An artificial rain enhancement operation condition intelligent identification method based on multi-source data fusion

Abstract

The application discloses an artificial rain enhancement operation condition intelligent identification method based on multi-source data fusion, comprising the following steps: using minute precipitation data of ground rain stations, combining precipitation processes in a quantitative detection target range of a radar, and constructing sample labels; using ground observation precipitation as a true value for judging the rain enhancement operation condition, and establishing positive and negative sample data sets; based on the established positive and negative sample data sets, constructing a rain enhancement operation condition judgment model based on a machine learning algorithm; wherein the machine learning algorithm comprises a random forest, LightGBM, a support vector machine SVM and a neural network. The data set of a specific region is collected, whether the specific region has the rain enhancement operation condition is judged through the rain enhancement operation condition judgment model, and an operation condition grade is output, thereby providing scientific guidance for the artificial rain enhancement operation. The method can accurately identify the artificial rain enhancement operation condition based on multi-source data, and improves the identification precision and the application range.

Description

Technical Field

[0001] This invention relates to the field of artificial rain enhancement operation technology, and in particular to an intelligent identification method for artificial rain enhancement operation conditions based on multi-source data fusion. Background Technology

[0002] Currently, the successful implementation of artificial rain enhancement operations mainly depends on two key factors: suitable cloud physics conditions (such as supercooled water content and cloud thickness) and accurate timing of operations. In recent years, with the continuous enrichment of multi-source observation methods such as radar, satellites, and numerical models, data fusion technology has provided a new technical path for the real-time comprehensive analysis of the macro and micro characteristics of cloud systems. The advantages of machine learning methods in mining massive multi-source data and modeling nonlinear relationships also provide a feasible solution to break through the traditional subjective experience-based analysis model.

[0003] Current methods for determining operational conditions primarily rely on forecasters' manual analysis of macro and micro parameters of cloud systems. This approach suffers from several technical limitations: First, the interpretation criteria are highly subjective, with different forecasters yielding significantly different assessments of the same cloud system, lacking objective and unified criteria. Second, it struggles to effectively integrate high-dimensional data from multiple sources, including radar, satellite, and numerical models, making it difficult to fully uncover the nonlinear relationships between numerous characteristic parameters through manual methods. Third, it lacks a systematic learning and verification mechanism for historical precipitation processes, preventing the solidification of judgment experience into reusable mathematical models and resulting in insufficient repeatability and generalization ability. These shortcomings hinder the further development of artificial rain enhancement operational condition assessment technology.

[0004] In view of this, the present invention is hereby proposed. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent identification method for artificial rain enhancement operation conditions based on multi-source data fusion, thereby solving the aforementioned technical problems in the prior art. The method described in this invention can accurately identify artificial rain enhancement operation conditions based on multi-source data, improving identification accuracy and applicability.

[0006] The objective of this invention is achieved through the following technical solution: A method for intelligent identification of artificial rain enhancement operation conditions based on multi-source data fusion, the method comprising: Step 1: Using minute-by-minute precipitation data from ground rain gauges, combined with radar quantitative detection of precipitation processes within the target area, construct sample labels; Step 2: Use ground-observed precipitation as the ground truth for determining rain enhancement operation conditions to establish a positive and negative sample dataset; Step 3: Based on the established positive and negative sample datasets, construct a rain enhancement operation condition discrimination model based on machine learning algorithms; wherein, the machine learning algorithms include random forest, LightGBM, support vector machine (SVM), and neural network.

[0007] Step 4: Collect datasets for specific areas, use a rain enhancement operation condition discrimination model to determine whether specific areas meet the conditions for rain enhancement operations, and output the operation condition level to provide scientific guidance for artificial rain enhancement operations.

[0008] Compared with existing technologies, the method provided by this invention can accurately identify artificial rain enhancement operation conditions based on multi-source data, thereby improving the identification accuracy and applicability. Attached Figure Description

[0009] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 This is a schematic diagram of the intelligent identification method for artificial rain enhancement operation conditions based on multi-source data fusion provided in an embodiment of the present invention. Detailed Implementation

[0011] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them, and do not constitute a limitation on the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the protection scope of the present invention.

[0012] First, the following explanations are provided for the terms that may be used in this article: The term "and / or" means that either or both can be achieved simultaneously. For example, X and / or Y means that it includes both "X" or "Y" as well as the three cases of "X and Y".

