2. According to the job screening conditions of the job cloud rain increase evaluation range, find the similar non-job cloud sub-matrix of the job cloud to predict the rain increase;

[0009] S3. Apply the pre-constructed general similarity matching search model to obtain the similarity similarity sorting list of job clouds in the similar non-job cloud submatrix;

[0010] S4. Track the cloud cluster trajectory of the first second calibration number in the similarity sorting list, and build a dedicated natural rainfall model for the operation cloud that needs to predict the increase in rainfall;

[0011] S5. Track the operation cloud trajectory that predicts the rainfall amount, and use the dedicated natural rainfall model built to calculate the natural rainfall amount of the operation cloud that predicts the rainfall amount under non-catalytic conditions.

[0012] Furthermore, the method further comprises:

[0013] S6. Track the trajectory of the operation cloud that predicts the increase in rainfall and obtain the rainfall of the operation cloud that predicts the increase in rainfall under catalytic conditions.

[0014] Further, the S1 , obtains the job cloud data to predict the rainfall amount and tries it to obtain the job cloud rainfall evaluation range, including:

[0015] S11. Obtain the three-dimensional puzzle data and rainfall data of the rain increase operation cloud area radar;

[0016] S12. The edges with the rocket operation point as the center, the calibration distance as the radius, and the operation azimuth angle are increased or decreased, and the operation azimuth angle are cut to obtain the operation sector area;

[0017] S13. Take an external rectangle parallel to the coordinate axis of the rectangular coordinate system;

[0018] S14. Adjust the external rectangle size to the training window size of the general similar match search model stored by the system;

[0019] S15. According to the adjusted sized external rectangle and calibrated height layer, the three-dimensional data area is cut out to obtain the operation cloud rain increase evaluation range.

[0020] Furthermore, the training window size is :

[0021] ;

[0022] in, is the total number of reasonable historical operation clouds, For the reasonable number of categories of different window sizes of historical job clouds, For the A reasonable number of historical job cloud categories, For the A reasonable size of the historical job cloud category window, For the A reasonable number of historical job cloud categories, For the A reasonable size of the historical job cloud category window, 1 to A positive integer, Indicates the rounding symbol.

[0023] Further, according to the job screening conditions of the job cloud to predict the rainfall, S2, finds a similar non-job cloud sub-matrix of the job cloud to predict the rainfall, including:

[0024] S21. Determine the operation screening conditions for the evaluation range of the operation cloud rain increase;

[0025] S22, according to the determined operational screening conditions of the operational cloud precipitation enhancement evaluation interval, searching for a non-operational cloud sub-matrix with the same operational screening conditions in cloud systems at other locations and historical non-operational cloud systems at the same time on the operational day;

[0026] S23. Reasonability evaluation and removal of the non-working cloud matrix is ​​carried out to obtain the reasonable non-working cloud matrix;

[0027] S24. The reasonable non-working cloud matrix is ​​filtered through the radar echo value to obtain a similar non-working cloud matrix;

[0028] Furthermore, the screening includes: respectively calculating the differences between the radar echo values ​​of the grid points of all non-operational clouds and the operational cloud whose rainfall increase is to be predicted in the similar non-operational cloud sub-matrix, taking the absolute values ​​and summing them; sorting the summation results, taking the absolute values ​​of the differences and the smallest first calibrated number samples as the similar non-operational cloud sub-matrix.

[0029] Further, the S3, applies a pre-constructed general similarity matching search model to obtain a similar cloud similarity sorting list of job clouds in the similar non-job cloud submatrix, including:

[0030] S31. Apply the pre-built job cloud and its general similarity matching search model for similar clouds, and extract the 512-dimensional feature vectors of the job cloud and the similar non-job cloud submatrix to be matched;

[0031] S32, calculating the Euclidean distance between the 512-dimensional feature vector of the operation cloud and the 512-dimensional feature vectors of all the non-operation clouds to be matched, and arranging the calculated Euclidean distance values ​​from small to large to obtain a similarity ranking list of similar clouds of the operation cloud.

[0032] Further, the construction process of the general similarity matching search model includes:

[0033] Set the input layer channel number of the ResNet network, the benchmark model of the Arcface algorithm that adapts to the open set classification problem, as the number of height layers in the historical job cloud rain increase evaluation interval, and load the pre-training parameters;

[0034] Freeze all parameters except the first layer, and use multiple historical rain-increasing job cloud data as model input, each of which is regarded as a class to train the input layer parameters of the model;

[0035] Random Gaussian noise is added to multiple historical rainfall operation cloud samples for sample augmentation. The saved model parameters are loaded, and the parameters of all layers of the model are trained using the sample augmented operation cloud data to obtain the trained general similarity matching retrieval model.

[0036] Further, the S4 tracks the cloud cluster trajectory of the first second calibration number in the similarity sorting list, and constructs a dedicated natural rainfall model for the operation cloud to predict the rainfall increase, including:

[0037] S41. Based on the improved optical flow method, track the cloud cluster trajectory of the first second calibration number in the similarity sorting list, and the tracking time is the front and back calibration time;

[0038] S42. Based on the tracked cloud location information of similar clouds, the radar echo data and rainfall data of the calibration time before and after the cloud group operation time are obtained.

