Flight area operation situation prediction method based on trainable fuzzy rule network

Abstract

The application provides a flight area operation situation prediction method based on a trainable fuzzy rule network, which firstly standardizes and time-aligns airport and meteorological data to construct 15-minute operation situation sequences. Then, 30-minute and 60-minute multi-scale feature sequences are generated, input into a three-way Transformer encoder to extract multi-scale memory representation, and after processing by a cross-scale fusion module, a self-recurrent decoder is used to predict future multi-step operation situation indicators. Finally, a trainable fuzzy rule residual head is introduced, fuzzy facts are constructed based on historical and known future features, a rule activation is calculated to generate a residual, which is superimposed with the main prediction to output the final prediction result and rule contribution explanation. Through multi-granularity time series modeling and a trainable fuzzy rule residual head, the application significantly improves the prediction accuracy and model explainability, and promotes the practical application of deep learning in the field of aviation operation management.

Description

Technical Field

[0001] This invention relates to the field of intelligent prediction of airport operation status and aviation operation management technology, and in particular to a method for predicting flight area operation status based on trainable fuzzy rule networks. Background Technology

[0002] Flight zone operational situation prediction is a core component of airport operations management and air-ground collaborative decision-making, directly impacting flight departure efficiency, ground taxiing safety, runway resource utilization, and overall delay control. With the continuous growth of air transport volume, airport operational loads during peak hours are constantly increasing, resulting in flight zone operational situations exhibiting strong time-varying and stochastic characteristics. Accurately predicting indicators such as takeoff traffic volume, average additional taxiing time, and ground queue status helps air traffic control departments formulate flow control and resource scheduling strategies in advance, improving operational efficiency and reducing delay risks. With improved data acquisition capabilities and the continuous growth in the scale of airport operational and meteorological data, machine learning and deep learning methods (such as LSTM and Transformer) are gradually becoming the mainstream solutions for operational situation prediction.

[0003] For example, Lim et al. proposed a machine learning-based method for predicting aircraft pushback time in their paper "Variable Taxi-OutTime Prediction Based on Machine Learning with Interpretable Attributes." This method decomposes the aircraft pushback process into two parts: the taxiing time from the parking stand to the takeoff queue and the waiting time within the queue. By training a gradient boosting decision tree model with different feature sets, they generated pushback time predictions with higher accuracy than traditional methods. However, because the operational patterns at different time scales are difficult to simultaneously characterize in a single-scale model, short-term fluctuations and medium-term trends often exhibit modeling biases. Furthermore, the models generally lack interpretability and struggle to output rules that provide clear meaning for the prediction results. There is an urgent need for a flight zone operational status prediction method that can integrate multi-granularity time-series information, provide interpretable rule outputs, and adapt to uncertain environments to improve prediction accuracy, enhance interpretability, and increase the feasibility of practical applications. Summary of the Invention

[0004] Purpose of the invention: The technical problem to be solved by the present invention is to address the shortcomings of the prior art by providing a flight zone operation status prediction method based on a trainable fuzzy rule network. By constructing multi-granularity temporal input, cross-scale fusion representation and trainable fuzzy rule residual head, the invention achieves accurate prediction and interpretable output of flight zone operation status.

[0005] The method includes the following steps:

[0006] Step S1: Read historical airport operation data and meteorological data, complete field standardization, time alignment, and anomaly cleaning. Using 15-minute intervals as the basic time window, statistically analyze flight schedules and actual events to construct ground operation status sequence features. Align and merge meteorological features to obtain the basic feature sequence. , where T is the number of time steps and F is the feature dimension of the basic feature sequence; Represent the space of real numbers;

[0007] Step S2: Based on the history window length and prediction step size Constructing supervised learning samples: obtaining historical input sequences by slicing from the feature sequence. Future known input Target output Where N is the number of samples, K is the dimension of the known future input, and the prediction targets are three situation indicators: the number of takeoffs, the average additional taxiing time, and the ground queue length in the future P steps.

[0008] Step S3: Perform multi-granularity aggregation on the 15-minute historical feature sequence according to the running attributes of the features. For count-type features, sum them within the corresponding aggregation interval. For state-type features, retain the end value of the corresponding aggregation interval. For continuous features, average them within the corresponding aggregation interval to obtain the 30-minute historical feature sequence and the 60-minute historical feature sequence.

