An artificial cloud seeding effect set evaluation method based on radar echo intelligent extrapolation

Abstract

This invention discloses an ensemble evaluation method for artificial cloud seeding effects based on intelligent extrapolation of radar echoes. The method includes acquiring sample radar data, selecting multiple intelligent extrapolation models, training them using preprocessed data, and constructing a model library; constructing an EES ensemble evaluation algorithm, calling the model library to infer the echo results at the time before the operation, and scoring the model performance; selecting high-scoring models to predict the echoes during the operation period, fusing and correcting them to obtain the ensemble prediction results; comparing the ensemble predictions with real-time monitoring data, extracting key physical parameter change characteristics, and quantitatively evaluating the operation effect. This invention solves the heterogeneous adaptation problem of different intelligent extrapolation models for cloud seeding effect evaluation, improves extrapolation accuracy by combining the advantages of multiple models, enhances the accuracy and reliability of the verification results, and provides technical support for the large-scale application of intelligent extrapolation technology in effect evaluation.

Description

Technical Field

[0001] This invention relates to the field of atmospheric cloud seeding effect ensemble evaluation technology, specifically a method for ensemble evaluation of artificial cloud seeding effect based on radar echo intelligent extrapolation. Background Technology

[0002] Verifying the effectiveness of weather modification operations is a crucial aspect of related scientific research and operational work, significantly contributing to the objective evaluation of operational benefits, improvement of technical capabilities, and advancement of cloud precipitation physics theory. Establishing a scientific, standardized, and quantifiable method for effectiveness evaluation has always been a core issue urgently needing resolution in cloud seeding operations. Existing evaluation methods mainly include statistical testing, physical testing, and numerical simulation. However, these methods often require idealization and simplification of the cloud precipitation process in practical applications, limiting their quantitative evaluation accuracy in complex real-world scenarios. In recent years, deep learning-based intelligent extrapolation technology using radar echoes has provided a new technical approach for evaluating operational effectiveness. This method constructs a high-precision radar echo extrapolation model to predict echo sequences under natural evolution and compares them with actual observation data, thereby quantitatively identifying and separating the impacts of weather modification operations. However, the reliability of this type of evaluation method highly depends on the accuracy and stability of the extrapolation model itself. Currently, various deep learning model structures have been developed in the field of radar echo extrapolation. Different models have their own characteristics in characterizing cloud precipitation dynamics and microphysical processes, but significant performance differences and prediction uncertainties also exist. The heterogeneity of such models leads to insufficient credibility and poor universality of the evaluation conclusions obtained based on a single deterministic extrapolation model, making it difficult to form a unified and reliable evaluation standard in actual operations. This limits the large-scale application of radar echo intelligent extrapolation technology in the evaluation of the effects of artificial weather modification.

[0003] CN116610959A discloses a method and system for verifying the effectiveness of cloud seeding operations based on UNET-GRU deep learning. This technical solution introduces a deep learning model to achieve quantitative prediction and evaluation of rainfall and radar echoes in the operation's impact area. However, this method has the following limitations: First, its evaluation results heavily rely on the structure and performance of the single UNET-GRU model used, failing to cover the uncertainties introduced by differences in model structure and performance, resulting in insufficient operational applicability and statistical robustness of the evaluation conclusions. Second, this method only uses rainfall and four basic parameters of Doppler weather radar as input factors, and only analyzes through residual value comparison, without conducting further physical feature extraction and spatiotemporal evolution analysis on the model extrapolation results, resulting in insufficient information mining depth and limited evaluation precision. CN114661700A discloses an AI-based method for verifying the effectiveness of artificial weather modification operations. This technical solution uses an artificial intelligence model to predict natural weather processes, but it is essentially still a physical verification method for specific convective cloud cases. The evaluation process relies heavily on subjective selection of convective cloud processes and the selection of typical cases, lacking a systematic verification framework. Furthermore, the AI ​​model used is a single deterministic architecture, failing to consider the performance differences of different model structures in representing cloud physical processes, resulting in a lack of statistical robustness in the evaluation results.

