Flight deviation risk prediction method based on mantis shrimp algorithm optimization combination model
By constructing a combined prediction model and optimizing the model using an improved mantis shrimp algorithm, the problem of poor nonlinear fitting of the LSTM model in flight deviation risk prediction was solved, achieving higher accuracy and more stable risk prediction.
Patent Information
- Application Number
- CN202511693895.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-18
- Publication Date
- 2026-02-13
AI Technical Summary
Existing LSTM models suffer from poor nonlinear fitting and difficulty in parameter tuning and optimization in predicting flight deviation risk, which affects prediction accuracy and makes it difficult to effectively quantify the risk of flight attitude and altitude deviation.
A combined prediction model was constructed, which included an autoregressive moving average model, an LSTM model, a Transformer model, and an XGBoost model. An improved mantis shrimp algorithm was used to optimize the model and perform parameter combination optimization. The model was trained using the flight deviation average risk dataset to optimize the model parameters and improve prediction accuracy.
It improves the accuracy and robustness of flight deviation risk prediction, and can automatically search for key hyperparameter combinations in complex models, thereby enhancing the accuracy and stability of prediction.
Smart Images
Figure CN121526323A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the field of flight deviation risk assessment, and in particular to a flight deviation risk prediction method based on a mantis shrimp algorithm optimized combination model. BACKGROUND
[0002] The most critical and main flight deviation of the aircraft is flight attitude deviation and flight height deviation. Based on the risk calculation theory in the civil aviation safety management system (SMS), how to quantitatively evaluate the flight attitude deviation and flight height deviation is currently a technical problem. Under the conditions of proper crew operation, management and equipment failure, the most important influencing factor is environmental factors, mainly including severe weather (such as wind shear, strong crosswind, thunderstorm, etc.), air space terrain interference (such as mountain ridges, high-rise buildings, and obstacle interference), electromagnetic interference (electromagnetic signal interference with navigation equipment and affecting positioning accuracy), etc. Environmental factors have seasonal time series. The time series data of the flight deviation event is statistically analyzed to form a flight deviation risk time series dataset. The long and short dependence of the flight deviation risk time series dataset is captured through the LSTM model, which can effectively predict the flight deviation risk of the aircraft. However, how to quantitatively evaluate the flight deviation risk of the time unit is a technical problem of the application of the LSTM model. At the same time, the single LSTM model has problems such as poor nonlinear fitting, difficulty in parameter optimization, etc., which seriously affects the risk prediction accuracy of the LSTM model and seriously restricts the effective implementation of the flight deviation risk prediction of the aircraft, which is a technical problem to be solved at present. SUMMARY
[0003] The purpose of the present application is to provide a flight deviation risk prediction method based on a mantis shrimp algorithm optimized combination model. The flight deviation average risk of each time period is calculated based on the historical flight deviation event data before the research time, and the flight deviation average risk dataset is collected in time sequence. A combination prediction model including an autoregressive moving average model, an LSTM model, a Transformer model and an XGBoost model is creatively constructed. The flight deviation average risk dataset is used as a sample for time series data prediction training of the combination prediction model. The mantis shrimp algorithm model is improved to perform global search and iterative optimization of the combination prediction model in the parameter space, and finally the optimal parameter combination with the minimum prediction error is obtained. The flight deviation average risk of the future continuous time periods can be predicted.
[0004] The purpose of the present application is achieved by the following technical solutions:
[0005] A flight deviation risk prediction method based on a mantis shrimp algorithm optimized combination model, the method comprising:
[0006] S1, construct a flight deviation risk assessment index system composed of attitude deviation and height deviation, obtain historical flight deviation event data before the research time, divide the data in sequence according to time period, calculate the flight deviation average risk of each time period based on the flight deviation risk assessment index system, and sequentially collect the flight deviation average risk of each time period to obtain a flight deviation average risk data set according to time sequence;
[0007] S2, construct a combination prediction model including an autoregressive moving average model, an LSTM model, a Transformer model and an XGBoost model, the combination prediction model is trained by time series data prediction using the flight deviation average risk data set, and the combination prediction model is optimized by an improved mantis shrimp algorithm model.