[0039] S43. Build a dedicated natural rainfall model for the operation cloud that needs to predict the increase in rainfall.

[0040] Further, the dedicated natural rainfall model of the job cloud is a fusion 3D-Unet algorithm model based on attention mechanism.

[0041] The beneficial effects of the present invention are:

[0042] 1. Based on the fact that similar matching of operating clouds and non-operating clouds belongs to an open set classification problem, the present invention proposes to apply the Arcface algorithm adapted to the open set classification problem to the feature extraction of operating clouds and non-operating clouds through transfer learning, adapting to the problems in the field of artificial rainfall, and effectively improving the performance, accuracy and analysis efficiency of the model;

[0043] 2. In the preliminary preparation work for the evaluation, the present invention builds a general similarity matching retrieval model in advance, and uses the historical operation cloud data after the unified training window as the data set for training the general similarity matching retrieval model, so that the model can better understand and learn the characteristics of the rain-increasing cloud layer, so that the model can be directly applied in any subsequent formal evaluation work, which greatly improves the matching retrieval efficiency of the operation cloud to be predicted for rainfall and its similar clouds, thereby improving the efficiency of the entire evaluation work;

[0044] 3. The present invention improves the original optical flow method, which is mainly used for trend tracking of image motion. For cloud tracking, since the area where the operating cloud affects the rainfall is limited, the tracking of the original optical flow method may cause the operating area to diffuse and aggregate, thereby failing to accurately calculate the rainfall in the area affected by the cloud. To alleviate this problem, the tracking method of the original optical flow method is improved and adjusted, using single pixel tracking, that is, each point of the operating cloud is tracked separately, and the method with the most pixels in a fixed area is selected as the result of tracking at a single moment, which helps to fix the size of the cloud and reduce the impact of clouds in other areas on rainfall.

[0045] 4. The present invention proposes a short-term rainfall and radar echo prediction algorithm that supports small data set training, and a fused 3D-Unet algorithm model based on the attention mechanism; the 3D-Unet algorithm based on the attention mechanism is a Unet algorithm based on three-dimensional convolution, and an attention mechanism is added thereto; the rainfall and radar echo are respectively input into two independent 3D-Unet models based on the attention mechanism, and finally a convolution layer is used to fuse the inputs of the two independent models, so as to obtain a better final output prediction result.

[0046] 5. The present invention utilizes multi-dimensional radar data and deep learning technology to find the comparison cloud that is most similar to the catalyzed operating cloud, and then constructs a special natural rainfall model for the operating cloud. By tracking the trajectory of the operating cloud and its similar clouds, the natural rainfall of the operating cloud under non-catalytic conditions and the rainfall under catalytic conditions are calculated, and the effect of rocket rainfall enhancement operations is evaluated, which effectively eliminates the influence of natural precipitation variability and improves the evaluation effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 A flow diagram of a rocket rain-enhancing operation effect evaluation method of the present invention;

[0048] Figure 2 A schematic diagram of the operation sector area of ​​the operation cloud according to an embodiment of the present invention;

[0049] Figure 3 An external rectangular diagram of a working sector area according to an embodiment of the present invention;

[0050] Figure 4 A schematic diagram of the ArcFace algorithm for one embodiment of the present invention;

[0051] Figure 5 A flow chart of deep learning feature extraction and similar matching retrieval algorithm for one embodiment of the present invention;

[0052] Figure 6A 3D-Unet grid structure diagram based on attention mechanism for one embodiment of the present invention;

[0053] Figure 7 A 3D-Unet grid structure diagram based on attention mechanism for a fused attention mechanism of one embodiment of the present invention;

[0054] Figure 8 Comparative display diagram of natural rainfall prediction results of natural rainfall model of one embodiment of the present invention. DETAILED DESCRIPTION

[0055] In order to make the object, technical solution and advantages of the present invention clearer, the following will be further explained in conjunction with the accompanying drawings and examples.

[0056] It should be noted that, in this article, relational terms such as "first" and "second" are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0057] The method of the present invention is divided into two parts: the first part is the preliminary preparation work for the evaluation of the effect of the rocket rain-enhancing operation, and the second part is the subsequent formal evaluation work for the effect of the rocket rain-enhancing operation.

[0058] First, we will conduct the first part: preliminary preparations for the evaluation of the effect of the rocket rain-enhancing operation.

[0059] The most important thing in the preliminary preparation work is to build a general deep learning similarity matching retrieval model in advance. The constructed deep learning similarity matching retrieval model is used to greatly improve the matching retrieval efficiency and matching accuracy between the job cloud and its similar clouds in each subsequent formal evaluation.

