A drought disaster weather prediction method based on semi-supervised ensemble learning

Abstract

The present invention relates to the field of semi-supervised and integrated learning technology, and specifically to a drought disaster weather prediction method based on semi-supervised integrated learning. In a semi-supervised learning algorithm based on self-training, a dynamic sample data selection algorithm is constructed, and sample data can be dynamically selected according to the confidence level of the model's prediction of unlabeled data during the training sample process, effectively improving the model training efficiency while ensuring a high accuracy rate. The proposed selective integrated learning pruning algorithm based on target maximization cuckoo optimization can meet the accuracy of base learners and the differences between base learners to the greatest extent, adopts multi-angle selective integration, has better generalization performance, and has faster algorithm operation efficiency and higher model accuracy. The present invention can be used for the prediction of drought disaster weather in the meteorological field.

Description

Technical Field

[0001] The invention relates to the field of semi-supervised and integrated learning, and in particular to a drought disaster weather prediction method based on semi-supervised integrated learning. Background Art

[0002] Drought weather in meteorological disasters has a serious impact on the natural environment, human life, and social economy. Effective prediction of meteorological drought plays an important role in the natural resource conditions of the basin, regional water resources planning and management, and alleviating the harmful effects of drought. It also helps relevant departments to optimize the operation of the water resources system and make corresponding drought prevention and disaster reduction measures and decision-making analysis. Therefore, how to make drought disaster weather forecasts more objective, quantitative, and accurate has become a crucial content in drought disaster weather forecasting business. At present, semi-supervised ensemble learning methods are often used in drought disaster weather forecasting.

[0003] The main problem of semi-supervised learning is the problem of using unlabeled sample data. At present, domestic and foreign research mainly focuses on the algorithm for removing false labels. The existing algorithm for removing false labels fixes the number of unlabeled samples to be selected when selecting unlabeled sample data, and does not consider the low confidence level in the early stage of model training. In addition, the fixed number of unlabeled samples will lead to a long model training time. The main problem of ensemble learning is the selective ensemble problem. At present, domestic and foreign research focuses on measurement and optimization algorithms. The existing measurement algorithms only use single angles such as sorting for measurement, and cannot comprehensively consider the overall base learner, resulting in poor generalization performance. In addition, when some optimization algorithms are selectively integrated, the convergence speed is very slow and the model training time is very long. Summary of the invention

[0004] Purpose of the Invention

[0005] The present invention proposes a drought disaster weather prediction method based on semi-supervised ensemble learning, which aims to solve the problems of low confidence level, slow training speed and poor generalization performance in the current drought disaster weather prediction based on semi-supervised ensemble learning.

[0006] Technical Solution

[0007] A drought disaster weather prediction method based on semi-supervised ensemble learning is characterized by performing the following steps:

[0008] Step (1): Establish a data sample set for predicting drought disaster weather;

[0009] Step (2): Perform data balancing on drought disaster weather data, and perform feature construction and feature selection;

[0010] Step (3): using self-training semi-supervised learning to use the labeled sample data and unlabeled sample data in the drought disaster weather dataset, and constructing a dynamic sample data selection algorithm in the self-training process to label the unlabeled sample data with pseudo labels;

[0011] Step (4): construct an objective function based on the Kappa measure and the standardized mutual information, and measure the difference and accuracy of the base learner used to train the drought disaster weather data;

[0012] Step (5): Use the cuckoo algorithm to optimize the objective function, construct a selective ensemble learning pruning algorithm based on target maximization cuckoo optimization, perform selective integration of ensemble learning according to the constructed algorithm, use the constructed algorithm to find the optimal subset for the initial base learner set, and use the optimal subset as the base learner of Stacking ensemble learning to train the drought disaster weather prediction model.

