A weather classification method of a hybrid neural network model after feature selection and clustering analysis

Abstract

The present invention discloses a weather classification method based on a hybrid neural network model after feature selection and clustering analysis. First, the maximum information coefficient (MIC) is used to select features, and features with an MIC lower than 0.5, that is, features with low correlation with other features, are removed. Then, the Bayesian information criterion (BIC) coefficient is used to estimate the optimal number of clusters k of the clustering model, and the Gaussian mixture model clustering is performed on the data set to screen similar samples. Next, an MLP classification neural network and an MLNN classification neural network are respectively constructed. Finally, the AdaBoost adaptive boosting algorithm in ensemble learning is used to sequentially train the two models and generate a hybrid model of MLP and MLNN neural networks. The present invention combines feature selection, clustering analysis and hybrid neural networks to process weather data, reducing the training time. By training a multi-layer perceptron and a morphological-linear neural network and incorporating the idea of ensemble learning, and applying the AdaBoost adaptive boosting algorithm during model training and hybridization, the accuracy of weather classification is improved.

Description

Technical Field

[0001] The present invention belongs to the fields of meteorological prediction and neural network data analysis, and particularly relates to a weather classification method of a hybrid neural network model after feature selection and clustering analysis. Background Art

[0002] Currently, there is a high demand for the applications of weather classification recognition and weather forecasting in many fields such as agriculture, air traffic services, floods, energy, and environmental control. In today's rapid development of various industries, predicting the weather in advance is crucial for protecting life and property safety. There are many modes of first performing feature selection on high-dimensional radar data through a similarity coefficient and then analyzing and studying the data set [1]. The method of screening features through the similarity between features is widely used in fields such as information retrieval, data mining, machine learning, and signal processing. Currently, the indicators widely used in research to describe the similarity between features include the Pearson correlation coefficient [2], the generalized Jaccard similarity coefficient [1], etc. However, most of the similarity indicators used can only one-sidedly describe the linear relationship between features, which will result in some high-quality features being discarded during feature selection.

[0003] For large meteorological radar data sets, dividing data samples into different clusters through a clustering method can effectively improve the accuracy of subsequent classification or prediction [1]. The goal of clustering is to group similar instances into clusters. Clustering is an important tool in aspects such as data analysis, customer segmentation, recommendation systems, search engines, image segmentation, semi-supervised learning, and dimensionality reconstruction. Currently, there are many studies on weather data or images through clustering methods. Cho et al. [3] defined the typical weather patterns in the East Asian region and objectively analyzed the weather patterns using k-means clustering. Kumpf [4] et al. observed those that stably belong to a given cluster by analyzing the sensitivity of clustering results related to the changes in the selected region, so as to identify similar cluster members, and provided a quantitative and visual representation of clustering robustness. Since weather data contains many features, the data cannot be processed quickly and effectively. In order to improve the prediction accuracy in the shortest time, Pooja et al. [1] proposed the technique of combination mapping expectation clustering based on the Tanimoto correlation and linear program enhanced classification (TC-CMECLPBC), which has obvious improvements in terms of enhancing the time of PA and the false positive rate (FPR). However, in the processing of large meteorological radar data, the application of neural network models in dealing with weather classification problems is less and not mature enough.

[0004] Recent research has shown that hybrid neural networks tend to perform better on large datasets compared to general neural networks and traditional numerical models. Among them, Tanzila Saba et al. [5] constructed a hybrid model of MLP and RBF, which outperformed the hybrid neural model and single feedforward neural network in predicting weather categories. Gerardo Hernandez et al. [6] proposed a morphological-linear neural network (MLNN) in their paper. The model combines two different types of neural layers, including a hidden layer of morphological neurons and an output layer of classical perceptrons. Experiments on 25 datasets showed that the MLNN model has an advantage over the MLP model and the DMN model in terms of classification accuracy. However, for hybrid neural network models, the model hybridization methods are relatively single, and the advantages of the models cannot be fully exploited.