[0060] The deep learning similarity matching search model is obtained through a large number of historical rain-enhancing job cloud data set training. The construction process of the deep learning similarity matching search model includes:

[0061] P1. Obtain the historical rain-enhancing job cloud dataset for training deep learning similar match search models, including:

[0062] P11. Obtain all historical rain-increasing operation cloud area radar three-dimensional puzzle data (the elements are radar echo values) and rainfall data;

[0063] P12. Preprocess the data time alignment of radar three-dimensional puzzle data and rainfall data;

[0064] Data time alignment; first detect the missing conditions of radar three-dimensional puzzle data and rainfall data separately to form a list of radar data missing and a list of missing rainfall data , take the union of these two missing lists to get the overall missing list and remove from the original radar 3D puzzle data and rainfall data The data at the corresponding moment is the data actually used in the future;

[0065] P13. Crop data for each historical rain-increasing operation cloud data, including:

[0066] Due to the limited range of action of rocket operation catalysts, compared with the previous large-area method of analyzing the entire large cloud, this method considers grasping the local area of ​​the operation to improve the accuracy of artificial rainmaking effect evaluation; further, considering the various possible movements of the cloud containing catalysts in the high altitude due to the influence of wind direction and wind speed, the operation point position is taken as the center of the circle, 10km as the radius, and the operation azimuth angle ±45º (the sides of the operation azimuth angle increase or decrease by 45º are the two sides) is used as the operation cloud system interval, the operation sector area of ​​each historical operation cloud is determined, and the coordinates of the sector vertex and the outermost point of the arc are recorded;

[0067] Since the unified coordinate system of the data is a rectangular coordinate system, different historical operation clouds have different orientations of the fan-shaped areas due to the different azimuths of rocket launches. For example, Guangdong Province carried out about a thousand rain-making operations between 2018 and 2022, and the operation sites were distributed in various meteorological stations in various cities. Assume that there are 500 operation sites, site 1 and site 2 correspond to different operation point locations (different longitudes and latitudes), and the operation azimuths of the rocket launches of the two are also different due to the operational requirements of the rainmaking operation; if site 1 has carried out multiple operations in the past five years, there are also inconsistencies in the operation azimuths at different times, which are not conducive to the analysis and comparison of subsequent data; and, in the process of rocket spreading catalysts, due to the effect of high-altitude winds, each operation cloud may also move at different speeds and directions. Considering the above two factors, the circumscribed rectangle of the operation sector area parallel to the coordinate axis is taken as the operation area; the specific method is: compare the coordinates of the operation point, the sector vertex and the outermost point of the arc horizontally and vertically, take the outermost coordinate values ​​and arrange and combine them, so as to determine the coordinates of the four vertices of the circumscribed rectangle of the operation sector area after adjustment; obtain the circumscribed rectangle of the historical operation sector area parallel to the coordinate axis of the rectangular coordinate system.

[0068] As attached Figure 3As shown, the figures show that when the rocket launches azimuth angle is 45 o When the result is an external rectangle A of the operation fan area of ​​the obtained operation cloud, and when the rocket launch azimuth angle is 15 o In the two cases, the external rectangle B in the operation fan area of ​​the obtained operation cloud. Of course, when the rocket launch azimuth angle is different, the obtained external rectangle is different.

[0069] After the external rectangle is determined, the data of each historical work cloud is cut according to the 2-3.5km height layer (according to a large amount of practical experience, the radar echo observation effect of this height layer is the best).

[0070] P14. Reasonable evaluation of the data of each cropped historical job cloud to obtain a reasonable job cloud list;

[0071] According to the rationality evaluation principle of ground rocket operations: "there is a radar echo at the altitude layer of 2-3.5km in the operation cloud system interval, the echo value is between 5 and 45dBZ, and the total number of echo values ​​that meet this condition is greater than or equal to 10", the rationality of the clipped historical operation cloud data is evaluated to obtain a reasonable historical operation cloud list;

[0072] P15. Unify the size of the circumscribed rectangle of the fan-shaped area of ​​all historical operation clouds in the reasonable historical operation cloud list. The size of this unified circumscribed rectangle is set as the training window size, and the training window size is stored in the system for subsequent evaluation.

[0073] Record the single-layer operation area size (also called window size) of each reasonable historical operation cloud in the reasonable historical operation cloud list, and count the number of operation clouds with different window sizes. The weight is the ratio of the number of operation clouds with this window size to the total number of reasonable historical operation clouds. The weighted window size obtained by multiplying the window size by the integer is the unified bounding rectangle size. This unified bounding rectangle size is defined as the training window size, and the training window size is stored in the system for subsequent evaluation.

[0074] The training window size is :

[0075] ;

[0076] in, is the total number of reasonable historical operation clouds, For the reasonable number of categories of different window sizes of historical job clouds, For the A reasonable number of historical job cloud categories, For the A reasonable size of the historical job cloud category window, For the A reasonable number of historical job cloud categories, For the A reasonable size of the historical job cloud category window, 1 to A positive integer, Indicates the rounding symbol.

[0077] Training window size After the determination, the bounding rectangle windows of different sizes of each historical operation cloud need to be unified. There are two cases: 1. For the bounding rectangle size larger than the training window size, the training window size is used to slide on the bounding rectangle of each historical operation cloud, and the window position with the largest intersection with the fan is selected; 2. For the bounding rectangle size smaller than the training window size, the original smaller one should be placed in the middle of the subsequent expansion window, and the expansion should be uniform in the left and right directions or in the top and bottom directions. For example, (16, 16) is smaller than (18, 16), so the previous coordinate direction is expanded by 1.