[0013] In the step (1), the drought disaster weather data specifically include the minimum wind speed at 10 meters, the relative humidity at 2 meters, the temperature range at 2 meters, the wind speed at 10 meters, the temperature at 2 meters, the minimum wind speed at 50 meters, the maximum temperature at 2 meters, the wind speed at 50 meters, the earth's skin temperature, the wind speed range at 50 meters, the maximum wind speed at 50 meters, the maximum wind speed at 10 meters, the wind speed range at 10 meters, the surface pressure, the dew point / freezing point at 2 meters, the minimum temperature at 2 meters, the wet bulb temperature at 2 meters, the precipitation, and the historical real-time data of the drought level.

[0014] In the step (2), the drought disaster weather data is subjected to data balancing processing, and an algorithm combining an oversampling algorithm for synthesizing minority classes with an undersampling algorithm for deleting majority class noise samples is adopted. The oversampling and undersampling techniques are combined for mixed resampling to increase the number of minority class sample data in the data set and reduce the number of majority class sample data in the data set.

[0015] In the step (2), the drought disaster weather data is feature constructed. Among the 18 meteorological attributes of the drought disaster weather data, each attribute feature involves 90 days of data. The mean of the 30-day, 60-day, and 90-day data of each attribute is counted, and the maximum value, minimum value, and median of each attribute data are classified and summarized to obtain a new feature set.

[0016] In the step (2), feature selection is performed on the drought disaster weather data, and a gradient boosting algorithm is used to obtain a weight score of each attribute feature of the drought disaster weather data, and features with a weight greater than an average weight are selected; a correlation coefficient is used to calculate the linear correlation between each attribute feature of the data and the label, and the correlation coefficient value is between -1 and 1, and a value of 0 indicates no linear correlation. The absolute values ​​of the correlation coefficients are sorted and the top 50% of the attribute features are selected; finally, the two attribute feature sets selected based on the gradient boosting weight and the correlation coefficient are combined to obtain a new attribute feature set;

[0017] In step (3), a dynamic sample data selection algorithm is constructed. The algorithm dynamically selects the optimal unlabeled sample data according to the change of the number of iterations. The process is as follows:

[0018] The labeled sample data in the drought disaster weather data is trained using the K nearest neighbor supervised learning method. In each subsequent round of training, the trained model is repeatedly used to predict the unlabeled sample data and add pseudo labels. After the prediction is completed, in the next round, the optimal unlabeled sample data selected by the dynamic sample data selection algorithm of formula (1) is added to the current training sample data set together with its pseudo label, and the result generated by itself is used to train itself again until all the unlabeled sample data are marked with pseudo labels.

[0019]

[0020] Where S is the total amount of unlabeled sample data selected in each iteration, D is the set of drought disaster weather dataset categories, g is the specific category of drought disaster weather dataset, Itea is the current iteration number, MaxItea is the maximum iteration number, Num g It is the number of unlabeled sample data of the i-th category under the current iteration number Itea.

[0021] In step (4), the objective function is constructed based on the Kappa measure and the standardized mutual information, and the process is as follows:

[0022] The objective function is constructed using Kappa measure and standardized mutual information. Kappa measure is used to measure the differences between base learners in selective ensemble learning, and standardized mutual information is used to measure the accuracy of base learners in selective ensemble learning. Kappa measure is calculated using the predicted label values ​​of two base learners, and standardized mutual information is calculated using the predicted label value of a base learner and the true label value of the current sample data. The objective function constructed by combining the two is formula (7).

[0023]

[0024] Where G is the set of predicted label values ​​of the initial base learners such as decision tree, logistic regression, and K nearest neighbor, and h m ,h n The predicted label values ​​of two different base learners in the initial base learner prediction label value set G, t is the number of base learners, u 2 is the true label value of the sample data, K(h m ,h n ) is the Kappa measurement calculation formula, NMI(h m ,u 2 ) is the calculation formula of normalized mutual information.

