Wind shear risk assessment prediction method based on improved PKO optimization combination model

Abstract

The invention discloses a wind shear risk assessment prediction method based on an improved PKO optimization combination model, and the method comprises the steps: S1, constructing a wind shear risk assessment index system, obtaining historical wind shear event data, dividing the historical wind shear event data into a plurality of time span groups, and calculating the span group wind shear average risk of each time span group, collecting according to a time sequence to form a wind shear average risk data set; and S2, constructing a combination model, inputting the wind shear average risk data set into the combination model to carry out model training, carrying out model parameter optimization processing by adopting an improved emerald optimization algorithm, and outputting a span group wind shear average risk after the current time by the combination model according to a time sequence. According to the method, the trend characteristic, the seasonal characteristic, the time sequence characteristic and the nonlinear relation of the historical time sequence wind shear average risk data set can be better captured, the future span group wind shear average risk prediction after the current time can be realized, and technical support is provided for aviation risk trend analysis.

Description

Technical Field

[0001] The invention relates to the field of wind shear risk model prediction, and in particular to a wind shear risk assessment and prediction method based on an improved PKO optimization combination model. Background Art

[0002] Wind shear refers to the sudden change of wind speed and direction at different altitudes or horizontal directions in the atmosphere, usually occurring at low altitudes (such as near the ground or near airport runways). Wind shear poses a serious threat to aviation safety, especially during takeoff and landing, which may lead to loss of flight control or even accidents. In recent years, with the increase in air transport volume, the prediction and management of wind shear risks have become a research hotspot. Domestic and foreign scholars have conducted extensive research on wind shear risk prediction. Existing research mostly focuses on the prediction of a single model, and lacks exploration of the optimization of multiple model combinations. Summary of the invention

[0003] The purpose of the present invention is to provide a wind shear risk assessment and prediction method based on an improved PKO optimization combination model, innovatively construct a wind shear risk assessment index system including event level, event cause and geographic information primary index items, obtain historical wind shear event data and divide them into several time span groups according to a natural year, calculate the average wind shear risk of each time span group based on the wind shear risk assessment index system, and form an original wind shear average risk data set according to time series aggregation; construct a combined model including an exponential smoothing model, an autoregressive sliding average model and an XGBoost model, use the wind shear average risk data set as a sample for model training to capture the trend, seasonality, time series characteristics and nonlinear relationship in the wind shear data, use the improved Pied Kingfisher optimization algorithm model to perform model parameter optimization processing, and be able to predict the average wind shear risk of the span group of several consecutive time span groups after the current time.

[0004] The purpose of the present invention is achieved through the following technical solutions:

[0005] A wind shear risk assessment and prediction method based on an improved PKO optimization combination model comprises:

[0006] S1. Construct a wind shear risk assessment index system, obtain historical wind shear event data of several years before the current time, divide the data into several time span groups according to one natural year, calculate the average wind shear risk of each time span group based on the wind shear risk assessment index system for the historical wind shear event data, and aggregate the average wind shear risk of the span group according to the time series to form a wind shear average risk data set;

[0007] S2. Construct a combined model including an exponential smoothing model, an autoregressive moving average model, and an XGBoost model, and input the wind shear average risk data set into the combined model for model training; during the model training process, use an improved pied kingfisher optimization algorithm model to optimize the model parameters, and the combined model outputs the span group wind shear average risk of a continuous number of time span groups after the current time according to the time series.

[0008] To better implement the present invention, the improved pied kingfisher optimization algorithm model includes an initialization stage, a perching and hovering strategy stage, a diving strategy stage, and a symbiosis stage. The expression in the initialization stage is as follows:

[0009] X n，n =LB n +rand·(UB n -LB n ), where X m，n is the position of the pied kingfisher individual m in the nth dimension, UB n and LB n are respectively the lower and upper bounds of the search range in the nth dimension, and rand is a random number of the model between [0, 1];

[0010] In the perching and hovering strategy stage, a dynamic search range adjustment formula is introduced to adjust the search range according to the number of iterations and the change in fitness, and at the same time, a dynamic step size adjustment formula is introduced to adjust the step size according to the difference in population fitness;

[0011] The expression of the improved diving strategy in the diving strategy stage is as follows:

[0012] X m (t1 + 1)=X m (t1)+α 改 ·(Best_Fitness - PKO_Fitness(m))·HA 改

[0013] X m (t1 + 1), X m (t1) are respectively the states or positions of the pied kingfisher individual m at time t1 + 1 and time t1; Best_Fitness is the best fitness value in all time iterations, PKO_Fitness(m) is the fitness value of the pied kingfisher individual m, α 改 is the improved control parameter of the model, and HA 改 is the improved hunting ability of the model;

[0014] The expression in the symbiosis stage is as follows:

[0015] X m (t1 + 1)=X m (t1)+PE改 ·(X m -X n );PE 改 represents an improved predation efficiency, where X m and X n are the states or positions of individual random pied kingfishers m and n, respectively.