[0005] References:

[0006] [1]. Pooja S.B, R.V. Siva Balan, Anisha M, M.S. Muthukumaran, Jothikumar R. Techniques. Tanimoto correlated feature selection system and hybridization of clustering and boosting ensemble classification of remote sensed big data for weather forecasting[J]. Computer Communications, 2020, 151: 266 - 274

[0007] [2] Leily Sheugh, Sasan H. Alizadeh. A note on pearson correlation coefficient as a metric of similarity in recommender system[J]. AI&Robotics(IRAN OPEN), 2015, 15: 458 - 106

[0008] [3]Cho, Young-Jun, Lee, Hyeon-Cheol, Lim, Byunghwan, Kim, Seung-Bum. Classification of Weather Patterns in the East Asia Region using the K-means Clustering Analysis[J]. Atmosphere, 2019, 29(4): 451-461

[0009] [4]Alexander Kumpf, Bianca Tost, Marlene Baumgart, Michael Riemer, Rüdiger Westermann, Marc Rautenhaus. Visualizing Confidence in Cluster-Based Ensemble Weather Forecast Analyses. IEEE, 2017, 109-119

[0010] [5]Saba T, Rehman A, AlGhamdi J S. Weather forecasting based on hybrid neural model[J]. Applied Water Science, 2017, 7(7): 3869-3874.

[0011] [6]Hernández, Gerardo, Erik Zamora, and Humberto Sossa. "Morphological-linear neural network." 2018 IEEE International Conference on Fuzzy Systems (FUZZ-IEEE). IEEE, 2018. Summary of the Invention

[0012] In view of the above problems, the present invention proposes a weather classification method for a hybrid neural network model after feature selection and clustering analysis. This method helps the hybrid neural network model process radar weather data to achieve weather classification by adding feature selection and clustering analysis methods.

[0013] Technical Solution: To achieve the above invention purpose, the present invention adopts the following technical solution: A weather classification method for a hybrid neural network model after feature selection and clustering analysis, the method comprising the following steps:

[0014] Step 1) Establish a dataset of meteorological observation elements, calculate the MIC (Maximal Information Coefficient) between pairs of weather data features, and screen the weather features according to the obtained MIC matrix to obtain a relevant weather feature dataset;

[0015] Step 2) Conduct a clustering analysis on the weather feature dataset, use the Bayesian Information Criterion to estimate the optimal number of clusters, and perform Gaussian mixture model clustering on the dataset according to the estimated number of clusters to obtain a set of weather sample clusters;

[0016] Step 3) Build an MLP (Multi-Layer Perceptron) model and an MLNN (Morphological-Linear Neural Network) model, and set the hyperparameters therein;

[0017] Step 4) Use the set of weather sample clusters to sequentially train the models in Step 3) according to the AdaBoost (Adaptive Boosting) algorithm, obtain the combined weights at the same time, and perform weighted combination on the probabilities of the weather types predicted by the two models to finally obtain the weather types.

[0018] Furthermore, the specific method of Step 1 is as follows:

[0019] Step 1.1: Establish a dataset of 8 meteorological observation elements including the average air pressure, daily maximum temperature, daily minimum temperature, daily average relative humidity, wind speed, total cloud cover, daily average air quality index, and sunshine duration after preprocessing and standardization and the result set of the type of the current weather where x represents the meteorological element value, {1,..., n} represents the serial number of the weather feature, n represents the number of weather fact samples, 8 represents the number of meteorological features, and y represents the weather type data of each meteorological data sample collection day;

[0020] Step 1.2: Calculate the MIC between pairs of weather data features. The calculation formula of the MIC for each pair of features <x j , x k > is as follows:

[0021]

[0022]

[0023] In formula (1), I(x j , x k ) represents the mutual information coefficient between features x j and x k , p(X j , X k ) represents the joint probability between feature variables x j and x k , X jand X k respectively represent the sample value sets of feature x j and x k i.e., p(X j ) and p(X k ) respectively represent the probability functions on X j and X k ;

[0024] In formula (2), mic(x j , x k ) represents the maximum information coefficient between feature variables x j and x k . The variables (x j , x k ) are regarded as scatter points projected onto a two-dimensional space to create a scatter plot. a and b are the numbers of divided regions in the directions of x j and x k respectively. B is a parameter variable, and its size is set to the 0.6th power of the sample size n, i.e., n 0.6 ;

[0025] Finally, the maximum information coefficient matrix of MIC between features is calculated m jk represents the maximum information coefficient of MIC between feature x j and x k , indicating the correlation between these two features;

[0026] Step 1.3: Through the maximum information coefficient matrix of MIC obtained in Step 1.2, select the features of the meteorological data set, and retain the features with the maximum information coefficient of MIC greater than 0.5 in the meteorological data set. That is, if m jk ≤0.5, then remove feature x j and x k . Finally, the filtered data set and its result set g represents the total number of features retained after screening.