[0025] In the step (5), a selective ensemble learning pruning algorithm based on target maximization cuckoo optimization is constructed, and the process is as follows:

[0026] The initial base learner set is randomly selected to obtain multiple base learner subsets, and each base learner subset is used as a set of solutions to the objective function. Then, the objective function is used as the fitness function of the cuckoo algorithm and the fitness value of each base learner subset is calculated. The base learner subsets are sorted according to the calculated fitness values, and the base learner subsets with poor fitness values ​​are discarded. The Levy flight algorithm is used to generate new base learner subsets. After repeated iterations, the base learner subset with the largest fitness value is found, and this base learner subset is used as the optimal solution of the objective function, which is the base learner subset used for selective integration. The objective function optimization calculation is used to obtain formula (8);

[0027]

[0028] Where T is the initial base learner set, P is an optimal subset found in the initial base learner set, and h p ,h q are two different base learners in P, r is the number of base learners, u 3 is the true label value of the sample data, K(h p ,h q ) is the Kappa measurement calculation formula, NMI(h p ,u 3 ) is the calculation formula of normalized mutual information.

[0029] Advantages and effects:

[0030] 1. The present invention constructs a dynamic sample data selection algorithm in a semi-supervised learning algorithm based on self-training, and can dynamically select sample data according to the confidence level of the model's prediction of unlabeled sample data during the sample data training process, thereby improving the model training efficiency while ensuring a high accuracy rate of the final result.

[0031] 2. The selective ensemble learning pruning algorithm based on target maximization cuckoo optimization can meet the accuracy of base learners and the differences between base learners to the greatest extent. The objective function is constructed using Kappa measure and NMI standardized mutual information for multi-angle pruning, which improves the generalization performance, operation efficiency and final result accuracy. It can be used for the prediction of drought disaster weather in the meteorological field. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 This is a general flow chart of the drought disaster weather prediction method based on semi-supervised ensemble learning described in the present invention;

[0033] Figure 2 The dynamic selection sample data change curve diagram of the present invention;

[0034] Figure 3 It is a line graph of fitness value changes of the selective ensemble learning pruning algorithm based on target maximization cuckoo optimization described in the present invention. DETAILED DESCRIPTION

[0035] A drought disaster weather prediction method based on semi-supervised ensemble learning, characterized by comprising the following steps:

[0036] Step (1): Establish a data sample set for predicting drought disaster weather, which specifically includes WS10M_MIN minimum wind speed 10 meters (m / s), QV2M 2 meters relative humidity (g / kg), T2M_RANGE 2 meters temperature range (℃), WS10M wind speed 10 meters (m / s), T2M 2 meters temperature (℃), WS50M_MIN minimum wind speed at 50 meters (m / s), T2M_MAX 2 meters maximum temperature (℃), WS50M 50 meters wind speed (m / s), TS earth skin temperature (℃), WS50M_RANGE 50 meters wind speed range (m / s), WS50M_MAX 50 meters maximum wind speed (m / s), WS10M_MAX 10 meters maximum wind speed (m / s), WS10M_RANGE 10 meters wind speed range (m / s), PS surface pressure (kPa), T2MDEW 2-meter dew point / freezing point (℃), T2M_MIN 2-meter minimum temperature (℃), T2MWET 2-meter wet bulb temperature (℃), PRECTOT precipitation (mm day-1), and historical real-time data of drought level;

[0037] Step (2): Perform data balancing on drought disaster weather data, and perform feature construction and feature selection;

[0038] Step (3): Use self-training semi-supervised learning to reasonably use the labeled sample data and unlabeled sample data in the drought disaster weather dataset, and construct a dynamic sample data selection algorithm in the self-training process to label the unlabeled sample data with pseudo labels;

[0039] Step (4): construct an objective function based on Kappa measure and normalized mutual information (NMI), and measure the difference and accuracy of the base learner used to train drought disaster weather data;

[0040] Step (5): Use the cuckoo algorithm to optimize the objective function, and then construct a selective ensemble learning pruning algorithm based on target maximization cuckoo optimization. According to the constructed algorithm, selective integration is performed on the ensemble learning. The constructed algorithm is used to find the optimal subset for the initial base learner set, and the optimal subset is used as the base learner of the Stacking ensemble learning to train the drought disaster weather prediction model.