[0008] S3, the combination prediction model outputs a flight deviation average risk sequence of several time periods after the research time according to time sequence.
[0009] In order to better realize the present application, the improved mantis shrimp algorithm model includes an initialization stage, a search stage, a predation stage and a reproduction stage. In the initialization stage, half of the population is randomly initialized in the search space, then another half of the population is generated by elite reverse learning, and the two parts of the population are combined and the optimal individual set is selected as the initial population. In the search stage, the initial population is used to simulate the individual random search and leader following mechanism of mantis shrimp when searching for prey to explore in the search space. In the predation stage, individuals with good fitness and far away from the current individual are selected as guides, and an adaptive step size is used. In the reproduction stage, the population diversity is increased by simulating the reproduction behavior of mantis shrimp.
[0010] Preferably, the improved mantis shrimp algorithm model generates an expression in the initialization stage by elite reverse learning as follows:
[0011] wherein The new individual generated by elite reverse learning is is the individual with the best fitness in the randomly initialized half population, is a model parameter in the range of [0, 1] for adjusting the position of the reverse solution, , are the lower and upper bounds of the search space respectively; the initial population is N individuals selected from the two parts of the population; the individual random search expression of the improved mantis shrimp algorithm model in the search stage is as follows:
[0012] wherein is the position of mantis shrimp individual i at t+1 iteration time, is the position of mantis shrimp individual i at t iteration time, , are the positions of two randomly selected individuals in the population at the tth iteration time, is the global optimal position of the population at the tth iteration time, , are random coefficients;
[0013] The expression of the improved mantis shrimp algorithm model in the search stage of the leader-following mechanism is as follows:
[0014] wherein is a randomly selected leader position in the high-quality solution group, is a random coefficient.
[0015] Preferably, the adaptive step size used by the improved mantis shrimp algorithm model is as follows: , is the maximum step size at the beginning of iteration, is the minimum step size at the later stage of iteration, is the total iteration time; the updated individual is updated according to the adaptive step size , and the position of the updated individual is as follows: , wherein is the position of the current individual, is the selected leader position.
[0016] Preferably, the autoregressive moving average model in the combined prediction model comprises an autoregressive model AR and a moving average model MA. The autoregressive moving average model first linearly combines the current time period and the past several time periods to predict the time period, then uses the prediction errors of the past several time periods to correct and constrain, and then captures the short-term dependence relationship of the time series. The improved mantis shrimp algorithm model performs parameter optimization processing on the autoregressive order of the autoregressive term of the autoregressive model AR and the moving average order of the moving average model MA, respectively.
[0017] Preferably, the LSTM model neuron in the combined prediction model is composed of an input gate, a forget gate and an output gate. The LSTM model is used to obtain the long-term dependence information of the flight deviation average risk data set. The improved mantis shrimp algorithm model performs parameter optimization processing on the number of hidden layer units, the learning rate and the dropout rate of the LSTM model. The Transformer model comprises an encoder, a decoder, a multi-head attention mechanism and a feedforward network. The Transformer model captures global relationships through parallel computing processing. The improved mantis shrimp algorithm model performs parameter optimization processing on the number of encoder layers, the number of attention heads and the dimension of the feedforward network of the Transformer model.
[0018] Preferably, the XGBoost model in the combined prediction model constructs several weak learners. Each weak learner uses the fitting error of the previous weak learner as the learning target to continue fitting. Then, all base learners are summed and the prediction result is output. The prediction expression is as follows:
[0019] ,in For the first Wheel pair samples The prediction results For the first Wheel pair samples The prediction results For the first The base learner trained in rounds on samples The predicted output; the objective function expression of the XGBoost model is as follows:
[0020] ,in The objective function of the XGBoost model; For the prediction results Compared with the actual results The loss function; For regularization; the improved mantis shrimp algorithm model optimizes the parameters of the XGBoost model, including the maximum tree depth, learning rate, and sampling ratio per tree.