[0027] Furthermore, the specific method of Step 2 is as follows:

[0028] Step 2.1: Before performing cluster analysis on the filtered meteorological data set, first use the Bayesian information criterion BIC to estimate the optimal number of clusters t of the cluster model, where t ∈ {t|0 < t ≤ 10, t ∈ N +}. Calculate its BIC estimated value for each t value. The formula is as follows:

[0029]

[0030] Formula (3) where n is the number of samples in the feature dataset, d is the parameter learned by the model, is the maximum value of the likelihood function of the model, and the t value when the BIC estimate is the lowest is selected as the number of clusters for clustering;

[0031] Step 2.2: Use the number of clusters t of the clustering model obtained in Step 2.1 as the clustering hyperparameter of the Gaussian mixture model, and for the filtered dataset perform clustering analysis to obtain the clustered sample clusters D = {s 1 , s 2 ,... s t}, where s j represents different clusters, with a total of t clusters, and each cluster contains several weather data samples.

[0032] Furthermore, the method of Step 3 is specifically as follows:

[0033] Step 3.1: Build an MLP classification neural network, where the number of neurons in the input layer is the number of input features g, the number of hidden layers is 5, the number of neurons in each hidden layer is 50, the learning rate is set to 0.1, the activation function of the output layer is the softmax activation function, the number of neurons in the output layer is the number of weather types h, and dropout training with a probability of p = 0.3 is added to the input neurons and hidden layer neurons, and finally the probability of the predicted weather type is output where f j represents the probability that the weather type of this sample data is the j-th type;

[0034] Step 3.2: Build an MLNN classification neural network, where the number of neurons in the input layer is the number of input features g, the number of hidden layers is 1, consisting of 150 morphological-linear neurons, the learning rate is set to 0.05, the activation function of the output layer is the softmax activation function, the number of neurons in the output layer is the number of weather types h, and dropout training with a probability of p = 0.3 is added to the input neurons and hidden layer neurons, and finally the probability of the predicted weather type is output where f j represents the probability that the weather type of this sample data is the j-th type.

[0035] Furthermore, the method of Step 4 is specifically as follows:

[0036] Step 4.1: Using the idea of ensemble learning, use the AdaBoost adaptive boosting algorithm, and use the clustering result D = {s 1 , s 2 ,... s tTrain the neural network built in Steps 3.1 and 3.2 according to different clusters in sequence. The AdaBoost adaptive boosting algorithm is as follows:

[0037] Calculate the weighted error rate e of the k-th classifier G k (x) during training k , and the formula is as follows:

[0038]

[0039] Calculate the weight coefficient α of the k-th classifier G k (x) through the weighted error rate e k , and the formula is as follows: k

[0040]

[0041] In formula (4), K represents the number of classifiers. In the present invention, K takes the value of 2. G k (x) represents the k-th classifier, the value set of k is {1, 2}, n represents the number of samples in the training set, i represents the i-th training sample, and w ki represents the training weight of the k-th trainer for the i-th training sample. Each training sample is initially given the same weight value, that is y i represents the result of the i-th training sample. I(G k (x i )≠y i ) represents the judgment on whether the classifier result G k (x i ) and the result y of the training sample i are different. When G k (x i )≠y i , the value of I(G k (x i )≠y i ) is 1, otherwise it is 0;

[0042] First, train the MLP multi-layer perceptron, then strengthen the training set through the training results, then train the MLNN morphological-linear neural network, and calculate the hybrid weights α k of the two models. Finally, obtain the trained MLP multi-layer perceptron model G 1 (x) and the MLNN morphological-linear neural network model G 2 (x) and their respective hybrid weights α 1 and α 2 ;

[0043] Step 4.2: According to the single classifier weight α obtained in Step 4.1 k , integrate and mix the two already trained classifiers G k (x), and the mixing formula is as follows:

[0044] F 混 =a 1 F 1 +a 2 F 2 (6)

[0045] In formula (6), α 1 and α 2 are the model mixing weights calculated in Step 4.1, F 1 and F 2 represent the prediction results of the two models, and the final mixing result is wherein, f k represents the probability that the weather type of the sample data is the k-th type, and select the weather type with the highest probability in as the weather type predicted by the final mixing model.