[0013] The terms "comprising," "including," "containing," "having," or other similar semantic descriptions should be interpreted as non-exclusive inclusion. For example, including a technical feature element (such as raw material, component, ingredient, carrier, dosage form, material, size, part, component, mechanism, device, step, process, method, reaction conditions, processing conditions, parameter, algorithm, signal, data, product or article of manufacture, etc.) should be interpreted as including not only the expressly listed technical feature element, but also other technical feature elements that are not expressly listed and are well-known in the art.

[0014] The term "composed of" excludes any technical features not expressly listed. When used in a claim, it closes the claim to exclude all technical features other than those expressly listed, except for associated conventional impurities. If the term appears only in a clause of a claim, it limits the claim to the elements expressly listed in that clause; elements recited in other clauses are not excluded from the overall claim.

[0015] The technical solution provided by this invention will be described in detail below. Contents not described in detail in the embodiments of this invention belong to prior art known to those skilled in the art. For example... Figure 1 The diagram shows a flowchart of an intelligent identification method for artificial rain enhancement operation conditions based on multi-source data fusion provided in an embodiment of the present invention. The method includes: Step 1: Using minute-by-minute precipitation data from ground rain gauges, combined with radar quantitative detection of precipitation processes within the target area, construct sample labels; In this step, cloud physics factors most relevant to precipitation formation are used as key identification factors for rain enhancement operation conditions. A total of 56 variables, including radar parameters, satellite parameters, and numerical model variables, are selected as influencing factors for rain enhancement operation conditions. The basic variables include longitude, latitude, precipitation, and labels; Radar parameters include combined reflectivity factor, echo top height, and vertically accumulated liquid water. Satellite parameters include cloud top height, cloud top temperature, supercooled layer thickness, effective particle radius, optical thickness, and liquid water path. There are 43 numerical model variables, including cloud bands, vertically accumulated liquid water, vertically accumulated supercooled water, cloud top height, cloud top temperature, cloud base height, cloud base temperature, and the U-wind speed, V-wind speed, vertical velocity, water vapor mixing ratio, cloud water mixing ratio, rainwater mixing ratio, temperature, and relative humidity of the five isobaric surfaces at 800hPa, 750hPa, 700hPa, 650hPa, and 600hPa.

[0016] Step 2: Use ground-observed precipitation as the ground truth for determining rain enhancement operation conditions to establish a positive and negative sample dataset; In this step, the start and end times of the precipitation process are used as the basis. For continuous precipitation periods where more than 50% of the stations within the radar coverage area experience precipitation and the precipitation lasts for more than 10 minutes, the corresponding multi-source remote sensing and model features are extracted as positive samples. Data with no precipitation but present radar echoes in the half hour before the start of a continuous precipitation period are used as negative samples. The effective time set H that satisfies the condition that more than 50% of the stations within the radar coverage area experience precipitation is: ; Let represent the precipitation at the meteorological station in the i-th country at hour h; This is an indicator function; it takes the value 1 if the condition inside the parentheses is true, and 0 otherwise. Let t be the minute time sequence number, and let t be the minute precipitation. Precipitation indicator function per minute for: ; Let L be the set of all minutes within a continuous precipitation interval of at least 10 minutes, represented as: ; The set of precipitation periods P is composed of connected components extracted from L after expansion at 3-minute intervals (precipitation intervals less than 3 minutes are considered continuous precipitation), represented as: ; This indicates the extraction of the connected components of the set, i.e., all maximal consecutive integer intervals; The set of positive samples S is represented as: ; In the specific implementation, the positive and negative sample datasets are randomly divided into 80% training set and 20% test set. The SMOTE (Synthetic Minority Over-sampling Technique) oversampling technique is used to address the imbalance problem in the training set. The specific process is as follows: For each minority class sample x i Calculate the Euclidean distance between it and the other minority class samples in the feature space, and select K nearest neighbors (K=5 in this invention); randomly select a nearest neighbor sample x. i-nn A new sample x is randomly generated along the direction of the line connecting the two. new The calculation formula is: ; The StandardScaler method is used to standardize the features. The 56 extracted feature parameters differ significantly in magnitude. To avoid these differences affecting model convergence and parameter optimization, this invention standardizes all features. The Z-score standardization method is used, that is, for each feature... Convert it to a standard normal distribution with a mean of 0 and a standard deviation of 1. The calculation formula is as follows: ; In the formula, Features The mean on the training set; Features Standard deviation on the training set; The standardization process is based on the mean and standard deviation calculated from the training set, while the test set uses the same... and The data needs to be converted to ensure consistency.