[0016] Preferably, the improved control parameter α of the improved pied kingfisher optimization algorithm model of the present invention 改 takes the step size S(t1) at the iteration of time t1 in the dynamic step size adjustment formula,

[0017] where S max is the maximum step size; ΔF(t1) is the improvement amount of the optimal solution at the iteration of time t1; γ is the step size adjustment sensitivity parameter, and δ is the improvement amount threshold;

[0018] The improved predation efficiency α in the symbiotic stage 改 takes the step size S(t1) at the iteration of time t1 in the dynamic step size adjustment formula.

[0019] Preferably, the improved hunting ability HA of the improved pied kingfisher optimization algorithm model of the present invention 改 takes the search range R(t1) at the iteration of time t1 in the dynamic search range adjustment formula,

[0020] where R max is the maximum search range, T is the maximum iteration time, α is the attenuation rate control parameter, and β is the attenuation starting point control parameter.

[0021] Preferably, the exponential smoothing model in the combined model of the present invention is weighted averaged through the wind shear average risk data set and used to predict future values. The weighted average expression of the exponential smoothing model is as follows:

[0022] Y = level + slope × t2 + s_t2 + irregular_t2

[0023] Y represents the predicted value at the iteration of time t2; level is the constant level term; slope is the trend term; s_t2 is the seasonal effect at the iteration of time t2, and irregular_t2 is the random term at the iteration of time t2.

[0024] Preferably, the autoregressive moving average model of the present invention is a time series prediction model combined by the autoregressive model AR, the moving average model MA, and the differencing method model and is used to capture the short-term dependence relationship of the time series.

[0025] Preferably, the XGBoost model of the present invention classifies each time-span group based on the wind shear average risk data set. The XGBoost model constructs a number of weak learners. Each weak learner continues to fit with the fitting error of the previous weak learner as the learning target, and then all the base learners are accumulated to obtain a better classification performance than a single model. The classification expression is as follows:

[0026] is the prediction result of sample i after the t3 - th tree iteration of the XGBoost model; is the prediction result of sample i after the (t3 - 1)-th tree iteration of the XGBoost model; f i (x i ) is the prediction result of sample i of the t3 - th tree.

[0028] Preferably, the objective function Obj of the XGBoost model of the present invention (t3) has the following expression:

[0029] where is the loss function corresponding to the real result and the prediction result, is the regularization term.

[0030] Preferably, the wind shear risk assessment index system of the present invention includes three first - level index items including event level, event cause, and geographical information. The event level includes second - level index items such as accident, symptom, and general event. The accident, symptom, and general event of the event level determine the risk weights according to the reciprocal method of Heinrich's law, and the accident, symptom, and general event of the event level record data according to the occurrence times; The event cause includes second - level index items such as warning, signal interference, ground support, aircraft impact, machinery, management, flight crew, and weather accident. The second - level index items of the event cause record data according to the number of risks caused; Geographical information includes second - level index items such as accident height and airport altitude. Among them, the accident height includes third - level index items such as landing stage, approach, cruise stage, climb to cruise stage, initial climb, and ground. The third - level index items of the accident height record data according to the number of risks caused; The airport altitude includes third - level indexes such as high - altitude plateau, plateau, and plain. The third - level indexes of the airport altitude assign data and risk weights according to the attribution classification; The second - level index items of the event cause and the third - level index items of the accident height determine the corresponding risk weights according to the Delphi method; Calculate the span - group wind shear average risk of each time - span group based on the wind shear risk assessment index system. The span - group wind shear average risk is obtained by adding the products of the data of the index items in the span - group and the corresponding risk weights; When aggregating the span - group wind shear average risk in method S1, it is normalized and then stored in the wind shear average risk data set.

[0031] Preferably, the time span groups are divided by day, and the time span of the time span groups is N1 days; or the time span groups are divided by month, and the time span of the time span groups is one natural month.