[0078] P16. Use all the obtained historical jobs cloud unified training window data after the data set trained by deep learning similar match search model.

[0079] Because the historical operational cloud dataset contains operational cloud rain-enhancing cloud layer characteristics under a variety of weather and climate conditions, these characteristics can provide the model with diverse learning materials, helping the model to learn more comprehensive rain-enhancing cloud layer characteristics; and by using rain-enhancing operational cloud data at different times, locations and weather conditions, the model can learn to identify and process different types of rain-enhancing cloud layer characteristics, thereby improving the generalization ability of the model and enabling it to better adapt to unseen rain-enhancing cloud layer data; at the same time, it can better understand the characteristics of rain-enhancing operational clouds, thereby providing more accurate results in subsequent prediction and classification tasks, improving the performance and accuracy of the model, and improving analysis efficiency.

[0080] P2. After transfer learning and training with data sets, the Arcface algorithm that adapts to the open set classification problem is obtained, which is specifically:

[0081] Based on the characteristics that similar matching of operating clouds and non-operating clouds belongs to the open set classification problem, it is proposed to apply the Arcface algorithm adapted to the open set classification problem to the feature extraction of operating clouds and non-operating clouds through transfer learning. The use of algorithms in other fields to adapt to the problems in the field of artificial rainfall is helpful to achieve cross-domain, open up research ideas and obtain more accurate results. The Arcface algorithm uses the ResNet network as the benchmark model, as shown in the attached figure. Figure 4 As shown, Arcface loss is used as a loss function, and its formula is as follows:

[0082] ;

[0083] The difference in research data means that the initial conditions are different. The difference in initial conditions is likely to show a huge difference in results after the deep learning model goes around in circles. Therefore, the present invention adopts a gradual migration strategy to adapt the data; first, the number of channels in the input layer is changed to the number of layers of the cloud cluster, the pre-trained parameters are loaded, all parameters except the first layer are frozen, and the job cloud that passes the rationality test is used as the model input. Each historical job cloud is regarded as a class, and the input layer parameters of the training model are trained;

[0084] Add random Gaussian noise to the job cloud samples that pass the rationality test ( =0.0065) to perform sample augmentation; load the model parameters saved in the above, use the Job Cloud data after sample augmentation to train the parameters of all layers of the model, and obtain the trained deep learning similarity matching retrieval model.

[0085] Then the second part is carried out: the formal evaluation of the subsequent rocket rain-increasing operation effect.

[0086] After completing the preliminary preparations for the first part of the rocket rain enhancement operation effect evaluation, the trained deep learning similarity matching retrieval model can be used to evaluate the effect of a rocket rain enhancement operation at any time, that is, a quantitative evaluation of the rainfall increase of a certain operation cloud.

[0087] As attached Figure 1 As shown in the present invention, the present invention provides a method for evaluating the effect of a rocket rain-enhancing operation, which includes:

[0088] S1. Obtain the data of the operation cloud to predict the rainfall and conduct the rocket rainfall operation effect evaluation and obtain the operation cloud rainfall evaluation range, including:

[0089] S11. Obtain the three-dimensional puzzle data of the radar in the operation cloud area to predict the rainfall (the elements of which are radar echo values) and rainfall data;

[0090] S12. Use the rocket operation point as the center, 10km as the radius, and the operation azimuth angle to increase and decrease by 45º as both sides, and cut the operation fan-shaped area, as attached Figure 2 As shown;

[0091] S13. Take an external rectangle parallel to the coordinate axis of the rectangular coordinate system;

[0092] S14. Adjust the external rectangle size to the training window size W of the deep learning similar matching search model stored by the system;

[0093] S15. According to the adjusted external rectangle and the height layer of 2-3.5km, cut out the three-dimensional data area to obtain the operation cloud rain increase evaluation range for predicting the rainfall increase.

[0094] S2. According to the job screening conditions of the job cloud rain increase evaluation range, find the job cloud similar non-job cloud sub-matrix to predict the rain increase; including:

[0095] S21. Determine the operation screening conditions for the evaluation range of the operation cloud rain increase;

[0096] The operation screening conditions for the operational cloud precipitation evaluation interval include window size, altitude, time and terrain; among them, the window size is the training window size W, the altitude is 2-3.5km, the time is the current operation time and the similar period before and after the same period in history, and the terrain is similar terrain, for example, the coastal plain terrain in both eastern and western Guangdong.