[0017] Step 3: Based on the established positive and negative sample datasets, construct a rain enhancement operation condition discrimination model based on machine learning algorithms; wherein, the machine learning algorithms include random forest, LightGBM, support vector machine (SVM), and neural network. In this step, the random forest algorithm draws samples with replacement from the original training set through bootstrap sampling, constructs multiple decision trees, and randomly selects some features as candidate split features when splitting at the node of each decision tree. Let the original training set be... ,in It is a 56-dimensional feature vector. The labels are 1 for rain enhancement operations and 0 for rain enhancement operations. The Random Forest algorithm makes judgments by integrating multiple decision trees. The specific construction process is as follows: (1) Bootstrap sampling: N samples are randomly selected from the original training set D to construct the training subset of each tree. , where K is the total number of decision trees (K=100 in this invention); (2) Random feature selection: When splitting at each node of each decision tree, m candidate features are randomly selected from 56 features (in this invention, m are selected from the m candidate features). ), and perform the optimal partitioning among these m features; This dual randomness gives random forests strong resistance to overfitting and good generalization performance. The algorithm can handle high-dimensional data without feature dimensionality reduction, can output feature importance, and has good robustness to missing and outlier values.

[0018] The LightGBM algorithm is a gradient boosting decision tree algorithm that incorporates one-sided gradient sampling and mutually exclusive feature binding techniques. It employs a histogram-based decision tree learning algorithm and a depth-constrained leaf growth strategy. This algorithm boasts fast training speed, low memory usage, and high accuracy. The specific process is as follows: (1) Single-sided gradient sampling: Calculate the gradient of each sample, sort them in descending order of absolute gradient value, select the sample with large gradient as high gradient sample, randomly select samples from the remaining samples as low gradient samples, and multiply the low gradient samples by weights to compensate, thereby greatly reducing the number of samples while ensuring accuracy. (2) Histogram algorithm: Discretize the continuous feature values ​​into a fixed number of histograms (255 bins in this invention), traverse the histogram statistics to find the optimal split point, which significantly improves training efficiency and reduces memory usage; (3) Leaf growth strategy: A depth-constrained leaf-wise growth strategy is adopted. Each time, the node with the largest splitting gain is selected from all current leaf nodes for splitting. For the candidate splitting feature f and the splitting threshold... The formula for calculating the split gain ΔGain is: ; In the formula, G L G R H represents the sum of the first-order gradients of the left and right child nodes, respectively; L H R λ is the sum of the second-order gradients of the left and right child nodes, respectively; λ is the regularization coefficient (λ=0.1 in this invention); γ is the minimum loss reduction required for leaf node splitting; the splitting feature f that maximizes ΔGain is selected. ∗ and the dividing threshold θ ∗ To split; (4) Iterative training: Set the learning rate η=0.05, and the maximum depth d max =8, number of leaf nodes num_leaves=31, number of iteration rounds n_estimators=100; each iteration fits the residual of the previous round, gradually reducing the training error, and outputs the final discrimination result; Support Vector Machines (SVMs) map the original feature space to a high-dimensional space using kernel functions, searching for the optimal hyperplane that maximizes the margin between samples of different classes. This algorithm exhibits excellent generalization ability under small sample conditions and can handle nonlinear classification problems. The specific construction process is as follows: (1) Optimization objective: For the training set ,in Support Vector Machine (SVM) solves the following convex quadratic programming problem: ; The constraints are: ; In the formula, w is the hyperplane normal vector; b is the bias term; is a slack variable; C is the penalty coefficient (in this invention, C=1); These are high-dimensional features after kernel function mapping; (2) Kernel function selection: The radial basis function (RBF kernel) is used to handle nonlinear classification problems. The kernel function expression is: ; In the formula, The kernel parameter controls the width of the Gaussian kernel (in this invention, the kernel width is determined by the kernel parameter). ); (3) Solving the dual problem: Introducing Lagrange multipliers The original problem is transformed into the dual problem: ; The constraints are: ; (4) Decision-making: Solve the dual problem to obtain the optimal Lagrange multipliers. For new samples Discriminant function for: ; when It was determined that conditions were suitable for rain enhancement operations. It was determined that there were no conditions for rain enhancement operations. The neural network uses a multilayer perceptron architecture, including an input layer, hidden layers, and an output layer. It adjusts the connection weights between neurons using backpropagation and a gradient descent optimizer. The specific construction process is as follows: (1) Forward propagation: Let the input features be The output of the l-th hidden layer for: ; In the formula, and These are the weight matrix and bias vector of the l-th layer, respectively; , The activation function is ReLU; in its specific implementation, this invention uses the ReLU activation function: ; (2) Output layer: The output layer uses the Softmax function to convert the output into a probability distribution, as follows: ; in This is the output value of the c-th node in the output layer; the final judgment result is the category with the higher probability. (3) Loss function: The cross-entropy loss function is adopted, which is expressed as: ; In the formula, Let be the true label (0 or 1) of the i-th sample in class c; N is the total number of samples. (4) Backpropagation and optimization: The Adam optimizer is used to iteratively update the weights. The learning rate η = 0.001, the first moment estimation decay rate β1 = 0.9, and the second moment estimation decay rate β2 = 0.999. The weight update formula is as follows: ; In the formula, These are network parameters, including all weights and biases; and These are the bias correction estimates for the first and second moments of the gradient, respectively; To prevent division by zero of constants; During training, each iteration calculates the gradient of the loss function with respect to each parameter using the current batch of samples. The error is then propagated layer by layer using the backpropagation algorithm. The Adam optimizer adaptively adjusts the learning rate of each parameter based on the gradient information, causing the loss function to gradually converge to its minimum value, thus completing network training. In the specific implementation, the number of iterations (epochs) is 100, and the batch size (batch_size) is 32.