[0041] In the step (2), the drought disaster weather data is subjected to data balancing processing, and an algorithm combining SMOTE (Synthetic Minority Oversampling Technique) and Tomek Link is used. The SMOTH algorithm is an oversampling algorithm for synthesizing minority classes, and the Tomek Link algorithm is an undersampling algorithm for deleting majority class noise samples. The oversampling and undersampling techniques are combined for mixed resampling to increase the number of minority class sample data in the data set and reduce the number of majority class sample data in the data set. The algorithm combining SMOTE and Tomek Link can effectively balance the training data set and improve the accuracy of model classification.

[0042] In the step (2), the drought disaster weather data is feature constructed. Among the 18 meteorological attributes of the drought disaster weather data, each attribute feature involves 90 days of data, and the mean of the 30-day, 60-day, and 90-day data of each attribute is counted, and the maximum value, minimum value, and median value of each attribute data are classified and summarized to obtain a new feature set. The new feature set can better reflect the overall distribution of the data set, and the features therein can be used to better mine data information and have higher value.

[0043] In the step (2), feature selection is performed on the drought disaster weather data, and the gradient boosting algorithm (ExtremeGradient Boosting, XGBoost) algorithm is used to obtain the weight score of each attribute feature of the drought disaster weather data, and the features greater than the average weight are selected. The correlation coefficient Spearman is used to calculate the linear correlation between each attribute feature of the data and the label. The Spearman correlation coefficient uses the rank size of the two variables for linear correlation analysis. Its value is between -1 and 1, and a value of 0 indicates no linear correlation. The absolute value of the Spearman correlation coefficient is sorted and the top 50% of the attribute features are selected. Finally, the two attribute feature sets selected based on the XGboost weight and the Spearman correlation coefficient are combined to obtain a new attribute feature set. Using XGBoost and Spearman in combination to select features effectively reduces the amount of calculation in the training process and improves the accuracy of model prediction.

[0044] In step (3), a dynamic sample data selection algorithm is constructed. The algorithm can dynamically select the optimal unlabeled sample data according to the change of the number of iterations. The process is as follows:

[0045] The labeled sample data in the drought disaster weather data is trained using the K-nearest neighbors (KNN) supervised learning method. In each subsequent round of training, the trained model is repeatedly used to predict the unlabeled sample data and add pseudo labels. After the prediction is completed, in the next round, the optimal unlabeled sample data selected by the dynamic sample data selection algorithm of formula (1) and its pseudo label are added to the current training sample data set, and the results generated by itself are used to train itself again until all unlabeled sample data are pseudo-labeled. Using the dynamic sample data selection algorithm can greatly shorten the total training time, and the accuracy of the trained model is higher, which effectively improves the training efficiency of self-training semi-supervised learning;

[0046]

[0047] Where S is the total amount of unlabeled sample data selected in each iteration, D is the set of drought disaster weather dataset categories, g is the specific category of drought disaster weather dataset, Itea is the current iteration number, MaxItea is the maximum iteration number, Num g It is the number of unlabeled sample data of the i-th category under the current iteration number Itea.

[0048] In step (4), the objective function is constructed based on the Kappa measure and the NMI standardized mutual information, and the process is as follows:

[0049] The objective function is constructed using Kappa measure and NMI standardized mutual information. Kappa measure is used to measure the differences between base learners in selective ensemble learning, and NMI standardized mutual information is used to measure the accuracy of base learners in selective ensemble learning. Kappa measure is calculated using the predicted label values ​​of two base learners, and NMI standardized mutual information is calculated using the predicted label value of a base learner and the true label value of the current sample data. The objective function constructed by combining the two is formula (7). Using Kappa measure and NMI standardized mutual information to construct the objective function from multiple angles makes the model more generalizable;

[0050]

[0051] Where G is the set of predicted label values ​​of the initial base learners such as decision tree, logistic regression, and K nearest neighbor, and h m ,h n The predicted label values ​​of two different base learners in the initial base learner prediction label value set G, t is the number of base learners, u 2 is the true label value of the sample data, K(h m ,h n ) is the Kappa measurement calculation formula, NMI(h m ,u 2 ) is the calculation formula of NMI normalized mutual information.

