A method for quickly predicting future airport throughput trends based on ridge regression algorithm

Abstract

This invention discloses a method for rapidly predicting future airport throughput trends based on the ridge regression algorithm, relating to the field of air transport operation management and data analysis technology. The method involves acquiring hourly departure and arrival passenger and cargo flow sequence data for each runway direction, terminal, and cargo area of ​​the airport over a historical period, while simultaneously acquiring the overall hourly throughput sequence data for the corresponding time period. Through data cleaning to remove outliers and standardization, a multi-channel flow time series set and an overall throughput benchmark sequence are obtained. This method for rapidly predicting future airport throughput trends based on the ridge regression algorithm improves the reliability, interpretability, and practical application value of airport throughput trend prediction in complex operational scenarios while ensuring prediction efficiency.

Description

Technical Field

[0001] This invention relates to the field of air transport operation management and data analysis technology, specifically to a method for rapidly predicting the future throughput trend of an airport based on the ridge regression algorithm. Background Technology

[0002] Existing airport throughput forecasting technologies are mostly based on historical statistical data to build models, with common methods including time series analysis, regression analysis, and some machine learning models. While these methods can depict long-term trends in airport throughput to some extent, they still have significant shortcomings in practical applications. On the one hand, some forecasting methods focus on modeling a single indicator of overall airport throughput, ignoring the complex flow structure within the airport, which consists of multiple business channels such as runway direction, terminal, and passenger / cargo types, making it difficult to reflect the mutual influence between different channels. On the other hand, when the number of input features is large and highly correlated, traditional regression models are easily affected by multicollinearity, leading to unstable forecast results and weakening the reliability of future trend predictions. From an operational perspective, airport throughput exhibits significant multi-level and multi-channel characteristics. Different channels, such as departures and arrivals, domestic and international, and passenger and cargo transport, show significant differences in temporal distribution, while all are constrained by the overall airport capacity. When a particular channel experiences concentrated traffic growth during a specific period, it often crowds out other channels and is limited by hard capacity constraints such as runway, terminal, and support resources. If the forecasting method fails to comprehensively consider the correlation, competition, and overall capacity boundaries between channels, and simply superimposes or independently extrapolates various types of traffic, it is prone to significant deviations during peak periods or complex operational scenarios. Furthermore, while some complex forecasting models offer advantages in accuracy, their complex structure and high computational cost make them unsuitable for rapid updates and rolling forecasts in airport operations management, failing to meet the actual scheduling requirements for rapid forecasting and trend analysis. Summary of the Invention

[0003] The purpose of this invention is to provide a method for rapidly predicting the future throughput trend of an airport based on the ridge regression algorithm, thereby solving the problems existing in the prior art.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a method for rapidly predicting the future throughput trend of an airport based on the ridge regression algorithm, comprising S1, obtaining a multi-channel traffic time series set and an overall throughput benchmark sequence; S2. For the traffic time series of each channel, each channel is independently modeled, and an initial independent prediction value sequence of each channel is generated within a fixed future time window to suppress multicollinearity between features and retain the trend information of each channel's prediction results. S3. Organize the independent prediction values ​​of all channels into a multi-level throughput prediction feature set that includes the bottom channel vector, the intermediate level aggregate vector, and the top level overall vector. S4. Based on the multi-level throughput prediction feature set, calculate the correlation coefficient between the bottom-level channels, the weight distribution of the intermediate-level channels, and the overall influence intensity between the top-level intermediate levels to obtain the weight coefficient set. S5. Based on the set of weighting coefficients, combined with the priority ranking of the proportion of peak traffic periods in the channels and the capacity allocation bias caused by the differences in channel types, the independent predicted values ​​of each channel are weighted and adjusted to obtain the intermediate predicted value sequence of the channels. S6. Accumulate and summarize all intermediate predicted values ​​of all channels layer by layer upwards according to the hierarchical structure to obtain a preliminary overall predicted throughput sequence. Compare the preliminary overall predicted throughput sequence with the airport capacity hard constraint threshold. If the preliminary overall predicted throughput sequence exceeds the capacity hard constraint threshold, trigger the capacity constraint projection calculation to adjust the prediction result; otherwise, proceed to the next step. S7. After projection of the capacity constraint, the channel prediction results are normalized and weighted and fused based on the set of weight coefficients and the channel prediction contribution output by the ridge regression model to obtain the final airport future throughput trend prediction result that meets the airport capacity constraint conditions.