[0046] Beneficial effects: Compared with the prior art, the technical solution of the present invention has the following beneficial technical effects:

[0047] The present invention proposes a classification method for a hybrid neural network model combined with feature selection and clustering analysis. The feature extraction index used therein, namely the MIC maximum information coefficient, can effectively indicate the linear and non-linear correlation degrees between features. When removing features with poor correlation, it can also retain non-linearly correlated features that cannot be indicated by other correlation parameters, thereby reducing the computational cost and improving the classification performance. Different from the traditional method of directly fitting a classification model after feature processing, the present invention uses a clustering analysis method to analyze the data set after feature selection, clusters the weather data samples into different clusters, and then establishes a hybrid neural network classification model according to the clustering results, which can achieve a good auxiliary classification effect and effectively improve the performance of the classifier. The present invention selects two neural network models that perform best and have relatively simple structures in multi-classification tasks, namely the MLNN morphological-linear neural network and the MLP multi-layer perceptron, and mixes them through an ensemble learning method to finally obtain a neural network hybrid model. It retains the characteristics of a simple model, requires fewer hyperparameters to be specified compared to other neural network models, and has a shorter calculation time compared to complex hybrid neural network models. Using the AdaBoost adaptive boosting algorithm in ensemble learning to perform weighted training on the data set and weighted mixing of the classification models for the two neural networks can effectively make up for the shortcomings of the two independent models, increase the dependence between the two models, and obtain better training results. Description of the Drawings

[0048] Figure 1 It is a schematic diagram of a multi-layer perceptron (MLP) neural network;

[0049] Figure 2 It is a schematic diagram of a morphological-linear neural network (MLNN);

[0050] Figure 3 It is a schematic diagram of the implementation process of a hybrid neural network;

[0051] Figure 4 It is a flowchart of the method of the present invention. Specific implementation scheme

[0052] The following further describes the implementation of the technical solution in detail with reference to the accompanying drawings:

[0053] The present invention will be further described in detail with reference to the flowchart and implementation cases for the classification technology of the hybrid neural model based on maximum mutual information coefficient feature selection and clustering.

[0054] The present invention proposes a weather classification method for a hybrid neural network model after feature selection and clustering analysis, and the method includes the following steps:

[0055] Step 1) Establish a dataset of meteorological observation elements, calculate the maximum information coefficient (MIC) between the weather data features pairwise, and screen the weather features according to the obtained MIC matrix to obtain a relevant weather feature dataset;

[0056] Step 2) Perform clustering analysis on the weather feature dataset, estimate the optimal number of clusters using the Bayesian information criterion, and perform Gaussian mixture model clustering on the dataset according to the estimated number of clusters to obtain a set of weather sample clusters;

[0057] Step 3) Build a multi-layer perceptron (MLP) model and a morphological-linear neural network (MLNN) model, and set the hyperparameters therein;

[0058] Step 4) Use the set of weather sample clusters to sequentially train the models in Step 3 according to the AdaBoost adaptive boosting algorithm, and at the same time obtain the hybrid weights, and perform weighted mixing on the probabilities of the weather types predicted by the two models to finally obtain the weather types.

[0059] Furthermore, the specific method of Step 1 is as follows:

[0060] Step 1.1: Establish a dataset of 8 meteorological observation elements including the average air pressure, daily maximum temperature, daily minimum temperature, daily average relative humidity, wind speed, total cloud cover, daily average air quality index, and sunshine duration after preprocessing and standardization and the result set of the type of the current weather Among them, x represents the meteorological element value, {1,..., n} represents the serial numbers of weather characteristics, n represents the number of weather fact samples, 8 represents the number of meteorological characteristics, and y represents the weather type data on the day when each meteorological data sample is collected;

[0061] Step 1.2: Calculate the maximum information coefficient (MIC) between weather data characteristics pairwise. The calculation formula for the MIC of each pair of feature pairs <x j , x k > is as follows:

[0062]

[0063]

[0064] In formula (1), I(x j , x k ) represents the mutual information coefficient between features x j and x k . p(X j , X k ) represents the joint probability between feature variables x j and x k . X j and X k respectively represent the sample value sets of features x j and x k , that is p(X j ) and p(X k ) respectively represent the probability functions on X j and X k ;