[0052] In the step (5), a selective ensemble learning pruning algorithm based on target maximization cuckoo optimization is constructed, and the process is as follows:

[0053] The initial base learner set is randomly selected to obtain multiple base learner subsets. Each base learner subset is used as a set of solutions to the objective function. Then, the objective function is used as the fitness function of the cuckoo algorithm and the fitness value of each base learner subset is calculated. The base learner subsets are sorted according to the calculated fitness value, and the base learner subsets with poor fitness values ​​are discarded. The Levy flight algorithm is used to generate a new base learner subset. Levy flight can increase the global search ability of the algorithm in the process of finding the optimal solution. A larger step size can prevent the algorithm from falling into a local extreme value, and a smaller step size can accelerate the convergence of individuals, improve the efficiency of the algorithm and the accuracy of optimization. After repeated iterations, the base learner subset with the largest fitness value is found, and this base learner subset is used as the optimal solution of the objective function, which is the base learner subset for selective integration. The formula (8) is obtained by optimizing the objective function. Compared with other pruning algorithms, the selective ensemble learning pruning algorithm based on target maximization cuckoo optimization can effectively improve the training efficiency of selective ensemble learning, and the base learners selected by the algorithm have better generalization performance, faster convergence speed and higher accuracy.

[0054]

[0055] Where T is the initial base learner set, P is an optimal subset found in the initial base learner set, and h p ,h q are two different base learners in P, r is the number of base learners, u 3 is the true label value of the sample data, K(h p ,h q ) is the Kappa measurement calculation formula, NMI(h p ,u 3 ) is the calculation formula of NMI normalized mutual information.

[0056] The following is a specific implementation of the content of the present invention, and the content of the present invention is further explained through the specific implementation.

[0057] The following specific embodiments are described only to illustrate different aspects of the present invention and should not be construed as limiting the scope of the present invention.

[0058] The drought disaster weather forecast based on semi-supervised ensemble learning is to construct a prediction model based on the drought disaster weather data in the corresponding historical time period through semi-supervised ensemble learning, and use this model as the basis to analyze and predict drought disaster weather.

[0059] like Figure 1 As shown in the figure, a drought disaster weather prediction method based on semi-supervised ensemble learning is proposed. This method performs data balancing, feature construction and feature selection on drought disaster weather data, and constructs a dynamic sample data selection algorithm and a selective ensemble learning pruning algorithm based on target maximization cuckoo optimization. For self-training semi-supervised learning, a dynamic sample data selection algorithm is used in the training process. The algorithm can dynamically select the optimal unlabeled sample data according to the change of the number of iterations in self-training semi-supervised learning. In addition, selective ensemble learning is performed, and a selective ensemble learning pruning algorithm based on target maximization cuckoo optimization is used. The Kappa measure can be used to measure the differences between base learners, and the NMI standardized mutual information can be used to measure the accuracy of the base learners. The optimal base learner subset for ensemble learning is found through the cuckoo optimization algorithm. Finally, the optimal base learner subset is used as the base learner of ensemble learning to train the drought disaster weather prediction model and realize drought disaster weather prediction.