[0004] The above-mentioned technical solutions have not yet effectively solved the key problems in the quantitative evaluation of the effects of artificial weather modification operations based on artificial intelligence models. Although the currently used intelligent extrapolation methods have made some improvements in technical approach, they still generally suffer from limitations such as failing to fully consider the performance differences of different models and lacking a systematic approach to handling forecast uncertainties. In addition, existing methods either rely on subjective case selection or are limited by the performance of a single model, and have not yet established a universal evaluation method applicable to diverse operational scenarios, nor have they formed a unified and reliable evaluation standard. At the same time, these methods do not fully consider the problem of increased cloud seeding effect evaluation errors caused by excessive smoothing of forecast results during long-term extrapolation, thus limiting the evaluation effectiveness and application potential of intelligent extrapolation technology in actual weather modification operations. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for ensemble evaluation of the effects of artificial cloud seeding based on intelligent extrapolation of radar echoes. The method constructs an ensemble evaluation algorithm based on an intelligent extrapolation model of radar echoes to achieve dynamic optimization of the deep learning extrapolation model; it improves the accuracy and robustness of radar echo extrapolation through the fusion and correction of the optimized model; and it achieves accurate quantitative evaluation of the effects of artificial cloud seeding operations by extracting key physical parameters within the operational influence area, thereby effectively solving the evaluation adaptability problem caused by the performance differences of different intelligent extrapolation models of radar echoes.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0007] This invention includes the following steps:

[0008] S1: Acquire sample Doppler weather radar data, and construct a training dataset after preprocessing the radar data;

[0009] S2: Train the training dataset based on the radar echo intelligent extrapolation model;

[0010] S3: Construct an ensemble evaluation algorithm, which is based on at least one evaluation benchmark factor and multiple error penalty terms; call each extrapolation model in the model library to infer the extrapolation results of the natural state radar echo of the target area during a preset period before the start of artificial cloud seeding operations; use the ensemble evaluation algorithm to score and rank the extrapolation results of each model.

[0011] S4: Select the top-scoring models as preferred models, extrapolate the radar echo evolution process of the target area under natural conditions after the start of the operation, and fuse and correct the deviation of the output results of each preferred model to generate the ensemble extrapolation prediction result.

[0012] S5: Compare the extrapolated prediction results of the set with the measured radar echo observation data of the same period, extract the change characteristics of key physical parameters in the operation's influence area, and quantitatively evaluate the actual effect of the artificial cloud seeding operation based on these change characteristics.

[0013] Further, step S1 includes:

[0014] S1.1: Based on the sample Doppler weather radar data, filter the data in each consecutive multi-frame sequence that contains at least one frame with a precipitation area coverage exceeding a set threshold, and select data with relatively isolated convective cells as individual case data.

[0015] S1.2: Perform quality control on the individual case data obtained through screening;

[0016] S1.3: Preprocess the individual case data into a standardized training dataset with uniform spatial dimensions and numerical range.

[0017] Furthermore, in step S2, the radar echo intelligent extrapolation model is a time-series prediction model based on deep learning; the selected model is trained on a standardized training dataset, and the data of the previous historical period is used as the model input, and the data of the next future period is used as the model validation, so as to obtain multiple pre-trained extrapolation models, which constitute the model library.

[0018] Further, in step S3, constructing the set evaluation algorithm includes:

[0019] S3.1: Select multiple objective evaluation indicators and group them based on correlation analysis and clustering methods;

[0020] S3.2: Based on the objective evaluation index, perform a comprehensive performance ranking of each extrapolation model, determine at least one optimal index, and determine the evaluation benchmark factor based on the optimal index;

[0021] S3.3: Based on the evaluation benchmark factor, construct multiple error penalty terms; the error penalty terms include at least two of the following: a spatially weighted error term to enhance the sensitivity of prediction errors to key areas, a continuity constraint term to suppress non-physical mutations, and an asymmetric prediction penalty term that sets differentiated weights for over-prediction and under-prediction;

[0022] S3.4: Establish a time error accumulator for each error penalty term and process it using a time-series smoothing model;

[0023] S3.5: Combining the evaluation benchmark factor and each error penalty term after time-series smoothing, construct the set evaluation algorithm, wherein the output value of the set evaluation algorithm is a ratio function of the evaluation benchmark factor and the sum of each error penalty term.