[0019] Step 4: Collect datasets for specific areas, use a rain enhancement operation condition discrimination model to determine whether specific areas meet the conditions for rain enhancement operations, and output the operation condition level to provide scientific guidance for artificial rain enhancement operations.

[0020] The specific process in this step is as follows: (1) Data collection and standardization of specific regions Radar data (combined reflectivity factor CR, echo top height ET, vertical cumulative liquid water VIL), satellite data (cloud top height CTH, cloud top temperature CTT, supercooled layer thickness HSC, effective particle radius CER, optical thickness COT, liquid water path LWP) and model data (43 variables, including vertical integral, cloud parameters, and wind speed, water vapor mixing ratio, cloud water mixing ratio, temperature, relative humidity, etc. at 5 isobaric surfaces) were collected for a specific area at the time of identification. Together with the basic variables (latitude and longitude, precipitation PRE), a 56-dimensional feature vector was formed after spatial matching. For each feature of the current sample, the mean of the training set is used. and standard deviation Z-score standardization is performed, and the process is represented as follows: ; Obtain the standard eigenvector Znew =(Z1,Z2,…,Z 56 ) T ; (2) Model discrimination and job level output The standard eigenvector Z new Input the trained rain enhancement operation condition discrimination model (LightGBM, hyperparameters: learning rate 0.1, maximum depth 10, number of leaf nodes 50), and output the probability that the grid point belongs to the category of having operation conditions. , represented as: ; Based on probability The job condition level (SL) will be output according to the following rules, and represented as follows: ; When precipitation PRE cannot be obtained, a comprehensive judgment is made based on substitution factors (CR contribution rate of up to 30.8%, threshold ≥ 35 dBZ; VIL threshold ≥ 1.5 kg / m²; rh_700 threshold ≥ 70%), and the judgment rules are the same as above; After the grid point is identified, all grid points with operation condition level SL=1 are summarized, and the spatial distribution of rain enhancement operation condition levels and suitable operation range for a specific area are output.

[0021] In its specific implementation, the method further includes: To reflect the merits of the rain enhancement operation condition discrimination model, a confusion matrix was used to evaluate the classification results of the rain enhancement operation condition discrimination model; In the confusion matrix, the main diagonal parameters N and H represent the number of non-precipitation and precipitation events correctly identified by the model, respectively, while the secondary diagonal parameters M and F represent the number of incorrect identifications. Table 1 below shows the evaluation results of the confusion matrix. Table 1 Evaluation results of the confusion matrix

[0022] The performance of the rain enhancement operation condition discrimination model was evaluated using TS score, POD hit rate, FAR false alarm rate, and FNR false alarm rate. The definitions of each indicator are as follows: Hit Rate POD: ; In the formula, H represents the number of correct identifications and M represents the number of missed identifications; FAR (Failure Rate): ; In the formula, F represents the number of false alarms; Failure rate (FNR): ; TS rating: .

[0023] In practice, to improve model performance, grid search can be used for hyperparameter tuning. Let the hyperparameter space to be optimized be... , training set Divide the hyperparameter space into 5 equal subsets {D1, D2, ..., D5}; for each set of candidate parameters in the hyperparameter space... The k-th fold subset is used as the validation set, k=1,2,…,5, and the remaining four folds are used as the training set for model training; the TS score is used as the performance evaluation metric at the k-th fold, denoted as . The average of the 50% TS scores is taken as the hyperparameters for this group. Cross-validation performance metrics: ; Traversing the hyperparameter space Select the combination that makes the candidate combination available. The largest combination of hyperparameters is the optimal parameter. , represented as: ; In this invention, the search space of LightGBM is set as follows: learning rate η∈{0.01,0.05,0.1}, maximum depth d max The random forest has the following characteristics: ∈{6,8,10,12}, number of leaf nodes num_leaves∈{15,31,50,70}; the search space of the random forest is: number of trees K∈{50,100,150}, maximum depth d. max The SVM search space is: penalty coefficient C ∈ {8, 10, 12}, kernel parameter γ ∈ {0.01, 0.1, 1}. A grid search iterates through all the above combinations, using the average TS score from 5-fold cross-validation as the objective, to select the optimal hyperparameters for each model. Optimize the model.