[0009] 15-minute historical feature sequence 30-minute historical feature sequence With 60-minute historical feature sequence By inputting three non-shared parameter Transformer encoders respectively, a three-scale memory representation is obtained: , where D is the hidden dimension; and during the training phase, the 15-minute memory representation is downsampled and aligned so that the 15-minute memory representation is calculated with the 30-minute and 60-minute memory representations respectively to maintain the consistency of the running situation representation at different time granularities;

[0010] Step S4: The three-scale memory representations are fused through a cross-scale fusion module using gated attention fusion to obtain the fused memory representation. Then input it into the autoregressive Transformer decoder, combined with known future inputs. Predict future P-step output step by step The gated attention fusion first calculates scale weights based on the scale-level summaries of the three-scale memory representations, then weights and concatenates the three-scale memory representations according to the scale weights, and completes cross-scale information interaction through the attention fusion encoder; the autoregressive Transformer decoder uses a causal mask during the prediction process, making the third... Each prediction step uses only historical prediction prefixes and corresponding future known input prefixes to generate the main model prediction results;

[0011] Step S5: Construct a fuzzy fact vector based on the original historical features and known future features. ,in The dimension of the fuzzy fact vector is represented by an unstandardized historical operational information and future known information. This fuzzy fact vector characterizes the impact of flight area operational load, demand changes, weather disturbances, and delay transmission on future operational trends. Rule activation is calculated using trainable monotonic predicates and soft AND rules. The trainable monotonic predicates convert fuzzy facts into membership degrees with operational semantics, while the soft AND rules characterize the rule triggering strength under the combined influence of two or more operational factors, generating rule residuals. The final prediction result is obtained by superimposing the learnable scaling factor with the prediction result of the main model. During the training phase, the prediction loss is calculated based on the final prediction result and the target output. The prediction loss, the cross-scale consistency constraint obtained in step S3, and the rule sparsity constraint determined by the rule-averaged activation are then weighted and combined to form the overall training objective. Based on this overall training objective, the three-scale Transformer encoder, the cross-scale fusion module, the autoregressive Transformer decoder, and the trainable fuzzy rule residual head are jointly optimized. After training, the rule activation and rule contribution of the samples are extracted, and interpretable rule descriptions are generated by sorting them by contribution. Here, tanh is the hyperbolic tangent function.

[0012] Step S3 includes:

[0013] Step S3.1: Transfer the historical input sequence Multi-granularity aggregation based on time scale yields historical feature sequences at three scales. , , The 30-minute and 60-minute historical feature sequences are generated from 15-minute sequences according to preset aggregation rules. These preset aggregation rules generally fall into three categories: 1) summation of count-type or cumulative features within the aggregation window, such as planned departures, planned arrivals, pushbacks, takeoffs, landings, and cumulative flight delay reasons; 2) taking the value of the last 15-minute period within the aggregation window for status-type features, such as ground queue length, arrival backlog status, and cloud cover level; and 3) taking the average value within the aggregation window for continuous features, such as average additional taxiing time, cancellation rate, temperature, dew point, visibility, and wind speed. For example, a 30-minute sequence is obtained by aggregating two adjacent 15-minute periods; a 60-minute sequence is obtained by aggregating four adjacent 15-minute periods.

[0014] Step S3.2: Input the historical feature sequences at the three scales into three non-shared parameter Transformer encoders respectively, and obtain the following through linear mapping at the l-th layer:

[0015] (1),

[0016] in Indicates the input number The historical feature representation matrix of the layer Transformer encoder can be a historical feature sequence representation at a scale of 15 min, 30 min, or 60 min. , , These are query and key-value matrix, respectively. , , Both are projection matrices, and then the attention matrices are calculated separately;

[0017] Step S3.3: After L layers of iteration, the final memory representations of the three-scale encoders are obtained respectively. , , .

[0018] In step S3.2, the attention matrix is ​​calculated using the following formula:

[0019] (2),

[0020] (3),

[0021] in It is the matrix dimension. It is an attention matrix. It is the output of the l-th layer model; softmax is the normalized exponential function.