[0065] In formula (2), mic(x j , x k ) represents the maximum information coefficient between feature variables x j and x k . The variables (x j , x k ) are regarded as scatter points projected into a two-dimensional space to make a scatter plot. a and b are the numbers of divided regions in the directions of x j and x k respectively. B is a parameter variable, and its size is set to the 0.6th power of the sample size n, that is n 0.6 ;

[0066] Finally, the maximum information coefficient matrix of features is calculated m jk represents features x j and x kThe MIC maximum information coefficient between them indicates the correlation between these two features;

[0067] Step 1.3: Through the MIC maximum information coefficient matrix obtained in Step 1.2, select the features of the meteorological dataset, and retain the features in the meteorological dataset with a MIC maximum information coefficient greater than 0.5 related, that is, if m jk ≤0.5, then remove the features x j and x k , and finally obtain the filtered dataset and its result set g represents the total number of features retained after screening.

[0068] Furthermore, the specific method of Step 2 is as follows:

[0069] Step 2.1: Before performing cluster analysis on the filtered meteorological dataset, first use the Bayesian Information Criterion BIC to estimate the optimal number of clusters t of the clustering model, where t ∈ {t|0 < t ≤ 10, t ∈ N +}}, calculate its BIC estimate value for each t value, and the formula is as follows:

[0070]

[0071] Formula (3) where n is the number of samples in the feature dataset, d is the parameter learned by the model, is the maximum value of the likelihood function of the model, and select the t value when the BIC estimate value is the lowest as the number of clusters for clustering;

[0072] Step 2.2: Use the number of clusters t obtained in Step 2.1 as the clustering hyperparameter of the Gaussian mixture model, and perform cluster analysis on the filtered dataset to obtain the clustered sample clusters D = {s 1 , s 2 ,... s t}, where s j represents different clusters, a total of t clusters, and each cluster contains several weather data samples.

[0073] Furthermore, the method of Step 3 is as follows:

[0074] Step 3.1: Build an MLP classification neural network, where the number of neurons in the input layer is the number of input features g, the number of hidden layers is 5, the number of neurons in each hidden layer is 50, the learning rate is set to 0.1, the activation function of the output layer is the softmax activation function, the number of neurons in the output layer is the number of weather types h, and dropout training with a probability of p = 0.3 is added between the input neurons and the hidden layer neurons, and finally output the probability of the predicted weather type where f j represents the probability that the weather type of the sample data is the j-th type;

[0075] Step 3.2: Build an MLNN classification neural network, where the number of neurons in the input layer is the number of input features g, the number of hidden layers is 1, consisting of 150 morphological-linear neurons, the learning rate is set to 0.05, the activation function of the output layer is the softmax activation function, the number of neurons in the output layer is the number of weather types h, and dropout training with a probability of p = 0.3 is added to the input neurons and hidden layer neurons, and finally the probability of the predicted weather type is output where f j represents the probability that the weather type of the sample data is the j-th type.

[0076] Furthermore, the method in Step 4 is specifically as follows:

[0077] Step 4.1: Use the idea of ensemble learning to use the AdaBoost adaptive boosting algorithm, and use the clustering result D = {s 1 , s 2 ,... s t} in Step 2.2 to sequentially train the neural network built in Step 3.1 and Step 3.2 according to different cluster pairs. The AdaBoost adaptive boosting algorithm is as follows:

[0078] Calculate the weighted error rate e k of the k-th classifier G k (x) during training. The formula is as follows:

[0079]

[0080] Calculate the weight coefficient α k of the k-th classifier G k (x) through the weighted error rate e k . The formula is as follows:

[0081]

[0082] In formula (4), K represents the number of classifiers. In the present invention, K takes the value of 2. G k (x) represents the k-th classifier. The value set of k is {1, 2}. n represents the number of samples in the training set. i represents the i-th training sample. w ki represents the training weight of the k-th trainer for the i-th training sample. Each training sample is initially given the same weight value, that is y i represents the result of the i-th training sample. I(G k (x i ) ≠ y i)Indicates the result G of the classifier k (x i ) and the result y of the training sample i to determine whether they are different. When G k (x i ) ≠ y i , I(G k (x i ) ≠ y i ) has a value of 1; otherwise, it is 0;