[0097] Under the condition of the same window size, if no detailed screening is done (from the time dimension (data year, day, time), altitude dimension and similar terrain dimension), taking the five-year historical data of Guangdong Province as an example, it is estimated that there are about 21 billion non-operational cloud sub-matrices to be retrieved. Too many sub-matrices to be retrieved will consume a lot of time and computing resources, resulting in a slower retrieval speed, which greatly affects the retrieval efficiency and evaluation process. Therefore, the following preliminary screening of non-operational cloud data is considered:

[0098] Time screening: Considering that the operating conditions vary in different seasons and months, the candidate non-operating cloud data sets are the current operating time (non-operating cloud systems at other locations at the same time on the operating day) and the similar periods before and after the same period in history;

[0099] Altitude screening: On the one hand, the altitude in the rationality evaluation standard of the operational cloud is limited to 2-3.5km. Only by keeping the altitude consistent can we find similar clouds to the operational cloud to a greater extent; on the other hand, the altitude distribution of the 21-layer radar echo data is not uniform. The distance between each layer of the 2-3.5km altitude layer is 0.5km, while the distances of other altitude layers are 0.5km, 1km, and 2km. The situation of the cloud layer also varies with the distance interval; comprehensively considered, the altitude layer of the fixed non-operational cloud is consistent with the operational cloud, both of which are 2-3.5km, and the radar echo observation effect of this altitude layer is also the best;

[0100] Terrain screening; the same cloud group will not usually have the same rainfall on different terrains. The influence of terrain on precipitation is multifaceted, including the following aspects: Rising motion: When moist air flows encounter high ground such as mountains, they are forced to rise. This rising motion causes the air to cool, and water vapor condenses into cloud droplets and raindrops, thereby increasing the probability of precipitation. Therefore, the rainfall is usually greater on the windward slopes of the mountains. Rain shadow effect: The leeward slopes of the mountains tend to have less rainfall because the rising air flow has released most of the moisture, forming the so-called "rain shadow area". Geometry and roughness of terrain: The geometry and roughness of terrain can affect the flow pattern of airflow, which in turn affects the development of clouds and the precipitation process. Local circulation: Specific terrain may cause local circulation, such as valley winds, sea and land breezes, etc. These local circulations can affect the distribution of precipitation. Therefore, even the same cloud group will have different rainfall when it moves over different terrains due to the influence of the above factors. For example, the same cloud group may only produce light rainfall in a flat area, but may cause heavy rainfall in a mountainous area. Similarly, the same cloud may bring substantial rainfall on windward slopes, while there may be little rainfall on leeward slopes. Therefore, the screening of similar terrains is equally important.

[0101] S22, according to the determined operational screening conditions of the operational cloud precipitation enhancement evaluation interval, searching for a non-operational cloud sub-matrix with the same operational screening conditions in cloud systems at other locations and historical non-operational cloud systems at the same time on the operational day;

[0102] S23. Reasonability evaluation and removal of the non-working cloud matrix is ​​carried out to obtain the reasonable non-working cloud matrix;

[0103] Then, according to the rationality evaluation principle of ground rocket operations - there is a radar echo at the 2-3.5km altitude layer within the operating cloud system range, the echo value is between 5 and 45dBZ, and the total number of echo values ​​that meet this condition is >= 10, the rationality of the sample cloud data in the non-operating cloud sub-matrix is ​​evaluated, and the non-operating cloud sub-matrix that meets the above rationality evaluation conditions is then eliminated according to the echo value elimination condition. If the elimination condition is met, the sub-matrix is ​​eliminated, and finally a reasonable non-operating cloud sub-matrix is ​​obtained.

[0104] The echo value rejection condition is:

[0105] (1) The maximum echo value is less than 5 dBZ, and the minimum echo value is greater than 45 dBZ. This is to exclude sub-matrices with echo values ​​that are too large or too small, and ensure that the radar echo value is within a reasonable range. Because the cloud clusters corresponding to excessively large echo values ​​have a high water content, they may represent extreme weather conditions such as heavy rainfall or thunderstorms. Continuing the rain enhancement operation will be counterproductive. Excessively small echo values ​​may represent noise or no precipitation, which is not suitable for operation. Therefore, by limiting the reasonable range of echo values, these abnormal conditions can be excluded and the accuracy of the data can be improved.

[0106] (2) The sum of the submatrix echo values ​​is less than the minimum value of the sum of the operating cloud submatrix echo values; set the lower limit of the submatrix echo value to ensure that the submatrix contains enough valid data; first sum the echo values ​​of all operating cloud submatrices and take the minimum value as the screening threshold; if the sum of the echo values ​​of the non-operating cloud submatrix to be selected is less than the screening threshold, it means that the echo condition of the submatrix is ​​outside the echo condition set of all historical operating clouds and can be excluded from the consideration range; this can help exclude those submatrices with smaller sums of echo values ​​to ensure that the submatrix contains enough valid data and improve the quality and availability of data;

[0107] (3) The number of grid points with echo values ​​in the reasonable interval [5, 45] is less than 30% of the total number of grid points; exclude those sub-matrices with uneven echo value distribution or concentrated near extreme values; by limiting the proportion of the number of grid points with echo values ​​in the reasonable interval to the total number of grid points, it can ensure that the echo values ​​in the sub-matrix are distributed more evenly, avoid the influence of extreme values ​​on the overall data, and help improve the stability and reliability of the data;

[0108] After the above comprehensive screening of time, altitude layer and sub-matrix radar echo value conditions, outliers can be effectively eliminated, data quality can be improved, and data reliability and stability can be ensured, so as to better perform data analysis and processing. The number of unscreened sub-matrices is more than 1,000 times the number of candidate non-operating cloud sub-matrices after the final screening, which is equivalent to a retrieval efficiency improvement of more than 1,000 times.