[0022] Step S4 includes:

[0023] Step S4.1: To obtain the importance weights of different time scales for prediction, first perform a three-scale memory representation. , , Perform mean pooling along the time dimension to obtain scale-level vectors. , , ; scale-level vector , , The concatenated input is then fed into a gated network MLP, which outputs a gated score and obtains the weight vector w using the softmax function.

[0024] (4),

[0025] in , These represent the contribution ratios at the 15-minute, 30-minute, and 60-minute scales, respectively, for subsequent adaptive fusion.

[0026] Step S4.2: Scale the three-scale memories according to the gating weights to highlight the time scales that are more critical to the current sample, and then concatenate them along the time dimension to obtain the scale-fused candidate representation:

[0027] (5),

[0028] (6),

[0029] in This indicates a splicing operation. , , These are the weighted 15-minute scaled memory representations, 30-minute scaled memory representations, and 60-minute scaled memory representations, respectively. It is a spliced ​​representation of memory;

[0030] Step S4.3: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require the full context.] Input an attention-fusion encoder (TransformerEncoder), perform cross-scale mixing within the same attention space, and output a fused memory representation. Furthermore, the training process is stabilized through a normalization layer;

[0031] Step S4.4: At prediction step t, add the historical prediction prefix. With known features of the future The input to the decoder is obtained by concatenating the data. ;

[0032] Step S4.5: Apply a causal mask to the decoder input to ensure timing consistency, utilizing fused memory. Decode step by step to obtain each prediction. , Finally, the situation prediction sequence for the next P steps is obtained. .

[0033] In step S4.3, the memory representation is fused. The calculation formula is:

[0034] (7).

[0035] In step S4.4, the decoder input The calculation formula is:

[0036] (8).

[0037] In step S4.5, The calculation formula is:

[0038] (9),

[0039] in Indicates the first The output of each prediction step is the flight area operational situation prediction result, including target indicators such as the number of takeoffs, average additional taxi time, and ground queue length; Decoder represents the autoregressive Transformer decoder, which is used to progressively generate the situation prediction results for each future prediction step based on the decoder input and fused memory representation.

[0040] Step S5 includes:

[0041] Step S5.1: Construct a fuzzy fact vector based on original historical features and known future features. Each prediction step corresponds to a set of factual features, including queue status, cancellation rate, planned takeoffs and landings, time information, and weather features, which are used to express the key prior factors affecting changes in the operational status.

[0042] Step S5.2: Establish a monotonically trainable predicate for each feature dimension in the fact vector, and learn the threshold. With slope Obtain membership degree:

[0043] (10)

[0044] in The input value is a one-dimensional original feature in the fuzzy fact vector. The membership degree of the feature value belonging to the high-level fuzzy semantics. The membership degree of the feature value belonging to low-level fuzzy semantics. It is the sigmoid function;

[0045] Step S5.3: Combine the predicates according to the rule structure and activate the rules using a soft AND operation in product form:

[0046] (11),

[0047] in Indicates the first The membership degree corresponding to each fuzzy predicate ranges from 0 to 1. This represents the set of predicates contained in the j-th rule, and the rule activation. This indicates the trigger strength of the j-th rule in the current sample and prediction step;

[0048] Step S5.4: Activate the rule vector Generating regular residuals through linear mapping And introduce a learnable scaling factor. Controlling residual contribution, among which Represents the regular residual mapping matrix. The bias vector representing the regular residual mapping, and the final prediction result are:

[0049] (12),

[0050] in The prediction results of the main model;

[0051] Step S5.5: Extract rule activation from the trained fuzzy rule head The rules are weighted by the output, and the contribution of each rule to the prediction target is calculated. The rules are then sorted by contribution and output, generating an interpretable explanation that provides the basis for the prediction results.

[0052] The present invention also provides an electronic device, including a processor and a memory, the memory storing program code that, when executed by the processor, causes the processor to perform the steps of the method.

[0053] The present invention also provides a storage medium storing a computer program or instructions that, when the computer program or instructions are run on a computer, execute the steps of the method described.