[0032] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0033] (1) The present invention innovatively constructs a wind shear risk assessment index system including event level, event cause and geographic information primary index items, obtains historical wind shear event data and divides them into several time span groups according to a natural year, calculates the average wind shear risk of each time span group based on the wind shear risk assessment index system, and forms an original wind shear average risk data set according to time series; constructs a combined model including an exponential smoothing model, an autoregressive sliding average model and an XGBoost model, uses the wind shear average risk data set as a sample for model training to capture the trend, seasonality, time series characteristics and nonlinear relationship in the wind shear data, and uses the improved Pied Kingfisher optimization algorithm model to optimize the model parameters, which can predict the average wind shear risk of a span group for several consecutive time span groups after the current time.

[0034] (2) The present invention uses the improved kingfisher optimization algorithm model to optimize the hyperparameters of the combined model, ensuring the optimal model performance. In terms of global exploration capability, in the early iterations, the search range is close to R max , which can effectively explore the solution space and avoid falling into the local optimum; in terms of local development capabilities, in the later iterations, the search range is significantly reduced, and the solution can be fine-tuned to improve the convergence accuracy; in terms of convergence speed, this embodiment can find a better solution with the same number of iterations; and the prediction accuracy and robustness of the model wind shear risk are improved.

[0035] (3) The present invention can better capture the trend characteristics, seasonal characteristics, time series characteristics and nonlinear relationships of the historical time series wind shear average risk data set, and can realize the future span group wind shear average risk prediction after the current time, providing technical support for aviation risk trend analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 A method flow chart of the wind shear risk assessment and prediction method of the present invention;

[0037] Figure 2 The hierarchical structure of the wind shear risk assessment index system and the risk weight distribution of the lowest level index items are given as examples in the embodiments;

[0038] Figure 3 For the embodiment, the time span groups are divided according to natural months and a normalized monthly span wind shear average risk time series diagram is obtained;

[0039] Figure 4 Schematic diagram for comparing the predicted results and the actual true results of the training set and the test set for each example. Detailed implementation manners

[0040] The present invention will be further described in detail below in conjunction with the embodiments:

[0041] Embodiment

[0042] As Figure 1 shown, a wind shear risk assessment and prediction method based on an improved PKO optimization combination model, the method comprising:

[0043] S1. Construct a wind shear risk assessment index system, obtain historical wind shear event data for several years before the current time, divide it into several time span groups according to a natural year, the time span groups are divided by days, and the time span of the time span group is N1 days; or the time span groups are divided by months, and the time span of the time span group is a natural month. Calculate the span group wind shear average risk of each time span group based on the wind shear risk assessment index system for the historical wind shear event data, and collect the span group wind shear average risk in time series to form a wind shear average risk data set.

[0044] In some embodiments, the preferred wind shear risk assessment index system of the present invention includes three first-level index items including event level, event cause, and geographic information, such as Figure 2As shown, the secondary indicator items of the event level include accidents, incidents, and general events. The data of accidents, incidents, and general events at the event level are recorded according to the number of occurrences. The risk weights of accidents, incidents, and general events at the event level are determined by the reciprocal method of Heinrich's Law. According to the reciprocal method of Heinrich's Law, there are 29 incidents and 300 incident precursors (incident precursors are defined as general events) corresponding to one accident. Therefore, according to the reciprocal method of Heinrich's Law, the risk weights of accidents, incidents, and general events are determined. The risk weight of general events is preset to 1 according to the reciprocal method of Heinrich's Law, the risk weight of incidents is preset to 10.34 according to the reciprocal method of Heinrich's Law, and the risk weight of accidents is preset to 300 according to the reciprocal method of Heinrich's Law. The secondary indicator items of the event cause include warning, signal interference, ground support, aircraft impact, machinery, management, flight crew, and weather accident. The data of the secondary indicator items of the event cause are recorded according to the number of risks caused. The geographical information includes secondary indicator items such as the height of the incident and the airport altitude. Among them, the height of the incident includes tertiary indicator items such as the landing phase, approach, cruise phase, climb to cruise phase, initial climb, and ground. The data of the tertiary indicator items of the height of the incident are recorded according to the number of risks caused. The airport altitude includes tertiary indicators such as high plateau, plateau, and plain. The tertiary indicators of the airport altitude are assigned data and risk weights according to the attribution classification. The risk weights corresponding to the secondary indicator items of the event cause and the tertiary indicator items of the height of the incident are determined by the Delphi method. Calculate the average risk of wind shear for each time span group based on the wind shear risk assessment index system. The average risk of wind shear for the span group is obtained by summing the product of the data of the indicator items within the span group and the corresponding risk weights (as shown in Figure 2 As shown, this embodiment exemplifies the risk weights of each bottom-level indicator item).