[0005] Preferably, step S1 includes acquiring historical airport operation records, aggregating the historical operation records by hourly window to generate hourly traffic sequences for each region and an overall hourly throughput sequence for the airport; detecting outliers in the hourly traffic sequences for each region and the overall hourly throughput sequence for the airport, and interpolating to correct the gaps after removing outliers to obtain corrected traffic sequences without outliers; standardizing the corrected traffic sequences without outliers to generate standardized traffic value sequences; and obtaining a multi-channel traffic time series set and an overall throughput benchmark sequence based on the standardized traffic value sequences.

[0006] Preferably, step S2 includes acquiring the traffic time series for each channel, constructing a multi-factor feature vector based on a historical time window, wherein the multi-factor feature vector includes periodic features, peak features, and trend features; using the ridge regression algorithm to independently model the multi-factor feature vector, and suppressing multicollinearity among features in the multi-factor feature vector through L2 regularization constraints to obtain ridge regression model parameters; generating an initial independent prediction value sequence for each channel within a fixed future time window based on the ridge regression model parameters; and outputting the initial independent prediction value sequence, retaining the trend information of the prediction results for each channel within the fixed future time window.

[0007] Preferably, step S3 includes obtaining the initial independent predicted value sequence of each channel, constructing a structural mapping matrix describing the channel affiliation by combining the runway number, terminal area, and passenger and cargo identification; reorganizing and accumulating the initial independent predicted value sequence according to the structural mapping matrix to generate the bottom-level channel vector and the intermediate-level aggregated vector; globally summarizing the intermediate-level aggregated vector to obtain the top-level overall vector, and concatenating the top-level overall vector, the intermediate-level aggregated vector, and the bottom-level channel vector to output a multi-level throughput prediction feature set.

[0008] Preferably, step S4 includes obtaining the bottom-level channel vector sequence in the multi-level throughput prediction feature set, constructing a cross-correlation matrix for the bottom-level channel vector sequence; calculating the traffic proportion value based on the cross-correlation matrix and the intermediate-level aggregated vector to obtain the proportion weight distribution sequence of the intermediate-level channels; performing regression analysis on the proportion weight distribution sequence and the top-level overall vector to determine the overall influence intensity value between each intermediate level of the top level; and integrating the overall influence intensity value, the proportion weight distribution sequence and the cross-correlation matrix to construct a feature importance mapping table, and extracting a set of weight coefficients to characterize the differences in the influence of different channels on the overall throughput prediction result.

[0009] Preferably, step S5 includes acquiring historical traffic load data and independent predicted values ​​for each physical channel, generating a traffic peak period percentage sequence; constructing a comprehensive competition interaction matrix based on the traffic peak period percentage sequence, the capacity allocation bias value determined by the channel type difference, and the channel dynamic influence coefficient; applying the comprehensive competition interaction matrix to the independent predicted values ​​to extract a channel competition correction set; and combining the channel competition correction set with the independent predicted values ​​for weighted adjustment to obtain a channel intermediate prediction value sequence considering the competition relationship between channels.

[0010] Preferably, step S6 includes obtaining the intermediate predicted value sequence of the channel, merging and summing the intermediate predicted value sequence of the channel according to the preset channel hierarchical structure mapping relationship to obtain a preliminary overall predicted throughput sequence; comparing the preliminary overall predicted throughput sequence with the airport capacity hard constraint threshold, and generating a throughput overflow difference sequence if there is an overflow.

[0011] Preferably, step S6 further includes constructing a multidimensional capacity constraint projection vector based on the throughput overflow difference sequence, and inversely decomposing the multidimensional capacity constraint projection vector to obtain a channel projection correction set; applying the channel projection correction set to adjust the channel intermediate prediction value sequence, and outputting the final overall predicted throughput sequence that satisfies the airport capacity hard constraint threshold.

[0012] Preferably, step S7 includes obtaining a corrected channel prediction value sequence, which is obtained by superimposing the channel projection correction set onto the original prediction value; and decomposing the corrected channel prediction value sequence according to the regression coefficient matrix of the ridge regression model to obtain a set of channel prediction contribution values.

[0013] Preferably, step S7 further includes performing calculations on the preset set of weight coefficients and the set of channel prediction contribution values ​​to generate a normalized weighted factor vector; using the normalized weighted factor vector to perform weighted fusion on the corrected channel prediction value sequence, and outputting the final airport future throughput trend prediction result that meets the airport capacity constraint conditions.