[0109] S24. The reasonable non-working cloud matrix is ​​filtered through the radar echo value to obtain a similar non-working cloud matrix;

[0110] The difference between the echo values ​​of all non-work clouds and the work cloud grid point radar in the reasonable non-work cloud matrix is ​​calculated separately, and the absolute value is taken and summed;

[0111] For the summing results, take the absolute value of the difference and the smallest first 10,000 samples as similar non-job cloud matrix.

[0112] S3. Through the trained deep learning similarity matching retrieval model, the similarity similarity sorting list of the job clouds in the similar non-job cloud submatrix is ​​obtained, as attached Figure 5 As shown, including:

[0113] S31. Through the trained deep learning similar matching search model, the 512-dimensional feature vectors of the non-job cloud to be matched in the job cloud and the similar non-job cloud submatrix are extracted respectively;

[0114] The trained model extracts the features of the job cloud and the non-job cloud to be matched, and the final output of the model is the 512-dimensional feature vector of the job cloud and the non-job cloud to be matched;

[0115] S32, calculating the Euclidean distance between the 512-dimensional feature vector of the operation cloud and the 512-dimensional feature vectors of all the non-operation clouds to be matched, and arranging the calculated Euclidean distance values ​​from small to large to obtain a similarity ranking list of similar clouds of the operation cloud.

[0116] Compute the Euclidean distance between the job cloud feature vector and the non-job cloud feature vector to be matched:

[0117] ;

[0118] in, is the feature vector of the job cloud, is the feature vector of the non-operation cloud to be matched, Corresponding to the job cloud and the non-job cloud feature vector to be matched respectively ; the smaller the distance value indicates the higher the similarity between the job cloud and the non-job cloud to be matched; the similarity similarity sorting list of the job clouds can be obtained according to the European distance value.

[0119] S4. Track the cloud cluster trajectories of the top 100 in the similarity ranking list, and build a dedicated natural rainfall model for the operation clouds that want to predict the increase in rainfall; including:

[0120] S41. Based on the improved optical flow method, track the cloud cluster trajectories of the top 100 in the similarity sorting list, and the tracking time is one hour before and after;

[0121] Improved optical flow method: The original optical flow method is mainly used for trend tracking of image motion. For cloud tracking, because the area where the operating cloud affects the rainfall is limited, the tracking of the original optical flow method may cause the operating area to diffuse and aggregate, making it impossible to accurately calculate the rainfall in the area affected by the cloud. To alleviate this problem, the tracking method of the original optical flow method is improved and adjusted to adopt single-pixel tracking, that is, each point of the operating cloud is tracked separately, and the method with the most pixels in a fixed area is selected as the result of tracking at a single moment, which helps to fix the cloud size and reduce the impact of clouds in other areas on rainfall.

[0122] Based on the improved optical flow method, track the motion trajectories of the top 100 similar clouds that are most similar to the job cloud in the similarity sorting list of job clouds, with a tracking time of one hour before and after;

[0123] S42. Based on the cloud location information similar to the clouds obtained by tracking, the radar echo data and rainfall data with resolution of one hour and six minutes before and after the cloud operations moment are obtained.

[0124] S43. Establish a dedicated natural rainfall model for the operation cloud, including:

[0125] As attached Figure 6-7 As shown in the figure, a short-term rainfall and radar echo prediction algorithm supporting small data set training is proposed, and a fused 3D-Unet algorithm model based on attention mechanism is proposed; the 3D-Unet algorithm based on attention mechanism is a Unet algorithm based on three-dimensional convolution, and an attention mechanism is added to it; rainfall and radar echo are respectively input into two independent 3D-Unet models based on attention mechanism, and finally a convolution layer is used to fuse the inputs of the two independent models to obtain the final output;

[0126] The radar echo data and rainfall data of 11 times, including the hour before the similar cloud operation time and the operation time, are taken as input factors. The output result of the model is the grid rainfall prediction value one hour after the operation time. A special natural rainfall model for the operation cloud based on the fused 3D-Unet algorithm based on the attention mechanism is established.

[0127] S5. Track the operation cloud trajectory that predicts the rainfall increase, and use the built dedicated natural rainfall model to calculate the natural rainfall of the operation cloud under non-catalytic conditions.

[0128] Based on the improved optical flow method, the trajectory of the operation cloud one hour before the operation time plus the operation time is tracked. The radar echo data and rainfall data of 11 times in total, including one hour before the operation time of the operation cloud and the operation time, are used as input factors. The constructed special natural rainfall model is used. The output result of the model is the grid point natural rainfall prediction value one hour after the operation time. The natural rainfall of the operation cloud one hour after the operation time predicted by the special natural rainfall model is recorded as .

[0129] S6. Track the operation cloud trajectory that predicts the rainfall of the operation cloud under catalytic conditions, and conduct an evaluation of the rocket rainfall effect.