[0045] This embodiment statistically analyzes the wind shear events that occurred around the airport during the period from January 2017 to December 2024 (a total of 96 months). The time span group is exemplified by monthly division. The example time span of the time span group is a natural month. The average risk of wind shear for the span group is obtained by summing the product of the data of the indicator items within the span group and the corresponding risk weights (in some embodiments, the calculation formula for the average risk of wind shear for the span group is as follows: ∑ (data of indicator items within the span group × risk weight) is the sum of the products of the data of all indicator items within the span group and the corresponding risk weights. The total flight time within the span group time range is the total flight time statistically counted at the airport within the span group time range. In this embodiment, after summing the product of the data of the indicator items within the span group and the corresponding risk weights, it is then divided by the total flight time within the span group time range to obtain the average risk of wind shear per unit flight time within the span group. The average risk of wind shear for the span group is the average monthly risk of wind shear. Calculate the average monthly risk of wind shear as shown in Table 1 (the data in Table 1 is only for example, and some data is hidden, which does not affect the technical implementation):

[0046] Table 12017 Monthly average wind shear risk from January to December 2024

[0047] Table1 Monthly wind shear risk values from January 2017 to December2024

[0048]

[0049]

[0050] After normalization during the aggregation of the average wind shear risk of the span group in method S1, it is stored in the average wind shear risk dataset. In the example of historical wind shear event data around the airport from January 2017 to December 2024, the normalization method for the monthly average wind shear risk (i.e., the average wind shear risk of the span group) is as follows:

[0051]

[0052] where \(i = 1, 2,\cdots, P\), and \(P\) is the number of eigenvalues; \(y i '\), \(y i are the values before and after normalization respectively; \(\min(y i )\), \(\max(y i )\) are the minimum and maximum values of the \(i\)-th dimensional feature before normalization respectively. The time series plot of the normalized monthly average wind shear risk (i.e., the average wind shear risk of the span group) stored in the average wind shear risk dataset is as Figure 3 shown. As can be seen from Figure 3 , the monthly average wind shear risk is relatively high in April or May each year, and the overall curve has strong seasonal and time series characteristics.

[0053] S2. Construct a combined model including an exponential smoothing model, an autoregressive moving average model, and an XGBoost model, and input the average wind shear risk dataset into the combined model for model training; during the model training process, an improved pied kingfisher optimization algorithm model (the full English name of the pied kingfisher optimization algorithm is Pied Kingfisher Optimizer, abbreviated as PK0 in English, and the improved pied kingfisher optimization algorithm model is abbreviated as the improved PKO model) is used for model parameter optimization. The combined model outputs the average wind shear risk of consecutive several time span groups after the current time according to the time series.

[0054] In the exponential smoothing model (English name: HOLT-WINTERS) of the combined model of the present invention, weighted averaging is performed through the wind shear average risk data set and used to predict future values (using historical observations in the time series to weight the series, the weights decreasing in a geometric progression from near to far, and the predicted value being the weighted average of all historical averages). The weighted average expression of the exponential smoothing model is as follows:

[0055] Y = level + slope × t2 + s_t2 + irregular_t2

[0056] Y represents the predicted value at the iteration of time t2; level is the constant level term, controlled by the smoothing parameter; slope is the trend term, controlled by the slope parameter; s_t2 is the seasonal effect at the iteration of time t2, controlled by the smoothing parameter; irregular-t2 is the random term at the iteration of time t2. In this embodiment, the Holt-Winters triple exponential smoothing model is used to capture the trend and seasonal components in the wind shear data; the parameter configuration of the exponential smoothing model is: alpha = 0.05: the level smoothing coefficient, controlling the influence of the current observation on the prediction. beta = 0.02: the trend smoothing coefficient, controlling the change of the trend. gamma = 0.02: the seasonal smoothing coefficient, controlling the change of the seasonality. season_length = 12: the seasonal cycle length, assumed to be 12 months.