[0024] To more clearly demonstrate the technical solution and its effects provided by this invention, the technical effects of the method described in this embodiment are described in detail below with specific examples. In this example, six datasets are involved, divided into two categories according to geographical region: typical grassland and desert grassland, and into three types according to feature source: radar, radar + pattern, and radar + pattern + satellite. The total number of positive and negative samples in each dataset ranges from 318 to 6546 pairs. Four machine learning algorithms were trained and evaluated using datasets containing all features. The TS scores of all models were above 0.98, and the TS scores of five datasets reached 1.000, indicating that the method of this invention can effectively determine the conditions for artificial rain enhancement operations. In terms of model comparison, LightGBM performed the best, achieving the highest TS scores in five of the six datasets. The best model and performance indicators for each dataset are shown in Table 2. Table 2 Performance Comparison of Full Models Including PRE Features

[0025] To explore the key meteorological factors influencing operational condition assessment, the feature importance of the outputs of the random forest and LightGBM models was analyzed. Precipitation (PRE) was the most critical factor in assessing operational conditions, ranking first across all datasets with an average contribution rate of 86.70%, reaching as high as 94.41% in the typical grassland-radar dataset. As dataset complexity increased, the contribution rate of PRE decreased from 94.41% to 72.71%, while the importance of humidity field factors (rh_800, rh_750) and temperature field factors (tc_800) gradually became apparent, indicating that a comprehensive assessment of multiple factors is necessary under complex conditions. Radar parameter CR ranked second in the desert grassland dataset, with a contribution rate exceeding 7.07%, indicating that radar echo intensity has significant reference value. VIL and CR ranked among the top three in multiple datasets, demonstrating that cloud liquid water content and radar echo intensity make significant contributions to operational condition assessment.

[0026] Table 3. Ranking of feature importance for each dataset (top three)

[0027] To verify the core role of precipitation (PRE) and reveal the true contributions of other factors, a comparative experiment was conducted with PRE features removed. The experimental results are shown in Table 4. After removing PRE, the model performance of all datasets decreased, with an average decrease of 17.1% and a maximum decrease of 40.2% (typical grassland-radar dataset). The decrease in performance of complex datasets was significantly smaller (3.6% for typical grassland-radar+model+satellite dataset, 6.0% for desert grassland-radar+model dataset, and 5.9% for desert grassland-radar+model+satellite dataset), indicating that the multi-factor comprehensive judgment system has stronger robustness.

[0028] In the absence of PRE (Presentation Reflection), the importance of factors such as CR (Constant Reflection), VIL (Volume Integral Liquid Water), rh_700, rh_800, and tc_800 significantly increased. Among them, CR was particularly prominent, becoming the primary factor in the typical grassland-radar dataset with a contribution rate of 32.20%, and also contributing the most in the desert grassland-radar dataset (27.68%), indicating that radar echo information is an important judgment criterion besides PRE. VIL (Vertical Integral Liquid Water) contributed 27.10% in the typical grassland-radar dataset, rh_700 (700 hPa humidity) contributed 17.93% in the typical grassland-radar+model dataset, and rh_800 (800 hPa humidity) became the primary factor in both the desert grassland-radar+model and radar+model+satellite datasets, with contribution rates of 6.19% and 4.27%, respectively, reflecting the significant impact of cloud liquid water content and mid-level humidity on the judgment of operational conditions.

[0029] Table 4. Model performance changes after removing PRE

[0030] The results of model tuning are shown in Table 5 below. Different models showed significant differences in parameter sensitivity after removing PRE. SVM tuning showed the most significant improvement, increasing the SVM performance from 0.743 to 0.7624 (an improvement of 0.1488) on the typical grassland-radar+ pattern dataset and from 0.8381 to 0.9330 (an improvement of 0.0949) on the desert grassland-radar+ pattern dataset, demonstrating its sensitivity to parameters and suitability for scenarios with ample data and a focus on precision. LightGBM performed stably across datasets, with an average TS of 0.821 by default and 0.826 after tuning, making it the most stable model overall. Random Forest also showed good stability, with an average improvement of 0.011 after tuning. Neural Network tuning achieved the best performance (0.7657) on the typical grassland-radar+ pattern dataset, but its performance declined in small sample datasets, with an overall average improvement of 0.010.