[0130] Based on the improved optical flow method, the movement trajectory of the operating cloud one hour after the operating time is tracked to obtain the rainfall of the operating cloud under catalytic conditions. The rainfall of the cloud one hour after the operating time is recorded as the rainfall after the operation. ;

[0131] The artificial rain rate calculation formula can be used to quantitatively evaluate the rain increase effect of each operation, and conduct rocket rain increase operation effect evaluation, including:

[0132] Artificial rainfall Effect evaluation: ;

[0133] Artificial rainfall rate Effect evaluation: ;

[0134] in, For the rainfall of the operation cloud one hour after the operation time under catalytic conditions, For the work cloud, the natural rainfall is one hour after the operation time under non-catalytic conditions.

[0135] Figure 8 The figure shows a comparison chart of the prediction results of the natural rainfall model in the present invention. There are four columns from left to right, and each figure corresponds to the first moment after the operation time to the tenth moment from top to bottom and from left to right. The first and third columns correspond to the true values, and the second and fourth columns correspond to the predicted value results of the model. At the same time, the grid average rainfall value corresponding to a certain moment has been marked. From the four columns, the rainfall distribution of the model prediction results is close to the true value, and the model result captures the rainfall position in the lower right corner. From the numerical point of view, the numerical difference between the predicted value and the true value at most times does not exceed 0.01mm; at the same time, the true value result as a whole shows a trend of gradually decreasing rainfall but with small fluctuations in the middle, which is also the change form in the model prediction results. It shows that the natural rainfall model of the fused 3D-Unet structure based on the attention mechanism can not only grasp the rainfall distribution, but also better approach the numerical size and change trend of the true value.

[0136] The rainfall results predicted by the fused 3D-Unet-based attention mechanism-based natural rainfall model are used as the natural rainfall of the operating cloud under non-catalytic conditions, and are used together with the rainfall results obtained by tracking the operating cloud through the optical flow method to evaluate the artificial rainfall effect of the operating cloud to predict the rainfall increase.

[0137] In general, the natural rainfall model of the fused 3D-Unet structure based on the attention mechanism has a high degree of grasp of the distribution trend and numerical situation of rainfall. The results of using this model can better represent the natural rainfall situation of the operating cloud.

[0138] The quantitative artificial rainmaking effect evaluation method based on the deep learning model provided by the present invention is suitable for quantitative rainmaking effect evaluation of artificial rainmaking operations. By using deep learning feature extraction and similar matching retrieval algorithms to find historical non-operation clouds with weather conditions similar to those of artificial rainmaking operations, it is possible to avoid the problem of difficulty in quantitatively evaluating the artificial rainmaking effect due to subjective selection of the operation area and the comparison area of ​​the artificial rainmaking operation or the instability of the two areas; and by introducing an improved optical flow method to dynamically track cloud cluster trajectories and establish a fused 3D-Unet-based natural rainfall model based on the attention mechanism, a more accurate artificial rainmaking rate can be obtained.

[0139] Optionally, in this embodiment, a person of ordinary skill in the art may understand that all or part of the steps in the various methods of the above embodiments may be completed by instructing hardware related to the terminal device through a program, and the program may be stored in a computer-readable storage medium, and the storage medium may include: a flash drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, etc.

[0140] If the integrated units in the above embodiments are implemented in the form of software functional units and sold or used as independent products, they can be stored in the above computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for enabling one or more computer devices (which can be personal computers, servers or network devices, etc.) to execute all or part of the steps of the methods of various embodiments of the present invention.

[0141] In the several embodiments provided in the present application, it should be understood that the disclosed client can be implemented in other ways. Among them, the device embodiments described above are only schematic, for example, the division of units is only a logical function division, and there may be other division methods in actual implementation, for example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, and the indirect coupling or communication connection of units or units can be electrical or other forms.

[0142] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0143] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0144] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

[0145] The present invention and its embodiments are described above, and such description is not restrictive. The drawings show only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if ordinary technicians in the field are inspired by it, without departing from the purpose of the invention, they can design a structure and embodiment similar to the technical solution without creativity, which should belong to the protection scope of the present invention.