[0024] Furthermore, the spatial weighted error term is obtained by spatially modulating the mean absolute error by fusing at least one of the gradient features, Gaussian texture features, and multi-scale contrast features of the image; the continuity constraint term is obtained by calculating the difference between the predicted result and the true result on the second derivative; and the asymmetric prediction penalty term is obtained by introducing an intensity weighting function and assigning different penalty coefficients to over-prediction and under-prediction.

[0025] Further, in step S4, at least three models with the highest scores are selected as the preferred models, and a weighted fusion method based on the scores of the set evaluation algorithm is used for fusion processing.

[0026] Further, in step S5, the key physical parameters include at least one of echo intensity, echo area, vertical cumulative liquid water content, precipitation intensity, and quantitative precipitation estimation; the quantitative assessment of the actual effect based on the change characteristics includes calculating the echo intensity residual value sequence and analyzing the temporal changes of the regional average value of each physical quantity and the area ratio of different intensity intervals within the affected area.

[0027] On the other hand, a set evaluation system for the effect of artificial cloud seeding based on intelligent extrapolation of radar echoes, used to implement the aforementioned set evaluation method for the effect of artificial cloud seeding based on intelligent extrapolation of radar echoes, includes:

[0028] The data preprocessing module (100) acquires sample Doppler weather radar data and performs case screening, quality control and preprocessing on the data;

[0029] The model library management module (200) calls the model library trained using a standardized Doppler weather radar case dataset;

[0030] The algorithm scoring module (300) combines the radar echo data of the spatiotemporal region of the target to be evaluated with the input of the operation information, calls each extrapolation model in the model library to infer the extrapolation result of the target region under the natural state 1 hour before the start of the operation, and scores the performance of the extrapolation result of each model based on the EES set evaluation algorithm.

[0031] The ensemble correction module (400) selects multiple models with the highest scores, uses a weighted fusion method based on EES algorithm indicators to fuse the output results of the selected models, corrects the bias of the fusion results, and generates the final ensemble extrapolation prediction results for effect evaluation.

[0032] The effect verification module (500) compares the extrapolation prediction results of the set with the extrapolation data of the measured radar echo during the same period. Based on the obtained residual values, the time-series changes of physical quantities and the evolution characteristics of intensity distribution, it comprehensively and quantitatively evaluates the actual effect of artificial cloud seeding operations.

[0033] In summary, this invention provides a method and system for evaluating the ensemble effect of artificial cloud seeding based on intelligent extrapolation of radar echoes.

[0034] Compared with existing single-model-based performance evaluation methods, this invention constructs ensemble extrapolation predictions through dynamic optimization, fusion, and correction, effectively improving case adaptability and the physical rationality of prediction results, and significantly enhancing the method's generalization ability. Furthermore, by comparing ensemble extrapolation with measured data, and integrating residual analysis, temporal changes in physical parameters, and intensity distribution characteristics, it achieves accurate quantitative evaluation of artificial cloud seeding effects. This invention solves the adaptability problem arising from the performance differences of different radar echo intelligent extrapolation models in artificial cloud seeding effect evaluation. By integrating the advantages of various models, it improves the accuracy and physical consistency of extrapolation predictions, thereby significantly enhancing the accuracy and reliability of artificial cloud seeding effect verification. Attached Figure Description

[0035] Figure 1 This is a reference diagram of a method for verifying the effectiveness of cloud seeding operations according to an embodiment of the present invention;

[0036] Figure 2 This is a reference diagram of a method for verifying the effectiveness of cloud seeding operations according to an embodiment of the present invention;

[0037] Figure 3This is a block diagram illustrating the implementation of the EES ensemble evaluation algorithm based on the radar echo intelligent extrapolation model in this embodiment of the invention.