[0060] The semi-supervised learning method based on self-training can add a large amount of unlabeled data to a limited number of labeled samples for training and learning, and can improve the learning performance. The semi-supervised learning method based on self-training needs to iteratively train the model several times, continuously adopt the model prediction results and expand the size of the labeled training data set. At the beginning of the self-training semi-supervised learning training, the model has poor prediction confidence for unlabeled sample data, and as the number of iterations increases, the prediction ability of the model gradually increases. A dynamic sample data selection algorithm is proposed. The algorithm obtains the current iteration number Itea, the category set D of the drought disaster weather dataset, and calculates the number of unlabeled sample data Num of the gth category. g , and dynamically select the optimal unlabeled sample data S according to the maximum number of iterations MaxItea, the dynamic selection sample data algorithm is as shown in formula (1), and the dynamic selection sample data change diagram is as shown in Figure 2 As shown in the figure, the algorithm calculates and dynamically selects the sample data with the best labeling result in the previous round, and finally labels all the unlabeled sample data with pseudo labels. The dynamic sample data selection algorithm can greatly shorten the total training time, and the trained model has a high accuracy rate, which can effectively improve the training efficiency of self-training semi-supervised learning.

[0061]

[0062] As a kind of combinatorial optimization learning method, ensemble learning can not only combine multiple simple models to obtain a combination model with better performance, but also design a combination scheme for specific machine learning problems to obtain a more powerful solution, thereby constructing a more accurate drought disaster prediction model. Ensemble learning is also called a multi-classifier learning system. It is constructed by combining multiple classification learners to complete learning tasks, and weak models are used to obtain a strong model with good prediction effect. At the same time, selective integration is performed in ensemble learning, and a selective ensemble learning pruning algorithm based on target maximization cuckoo optimization is constructed. In the selective integration process, the algorithm considers the differences between base learners and satisfies the accuracy of base learners, and measures base learners from multiple angles. First, the objective function is constructed using the Kappa measure and NMI standardized mutual information. The Kappa measure is used to measure the differences between base learners, and the NMI standardized mutual information is used to measure the accuracy of base learners. Then the cuckoo algorithm is used to optimize the objective function and find the optimal subset of base learners for ensemble learning. The cuckoo algorithm has the advantages of easy implementation and strong search ability. Finally, the optimal base learner subset is used as the base learner of ensemble learning to train the drought disaster weather prediction model and realize drought disaster weather prediction.

[0063] In the process of constructing the objective function, we first calculate the total number of samples n, the number of real samples of the first category a 1 , the number of real samples of the second category a2 , the number of real samples of the cth category a c , the number of samples predicted by the first category b 1 , the number of samples predicted for the second category b 2 , the number of samples predicted by the cth category b c And the overall classification accuracy P 0 , and then calculate P using formula (2) e , and calculate the Kappa measure using formula (3).

[0064]

[0065]

[0066] Then calculate the joint distribution probability P(i, j) of base learners i and j, the probability distribution function P(i) of i, and the probability distribution function P(j) of j, and calculate the NMI normalized mutual information by formula (4).

[0067]

[0068] Then two different base learners predict the value h i ,h j , the true label value u of the sample data 1 , as well as the NMI normalized mutual information calculation formula and the Kappa measure calculation formula to construct formula (5).

[0069]

[0070] By using formula (5) to calculate any two different base learners in the initial base learner set, we can get the objective function formula (6), where H is the prediction value set of the initial base learners such as decision tree, logistic regression, and K nearest neighbor, and h k ,h l are the predicted values ​​of two different base learners.

[0071]

[0072] Formula (6) can be further solved to obtain formula (7), where G is the set of prediction values ​​of the initial base learners such as decision tree, logistic regression, and K nearest neighbor, and h m ,h n is the predicted value of two different base learners, t is the number of base learners, u 2 is the true label value of the sample data:

[0073]

[0074] Finally, the selective ensemble learning pruning algorithm with objective maximization cuckoo optimization is constructed through formula (7), as shown in formula (8), where T is the initial base learner set, P is an optimal subset found in the initial base learner set, and h p ,h q are two different base learners, u 3 is the true label value of the sample data, r is the number of base learners, and the algorithm is used to calculate the best individual fitness and average fitness value in each iteration. The fitness value change diagram is shown in Figure 3 .