[0057] The autoregressive moving average model (English name: ARIMA model) in the combined model of the present invention is a time series prediction model combined by the autoregressive model AR, the moving average model MA, and the differencing method model and is used to capture the short-term dependence relationship of the time series. The autoregressive moving average model is a hybrid model with p-order autoregression and q-order moving average, denoted as ARIMA(p,q). In practice, the data sequence is a non-stationary time series. Therefore, in this embodiment, the differencing d is the number of differencing times, denoted as ARIMA(p,d,q). After making it stationary, it will not change over time, and thus the future can be predicted based on the past behavior of the series. In the autoregressive moving average model, the future value of the series is expressed as a linear function of the lag terms and the current and lagged values of the random disturbance terms. The parameter configuration of the autoregressive moving average model in this embodiment is: automatically select the optimal order (p,q), where: p: the autoregressive (AR) order, ranging from 0 to 4; q: the moving average (MA) order, ranging from 0 to 4.

[0058] The autoregressive moving average model (i.e., the ARIMA model) is used to capture the autocorrelation and moving average components in the residuals of the Holt-Winters model. By traversing the (p,q) combinations, the optimal model is selected.

[0059]

[0060] In the XGBoost model of the combined model of the present invention (which optimizes the model performance through the gradient boosting framework, can handle non-linear relationships, supports feature engineering, and is suitable for complex data), classification processing is performed on each time-span group based on the wind shear average risk data set. The XGBoost model constructs several weak learners, and each weak learner continues to fit with the fitting error of the previous weak learner as the learning target, and then all the base learners are accumulated to obtain a better classification performance than a single model. The classification expression is as follows:

[0061] is the prediction result of sample i after the t3rd tree iteration of the XGBoost model; is the prediction result of sample i after the (t3 - 1)th tree iteration of the XGBoost model; f i (x i ) is the prediction result of sample i of the t3rd tree. The objective function obj (t3) of the XGBoost model is expressed as follows:

[0063] where is the loss function corresponding to the true result and the prediction result, is the regularization term. The parameter configuration of the XGBoost model in this embodiment is:

[0064] learning_rate: learning rate, with a range of [0.01, 0.1].

[0065] max_depth: maximum depth of the tree, with a range of [1, 15].

[0066] min_child_weight: minimum child node weight, with a range of [0.1, 2].

[0067] subsample: subsampling rate, with a range of [0.5, 0.9]. In this embodiment, an improved spotted kingfisher optimization algorithm model is used to optimize the above parameters.

[0068] In some embodiments, the improved spotted kingfisher optimization algorithm model includes an initialization stage, a perching and hovering strategy stage, a diving strategy stage, and a symbiosis stage. The expression of the initialization stage is as follows:

[0069] X m，n = LB n + rand·(UB n - LB n ), where X m，n is the position of the spotted kingfisher individual m in the nth dimension, UB n and LBn are respectively the lower bound and the upper bound of the n-dimensional search range, and rand is a random number of the model between [0, 1];

[0070] In the perching and hovering strategy stage (also known as the exploration stage), a dynamic search range adjustment formula is introduced to adjust the search range according to the number of iterations and the change of fitness, and at the same time, a dynamic step size adjustment formula is introduced to adjust the step size according to the difference in population fitness. The present invention introduces a dynamic search range adjustment formula to adjust the search range according to the number of iterations and the change of fitness, so as to avoid premature convergence. The present invention innovatively combines an adaptive weight and a non-linear attenuation mechanism, which can expand the search range in the early iteration to enhance the global exploration ability, and narrow the search range in the later iteration to improve the local development ability. The preferred dynamic search range adjustment formula of the present invention is as follows:

[0071] where R max is the maximum search range, T is the maximum iteration time, α is the attenuation rate control parameter, which controls the speed of search range attenuation, and β is the attenuation starting point control parameter, which determines the iteration ratio at which the search range starts to significantly attenuate.

[0072] In terms of the global exploration ability of this embodiment, in the early iteration, the search range is close to R max , which can effectively explore the solution space and avoid falling into local optima. In terms of the local development ability, in the later iteration, the search range is significantly reduced, which can finely adjust the solution and improve the convergence accuracy. In terms of the convergence speed, this embodiment can find a better solution under the same number of iterations. The innovative dynamic search range adjustment formula of the present invention realizes expanding the search range in the early iteration to enhance the global exploration ability and narrowing the search range in the later iteration to improve the local development ability by combining the adaptive weight and the non-linear attenuation mechanism.