[0054] Beneficial effects:

[0055] 1. This invention proposes a flight area operation status prediction method based on multi-granularity time series modeling. By encoding historical feature sequences at different time scales of 15 min, 30 min, and 60 min respectively, and combining them with a cross-scale fusion module to achieve joint modeling of multi-scale time series information, it can simultaneously characterize the short-term fluctuation characteristics and medium- and long-term change trends of the flight area operation status. Compared with traditional single-time scale prediction methods, it can effectively improve the prediction accuracy of key indicators such as takeoff number, average additional taxiing time, and ground queue length.

[0056] 2. This invention introduces a trainable fuzzy rule residual head. Based on the output of the main prediction model, it constructs fuzzy facts by combining historical operational features and known future features. It also uses trainable monotonic predicates and soft AND rules to calculate the rule activation strength and generate rule residuals to compensate and correct the prediction results of the main model. This enhances the model's ability to express complex nonlinear relationships, uncertainties, and fuzzy boundary scenarios, and improves the stability and robustness of flight zone operational situation prediction results.

[0057] 3. While outputting prediction results, this invention can further extract the rule activation degree and rule contribution information in the fuzzy rule network, and provide corresponding rule basis and influence sources for the prediction results, thus solving the problem of insufficient interpretability of existing deep learning prediction methods. At the same time, this invention integrates airport operation data, flight plan information and meteorological data to realize the prediction of future multi-step operation status, which can provide effective support for airport traffic management, operation scheduling and resource allocation, and has high engineering application value. Attached Figure Description

[0058] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments, and the advantages of the present invention in the above and / or other aspects will become clearer.

[0059] Figure 1 This is a diagram illustrating the overall framework of the method of this invention.

[0060] Figure 2 This is a flowchart of the method of the present invention. Detailed Implementation

[0061] See Figures 1 to 2 This invention provides a method for predicting the operational status of a flight area based on a trainable fuzzy rule network. Taking an airport operation dataset as an example, the method includes the following steps from loading the data to obtaining the situation prediction results:

[0062] S1: Read historical airport operation data and meteorological data, perform field standardization, time alignment, and anomaly cleaning, statistically analyze flight schedules and actual events, construct ground operation status sequence features, and align and merge meteorological features to obtain the basic feature sequence. , where T is the number of time steps and F is the feature dimension of the basic feature sequence;

[0063] S2: Construct supervised samples with historical window length H: obtain historical input by slicing the feature sequence. Future known input Target output Where N is the number of samples, K is the dimension of the known future inputs, and the prediction targets are three situational indicators: the number of takeoffs, the average additional taxiing time, and the ground queue length in the next P steps. The first sample... , and As shown in formulas (1), (2) and (3);

[0064] (1),

[0065] (2),

[0066] (3),

[0067] S3: Perform multi-granularity aggregation on the 15-minute historical features to obtain the 30-minute and 60-minute historical feature sequences: By inputting three non-shared parameter Transformer encoders respectively, a three-scale memory representation is obtained: Where D is the hidden dimension, and the first sample's... , and As shown in formulas (4), (5), and (6):

[0068] (4),

[0069] (5),

[0070] (6),

[0071] S4: The three-scale memory representations are fused through a cross-scale fusion module using gated attention fusion to obtain the fused memory representation. Then input the autoregressive Transformer decoder, combined with the known future input. Predict future P-step output step by step The future P-step output of the first sample As in formula (7):

[0072] (7),

[0073] S5: Constructing a fuzzy fact vector based on original historical features and known future features. in To represent the dimension of the fuzzy fact vector, we use trainable monotonic predicates and soft AND rules to calculate rule activation and generate rule residuals. The final prediction result is obtained by superimposing the learnable scaling factor with the prediction result of the main model. Based on the trained fuzzy rule head, the rule activation and rule contribution of the samples are extracted, and interpretable rule descriptions are generated by sorting them by contribution.

[0074] Furthermore, the specific steps of step S3 are as follows:

[0075] Step S3.1: Transfer the historical input sequence Multi-granularity aggregation based on time scale yields historical feature sequences at three scales. , , The 30-minute and 60-minute sequences were generated from the 15-minute sequences according to a preset aggregation rule.