[0083] First, train the MLP multi-layer perceptron, then strengthen the training set with the training results, then train the MLNN morphological-linear neural network, and calculate the mixing weights α k of the two models. Finally, obtain the trained MLP multi-layer perceptron model G 1 (x) and the MLNN morphological-linear neural network model G 2 (x) and their respective mixing weights α 1 and α 2 ;

[0084] Step 4.2: According to the single classifier weight α k obtained in Step 4.1, integrate and mix the two trained classifiers G k (x). The mixing formula is as follows:

[0085] F 混 = a 1 F 1 + a 2 F 2 (6)

[0086] In formula (6), α 1 and α 2 are the model mixing weights calculated in Step 4.1, F 1 and F 2 represent the prediction results of the two models. The final mixing result is where f k represents the probability that the weather type of the sample data is the k-th type. Select the weather type with the highest probability as the weather type predicted by the final mixing model.

[0087] The specific implementation described above further details the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above is only the specific implementation of the present invention and is not intended to limit the scope of the present invention. Any equivalent changes and modifications made by those skilled in the art without departing from the concept and principles of the present invention shall fall within the scope of protection of the present invention.

Claims

1. A weather classification method based on a hybrid neural network model after feature selection and clustering analysis, characterized in that, the method comprises the following steps: Step 1) Establish a meteorological observation element data set, calculate the maximum information coefficient (MIC) between the weather data features pairwise, and screen the weather features according to the obtained MIC maximum information coefficient matrix to obtain a relevant weather feature data set; Step 2) Conduct clustering analysis on the weather feature data set, use the Bayesian information criterion to estimate the optimal number of clusters, and perform Gaussian mixture model clustering on the data set according to the estimated number of clusters to obtain a weather sample cluster set; Step 3) Build a multi-layer perceptron (MLP) model and a morphological-linear neural network (MLNN) model, and set the hyperparameters therein; Step 4) Use the weather sample cluster set to sequentially train the models in Step 3) according to the AdaBoost adaptive boosting algorithm, and at the same time obtain the hybrid weights, and perform weighted mixing on the probabilities of the weather types predicted by the two models to finally obtain the weather types.

2. The weather classification method based on a hybrid neural network model after feature selection and clustering analysis according to claim 1, characterized in that, the specific method of Step 1 is as follows: Step 1.1: Establish a dataset of 8 meteorological observation elements, namely the average air pressure, daily maximum temperature, daily minimum temperature, daily average relative humidity, wind speed, total cloud cover, daily average air quality index, and sunshine duration, after preprocessing standardization and the result set of the type of the weather at that time Among them, x represents the meteorological element value, {1,..., n} represents the serial number of the weather characteristics, n represents the number of weather fact samples, 8 represents the number of meteorological characteristics, and y represents the weather type data of the day when each meteorological data sample is collected; Step 1.2: Calculate the MIC (Maximal Information Coefficient) between weather data features pairwise. The calculation formula for the MIC of each pair of feature pairs <x j , x k > is as follows: In formula (1), I(x j ,x k ) represents the mutual information coefficient between features x j and x k . p(X j ,X k ) represents the joint probability between feature variables x j and x k . X j and X k respectively represent the sample value sets of features x j and x k , that is p(X j ) and p(X k ) respectively represent the probability functions on X j and X k ; In formula (2), mic(x j , x k ) represents the maximum information coefficient between the feature variables x j and x k . The variables (x j , x k ) are regarded as scatter points projected onto a two-dimensional space to create a scatter plot. a and b are the numbers of divided regions in the directions of x j and x k respectively. B is a parameter variable, and its size is set to the 0.6th power of the sample size n, that is, n 0.6 ; Finally, the maximum information coefficient matrix of MIC between features is calculated m jk represents feature x j and x k the maximum information coefficient of MIC between them, indicating the correlation between these two features; Step 1.3: Select the features of the meteorological dataset through the MIC maximum information coefficient matrix obtained in Step 1.2, and retain the features in the meteorological dataset where the MIC maximum information coefficient is greater than 0.5, that is, if m jk ≤0.5, then remove the features x j and x k , and finally obtain the filtered dataset and its result set g represents the total number of features retained after screening.