[0073] This embodiment introduces a dynamic step size adjustment formula to adjust the step size according to the difference in population fitness in the perching and hovering strategy stage, so as to improve the search efficiency; the dynamic adjustment of the step size in this embodiment is the key to improving the search efficiency and accuracy. This embodiment gives an innovative dynamic step size adjustment formula, which can dynamically adjust the step size according to the quality of the current optimal solution: if the improvement of the current optimal solution is small, the step size is reduced for fine search; if the improvement is large, the step size is increased to quickly approach the global optimum. The dynamic step size adjustment formula is as follows: where S maxis the maximum step size; ΔF(t1) is the improvement of the optimal solution at the iteration of time t1, and ΔF(t) = F(t - 1) - F(t), where (t) is the optimal fitness value at the iteration of time t1; γ is the step size adjustment sensitivity parameter (controlling the response speed of the step size to the improvement), and δ is the improvement threshold (used to judge whether the improvement of the current optimal solution is significant). In this embodiment, an innovative dynamic step size adjustment formula is introduced in the perching and hovering strategy phase, which can converge quickly (when the improvement is large, the step size is close to the maximum value, and it can quickly approach the global optimum), and has the advantages of fine search (when the improvement is small, the step size is close to the minimum value, and it can finely adjust the solution to improve the convergence accuracy), high stability (the model performance is more stable, and the variance of the fitness value is significantly reduced), etc.

[0074] The perching and hovering strategy expressions in the perching and hovering strategy phase of the present invention are as follows:

[0075] X i (t + 1) = X i (t) + α 改 ·T 改 where α 改 is the improvement control parameter, taking the step size S(t1) at the iteration of time t1 in the dynamic step size adjustment formula, that is T 改 is the dynamic parameter in the perching and hovering strategy, taking the search range R(t1) at the iteration of time t1 in the dynamic search range adjustment formula, that is

[0076] The improved diving strategy expression in the diving strategy phase is as follows:

[0077] X m (t1 + 1) = X m (t1) + α 改 ·(Best_Fitness - PKO_Fitness(m))·HA 改

[0078] X m (t1 + 1), X m (t1) are the states or positions of the pied kingfisher individual m at the iterations of time t1 + 1 and time t1 respectively; Best_Fitness is the best fitness value in all time iterations, PKO_Fitness(m) is the fitness value of the pied kingfisher individual m, and α 改 is the improvement control parameter of the model, and HA 改 is the improved hunting ability of the model. In some embodiments, the improved hunting ability HA of the improved pied kingfisher optimization algorithm model 改 takes the search range R(t1) at the iteration of time t1 in the dynamic search range adjustment formula, that is Improved control parameter α for the improved pied kingfisher optimization algorithm model 改 Take the step size S(t1) at the iteration of time t1 in the dynamic step size adjustment formula. The improved diving strategy expression in the diving strategy stage is as follows:

[0079] where S max is the maximum step size; ΔF(t1) is the improvement amount of the optimal solution at the iteration of time t1; γ is the step size adjustment sensitivity parameter, δ is the improvement amount threshold; where R max is the maximum search range, T is the maximum iteration time, α is the attenuation rate control parameter, and β is the attenuation starting point control parameter.

[0080] The expression in the symbiosis stage is as follows:

[0081] X m (t1 + 1) = X m (t1) + PE 改 ·(X m - X n ); PE 改 represents the improved predation efficiency, and X m , X n are the states or positions of the random pied kingfisher individuals m and n respectively. In some embodiments, the improved predation efficiency α 改 in the symbiosis stage takes the step size S(t1) at the iteration of time t1 in the dynamic step size adjustment formula.

[0082] This embodiment's combined model also adopts an early stopping mechanism. When the fitness change is small, the iteration is stopped in advance to prevent the model from overfitting. The core idea of the early stopping mechanism is to monitor the performance of the validation set (such as the loss function or accuracy) during the model training process. When the performance of the validation set no longer improves, the training is stopped in advance to avoid the model overfitting on the training set. The early stopping mechanism is implemented through the following steps:

[0083] (1) Monitor the performance of the validation set

[0084] In each iteration, calculate the performance metrics of the validation set (such as the loss function, accuracy, or correlation coefficient). In your code, the performance of the validation set is measured by calculating the correlation coefficient of the test set (corr(Y_test, predictions)).