[0038] Figure 4 This is a schematic diagram of the structural composition of the artificial cloud seeding effect verification system according to an embodiment of the present invention. Detailed Implementation

[0039] To make the objectives, technical solutions, and advantages of this invention clearer, the embodiments of this invention will be described in detail below with reference to the accompanying drawings and examples. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without inventive effort are within the scope of protection of this invention.

[0040] like Figures 1 to 4 As shown, this invention provides a method and system for evaluating the ensemble effect of artificial cloud seeding based on intelligent extrapolation of radar echoes. The method is implemented through the system and specifically includes the following steps:

[0041] S1: Acquire sample Doppler weather radar data, and perform case screening, quality control and preprocessing on the data;

[0042] S1.1: Based on the sample Doppler weather radar data, filter each sequence of 40 consecutive frames to include at least 2 frames of data with precipitation area coverage exceeding 5% of the radar scanning range, and further select relatively isolated convective cells as individual data.

[0043] S1.2: For sample Doppler weather radar data with a resolution of 6 minutes, perform quality control on the individual case data obtained after screening;

[0044] S1.3: Preprocess the original Doppler weather radar data into a standardized training dataset; the original data covers an area of ​​230km×230km and has a spatial resolution of 256×256 pixels; during the dataset construction process, all training set images are uniformly sampled to 64×64 pixels, and the combined reflectivity values ​​are normalized to convert them to a pixel value range of 0–255.

[0045] S1.4: Based on the processing in steps S1.1 to S1.3, a standardized Doppler weather radar case dataset is obtained.

[0046] S2: Select multiple intelligent extrapolation models of radar echoes, train each model using preprocessed radar data, and build a trained model library.

[0047] S2.1: Select a variety of radar echo intelligent extrapolation models, including but not limited to ConvLSTM, PredRNN, PhyDNet, SimVP, SwinLSTM, TAU, E3D-LSTM, Mau, MIM, PredRNN v2 and REE-TTT;

[0048] S2.2: Train the selected extrapolation model on the standardized Doppler weather radar case dataset, and track the echo evolution characteristics of no less than 30 consecutive time steps during the training process;

[0049] S2.3: Obtain the extrapolation sequence of each model for each training instance; where the total duration of each instance sequence is 4 hours, the first 20 frames are used as model input, and the last 20 frames are used as model validation;

[0050] S2.4: Based on the processing in steps S2.1 to S2.3, obtain multiple extrapolation models pre-trained on the standardized Doppler weather radar case dataset, which constitute the model library.

[0051] S3: Construct an EES ensemble evaluation algorithm based on a radar echo intelligent extrapolation model; call the model library to infer the radar echo extrapolation results of the target area in its natural state at the moment before operation; and score the performance of the extrapolation results of each model based on the ensemble evaluation algorithm.

[0052] S3.1: Select multiple objective evaluation indicators in the field of radar echo extrapolation, including but not limited to CSI, POD, FAR, ETS, HSS, NCSI, PSD, ACC, Precision, Recall, F1-score, FSS, SSIM, CRPS, MAE, MSE, PSNR, LPIPS, PCC, GDL, FVD, etc.

[0053] S3.2: The Kendall rank correlation coefficient is used to perform correlation analysis on the objective evaluation indicators, and clustering is performed based on the analysis results;

[0054] S3.3: Using the aforementioned objective evaluation indicators as the classification basis, the CRITIC-Borda weighted counting method is used to rank the overall performance of each extrapolation model in terms of effectiveness;

[0055] S3.4: The coefficient of variation is used to quantify the fluctuation characteristics of each evaluation index in different models. The coefficient of variation is defined as the ratio of the standard deviation to the mean of the corresponding index.