[0031] Table 5 Comparison of Model Tuning Results (without PRE)

[0032] Comparing the optimal parameters with and without PRE, model complexity generally increases after PRE removal: the number of trees in Random Forest increases from 50 to 100-150, and the maximum depth increases from 8-10 to 12; the learning rate of LightGBM increases from 0.05 to 0.1, the maximum depth increases from 8 to 10-12, and the number of leaf nodes increases from 31 to 50; the gamma of SVM changes from adaptive 'scale' to a fixed value of 0.1; the neural network structure remains largely unchanged. This change reflects the adaptive adjustments made by the model to learn alternative nonlinear relationships between features.

[0033] Based on the performance of different models across various datasets, the optimal model is recommended for each dataset. Specifically: LightGBM achieved five best-in-class rankings across six datasets. It performed optimally on the three typical grassland datasets, as well as the desert steppe-radar+pattern and desert steppe-radar+pattern+satellite datasets (both tuned and default). NeuralNetwork_Tuned performed best on the typical grassland-radar+pattern dataset (0.7657). Overall, LightGBM demonstrated stable performance on most datasets, making it a suitable first-choice algorithm for determining the conditions of artificial shadowing operations. After tuning, SVM achieved near-optimal results on the typical grassland-radar+pattern dataset (0.7624) and the desert steppe-radar+pattern dataset (0.9330), demonstrating its potential application in scenarios with abundant data and rich features.

[0034] Table 6. Optimal Model Recommendations for Each Dataset (Without PRE)

[0035] It is worth noting that the contents not described in detail in the embodiments of the present invention belong to the prior art known to those skilled in the art.

[0036] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims. The information disclosed in the background section is intended only to enhance the understanding of the overall background technology of the present invention and should not be construed as an admission or implication in any way that such information constitutes prior art known to those skilled in the art.

Claims

1. A method for intelligent identification of artificial rain enhancement operation conditions based on multi-source data fusion, characterized in that, The method includes: Step 1: Using minute-by-minute precipitation data from ground rain gauges, combined with radar quantitative detection of precipitation processes within the target area, construct sample labels; Step 2: Use ground-observed precipitation as the ground truth for determining rain enhancement operation conditions to establish a positive and negative sample dataset; Step 3: Based on the established positive and negative sample datasets, construct a rain enhancement operation condition discrimination model based on machine learning algorithms; wherein, the machine learning algorithms include random forest, LightGBM, support vector machine (SVM), and neural network. Step 4: Collect datasets for specific areas, use a rain enhancement operation condition discrimination model to determine whether specific areas meet the conditions for rain enhancement operations, and output the operation condition level to provide scientific guidance for artificial rain enhancement operations.

2. The intelligent identification method for artificial rain enhancement operation conditions based on multi-source data fusion according to claim 1, characterized in that, In step 1, Based on the cloud physics factors most relevant to precipitation formation as the key identification factors for rain enhancement operation conditions, a total of 56 variables, including radar parameters, satellite parameters, and numerical model variables, were selected as influencing factors for rain enhancement operation conditions. The basic variables include longitude, latitude, precipitation, and labels; Radar parameters include combined reflectivity factor, echo top height, and vertically accumulated liquid water. Satellite parameters include cloud top height, cloud top temperature, supercooled layer thickness, effective particle radius, optical thickness, and liquid water path. There are 43 numerical model variables, including cloud bands, vertically accumulated liquid water, vertically accumulated supercooled water, cloud top height, cloud top temperature, cloud base height, cloud base temperature, and the U-wind speed, V-wind speed, vertical velocity, water vapor mixing ratio, cloud water mixing ratio, rainwater mixing ratio, temperature, and relative humidity of the five isobaric surfaces at 800hPa, 750hPa, 700hPa, 650hPa, and 600hPa.