[0021] Preferably, in method S1, the average risk of flight deviation is calculated based on historical flight deviation event data for different time periods. The calculation expression is as follows:
[0022] ,in This represents the risk of attitude deviation for the corresponding time period. This represents the risk of significant deviation within a given time period. The weights of the flight phases corresponding to the time periods. The environmental impact factors corresponding to the time period are stored in the flight deviation risk assessment index system, which contains the weights of flight stages according to the flight stages. and environmental impact factors classified according to environmental impact data. The flight time for a time period is the sum of the flight times of historical flight deviation events within that time period.
[0023] Preferably, the attitude deviation risk The following formula is used to calculate the result based on historical flight deviation event data:
[0024] , The risk of high deviation Based on the historical flight deviation event data, the flight deviation average risk is calculated according to the following formula: Wherein is the actual pitch angle, is the preset standard pitch angle, is the preset threshold, is the actual slope, is the preset standard slope, is the preset threshold, and a and β are weight coefficients, is the actual height, is the preset specified height.
[0025] Preferably, the time period is divided by day, and the time length of the time period is N1 days; or the time period is divided by month, and the time length of the time period is a natural month.
[0026] Compared with the prior art, the present application has the following advantages and beneficial effects:
[0027] (1) The present application calculates the flight deviation average risk of each time period based on the historical flight deviation event data before the research time and aggregates the flight deviation average risk data set in time sequence, creatively constructs a combination prediction model including an autoregressive moving average model, an LSTM model, a Transformer model and an XGBoost model, performs time series data prediction training of the combination prediction model with the flight deviation average risk data set as a sample, improves the mantis shrimp algorithm model to perform global search and iterative optimization of the combination prediction model in the parameter space, and finally optimizes the optimal parameter combination that minimizes the prediction error, which can predict the flight deviation average risk of future continuous time periods.
[0028] (2) The present application uses the improved mantis shrimp algorithm model to optimize the model parameters of the combination prediction model, which can efficiently and automatically search for the key super parameter combination in the complex combination prediction model; in terms of convergence speed, a better solution can be found under the same number of iterations, which improves the prediction accuracy and robustness of the flight deviation average risk. BRIEF DESCRIPTION OF DRAWINGS
[0029] Figure 1 is the flow chart of the method of the present application;
[0030] Figure 2 is the flight deviation risk time series graph of the flight deviation average risk data set selected in the embodiment;
[0031] Figure 3 is the comparison graph of the flight deviation risk true value and the flight deviation risk predicted value test verification using the test set in the embodiment. DETAILED DESCRIPTION
[0032] The application will be further described in detail below in combination with examples.
[0033] Embodiment
[0034] As shown in the figure, a flight deviation risk prediction method based on a praying mantis shrimp algorithm optimized combination model, the method comprises: Figure 1
[0035] S1, construct a flight deviation risk assessment index system composed of attitude deviation and altitude deviation, obtain historical flight deviation event data before the research time (the data source of the historical flight deviation event data includes QAR data and other data of the aircraft), divide the data in sequence according to time period, calculate the flight deviation average risk of each time period based on the flight deviation risk assessment index system, and collect the flight deviation average risk of each time period in time sequence to obtain a flight deviation average risk data set.