[0076] Step S3.2: Input the feature sequences of the three scales into three non-shared parameter Transformer encoders respectively, and obtain the following through linear mapping at the l-th layer:

[0077] (8),

[0078] in , , It is a query, a key-value matrix. , , These are all projection matrices; then, the attention matrices are calculated separately:

[0079] (9),

[0080] in It is the matrix dimension. It is the attention matrix:

[0081] (10)

[0082] in It is the output of the l-th layer model;

[0083] Step S3.3: After L layers of iteration, the final memory representations of the three-scale encoders are obtained respectively. , , .

[0084] Furthermore, the specific steps of step S4 are as follows:

[0085] Step S4.1: To obtain the importance weights of different time scales for prediction, first perform a three-scale memory representation. , , Perform mean pooling along the time dimension to obtain a scale-level vector. , , The three-scale vectors are concatenated and input into a gated network (MLP) to output the gate score. The weight vector is then obtained using the softmax function.

[0086] (11),

[0087] in , representing the contribution ratio of the three scales, used for subsequent adaptive fusion, vector As shown in formula (12):

[0088] (12),

[0089] Step S4.2: Scale the three-scale memories according to the gating weights to highlight the time scales that are more critical to the current sample, and then concatenate them along the time dimension to obtain the scale-fused candidate representation:

[0090] (13)

[0091] (14)

[0092] in This indicates a splicing operation. , , This is a weighted three-scale memory representation. It is a spliced ​​representation of memory;

[0093] Step S4.3: To explicitly model the interaction relationships between memories at different scales, Input an attention-fusion encoder (TransformerEncoder), perform cross-scale mixing within the same attention space, and output a fused memory representation. :

[0094] (15)

[0095] Furthermore, the training process is stabilized through normalization layers, ensuring that the fused representation retains both scale differences and global consistency. This is achieved through the fusion of memory representations. As shown in formula (16):

[0096] (16)

[0097] Step S4.4: At prediction step t, add the historical prediction prefix. With known features of the future By concatenating the data, we obtain the decoder input:

[0098] (17)

[0099] The input includes both the predicted trend of operational status evolution and prior known planning information;

[0100] Step S4.5: Apply a causal mask to the decoder input to ensure timing consistency, utilizing fused memory. Decode step by step to obtain each prediction:

[0101] (18)

[0102] Finally, obtain the situation prediction sequence for the next P steps. .

[0103] Furthermore, the specific steps of step S5 are as follows:

[0104] Step S5.1: Construct a fuzzy fact vector based on original historical features and known future features. Each prediction step corresponds to a set of factual features, including queue status, cancellation rate, planned takeoffs and landings, time information, and weather characteristics, used to express the key prior factors affecting changes in operational status. The fuzzy fact vector of the first sample... As shown in formula (19):

[0105] (19)

[0106] Step S5.2: Establish a monotonically trainable predicate for each feature dimension in the fact vector, and learn the threshold. With slope Obtain membership degree:

[0107] (20)

[0108] in For the sigmoid function, Used to control the steepness of the fuzzy membership function, thereby avoiding the non-differentiability problem caused by hard thresholding;

[0109] Step S5.3: Combine multiple predicates according to the rule structure, and activate the rules using a soft AND operation in product form:

[0110] (twenty one),

[0111] in This represents the set of predicates contained in the j-th rule, and the rule activation. This indicates the trigger strength of the rule on the current sample and in the prediction step; rule activation. As shown in formula (22):

[0112] (twenty two),

[0113] Step S5.4: Generate rule residuals from the rule activation vectors through linear mapping. And introduce a learnable scaling factor. Controlling for residual contributions, the final prediction result is:

[0114] (12),

[0115] in The main model's prediction results Limit the contribution of rule residuals to a stable range;

[0116] Step S5.5: Extract rule activation from the trained fuzzy rule head The weight coefficients from the rules to the output are used to calculate the contribution of each rule to the prediction target. The rules are then sorted by contribution and output, generating an interpretable description that provides the rule basis for the prediction result.

[0117] As shown in Table 1, the test set of the dataset was put into the model to obtain the situation prediction results. The performance of the situation prediction results was evaluated using the real situation labels. The root mean square error of the ground queue length situation index was 5.54 and the mean absolute error was 3.48; the root mean square error of the ground congestion situation index was 2.42 and the mean absolute error was 1.67, which are significantly better than the traditional situation prediction method.