3. The intelligent identification method for artificial rain enhancement operation conditions based on multi-source data fusion according to claim 1, characterized in that, In step 2, specifically based on the start and end time of the precipitation process, for continuous precipitation periods where more than 50% of the stations within the radar coverage area experience precipitation and the precipitation lasts for more than 10 minutes, the corresponding multi-source remote sensing and model features are extracted as positive samples. Data with no precipitation but present radar echoes in the half hour before the start of a continuous precipitation period are used as negative samples. The effective time set H that satisfies the condition that more than 50% of the stations within the radar coverage area experience precipitation is: ； Let represent the precipitation at the meteorological station in the i-th country at hour h; This is an indicator function; it takes the value 1 if the condition inside the parentheses is true, and 0 otherwise. Let t be the minute time sequence number, and let t be the minute precipitation. Precipitation indicator function per minute for: ； Let L be the set of all minutes within a continuous precipitation interval of at least 10 minutes, represented as: ； The set of precipitation periods P is composed of connected components extracted from L after dilation at 3-minute intervals, and is represented as: ； This indicates the extraction of the connected components of the set, i.e., all maximal consecutive integer intervals; The set of positive samples S is represented as: 。 4. The intelligent identification method for artificial rain enhancement operation conditions based on multi-source data fusion according to claim 1, characterized in that, In step 2, the positive and negative sample datasets are randomly divided into 80% training set and 20% test set. The SMOTE oversampling technique is used to address the imbalance problem in the training set. The specific process is as follows: For each minority class sample x i Calculate the Euclidean distance between it and the other minority class samples in the feature space, and select K nearest neighbors; randomly select a nearest neighbor sample x. i-nn A new sample x is randomly generated along the direction of the line connecting the two. new The calculation formula is: ； The Z-score normalization method is used, that is, for each feature... Convert it to a standard normal distribution with a mean of 0 and a standard deviation of 1. The calculation formula is as follows: ； In the formula, Features The mean on the training set; Features Standard deviation on the training set; The standardization process is based on the mean and standard deviation calculated from the training set, while the test set uses the same... and The data needs to be converted to ensure consistency.

5. The intelligent identification method for artificial rain enhancement operation conditions based on multi-source data fusion according to claim 1, characterized in that, In step 3, The Random Forest algorithm draws samples with replacement from the original training set through bootstrap sampling, constructs multiple decision trees, and randomly selects some features as candidate split features when splitting at the node of each decision tree. Let the original training set be... ,in It is a 56-dimensional feature vector. The labels are 1 for rain enhancement operations and 0 for rain enhancement operations. The random forest algorithm makes judgments by integrating multiple decision trees. The specific construction process is as follows: (1) Bootstrap sampling: N samples are randomly selected from the original training set D to construct the training subset of each tree. Where K is the total number of decision trees; (2) Random feature selection: When splitting at each node of each decision tree, m candidate features are randomly selected from 56 features, and the optimal split is performed among these m features; The LightGBM algorithm is a gradient boosting decision tree algorithm that introduces one-sided gradient sampling and mutually exclusive feature binding techniques. It employs a histogram-based decision tree learning algorithm and a depth-constrained leaf growth strategy. The specific process is as follows: (1) Single-sided gradient sampling: Calculate the gradient of each sample, sort them in descending order of absolute gradient value, select the sample with large gradient as high gradient sample, randomly select samples from the remaining samples as low gradient samples, and multiply the low gradient samples by weights to compensate. (2) Histogram algorithm: Discretize the continuous feature values ​​into a fixed number of histograms, and traverse the histogram statistics to find the optimal split point; (3) Leaf growth strategy: A depth-constrained leaf-wise growth strategy is adopted. Each time, the node with the largest splitting gain is selected from all current leaf nodes for splitting. For the candidate splitting feature f and the splitting threshold... The formula for calculating the split gain ΔGain is: ； In the formula, G L G R H represents the sum of the first-order gradients of the left and right child nodes, respectively; L H R Let f be the sum of the second-order gradients of the left and right child nodes, respectively; λ be the regularization coefficient; γ be the minimum loss reduction required for leaf node splitting; and f be the splitting feature that maximizes ΔGain. ∗ and the dividing threshold θ ∗ To split; (4) Iterative training: Set the learning rate η=0.05, and the maximum depth d max =8, number of leaf nodes num_leaves=31, number of iteration rounds n_estimators=100; each iteration fits the residual of the previous round, gradually reducing the training error, and outputs the final discrimination result; Support Vector Machine (SVM) maps the original feature space to a high-dimensional space using kernel functions, searching for the optimal hyperplane that maximizes the margin between samples of different classes. The specific construction process is as follows: (1) Optimization objective: For the training set ,in Support Vector Machine (SVM) solves the following convex quadratic programming problem: ； The constraints are: ； In the formula, w is the hyperplane normal vector; b is the bias term; C is a slack variable; C is a penalty coefficient; These are high-dimensional features after kernel function mapping; (2) Kernel function selection: The radial basis function is used to handle nonlinear classification problems. The kernel function expression is: ； In the formula, The kernel parameter controls the width of the Gaussian kernel; (3) Solving the dual problem: Introducing Lagrange multipliers The original problem is transformed into the dual problem: ； The constraints are: ； (4) Decision-making: Solve the dual problem to obtain the optimal Lagrange multipliers. For new samples Discriminant function for: ； when It was determined that conditions were suitable for rain enhancement operations. It was determined that there were no conditions for rain enhancement operations. The neural network uses a multilayer perceptron architecture, including an input layer, hidden layers, and an output layer. It adjusts the connection weights between neurons using backpropagation and a gradient descent optimizer. The specific construction process is as follows: (1) Forward propagation: Let the input features be The output of the l-th hidden layer for: ； In the formula, and These are the weight matrix and bias vector of the l-th layer, respectively; , For activation functions; (2) Output layer: The output layer uses the Softmax function to convert the output into a probability distribution, as follows: ； in This is the output value of the c-th node in the output layer; the final judgment result is the category with the higher probability. (3) Loss function: The cross-entropy loss function is adopted, which is expressed as: ； In the formula, Let be the true label of the i-th sample in class c; N is the total number of samples; (4) Backpropagation and optimization: The Adam optimizer is used to iteratively update the weights. The learning rate η = 0.001, the first moment estimation decay rate β1 = 0.9, and the second moment estimation decay rate β2 = 0.