3. The weather classification method based on a hybrid neural network model after feature selection and clustering analysis according to claim 2, characterized in that, the specific method of Step 2 is as follows: Step 2.1: Before performing cluster analysis on the filtered meteorological data set, first use the Bayesian Information Criterion (BIC) to estimate the optimal number of clusters t of the clustering model, where t ∈ {t|0 < t ≤ 10, t ∈ N +}, calculate its BIC estimated value for each t value, and the formula is as follows: Formula (3) where n is the number of samples in the feature dataset and d is the parameter learned by the model, is the maximum value of the likelihood function of the model, and the t value with the lowest BIC estimate is selected as the number of clusters for clustering; Step 2.2: Use the number of clusters t obtained in Step 2.1 as the clustering hyperparameter of the Gaussian mixture model, and apply it to the filtered dataset for clustering analysis to obtain the clustered sample clusters D = {s 1 , s 2 ,... s t}, where s j represents different clusters, with a total of t clusters, and each cluster contains several weather data samples.

4. The weather classification method based on a hybrid neural network model after feature selection and clustering analysis according to claim 3, characterized in that, the method of Step 3 is specifically as follows: Step 3.1: Build an MLP classification neural network, where the number of neurons in the input layer is the number of input features g, the number of hidden layers is 5, the number of neurons in each hidden layer is 50, the learning rate is set to 0.1, the activation function of the output layer is the soft max activation function, the number of neurons in the output layer is the number of weather types h, and dropout training with a probability of p = 0.3 is added to the input neurons and hidden layer neurons, and finally the probability of the predicted weather type is output where, f j represents the probability that the weather type of the sample data is the jth type; Step 3.2: Build an MLNN classification neural network, where the number of neurons in the input layer is the number of input features g, the number of hidden layers is 1, which consists of 150 morphological-linear neurons, the learning rate is set to 0.05, the activation function of the output layer is the softmax activation function, the number of neurons in the output layer is the number of weather types h, and dropout training with a probability of p = 0.3 is added to the input neurons and hidden layer neurons, and finally the probability of the predicted weather type is output where f j represents the probability that the weather type of the sample data is the j-th type.

5. The weather classification method based on a hybrid neural network model after feature selection and clustering analysis according to claim 4, characterized in that, the method of Step 4 is specifically as follows: Step 4.1: Using the idea of ensemble learning, the AdaBoost adaptive boosting algorithm is used to sequentially train the neural network constructed in Steps 3.1 and 3.2 according to different cluster pairs using the clustering result D = {s 1 , s 2 ,... s t}. The AdaBoost adaptive boosting algorithm is as follows: Calculate the k-th classifier G k (x) weighted error rate e during training k , the formula is as follows: Through the weighted error rate e k Calculate the weight coefficient α k of the k-th classifier G k (x), and the formula is as follows: In formula (4), K represents the number of classifiers, and the value of K is 2. G k (x) represents the k-th classifier, and the value set of k is {1, 2}. n represents the number of samples in the training set, i represents the i-th training sample, and w ki represents the training weight of the k-th trainer for the i-th training sample. Each training sample is initially assigned the same weight, that is, y i represents the result of the i-th training sample. I(G k (x i )≠y i ) represents the judgment of whether the classifier result G k (x i ) and the result y of the training sample i are different. When G k (x i )≠y i , the value of I(G k (x i )≠y i ) is 1, otherwise it is 0; First, train the MLP (Multi-Layer Perceptron), then reinforce the training set with the training results, then train the MLNN (Morphological-Linear Neural Network), and calculate the hybrid weights α of the two models k , and finally obtain the trained MLP (Multi-Layer Perceptron) model G 1 (x) and the MLNN (Morphological-Linear Neural Network) model G 2 (x) and their respective hybrid weights α 1 and α 2 ; Step 4.2: According to the single classifier weight α obtained in Step 4.1 k , integrate and mix the two already trained classifiers G k (x), and the mixing formula is as follows: F 混 = a 1 F 1 + a 2 F 2 (6) In formula (6), α 1 and α 2 are the model mixing weights calculated in step 4.1, F 1 and F 2 represent the prediction results of two models, and the final mixing result is where f k represents the probability that the weather type of this sample data is the k-th type, and the weather type with the highest probability in is selected as the weather type predicted by the final mixing model.

Images