[0118] Table 1

[0119]

[0120] This invention provides a method for predicting flight zone operational status based on a trainable fuzzy rule network. Many methods and approaches exist for implementing this technical solution; the above description is merely a preferred embodiment of the invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications should also be considered within the scope of protection of this invention. All components not explicitly stated in this embodiment can be implemented using existing technologies.

Claims

1. A method for predicting flight zone operational status based on trainable fuzzy rule networks, characterized in that, Includes the following steps: Step S1: Read historical airport operation data and meteorological data, complete field standardization, time alignment, and anomaly cleaning. Using 15-minute intervals as the basic time window, statistically analyze flight schedules and actual events to construct ground operation status sequence features. Align and merge meteorological features to obtain the basic feature sequence. , where T is the number of time steps and F is the feature dimension of the basic feature sequence; Represent the space of real numbers; Step S2: Based on the history window length and prediction step size Constructing supervised learning samples: obtaining historical input sequences by slicing from the feature sequence. Future known input Target output Where N is the number of samples, K is the dimension of the known future input, and the prediction targets are three situation indicators: the number of takeoffs, the average additional taxiing time, and the ground queue length in the future P steps. Step S3: Perform multi-granularity aggregation on the 15-minute historical feature sequence according to the operational attributes of the features. Specifically, sum the count-type features within their corresponding aggregation intervals, retain the end values ​​of the corresponding aggregation intervals for state-type features, and average the continuous features within their corresponding aggregation intervals to obtain the 30-minute and 60-minute historical feature sequences. Then, process the 15-minute historical feature sequence... 30-minute historical feature sequence With 60-minute historical feature sequence By inputting three non-shared parameter Transformer encoders respectively, a three-scale memory representation is obtained: , where D is the hidden dimension; and during the training phase, the 15-minute memory representation is downsampled and aligned so that the 15-minute memory representation is calculated with the 30-minute and 60-minute memory representations respectively to maintain the consistency of the running situation representation at different time granularities; Step S4: The three-scale memory representations are fused through a cross-scale fusion module using gated attention fusion to obtain the fused memory representation. Then input it into the autoregressive Transformer decoder, combined with known future inputs. Predict future P-step output step by step The gated attention fusion first calculates scale weights based on the scale-level summaries of the three-scale memory representations, then weights and concatenates the three-scale memory representations according to the scale weights, and completes cross-scale information interaction through the attention fusion encoder; the autoregressive Transformer decoder uses a causal mask during the prediction process, making the third... Each prediction step uses only historical prediction prefixes and corresponding future known input prefixes to generate the main model prediction results; Step S5: Construct a fuzzy fact vector based on the original historical features and the known future features. ,in The dimension of the fuzzy fact vector is represented by a formula. This fuzzy fact vector is extracted from unstandardized historical operational information and future known information. Rule activation is calculated using trainable monotonic predicates and soft AND rules. The trainable monotonic predicates are used to convert fuzzy facts into membership degrees with operational semantics, and the soft AND rules are used to characterize the rule triggering strength under the combined effect of two or more operational factors, generating rule residuals. The final prediction result is obtained by superimposing the learnable scaling factor with the prediction result of the main model. During the training phase, the prediction loss is calculated based on the final prediction result and the target output. The prediction loss, the cross-scale consistency constraint obtained in step S3, and the rule sparsity constraint determined by the rule average activation are weighted and combined to form the overall training objective. Based on the overall training objective, the three-scale Transformer encoder, the cross-scale fusion module, the autoregressive Transformer decoder, and the trainable fuzzy rule residual head are jointly optimized. After training, the rule activation and rule contribution of the samples are extracted and sorted by contribution to generate interpretable rule descriptions.

2. The method according to claim 1, characterized in that, Step S3 includes: Step S3.1: Transfer the historical input sequence Multi-granularity aggregation based on time scale yields historical feature sequences at three scales. , , The 30-minute and 60-minute historical feature sequences are generated from the 15-minute sequences according to a preset aggregation rule. Step S3.2: Input the historical feature sequences at the three scales into three non-shared parameter Transformer encoders respectively, and obtain the following through linear mapping at the l-th layer: （1）， in Indicates the input number The historical feature representation matrix of the layer Transformer encoder; , , These are query and key-value matrix, respectively. , , Both are projection matrices, and then the attention matrices are calculated separately; Step S3.3: After L layers of iteration, the final memory representations of the three-scale encoders are obtained respectively. , , .