[0085] (2) Judge whether the performance improves

[0086] After each iteration, check whether the performance on the validation set has improved. If the performance has not improved, or the improvement is very small, it is considered that the model has reached the optimal state and the training can be stopped. When the above conditions are met, the code will execute the break statement to jump out of the loop and stop the training. In the early stopping mechanism, strategies for dynamically adjusting the search range and step size are combined. The purpose of these strategies is to dynamically adjust the search range and step size during training according to the current performance, in order to accelerate the convergence speed or avoid getting stuck in a local optimum. These adjustment strategies, when used in combination with the early stopping mechanism, can further improve the training efficiency and performance of the model. The core idea of the early stopping mechanism is to determine whether the model has reached the optimal state by monitoring the performance of the validation set. If the performance of the validation set no longer improves, or the improvement is very small, stop the training early to avoid overfitting. In your code, the early stopping mechanism is implemented by checking the standard deviation of the fitness and combines strategies for dynamically adjusting the search range and step size to further improve the training efficiency and performance of the model.

[0087] To further verify the prediction performance of the combined model of the present invention, three statistical indicators are selected to evaluate the prediction results, namely, root mean square error (RMSE), mean square error (MSE), mean absolute error (MAE), and coefficient of determination (R2). RMSE, MSE, and MAE can evaluate the degree of data variation, and the smaller their values, the higher the accuracy of the prediction model in describing the experimental data. R 2 It is used to evaluate the fitting effect of the model. The closer its value is to 1, the better the fitting effect. In the example of historical wind shear event data around the airport from January 2017 to December 2024, in this embodiment, the training set and the test set are divided for model prediction respectively. After the training set and the test set are predicted respectively, the comparison results between the prediction results and the actual real results are as Figure 4 shown, where the Train RMSE of the training set is 0.3278, the Train MAE is 0.2249, and the Train Correlation is 0.9541. The Train RMSE of the test set is 0.2707, the Train MAE is 0.2164, and the Train Correlation is 0.9795; combined Figure 4 with the above indicators, the prediction accuracy of the combined model of the present invention is relatively high.

[0088] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, and improvements made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A wind shear risk assessment and prediction method based on an improved PKO optimization combination model, characterized in that: The method includes: S1. Construct a wind shear risk assessment index system, obtain historical wind shear event data for several years before the current time, divide it into several time-span groups according to a natural year, calculate the span-group wind shear average risk for each time-span group based on the wind shear risk assessment index system, and collect the span-group wind shear average risks in time series to form a wind shear average risk data set; S2. Construct a combined model including an exponential smoothing model, an autoregressive moving average model, and an XGBoost model, and input the wind shear average risk data set into the combined model for model training; during the model training process, use an improved pied kingfisher optimization algorithm model for model parameter optimization processing, and the combined model outputs the span-group wind shear average risks for several consecutive time-span groups after the current time in time series.

2. The wind shear risk assessment and prediction method based on the improved PKO optimization combination model according to claim 1, characterized in that: The improved pied kingfisher optimization algorithm model includes an initialization stage, a perching and hovering strategy stage, a diving strategy stage, and a symbiosis stage. The expression in the initialization stage is as follows: X m，n = LB n + rand·(UB n - LB n ), where X m，n is the position of the individual m of the pied kingfisher in n dimensions, UB n and LB n are the lower and upper bounds of the search range in n dimensions respectively, and rand is a random number of the model between [0, 1]; In the perching and hovering strategy stage, a dynamic search range adjustment formula is introduced to adjust the search range according to the number of iterations and fitness changes, and a dynamic step size adjustment formula is introduced to adjust the step size according to the differences in population fitness; the expression of the improved diving strategy in the diving strategy stage is as follows: X m (t1 + 1) = X m (t1) + α 改 ·(Best_Fitness - PKO_Fitness(m))·HA 改 X m (t1 + 1), X m (t1) are the states or positions of the individual m of the pied kingfisher at iterations of time t1 + 1 and time t1 respectively; Best_Fitness is the best fitness value in all time iterations, PKO_Fitness(m) is the fitness value of the individual m of the pied kingfisher, and α 改 is the improved control parameter of the model, and HA 改 is the improved hunting ability of the model; The expression in the symbiosis stage is as follows: X m (t1 + 1)= X m (t1)+ PE 改 ·(X m - X n ); PE 改 represents an improved predation efficiency, where X m and X n are the states or positions of individual random pied kingfishers m and n, respectively.