[0014] As can be seen from the above technical solution, the present invention has the following beneficial effects: This method for rapidly predicting future airport throughput trends based on the ridge regression algorithm decomposes airport throughput into multiple channel flows with clear business meanings and models and merges them within a multi-level structural framework. This effectively characterizes the correlations and competitive constraints between different channels, such as runway direction, terminal, and passenger / cargo types. On one hand, the ridge regression algorithm is used to independently model the flow of each channel, and L2 regularization is used to suppress multicollinearity, improving the stability of the prediction results and the ability to characterize trend changes under multi-factor characteristics. This avoids the high computational costs associated with complex models, meeting the needs of airport operation management for rapid prediction. On the other hand, by introducing multi-level weight calculation, dynamic channel influence adjustment, and a capacity hard constraint projection mechanism, the prediction results reflect the mutual constraints between channels and conform to the overall airport capacity limitations, thereby reducing the risk of prediction distortion and resource allocation deviations during peak hours. In summary, this invention improves the reliability, interpretability, and practical application value of airport throughput trend prediction in complex operational scenarios while ensuring prediction efficiency. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating the method for predicting future airport throughput trends according to the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] like Figure 1 As shown, this invention provides a technical solution: a method for rapidly predicting the future throughput trend of an airport based on the ridge regression algorithm, comprising: S1. Obtain the hourly departure and arrival international and domestic passenger and cargo traffic sequence data of each runway direction, each terminal and each cargo area of ​​the airport in the historical period. At the same time, obtain the overall hourly throughput sequence data of the airport in the corresponding time period. Remove outliers through data cleaning and standardize the data to obtain a multi-channel traffic time series set and an overall throughput benchmark sequence. S2. For the traffic time series of each channel, construct a multi-factor feature vector based on the historical time window, use the ridge regression algorithm to independently model each channel, generate the initial independent prediction value sequence of each channel within a fixed future time window, and suppress multicollinearity between features through L2 regularization constraint, while retaining the trend information of change in the prediction results of each channel. S3. Based on the three-layer structure of the airport's pre-divided runways, terminals, and passenger and cargo types, organize the independent predicted values ​​of all channels into a multi-layered throughput prediction feature set that includes bottom-level channel vectors, intermediate-level aggregated vectors, and top-level overall vectors. S4. Based on the multi-level throughput prediction feature set, calculate the correlation coefficient between the bottom-level channels, the weight distribution of the intermediate-level channels, and the overall influence strength between the top-level intermediate levels to obtain a set of weight coefficients that reflect the correlation of channel input features. S5. Based on the channel dynamic influence coefficient, combined with the priority ranking of the proportion of peak traffic periods of the channels and the capacity allocation bias caused by the differences in channel types, the independent predicted values ​​of each channel are weighted and adjusted to obtain a sequence of intermediate predicted values ​​of the channels that takes into account the competition between channels. S6. Accumulate and summarize all intermediate predicted values ​​of all channels layer by layer upwards according to the hierarchical structure to obtain a preliminary overall predicted throughput sequence. Compare the preliminary overall predicted throughput sequence with the airport capacity hard constraint threshold. If the preliminary overall predicted throughput sequence exceeds the capacity hard constraint threshold, trigger the capacity constraint projection calculation to adjust the prediction result; otherwise, proceed directly to the next step. S7. After projection of the capacity constraint, the channel prediction results are normalized and weighted and fused based on the set of weight coefficients and the channel prediction contribution output by the ridge regression model to obtain the final airport future throughput trend prediction result that meets the airport capacity constraint conditions.

[0018] This method is based on the hierarchical characteristics of airport throughput in terms of time and spatial structure, decomposing the overall airport throughput into multiple traffic channels with clear physical and operational meanings. First, historical hourly traffic data from different channels such as runway directions, terminals, and cargo areas are cleaned and standardized to eliminate abnormal fluctuations and dimensional differences, providing a stable data foundation for model training.

[0019] During the modeling phase, for each channel, a feature vector containing information from multiple time steps is constructed based on a historical time window, and the Ridge Regression algorithm is used for independent modeling. Ridge Regression effectively suppresses multicollinearity among traffic features by introducing an L2 regularization term into the loss function to constrain the feature coefficients, thus maintaining good stability and generalization ability even with a limited number of samples and strong feature correlations. The predicted values ​​for each channel obtained in this way retain the trend information of traffic changes over time.

[0020] Specifically, at the level of forecast result organization, based on the actual structure of airport operation management, the channel forecast results are mapped into a bottom-level channel vector, an intermediate-level aggregated vector, and a top-level overall vector, forming a multi-level throughput forecast feature set. Subsequently, by calculating the correlation coefficients between channels, the weight of different intermediate levels in the overall throughput, and the influence intensity of each level on the overall throughput, a set of weight coefficients reflecting the channel correlation is constructed.