3. The method according to claim 2, characterized in that, In step S3.2, the attention matrix is ​​calculated using the following formula: （2）， （3）， in It is the matrix dimension. It is an attention matrix. It is the output of the l-th layer model; softmax is the normalized exponential function.

4. The method according to claim 3, characterized in that, Step S4 includes: Step S4.1: To obtain the importance weights of different time scales for prediction, first perform a three-scale memory representation. , , Perform mean pooling along the time dimension to obtain scale-level vectors. , , ; scale-level vector , , The concatenated input is then fed into a gated network MLP, which outputs a gated score and obtains the weight vector w using the softmax function. （4）， in , These represent the contribution ratios at the 15-minute, 30-minute, and 60-minute scales, respectively. Step S4.2: Scale the three-scale memories according to the gating weights to highlight the time scales that are more critical to the current sample, and then concatenate them along the time dimension to obtain the scale-fused candidate representation: （5）， （6）， in This indicates a splicing operation. , , These are the weighted 15-minute scaled memory representations, 30-minute scaled memory representations, and 60-minute scaled memory representations, respectively. It is a spliced ​​representation of memory; Step S4.3: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require the full context.] Input an attention-fusion encoder (TransformerEncoder), perform cross-scale mixing within the same attention space, and output a fused memory representation. Furthermore, the training process is stabilized through a normalization layer; Step S4.4: At prediction step t, add the historical prediction prefix. With known features of the future The input to the decoder is obtained by concatenating the data. ; Step S4.5: Apply a causal mask to the decoder input to ensure timing consistency, utilizing fused memory. Decode step by step to obtain each prediction. , Finally, the situation prediction sequence for the next P steps is obtained. .

5. The method according to claim 4, characterized in that, In step S4.3, the memory representation is fused. The calculation formula is: （7）。 6. The method according to claim 5, characterized in that, In step S4.4, the decoder input The calculation formula is: （8）。 7. The method according to claim 6, characterized in that, In step S4.5, The calculation formula is: （9）， in Indicates the first The predicted flight zone operational status output for each prediction step; Decoder represents the autoregressive Transformer decoder.

8. The method according to claim 7, characterized in that, Step S5 includes: Step S5.1: Construct a fuzzy fact vector based on original historical features and known future features. Each prediction step corresponds to a set of factual features; Step S5.2: Establish a monotonically trainable predicate for each feature dimension in the fact vector, and learn the threshold. With slope Obtain membership degree: （10）， in The input value is a one-dimensional original feature in the fuzzy fact vector. The membership degree of the feature value belonging to the high-level fuzzy semantics. The membership degree of the feature value belonging to low-level fuzzy semantics. It is the sigmoid function; Step S5.3: Combine the predicates according to the rule structure and activate the rules using a soft AND operation in product form: （11）， in Indicates the first The membership degree corresponding to each fuzzy predicate ranges from 0 to 1. This represents the set of predicates contained in the j-th rule, and the rule activation. This indicates the trigger strength of the j-th rule in the current sample and prediction step; Step S5.4: Activate the rule vector Generating regular residuals through linear mapping And introduce a learnable scaling factor. Controlling residual contribution, among which Represents the regular residual mapping matrix. The bias vector representing the regular residual mapping, and the final prediction result are: （12）， in The main model prediction results; tanh is the hyperbolic tangent function; Step S5.5: Extract rule activation from the trained fuzzy rule head The rules are weighted by the output, and the contribution of each rule to the prediction target is calculated. The rules are then sorted by contribution and output, generating an interpretable explanation that provides the basis for the prediction results.

9. An electronic device, characterized in that, It includes a processor and a memory, the memory storing program code that, when executed by the processor, causes the processor to perform the steps of the method as described in any one of claims 1 to 8.

10. A storage medium, characterized in that, It stores a computer program or instructions that, when run on a computer, perform the steps of the method as described in any one of claims 1 to 8.

Images