[0075]

[0076] The specific steps of using the selective ensemble learning pruning algorithm of objective maximization cuckoo optimization for selective ensemble and establishing the drought disaster weather prediction model are as follows:

[0077] Step 1: Generate an initial set of base learners

[0078] Step 2: Randomly select the generated initial base learners to obtain multiple sets of base learner subsets p = (p 1 , ..., p n )

[0079] Step 3: Use formula (7) as the fitness function of the cuckoo algorithm and perform the following calculation on the base learner subset p: i Evaluate the fitness value f i . First, calculate the base learner subset p i The predicted values ​​of each base learner are then calculated using the fitness function to obtain the current base learner subset p i The fitness value f i .

[0080] Step 4: Use the Levy flight algorithm to learn the current base learner subset p i The base learners in are randomly replaced to obtain a new base learner subset p j , use the fitness function to train the base learner subset p j Calculate its fitness value f j If the base learner subset p i The fitness value f i Less than the base learner subset p j The fitness value f j , then the base learner subset p i Using the base learner subset p j Alternative.

[0081] Step 4: Sort all base learner subsets according to their fitness values. Through multiple iterations, the base learner subset corresponding to the maximum fitness value is obtained, which is the optimal solution for selective ensemble learning.

[0082] Step 5: Use the optimal solution base learner subset as the base learner of Stacking ensemble learning to train the drought disaster weather prediction model to achieve drought disaster weather prediction.

[0083] The drought disaster weather prediction capability and comprehensive disaster warning capability are evaluated based on the actual data of the US Drought Monitoring Agency's 90-day WS10M_MIN minimum wind speed 10 meters (m / s), QV2M 2 meters relative humidity (g / kg), T2M_RANGE 2 meters temperature range (℃), WS10M wind speed 10 meters (m / s). The proposed dynamic selection sample data algorithm can greatly shorten the total time of self-training semi-supervised learning, overcome the problem of poor confidence level, effectively improve the training efficiency of self-training semi-supervised learning, and quantitatively illustrate the advantages of this algorithm. The proposed selective ensemble learning pruning algorithm based on target maximization cuckoo optimization is significantly better than other selective ensemble algorithms, which can effectively improve the training efficiency of selective ensemble learning, and the base learner selected by the algorithm has better generalization performance, faster convergence speed, and higher accuracy.

Claims

1. A drought disaster weather prediction method based on semi-supervised ensemble learning, characterized in that: Follow these steps: Step (1): Establish a data sample set for predicting drought disaster weather; Step (2): Perform data balancing on drought disaster weather data, and perform feature construction and feature selection; Step (3): using self-training semi-supervised learning to use the labeled sample data and unlabeled sample data in the drought disaster weather dataset, and constructing a dynamic sample data selection algorithm in the self-training process to label the unlabeled sample data with pseudo labels; Step (4): construct an objective function based on Kappa measure and standardized mutual information, and the constructed objective function is formula (7); Where G is the set of predicted label values ​​of the initial base learners such as decision tree, logistic regression, and K nearest neighbor, and h m ,h n The predicted label values ​​of two different base learners in the initial base learner prediction label value set G, t is the number of base learners, u2 is the true label value of the sample data, K(h m ,h n ) is the Kappa measurement calculation formula, NMI(h m , u2) is the calculation formula of normalized mutual information; Step (5): Optimize the objective function using the cuckoo algorithm, construct a selective ensemble learning pruning algorithm based on target maximization cuckoo optimization, perform selective ensemble learning on the basis of the constructed algorithm, use the constructed algorithm to find the optimal subset for the initial base learner set, and use the optimal subset as the base learner of the stacking ensemble learning to train the drought disaster weather prediction model; In the step (1), the drought disaster weather data specifically include historical live data of the minimum wind speed at 10 meters, the relative humidity at 2 meters, the temperature range at 2 meters, the wind speed at 10 meters, the temperature at 2 meters, the minimum wind speed at 50 meters, the maximum temperature at 2 meters, the wind speed at 50 meters, the earth's skin temperature, the wind speed range at 50 meters, the maximum wind speed at 50 meters, the maximum wind speed at 10 meters, the wind speed range at 10 meters, the surface pressure, the dew point / freezing point at 2 meters, the minimum temperature at 2 meters, the wet bulb temperature at 2 meters, the precipitation, and the drought level; In the step (5), a selective ensemble learning pruning algorithm based on target maximization cuckoo optimization is constructed, and the process is as follows: The initial base learner set is randomly selected to obtain multiple base learner subsets, and each base learner subset is used as a set of solutions to the objective function. The objective function is then used as the fitness function of the cuckoo algorithm and the fitness value of each base learner subset is calculated. The base learner subsets are sorted according to the calculated fitness values, and the base learner subsets with poor fitness values ​​are discarded. The Levy flight algorithm is used to generate new base learner subsets. After repeated iterations, the base learner subset with the largest fitness value is found, and this base learner subset is used as the optimal solution to the objective function, which is the base learner subset used for selective integration.