3. The wind shear risk assessment and prediction method based on the improved PKO optimization combination model according to claim 2, wherein: Improved control parameter α of the improved pied kingfisher optimization algorithm model 改 Take the step size S(t1) at the iteration of time t1 in the dynamic step size adjustment formula where S max is the maximum step size; ΔF(t1) is the improvement of the optimal solution at time t1 iteration; γ is the step size adjustment sensitivity parameter, and δ is the improvement threshold; Improved predation efficiency α in the symbiotic stage 改 Take the step size S(t1) at the iteration of time t1 in the dynamic step size adjustment formula.

4. The wind shear risk assessment and prediction method based on the improved PKO optimization combination model according to claim 2, characterized in that: Improved Hunting Ability HA of the Improved Kingfisher Optimization Algorithm Model for Spotted Kingfishers 改 Take the search range R(t1) at the iteration of time t1 in the dynamic search range adjustment formula, Where R max is the maximum search range, T is the maximum iteration time, α is the decay rate control parameter, and β is the decay starting point control parameter.

5. The wind shear risk assessment and prediction method based on the improved PKO optimization combination model according to claim 1, characterized in that: The exponential smoothing model in the combined model performs weighted averaging through the wind shear average risk data set and is used to predict future values. The weighted averaging expression of the exponential smoothing model is as follows: Y = level + slope × t2 + s_t2 + irregular_t2 Y represents the predicted value at the iteration of time t2; level is the constant level term; slope is the trend term; s_t2 is the seasonal effect at the iteration of time t2, and irregular_t2 is the random term at the iteration of time t2.

6. The wind shear risk assessment and prediction method based on the improved PKO optimization combination model according to claim 1, characterized in that: The autoregressive moving average model is a time series prediction model combined by an autoregressive model AR, a moving average model MA, and a differencing method model and is used to capture the short-term dependencies of the time series.

7. The wind shear risk assessment and prediction method based on the improved PKO optimization combination model according to claim 1, wherein: The XGBoost model classifies each time-span group based on the wind shear average risk data set. The XGBoost model constructs several weak learners, and each weak learner takes the fitting error of the previous weak learner as the learning target and continues to fit. Then, all the base learners are accumulated to obtain a better classification performance than a single model. The classification expression is as follows: is the prediction result of sample i after the t3 - th tree iteration of the XGBoost model; is the prediction result of sample i after the (t3 - 1)-th tree iteration of the XGBoost model; f i (x i ) is the prediction result of sample i of the t3 - th tree.

8. The wind shear risk assessment and prediction method based on the improved PKO optimization combination model according to claim 7, characterized in that: The objective function Obj of the XGBoost model (t3) The expression is as follows: Among them is the loss function corresponding between the true result and the predicted result, is the regularization term.

9. The wind shear risk assessment and prediction method based on the improved PK0 optimization combination model according to claim 1, characterized in that: The wind shear risk assessment index system includes three first-level index items, namely event level, event cause, and geographical information. The event level includes second-level index items such as accident, symptom, and general event. The accident, symptom, and general event of the event level determine the risk weight according to the reciprocal method of Heinrich's law, and the accident, symptom, and general event of the event level record data according to the occurrence times; the event cause includes second-level index items such as alarm, signal interference, ground support, aircraft influence, machinery, management, flight crew, and weather accident. The second-level index items of the event cause record data according to the number of risks caused; the geographical information includes second-level index items such as the height of the incident and the airport elevation. Among them, the height of the incident includes third-level index items such as landing phase, approach, cruise phase, climb to cruise phase, initial climb, and ground. The third-level index items of the height of the incident record data according to the number of risks caused; the airport elevation includes third-level indexes such as high plateau, plateau, and plain. The third-level indexes of the airport elevation assign data and risk weights according to the attribution classification; The second-level index items of the event cause and the third-level index of the height of the incident determine the corresponding risk weights according to the Delphi method; Based on the wind shear risk assessment index system, calculate the span group wind shear average risk of each time span group. The span group wind shear average risk is obtained by summing the product of the data of the index items in the span group and the corresponding risk weights; when aggregating the span group wind shear average risk in method S1, it is normalized and then stored in the wind shear average risk dataset.

10. The wind shear risk assessment and prediction method based on the improved PKO optimization combination model according to claim 1, characterized in that: The time span group is divided by days, and the time span of the time span group is N1 days; or the time span group is divided by months, and the time span of the time span group is a natural month.

Images