[0036] In some embodiments, the flight deviation average risk is calculated based on the historical flight deviation event data according to time period, and the flight deviation average risk The calculation expression is as follows:
[0037] , wherein is the attitude deviation risk corresponding to the time period, is the altitude deviation risk corresponding to the time period, is the flight phase weight corresponding to the time period, is the environmental influence factor corresponding to the time period, the flight phase weight divided according to the flight phase and the environmental influence factor divided according to the environmental influence data are stored in the flight deviation risk assessment index system; in the flight deviation risk assessment index system, the flight phase weight is evaluated as follows: the flight phase weight =1.2 during take-off and landing, the flight phase weight =1.0 during climbing, the flight phase weight =0.6 during cruising……; the environmental influence factor is evaluated as follows: the environmental influence factor =1.2 in bad weather, and the environmental influence factor =1.0 in normal weather. The flight time of the time period is the total flight time of the historical flight deviation event data in the time period.
[0038] The preferred attitude deviation risk is calculated based on the historical flight deviation event data according to the following formula (i.e. attitude deviation degree quantification function):
[0039] ,
[0040] The risk of high deviation The following formula is used to calculate the result based on historical flight deviation event data: ,in The actual pitch angle (the actual pitch angle extracted from historical flight deviation event data). Preset the standard pitch angle (example as follows: set it to 0° during level flight, setting 0° means that the standard pitch angle during level flight is 0°). Set a preset threshold (examples are shown below: threshold is set to +25° or -10°). The actual slope (the actual slope extracted from historical flight deviation event data). Preset the standard bank angle (example below: set to 0° during level flight, setting 0° means that the standard bank angle during level flight is 0°). The preset threshold is set as follows (example: threshold is set to 45°), and α and β are the weighting coefficients (example: α=0.6, β=0.4, α and β are the weighting coefficients for the dominant influence of pitch angle and slope on attitude safety, respectively). The actual altitude (the actual altitude extracted from historical flight deviation event data). The preset height is specified, with 60 (in meters) as the threshold.
[0041] In some embodiments, the time period of the present invention can be divided by days, with a duration of N1 days; or the time period can be divided by months, with a duration of one calendar month. This embodiment will be described using the monthly division as an example. Figure 2 As shown, this embodiment acquires event data from a certain country from January 2017 to August 2025 (a total of 104 months divided by month) and conducts statistical analysis to obtain historical flight deviation event data for 104 time periods. The average flight deviation risk for each time period (divided by month, totaling 104 months) is then sequentially aggregated according to the time series to obtain the average flight deviation risk dataset. Preferably, the historical flight deviation event data from January 2017 to August 2025 is normalized and then stored in the historical flight deviation event dataset according to the time series. The normalization method is as follows:
[0042]
[0043] in , It is the total number of sample data; , These are the values before and after normalization, respectively. , min and max are the minimum and maximum values before normalization, respectively. The wind shear risk dataset is randomly divided into training set and test set with a ratio of 8:2.
[0044] S2, a combination prediction model comprising an autoregressive moving average model, an LSTM model, a Transformer model and an XGBoost model is constructed, the combination prediction model is trained for time series data prediction using the flight deviation average risk dataset, and the combination prediction model is subjected to model hyperparameter combination optimization processing using an improved mantis shrimp optimization algorithm model.
[0045] In some embodiments, the improved mantis shrimp optimization algorithm model (English full name: Mantis Shrimp Optimization Algorithm, abbreviated as MShOA, which is a meta-heuristic algorithm inspired by the hunting behavior of mantis shrimp) comprises an initialization stage, a search stage, a predation stage and a reproduction stage. In the initialization stage, half of the population is randomly initialized in the search space, then the other half of the population is generated using elite inverse learning, and then the two populations are combined and the optimal individual set is selected as the initial population. Preferably, the improved mantis shrimp optimization algorithm model generates an expression as follows in the initialization stage using elite inverse learning:
[0046] wherein the new individual generated by elite inverse learning, is the individual with the best fitness in the randomly initialized half population, is a model parameter in the range of [0, 1] for adjusting the position of the inverse solution, , are the lower and upper bounds of the search space, respectively; and the initial population is the N individuals selected from the two populations.