[0056] S3.5: Determine the optimal index based on the effectiveness ranking and locate its corresponding cluster group; select the index with the largest coefficient of dispersion from the group as the evaluation benchmark factor;

[0057] S3.6: Based on the aforementioned benchmark factor, construct a spatially weighted error term ( This is used to enhance the sensitivity to prediction errors in key catalytic regions; Item fusion Gaussian texture features ( ) and multi-scale contrast features ( This is achieved through spatial modulation of MAE, and its calculation formula is as follows: Among them, the characteristic coefficients α, β, and γ are optimized by grid search to take values ​​of 0.2, 0.5, and 0.3, respectively;

[0058] S3.7: Based on the aforementioned benchmark factor, construct a continuity constraint term ( By introducing a second-order continuity penalty based on the Navier-Stokes equation prior, non-physical abrupt changes and block artifacts in the prediction results are suppressed. The expression is as follows: ;

[0059] S3.8: Based on the aforementioned benchmark factor, construct an asymmetric prediction penalty term ( Differential penalty weights are set for over-prediction and under-prediction based on the echo intensity range, and an intensity weighting function is introduced. To enhance sensitivity to errors in areas with strong echoes, its expression is: ;in, and These are the penalty coefficients for over-prediction and under-prediction, respectively, which are set to 10 and 1 after optimization. Defined as ;

[0060] S3.9: Establish independent time error accumulators for each error penalty term, and use an exponential decay model for time series smoothing. The update formula is as follows: The attenuation factor η is 0.85.

[0061] S3.10: Combining the above, the EES ensemble evaluation algorithm based on the radar echo intelligent extrapolation model is constructed, and its expression is: ;

[0062] S3.11: Combining the operation information, call the extrapolation models in the model library, input the radar echo data of the spatiotemporal region of the target to be evaluated, and respectively infer the extrapolation results of the target region under its natural state 1 hour before the start of the operation;

[0063] S3.12: Based on the EES set evaluation algorithm, perform performance scoring on the extrapolation results obtained by each model 1 hour before the start of the operation;

[0064] S3.13: Based on the processing results of steps S3.1 to S3.12, determine the extrapolation performance ranking of each model 1 hour before the start of the job.

[0065] S4: Select several models with the highest scores, extrapolate the radar echo evolution process of the target area under the natural state after the operation, and fuse and correct the output results of the multiple preferred models to generate ensemble extrapolation prediction results.

[0066] S4.1: Select at least three top-scoring models from the first hour of the operation to extrapolate the radar echo evolution of the target area under natural conditions 2 hours after the start of the operation.

[0067] S4.2: A weighted fusion method based on EES algorithm indicators is used to fuse the output results of the multiple preferred models;

[0068] S4.3: Calculate the average deviation between the ensemble prediction results and the actual observation data within the first 12 minutes after the start of the operation, and systematically correct the fusion results based on this deviation;

[0069] S4.4: Based on the processing in steps S4.1 to S4.3, generate the final set extrapolation prediction results for effect evaluation.

[0070] S5: Compare the extrapolation prediction results of the set with the measured radar echo extrapolation data of the same period, and extract the change characteristics of key physical parameters in the operation impact area. The key physical parameters include at least one of echo intensity, echo volume (or area), vertical cumulative liquid water content, precipitation intensity and quantitative precipitation estimation. Based on the change characteristics, quantitatively evaluate the actual effect of artificial cloud seeding operation.

[0071] S5.1: Compare the extrapolation prediction results of the set with the extrapolation data of the measured radar echo during the same period, and calculate the echo intensity residual value sequence.

[0072] S5.2: Determine the spatial range of the operation's impact area and select radar physical quantities to characterize the physical properties of cloud precipitation; the physical quantities include at least one of combined reflectivity, vertical integral liquid water content, precipitation intensity, and quantitative precipitation estimation; extract the physical quantities from the ensemble prediction results and actual observation data corresponding to the impact area during the assessment period, calculate their regional average values, and obtain the temporal variation characteristics of each physical quantity in the impact area after the operation.