2. The drought disaster weather prediction method based on semi-supervised ensemble learning according to claim 1 is characterized in that: In the step (2), the drought disaster weather data is subjected to data balancing processing, and an algorithm combining an oversampling algorithm for synthesizing minority classes with an undersampling algorithm for deleting majority class noise samples is adopted. The oversampling and undersampling techniques are combined for mixed resampling to increase the number of minority class sample data in the data set and reduce the number of majority class sample data in the data set.

3. The drought disaster weather prediction method based on semi-supervised ensemble learning according to claim 1 is characterized in that: In the step (2), the drought disaster weather data is feature constructed. Among the 18 meteorological attributes of the drought disaster weather data, each attribute feature involves 90 days of data. The mean of the 30-day, 60-day, and 90-day data of each attribute is counted, and the maximum value, minimum value, and median of each attribute data are classified and summarized to obtain a new feature set.

4. The drought disaster weather prediction method based on semi-supervised ensemble learning according to claim 1 is characterized in that: In the step (2), feature selection is performed on the drought disaster weather data, a gradient boosting algorithm is used to obtain a weight score of each attribute feature of the drought disaster weather data, and features with a weight greater than the average weight are selected; The correlation coefficient is used to calculate the linear correlation between each attribute feature of the data and the label. The correlation coefficient value is between -1 and 1, and a value of 0 indicates no linear correlation. The absolute values ​​of the correlation coefficients are sorted and the top 50% of the attribute features are selected; finally, the two attribute feature sets selected based on the gradient boosting weight and the correlation coefficient are combined to obtain a new attribute feature set.

5. The drought disaster weather prediction method based on semi-supervised ensemble learning according to claim 1 is characterized in that: In step (3), a dynamic sample data selection algorithm is constructed. The algorithm dynamically selects the optimal unlabeled sample data according to the change of the number of iterations. The process is as follows: The labeled sample data in the drought disaster weather data is trained using the K nearest neighbor supervised learning method. In each subsequent round of training, the trained model is repeatedly used to predict the unlabeled sample data and add pseudo labels. After the prediction is completed, in the next round, the optimal unlabeled sample data selected by the dynamic sample data selection algorithm of formula (1) is added to the current training sample data set together with its pseudo label, and the result generated by itself is used to train itself again until all the unlabeled sample data are marked with pseudo labels. Where S is the total amount of unlabeled sample data selected in each iteration, D is the set of drought disaster weather dataset categories, g is the specific category of drought disaster weather dataset, Itea is the current iteration number, MaxItea is the maximum iteration number, Num g It is the number of unlabeled sample data of the i-th category under the current iteration number Itea.

Images