[0047] The search stage of the improved mantis shrimp optimization algorithm model simulates the individual random search and leader following mechanism of mantis shrimp when searching for prey in the search space using the initial population. The individual random search expression of the improved mantis shrimp optimization algorithm model in the search stage is as follows:
[0048] wherein is the position of mantis shrimp individual i at t+1 iteration time, is the position of mantis shrimp individual i at t iteration time, , are the positions of two randomly selected individuals in the population at t iteration time, is the global optimal position of the population at t iteration time, , are random coefficients;
[0049] The expression for the leader-following mechanism in the search phase of the improved mantis shrimp algorithm model is as follows:
[0050] ,in A leader position is randomly selected from the population of high-quality solutions. These are random coefficients.
[0051] The improved mantis shrimp algorithm model selects individuals with good fitness and a relatively large, balanced distance from the current individual during the predation phase (i.e., selecting a leader based on the Fitness Distance Balance (FDB) method) as the leader, and employs an adaptive step size. The adaptive step size used in the improved mantis shrimp algorithm model is... The expression is as follows: , This represents the maximum step size in the initial stage of the iteration. This represents the minimum step size in the later stages of the iteration. This is the total iteration time; after updating, individuals follow an adaptive step size. Update, the location of the updated individual The expression is as follows: ,in For the current position of the individual, The selected leader position. The breeding phase increases population diversity by simulating the reproductive behavior of mantis shrimp. The breeding phase includes mating and offspring generation, where mating is performed by ranking the population by fitness, with the first half considered "female" and the second half "male". Offspring generation occurs as follows: each parent pair produces two offspring through crossover. ;in: It is a weight coefficient that is randomly generated in the range [0.5,1].
[0052] In some embodiments, the autoregressive moving average model ARMA (which mainly extracts time series residual features) in the combined prediction model includes an autoregressive model AR and a moving average model MA. The autoregressive moving average model first linearly combines the current time period with several past time periods to predict the time period. Then, it uses the prediction error of several past time periods to perform constraint correction. Then, it captures the short-term dependence of the time series. The improved mantis shrimp algorithm model optimizes the parameters of the autoregressive order of the autoregressive term of the autoregressive model AR and the moving average order of the moving average model MA.
[0053] In some embodiments, the LSTM model (mainly capturing long-term dependence information) in the combined prediction model is composed of three parts of input gate, forget gate and output gate. The LSTM model is used to obtain the long-term dependence information of the flight deviation average risk data set, and the mantis shrimp algorithm model is used to optimize the parameters of the hidden layer unit number, learning rate and dropout rate of the LSTM model. The expression of the forget gate is as follows: The expression of the input gate and the candidate cell state is as follows:
[0054] , ;
[0055] The expression of the cell state update is as follows:
[0056] The expression of the output gate is as follows:
[0057] , ;
[0058] In the above formula, , are the input and hidden states corresponding to the time , , and are the forget gate, input gate and output gate states respectively; , and are the to-be-updated neuron and cell states respectively; , , , and , , , are the weight matrices and bias terms of each gate respectively; is the Sigmoid activation function in the hidden layer. represents the hidden state of the current time and the previous time step, represents the generated prediction result.
[0059] The Transformer model includes an encoder, a decoder, a multi-head attention mechanism, and a feedforward network. The Transformer model (mainly used to capture global features) captures global relationships through parallel computation (the self-attention mechanism in the encoder searches for global relationships in the input data, processes the output through a feedforward neural network to output a high-dimensional continuous sequence, and the decoder transforms the encoder's data into the target sequence, captures the output relationship between the decoder and the encoder, and then processes the output through a feedforward neural network to output the target sequence). The improved Mantis Shrimp algorithm model optimizes the parameters of the Transformer model, including the number of encoder layers, the number of attention heads, and the dimensions of the feedforward network.