[0073] S5.3: Calculate the characteristics of the change in volume or area ratio of each physical quantity in the affected area at different intensity intervals over time;

[0074] S5.4: Based on the residual values, temporal changes in physical quantities, and intensity distribution evolution characteristics obtained in steps S5.1 to S5.3, comprehensively and quantitatively evaluate the actual effect of artificial cloud seeding operations.

[0075] This embodiment selects a ground-based hail suppression operation in City A and a combined air-ground rain enhancement operation in City B as examples. The basic operational information of the two cases is listed in Table 1.

[0076] Table 1 Description of relevant information in the embodiments

[0077]

[0078] For the two typical operation cases, performance scores were calculated between the extrapolated sequences of each model and the actual observation data based on the extrapolation results one hour before the start of the operation. The relevant score results are listed in Tables 2 and 3.

[0079] Table 2 Comparison of performance indicators of various extrapolation models in Example 1 (16:00-17:00)

[0080]

[0081] Table 3. Performance Indicators Comparison of Extrapolation Models in Example 2 (13:00-14:00)

[0082]

[0083] Analysis of the evaluation results shown in Tables 2 and 3 indicates that, in the horizontal comparison of indicators, the optimal models selected by different evaluation indicators differ due to their different measurement dimensions. The EES ensemble evaluation algorithm of this invention can effectively solve the problem of multi-indicator selection, and its results can comprehensively reflect the overall selection trend of most indicators, which is more in line with subjective perception and judgment, and provides a reliable basis for subsequent multi-model integration. In the vertical model comparison, the ensemble prediction model obtained by the method of this invention performs stably and ranks among the top in most indicators, effectively integrating the characteristics of various advantageous models to obtain more accurate prediction results. Comparing Tables 2 and 3 further reveals that, although the ensemble prediction results for a certain city B in Table 3 are slightly lower than those for a certain city A (Table 2) which overlaps with the training area, this method still maintains a stable performance advantage in Table 3, showing good regional generalization ability. In summary, the ensemble evaluation method of this invention can stably generate prediction results that are superior to those of a single model, showing superior performance in all evaluation indicators, significantly improving the accuracy and reliability of artificial cloud seeding operation effect evaluation, and has broad practical application potential.

[0084] This embodiment compares and analyzes the key physical parameters extracted from the ensemble extrapolation prediction results and the actual observation data. Tables 4 and 5 list the temporal changes of the mean echo intensity and the area ratio of echoes with different intensities within the affected area. The start time recorded in the tables is the moment when the effective physical parameter grid point first appears in the affected area during the extrapolation period.

[0085] Table 4. Time-series changes in mean echo intensity and area proportion of echoes with different intensities within the affected area in Example 1 (17:00-19:00)

[0086]

[0087] According to the comparison results in Table 4, effective grid points began to appear in the affected area approximately half an hour after the operation ended. During the extrapolated period, the actual average echo intensity in the affected area decreased from 23.889 dBZ to 8.365 dBZ. The actual monitoring data was generally lower than the ensemble prediction results, and the difference between the two gradually widened over time. In addition, during the extrapolated period, the area proportion of non-precipitation echoes (-10 dBZ to 10 dBZ) in the affected area increased rapidly, while the area proportion of moderate intensity echoes (20 dBZ to 40 dBZ) decreased from 84.614% to 15.161%, which was also significantly lower than the ensemble prediction results for the same period. The ensemble assessment and quantitative analysis results using the method and system of this invention show that the ground-based catalytic operation in Example 1 has a positive catalytic effect on suppressing the development of local severe convective weather and reducing the probability of hail formation.