999. The weight update formula is as follows: ； In the formula, These are network parameters, including all weights and biases; and These are the bias correction estimates for the first and second moments of the gradient, respectively; To prevent division by zero of constants; During training, each iteration uses the current batch of samples to calculate the gradient of the loss function with respect to each parameter. The error is then propagated layer by layer through the backpropagation algorithm. The Adam optimizer is used to adaptively adjust the learning rate of each parameter based on the gradient information, so that the loss function gradually converges to the minimum value, thus completing the network training.

6. The intelligent identification method for artificial rain enhancement operation conditions based on multi-source data fusion according to claim 1, characterized in that, The process of step 4 is as follows: (1) Data collection and standardization of specific regions Radar data, satellite data, and pattern data of a specific area are collected during identification, and together with basic variables, a 56-dimensional feature vector is formed after spatial matching. For each feature of the current sample, the mean of the training set is used. and standard deviation Z-score standardization is performed, and the process is represented as follows: ； Obtain the standard eigenvector Z new =(Z1,Z2,…,Z 56 ) T ; (2) Model discrimination and job level output The standard eigenvector Z new Input the trained rain enhancement operation condition discrimination model and output the probability that the grid point belongs to the category of having operation conditions. , is represented as: ； Based on probability The job condition level (SL) will be output according to the following rules, and represented as follows: ； When precipitation PRE cannot be obtained, a comprehensive judgment is made based on alternative factors, and the judgment rules are the same as above. After the grid point is identified, all grid points with operation condition level SL=1 are summarized, and the spatial distribution of rain enhancement operation condition levels and suitable operation range for a specific area are output.

7. The intelligent identification method for artificial rain enhancement operation conditions based on multi-source data fusion according to claim 1, characterized in that, The method further includes: To reflect the merits of the rain enhancement operation condition discrimination model, a confusion matrix was used to evaluate the classification results of the rain enhancement operation condition discrimination model; In the confusion matrix, the main diagonal parameters N and H represent the number of non-precipitation and precipitation events correctly identified by the model, respectively, while the secondary diagonal parameters M and F represent the number of incorrect identifications. The performance of the rain enhancement operation condition discrimination model was evaluated using TS score, POD hit rate, FAR false alarm rate, and FNR false alarm rate. The definitions of each indicator are as follows: Hit Rate POD: ; In the formula, H represents the number of correct identifications and M represents the number of missed identifications; FAR (Failure Rate): ; In the formula, F represents the number of false alarms; Failure rate (FNR): ; TS rating: .

8. The intelligent identification method for artificial rain enhancement operation conditions based on multi-source data fusion according to claim 7, characterized in that, The method further includes: Hyperparameter tuning is performed using grid search, where the hyperparameter space to be optimized is denoted as . , training set Divide the hyperparameter space into 5 equal subsets {D1, D2, ..., D5}; for each set of candidate parameters in the hyperparameter space... The k-th fold subset is used as the validation set, k=1,2,…,5, and the remaining four folds are used as the training set for model training; the TS score is used as the performance evaluation metric at the k-th fold, denoted as . The average of the 50% TS scores is taken as the hyperparameters for this group. Cross-validation performance metrics: ； Traversing the hyperparameter space Select the combination that makes the candidate combination available. The largest combination of hyperparameters is the optimal parameter. , is represented as: ； Selecting the optimal hyperparameters Optimize the model.

Images