[0060] In some embodiments, the XGBoost model in the combined prediction model constructs several weak learners. Each weak learner uses the fitting error of the previous weak learner as the learning target to continue fitting. Then, all base learners are summed and the prediction result is output. The prediction expression is as follows:
[0061] ,in For the first Wheel pair samples The prediction results For the first Wheel pair samples The prediction results For the first The base learner trained in rounds on samples The predicted output; the objective function expression of the XGBoost model is as follows:
[0062] ,in The objective function of the XGBoost model; For the prediction results Compared with the actual results The loss function; For regularization; the improved mantis shrimp algorithm model optimizes the parameters of the XGBoost model, including the maximum tree depth, learning rate, and sampling ratio per tree.
[0063] The combined prediction model of this invention is a combined model composed of an autoregressive moving average model, an LSTM model, a Transformer model, and an XGBoost model. In this embodiment, the model hyperparameters of the combined prediction model (i.e., the key parameters of the autoregressive moving average model, LSTM model, Transformer model, and XGBoost model, with each model automatically weighted based on RMSE performance, and the prediction results of each model integrated using a weighted average method) include:
[0064] (1) Autoregressive order of the autoregressive moving average model ARMA (autoregressive order p in the example of the embodiment is in [1, 5]), moving average order (moving average order q in the example of the embodiment is in [1, 5]);
[0065] (2) Number of hidden layer units of the LSTM model (number of hidden layer units in the example of the embodiment is in [50, 200]), learning rate (learning rate in the example of the embodiment is in [0.001, 0.1]), dropout rate (dropout rate in the example of the embodiment is in [0, 0.5]);
[0066] (3) Number of encoder layers of the Transformer model (number of encoder layers in the example of the embodiment is in [2, 6]), number of attention heads (number of attention heads in the example of the embodiment is in [2, 8]), dimension of the feedforward network (dimension of the feedforward network in the example of the embodiment is in [128, 512]); (4) Maximum depth of the tree of the XGBoost model (maximum depth of the tree in the example of the embodiment is in [3, 10]), learning rate (learning rate in the example of the embodiment is in [0.01, 0.3]), sampling ratio of each tree (the sampling ratio of each tree is also called the subsampling ratio, and the subsampling ratio in the example of the embodiment is in [0.7, 1.0]). The improved mantis shrimp algorithm model searches for the optimal hyperparameter combination of the combined prediction model, takes the RMSE of the test set as the fitness function, updates the population in the search, predation (adaptive step and FDB guidance), and reproduction stages of the improved mantis shrimp algorithm model MShOA, iteratively optimizes to convergence, and outputs the globally optimal hyperparameter combination . The autoregressive moving average model ARMA, the LSTM model, and the Transformer model parameters in the optimal hyperparameter combination are used to retrain the models on the training set: the autoregressive moving average model ARMA is used to fit the time series data and outputs the residual sequence as a linear residual feature; the LSTM model is used to extract the hidden state vector at the last time step as a long-term dependence feature; and the Transformer model extracts the global attention feature vector of the encoder output. The original feature matrix is spliced with the ARMA residual, the LSTM hidden state, and the Transformer encoding feature to form an enhanced feature matrix X fused. The XGBoost model is trained based on the fused features, gradient boosting trees are used to realize accurate regression prediction, the XGBoost parameters in the optimal hyperparameter combination are used, the XGBoost model uses the enhanced feature matrix X fused to train the model, and outputs the prediction result. The combined prediction model and each model (each model in the combined prediction model) are evaluated in terms of effect by using evaluation indexes such as RMSE, MSE, MAE, and R². For example Figure 3As shown in the event data case of a certain country occurring from January 2017 to August 2025, the flight deviation risk true value and the flight deviation risk predicted value are tested and verified by using the test set, for example Figure 3 As shown in the event data case of a certain country occurring from January 2017 to August 2025, the flight deviation risk true value and the flight deviation risk predicted value are tested and verified by using the test set, for example Figure 3 It can be seen that the flight deviation risk predicted value is very close to the flight deviation risk true value in the time trend, and the deviation of the flight deviation risk predicted value and the flight deviation risk true value is close. The case evaluation indicators by using the test set are as follows: test set Testing RMSE: 0.11636, test set Testing MSE: 0.01354, test set Testing MAE: 0.085487, test set Testing R²: 0.81398.