[0088] Table 5. Time-series changes in mean echo intensity and area proportion of echoes with different intensities within the affected area in Example 2 (14:00-16:00)

[0089]

[0090] Example 2 describes a combined air-ground rain enhancement operation in City B. Aircraft seeding began early and spanned a long period, while ground operations were conducted in multiple locations in succession, resulting in a dispersed distribution of the affected area. Therefore, this example selects the period of combined operation where the aircraft's catalytic effect was weaker but ground operations had already begun as the period for extrapolation effect evaluation and analysis. The comparison results in Table 5 show that the actual average echo intensity increased from an initial 2.562 dBZ to 14.141 dBZ during the extrapolated period. After multiple rounds of ground operations, the actual monitored value began to exceed the ensemble predicted value and continued to rise. Meanwhile, the proportion of non-precipitation echoes within the affected area gradually decreased after the operation began. The area proportion of weak precipitation echoes (10–20 dBZ) peaked at 36.364% at 15:12. Subsequently, with the superposition of the combined air-ground influence, stronger echoes (above 30 dBZ) began to appear in the affected area at 15:54. Although echoes of a certain intensity appeared in the ensemble prediction results at this time, their intensity and range were significantly lower than the actual monitoring results. The quantitative analysis results of the ensemble assessment show that the combined air-ground operation in Example 2 promoted cloud water conversion and achieved the expected operational effect of artificial rain enhancement.

[0091] The quantitative analysis results of Examples 1 and 2 show that the ensemble evaluation method provided by this invention exhibits stable evaluation performance in experimental cases across different operation types and geographical regions. This method can effectively characterize the temporal variations of key physical parameters caused by catalytic operations and accurately identify systematic differences between actual and predicted results, thereby achieving an objective and quantitative evaluation of the effects of different weather modification operations. The evaluation system constructed by this invention does not rely on the spatiotemporal distribution characteristics of radar echo data in a specific region, nor is it limited by differences in operation methods. It possesses good applicability and robustness under various geographical environments and operation configurations, demonstrating strong generalization ability and operational potential.

[0092] This application constructs an ensemble evaluation method that integrates the advantages of multiple radar echo intelligent extrapolation models. It systematically corrects the ensemble results to address the relatively small catalytic impact in the early stages of operations, thereby consistently generating prediction results superior to those of a single model. Simultaneously, this invention transforms the extrapolated time-series results into a quantitative evaluation method using gridded statistical data of key physical parameters. This effectively identifies and reflects the catalytic effect of artificial cloud seeding operations, reducing the interference of prediction error growth caused by time-series accumulation on the effect verification results. It improves the accuracy and physical rationality of the effect evaluation, providing effective technical support for the large-scale operational application of radar echo intelligent extrapolation technology in the evaluation of weather modification effects.

[0093] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, the intent of this invention is also encompassed within these modifications and variations.

Claims

1. An artificial cloud seeding effect set evaluation method based on intelligent extrapolation of radar echoes, characterized by The steps include the following: S1: Acquire sample Doppler weather radar data, and construct a training dataset after preprocessing the radar data; S2: Train the training dataset based on the radar echo intelligent extrapolation model; S3: Construct an ensemble evaluation algorithm, which is based on at least one evaluation benchmark factor and multiple error penalty terms; call each extrapolation model in the model library to infer the extrapolation results of the natural state radar echo of the target area during a preset period before the start of artificial cloud seeding operations; use the ensemble evaluation algorithm to score and rank the extrapolation results of each model. S4: Select the top-scoring models as preferred models, extrapolate the radar echo evolution process of the target area under natural conditions after the start of the operation, and fuse and correct the deviation of the output results of each preferred model to generate the ensemble extrapolation prediction result. S5: Compare the extrapolated prediction results of the set with the measured radar echo observation data of the same period, extract the change characteristics of key physical parameters in the operation's impact area, and quantitatively evaluate the actual effect of the artificial cloud seeding operation based on these change characteristics; Step S1 includes: S1.1: Based on the sample Doppler weather radar data, filter the data in each consecutive multi-frame sequence that contains at least one frame with a precipitation area coverage exceeding a set threshold, and select data with relatively isolated convective cells as individual case data. S1.2: Perform quality control on the individual case data obtained through screening; S1.3: Preprocess the individual case data into a standardized training dataset with uniform spatial dimensions and numerical range; In step S3, constructing the set evaluation algorithm includes: S3.1: Select multiple objective evaluation indicators and group them based on correlation analysis and clustering methods; S3.2: Based on the objective evaluation index, perform a comprehensive performance ranking of each extrapolation model, determine at least one optimal index, and determine the evaluation benchmark factor based on the optimal index; S3.3: Based on the evaluation benchmark factor, construct multiple error penalty terms; the error penalty terms include at least two of the following: a spatially weighted error term to enhance the sensitivity of prediction errors to key areas, a continuity constraint term to suppress non-physical mutations, and an asymmetric prediction penalty term that sets differentiated weights for over-prediction and under-prediction; S3.4: Establish a time error accumulator for each error penalty term and process it using a time-series smoothing model; S3.5: Combining the evaluation benchmark factor and each error penalty term after time-series smoothing, construct the set evaluation algorithm, wherein the output value of the set evaluation algorithm is a ratio function of the evaluation benchmark factor and the sum of each error penalty term.