Claims

1. A method for evaluating the effect of rocket rainfall enhancement operations, characterized in that: It includes: S1. Obtain the operational cloud data for the rainfall increase to be predicted and cut it to obtain the operational cloud rainfall increase evaluation interval, including: S11. Obtain radar three-dimensional mosaic data and rainfall data of the cloud area for rain enhancement operation; S12, with the rocket operation point as the center, the calibration distance as the radius, and the increase and decrease calibration angles of the operation azimuth as the two sides, cutting out the operation sector area; S13, taking a circumscribed rectangle of the operation sector area parallel to the coordinate axes of the rectangular coordinate system; S14, adjusting the size of the circumscribed rectangle to the training window size of the general similarity matching retrieval model stored in the system; S15, cutting out the three-dimensional data area according to the circumscribed rectangle after the size adjustment and the calibrated height layer, and obtaining the operational cloud precipitation enhancement evaluation interval; S2. according to the operation screening conditions of the operation cloud rainfall enhancement evaluation interval, find the similar non-operation cloud sub-matrix of the operation cloud for which rainfall enhancement is to be predicted; The operation screening conditions of the operation cloud precipitation enhancement evaluation interval include window size, altitude layer, time and terrain; S3. Apply the pre-built general similarity matching retrieval model to obtain a similar cloud similarity ranking list of operational clouds for which rainfall increase is to be predicted in the similar non-operational cloud sub-matrix, including: S31, applying a pre-built general similarity matching retrieval model of an operating cloud and its similar clouds, respectively extracting feature vectors of the operating cloud for which rainfall increase is to be predicted and the non-operating cloud to be matched in the sub-matrix of similar non-operating clouds; The construction process of the general similarity matching retrieval model includes: The number of input layer channels of the ResNet network, a benchmark model of the Arcface algorithm adapted to the open set classification problem, is set to the number of height layers of the historical cloud rainfall enhancement evaluation interval, Arcface loss is used as the loss function, and the pre-training parameters are loaded; Freeze all parameters except the first layer, use multiple historical rainfall enhancement cloud data as model input, treat each historical rainfall enhancement cloud as a class, and train the model's input layer parameters; Add random Gaussian noise to multiple historical rain enhancement operation cloud samples for sample augmentation, load the saved model parameters, use the sample augmented operation cloud data to train the parameters of all layers of the model, and obtain the trained general similarity matching retrieval model; S32, calculating the Euclidean distance between the feature vector of the operational cloud for which the rainfall increase is to be predicted and the feature vectors of all the non-operational clouds to be matched, and arranging the calculated Euclidean distance values ​​from small to large to obtain a similarity ranked list of similar clouds of the operational cloud for which the rainfall increase is to be predicted; S4, tracking the cloud cluster trajectories of the second most marked number in the similar cloud similarity ranking list, and constructing a special natural rainfall model for the operational cloud for which rainfall increase is to be predicted, including: S41, tracking the trajectory of the second most calibrated cloud cluster in the similar cloud similarity sorting list based on the improved optical flow method, and the tracking time is the time before and after calibration; The improved optical flow method is: the tracking method of the original optical flow method is improved and adjusted, and single pixel tracking is adopted, that is, each point of the operating cloud is tracked separately, and the method with the most pixels in a fixed area is selected as the result of tracking at a single moment, which helps to fix the size of the cloud and reduce the impact of clouds in other areas on rainfall; S42, obtaining radar echo data and rainfall data of the calibration duration before and after the cloud operation time according to the cloud cluster position information of the similar cloud obtained by tracking; S43, constructing a special natural rainfall model for the operational cloud for which rainfall increase is to be predicted; The dedicated natural rainfall model of the Job Cloud is a fused 3D-Unet algorithm model based on the attention mechanism. The rainfall and radar echo are input into two independent 3D-Unet models based on the attention mechanism respectively. Finally, a convolutional layer is used to fuse the inputs of the two independent models to obtain the final output. S5. Track the trajectory of the operational cloud whose rainfall increase is to be predicted, and use the constructed special natural rainfall model to calculate the natural rainfall of the operational cloud under non-catalytic conditions.

2. The rocket rainfall enhancement operation effect evaluation method according to claim 1 is characterized in that: The method further comprises: S6. Track the trajectory of the operational cloud whose rainfall is to be predicted, obtain the rainfall of the operational cloud under catalytic conditions, and evaluate the effectiveness of the rocket rainfall enhancement operation.

3. The rocket rainfall enhancement operation effect evaluation method according to claim 1 is characterized in that: The training window size is : ； in, is the total number of reasonable historical operation clouds, To be reasonable, the number of categories with different window sizes for historical job clouds is For the A reasonable number of historical job cloud categories, For the A reasonable size of the historical job cloud category window, For the A reasonable number of historical job cloud categories, For the A reasonable size of the historical job cloud category window, 1 to A positive integer, Indicates the rounding symbol.

4. The rocket rainfall enhancement operation effect evaluation method according to claim 1 is characterized in that: The step S2, searching for a similar non-operational cloud sub-matrix of the operation cloud for which the rainfall increase amount is to be predicted according to the operation screening condition of the operation cloud rainfall increase evaluation interval, comprises: S21, determining the operation screening conditions of the operation cloud precipitation enhancement evaluation interval; S22, according to the determined operational screening conditions of the operational cloud precipitation enhancement evaluation interval, searching for a non-operational cloud sub-matrix with the same operational screening conditions in cloud systems at other locations and historical non-operational cloud systems at the same time on the operational day; S23, evaluating the rationality of the non-operation cloud sub-matrix and eliminating it to obtain a reasonable non-operation cloud sub-matrix; S24. Filter the reasonable non-operational cloud sub-matrix by radar echo values ​​to obtain a similar non-operational cloud sub-matrix.

5. The rocket rainfall enhancement operation effect evaluation method according to claim 4 is characterized in that: The screening process in S24 includes: Calculate the difference between the radar echo values ​​of all non-operational clouds in the similar non-operational cloud sub-matrix and the operational cloud grid points for predicting rainfall, take the absolute value and sum them; Sort the summation results, take the absolute value of the difference and the smallest first calibration number samples as the similar non-operation cloud sub-matrix.

Images