[0067] In this embodiment, the root mean square error RMSE of the test set is used as the fitness value, and the calculation expression is as follows: , wherein is the number of samples; is the true value; is the predicted value.
[0068] The function of the fitness function is to evaluate the advantages and disadvantages of each group of parameters and guide the algorithm to search for the optimal parameter combination in the search space. Through continuous iteration, the algorithm will gradually find the parameter combination that minimizes the RMSE, thereby optimizing the prediction performance of the model.
[0069] S3, the combined prediction model outputs the flight deviation average risk sequence of a plurality of time periods after the research time according to the time sequence.
[0070] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. Any modification, equivalent replacement and improvement made within the spirit and principle of the present application shall be included in the protection scope of the present application.
Claims
1. A method for predicting flight deviation risk based on a mantis shrimp algorithm-optimized combined model, characterized in that: The methods include: S1. Construct a flight deviation risk assessment index system consisting of attitude deviation and altitude deviation. Obtain historical flight deviation event data before the research time and divide the data into time periods. Calculate the average flight deviation risk for each time period based on the flight deviation risk assessment index system. Collect the average flight deviation risk for each time period in sequence according to the time series to obtain the average flight deviation risk dataset. S2. Construct a combined prediction model that includes an autoregressive moving average model, an LSTM model, a Transformer model, and an XGBoost model. The combined prediction model is trained using the flight deviation average risk dataset for time series data prediction. The combined prediction model uses an improved mantis shrimp algorithm model for model hyperparameter combination optimization. S3. The combined prediction model outputs the average risk sequence of flight deviations for several time periods after the research time according to the time series.
2. The flight deviation risk prediction method based on the mantis shrimp algorithm optimization combination model according to claim 1, characterized in that: The improved mantis shrimp algorithm model includes an initialization phase, a search phase, a predation phase, and a reproduction phase. In the initialization phase, half of the population is randomly initialized within the search space. Then, the other half is generated using elite back-learning. The two populations are then merged, and the optimal set of individuals is selected as the initial population. In the search phase, the initial population simulates the random search and leader-following mechanism of mantis shrimp when searching for prey, exploring the search space. In the predation phase, individuals with good fitness and a relatively large distance from the current individual are selected as leaders, and an adaptive step size is used. In the reproduction phase, population diversity is increased by simulating the reproductive behavior of mantis shrimp.
3. The flight deviation risk prediction method based on the mantis shrimp algorithm optimization combination model according to claim 2, characterized in that: The improved mantis shrimp algorithm model utilizes elite back-learning to generate the following expression during the initialization phase: ,in New individuals generated through elite reverse learning To randomly initialize the best-fit individual from half of the population, These are model parameters used to adjust the position of the inverse solution within the range of [0,1]. , These are the lower and upper bounds of the search space, respectively; the initial population consists of N individuals selected from the two populations; the individual random search expression of the improved mantis shrimp algorithm model in the search phase is as follows: ,in Let i be the position of mantis shrimp individual i at iteration time t+1. Let i be the position of mantis shrimp individual i at iteration time t. , These represent the positions of two randomly selected individuals from the population at iteration time t. This represents the globally optimal position for the population at iteration time t. , All coefficients are random. The expression for the leader-following mechanism in the search phase of the improved mantis shrimp algorithm model is as follows: ,in A leader position is randomly selected from the population of high-quality solutions. These are random coefficients.