2. The artificial cloud seeding effect set evaluation method based on radar echo intelligent extrapolation according to claim 1, characterized in that In step S2, the radar echo intelligent extrapolation model is a time-series prediction model based on deep learning; the selected model is trained on a standardized training dataset, and the data of the previous historical period is used as the model input, and the data of the next future period is used as the model validation, so as to obtain multiple pre-trained extrapolation models, which constitute the model library.

3. The method of claim 1, wherein, The spatial weighted error term is obtained by spatially modulating the mean absolute error by fusing at least one of the gradient features, Gaussian texture features, and multi-scale contrast features of the image; the continuity constraint term is obtained by calculating the difference between the predicted result and the true result on the second derivative; the asymmetric prediction penalty term is obtained by introducing an intensity weighting function and assigning different penalty coefficients to over-prediction and under-prediction.

4. The method of claim 1, wherein, In step S4, at least three models with the highest scores are selected as the preferred models, and a weighted fusion method based on the scores of the set evaluation algorithm is used for fusion processing.

5. The method of claim 1, wherein, In step S5, the key physical parameters include at least one of echo intensity, echo area, vertical cumulative liquid water content, precipitation intensity, and quantitative precipitation estimation; the quantitative assessment of the actual effect based on the change characteristics includes calculating the echo intensity residual value sequence and analyzing the temporal changes of the regional average value of each physical quantity and the area ratio of different intensity intervals within the affected area.

6. An artificial cloud seeding effect set evaluation system based on intelligent extrapolation of radar echoes, used to implement the artificial cloud seeding effect set evaluation method based on intelligent extrapolation of radar echoes according to any one of claims 1-5, characterized in that, include: The data preprocessing module (100) acquires sample Doppler weather radar data and performs case screening, quality control and preprocessing on the data; The model library management module (200) calls the model library trained using a standardized Doppler weather radar case dataset; The algorithm scoring module (300) combines the radar echo data of the spatiotemporal region of the target to be evaluated with the input of the operation information, calls each extrapolation model in the model library to infer the extrapolation result of the target region under the natural state 1 hour before the start of the operation, and scores the performance of the extrapolation result of each model based on the EES set evaluation algorithm. The ensemble correction module (400) selects multiple models with the highest scores, uses a weighted fusion method based on EES algorithm indicators to fuse the output results of the selected models, corrects the bias of the fusion results, and generates the final ensemble extrapolation prediction results for effect evaluation. The effect verification module (500) compares the extrapolation prediction results of the set with the extrapolation data of the measured radar echo during the same period. Based on the obtained residual values, the time-series changes of physical quantities and the evolution characteristics of intensity distribution, it comprehensively and quantitatively evaluates the actual effect of artificial cloud seeding operations.

Images