4. The flight deviation risk prediction method based on the mantis shrimp algorithm optimization combination model according to claim 2 or 3, characterized in that: The improved mantis shrimp algorithm model uses an adaptive step size. The expression is as follows: , This represents the maximum step size in the initial stage of the iteration. This represents the minimum step size in the later stages of the iteration. This is the total iteration time; after updating, individuals follow an adaptive step size. Update, the location of the updated individual The expression is as follows: ,in For the current position of the individual, The selected facilitator position.
5. The flight deviation risk prediction method based on the mantis shrimp algorithm optimization combination model according to claim 1, characterized in that: The combined prediction model includes the autoregressive moving average model (AR) and the moving average model (MA). The AR model first linearly combines the current time period with several past time periods to predict the time period. Then, it uses the prediction error of several past time periods for constraint correction. Finally, it captures the short-term dependence of the time series. The improved mantis shrimp algorithm model optimizes the parameters of the autoregressive order of the AR autoregressive term and the moving average order of the MA moving average model.
6. The flight deviation risk prediction method based on the mantis shrimp algorithm optimization combination model according to claim 1, characterized in that: The LSTM model neurons in the combined prediction model consist of three parts: input gate, forget gate, and output gate. The LSTM model is used to obtain long-term dependency information of the average risk dataset of flight deviation. The improved Mantis Shrimp algorithm model optimizes the parameters of the LSTM model, including the number of hidden layer units, learning rate, and dropout rate. The Transformer model includes an encoder, a decoder, a multi-head attention mechanism, and a feedforward network. The Transformer model captures global relationships through parallel computation. The improved Mantis Shrimp algorithm optimizes the parameters of the Transformer model, including the number of encoder layers, the number of attention heads, and the dimensions of the feedforward network.
7. The flight deviation risk prediction method based on the mantis shrimp algorithm optimization combination model according to claim 1, characterized in that: The XGBoost model in the combined prediction model constructs several weak learners. Each weak learner uses the fitting error of the previous weak learner as the learning target to continue fitting. Then, all base learners are summed and the prediction result is output. The prediction expression is as follows: ,in For the first Wheel pair samples The prediction results For the first Wheel pair samples The prediction results For the first The base learner trained in rounds on samples The predicted output; the objective function expression of the XGBoost model is as follows: ,in The objective function of the XGBoost model; For the prediction results Compared with the actual results The loss function; For regularization; the improved mantis shrimp algorithm model optimizes the parameters of the XGBoost model, including the maximum tree depth, learning rate, and sampling ratio per tree.
8. The flight deviation risk prediction method based on the mantis shrimp algorithm optimization combination model according to claim 1, characterized in that: In method S1, the average risk of flight deviation is calculated based on historical flight deviation event data for different time periods. The calculation expression is as follows: ,in This represents the risk of attitude deviation for the corresponding time period. This represents the risk of significant deviation within a given time period. The weights of the flight phases corresponding to the time periods. The environmental impact factors corresponding to the time period are stored in the flight deviation risk assessment index system, which contains the weights of flight stages according to the flight stages. and environmental impact factors classified according to environmental impact data. The flight time for a time period is the sum of the flight times of historical flight deviation events within that time period.
9. The flight deviation risk prediction method based on the mantis shrimp algorithm optimization combination model according to claim 8, characterized in that: The risk of attitude deviation The following formula is used to calculate the result based on historical flight deviation event data: ， The risk of high deviation The following formula is used to calculate the result based on historical flight deviation event data: ,in This is the actual pitch angle. To preset the standard pitch angle, For the preset threshold, This is the actual slope. To preset the standard slope, The threshold is set, and α and β are the weighting coefficients, respectively. This is the actual height. Specify the height for the preset.
10. The flight deviation risk prediction method based on the mantis shrimp algorithm optimization combination model according to claim 1, 8, or 9, characterized in that: The time period is divided into days, and the duration of the time period is N1 days; or the time period is divided into months, and the duration of the time period is one calendar month.