[0021] Based on this, and considering the peak proportion characteristics of each channel during different time periods and the capacity allocation bias caused by differences in passenger and freight channel types, the independent predicted values ​​of each channel are dynamically weighted and adjusted to characterize the competition and coupling relationships between channels. The weighted intermediate predicted values ​​of each channel are then aggregated hierarchically to form a preliminary overall predicted throughput sequence. Capacity constraint projection calculations are used to ensure that the predicted results do not exceed the airport's established hard constraint capacity range. Finally, by normalizing and weighting the predicted contributions of each channel, a future throughput trend prediction result that meets the airport's operational constraints is obtained.

[0022] In this embodiment, S1 includes acquiring historical airport operation records, aggregating the historical operation records by hourly window to generate hourly traffic sequences for each region and an overall hourly throughput sequence for the airport; detecting outliers in the hourly traffic sequences for each region and the overall hourly throughput sequence for the airport, and interpolating to correct the gaps after removing outliers to obtain corrected traffic sequences without outliers; standardizing the corrected traffic sequences without outliers to generate standardized traffic value sequences; and obtaining a multi-channel traffic time series set and an overall throughput benchmark sequence based on the standardized traffic value sequences.

[0023] In this implementation, fields directly related to throughput statistics are first extracted from the raw records generated by airport operations management. These fields include at least the time of the transaction, the transaction type identifier, the region identifier, the quantity value, the data source identifier, and the unique record identifier. The region identifier corresponds to the airport's pre-defined runway direction, terminal, and cargo area, forming a region mapping table. When establishing the region mapping table, the current airport operations zoning and transaction attribution criteria are compared item by item to bind each raw record to a unique region, preventing the same record from being counted repeatedly in multiple regions. The time base uses the airport's local standard time, with hourly windows divided by the hour. The hourly window boundaries are determined by rounding down to the hour, mapping any transaction time to the start time of its corresponding hourly window, ensuring strict alignment of all regional sequences on the timeline and guaranteeing the additivity of subsequent cross-regional aggregations.

[0024] When aggregating by hourly window, the original records are first deduplicated and validated for consistency. Deduplication uses the unique identifier of the record as the primary key, and the same primary key is retained only once within the same hourly window. Consistency validation includes non-negativity checks on quantity values ​​and validity checks on business types. The non-negativity threshold is set to 0 (threshold determination method: throughput statistics are based on the number of people and weight as the measurement objects, and negative numbers are not allowed in the business meaning; any quantity value less than 0 is marked as abnormal). In aggregation calculation, an hourly counter is established for each region: passenger departures, passenger arrivals, cargo departures, and cargo arrivals are accumulated within their respective hourly windows. The accumulation process is summarized using key-value combinations of the same region, the same hourly window, and the same business type. When there are multiple reports of the same flight or the same batch of goods within the same hourly window in the original records, the cumulative result after deduplication using the unique identifier of the record is used to avoid duplicate counting. The overall hourly throughput series of the airport is generated synchronously on the same time axis. The calculation of the overall series adopts the same aggregation method as the regional series. The overall series is accumulated according to the business type in each hourly window and then summarized into the overall throughput value to ensure that the overall series and the regional series are consistent in statistical method.

[0025] Outlier detection is carried out simultaneously on the regional hourly flow series and the overall hourly throughput series. The detection process uses multiple rules in series and merges the judgment results within the same hourly window.

[0026] In this embodiment, the first type of rule is the physical boundary rule: the lower threshold is 0 (the threshold is determined in the same way as the aforementioned non-negative check), and the upper threshold adopts the constraint value of the airport capacity limit at the hourly granularity (the threshold is determined by extracting the runway hourly take-off and landing capacity, terminal hourly passenger handling capacity and cargo area hourly handling capacity from the capacity benchmark table published by the airport operation support department, converting each capacity value that constitutes a constraint on the same statistical object to a unified throughput caliber, and taking the minimum value as the hourly upper limit threshold to ensure that any flow value exceeding the threshold is identified as an anomaly that does not conform to the capacity boundary).

[0027] In this embodiment, the second type of rule is the mutation rule: establish a background distribution for each sequence with the same hour, that is, group the samples with the same weekday attribute and hour attribute in the historical period into a group, and calculate the mean and standard deviation of the group; when the deviation of the observation value of a certain hour from the mean of the group exceeds the mutation threshold, it is marked as an anomaly. The mutation threshold is 3 times the standard deviation (threshold determination method: run anomaly detection with 2 times, 3 times and 4 times the standard deviation as candidate thresholds in the historical period, use the sequences after anomaly removal and interpolation correction for subsequent prediction model training and validation, and use the candidate threshold with the smallest mean of overall throughput prediction error in the validation period as the final threshold, to obtain 3 times the standard deviation).

[0028] In this embodiment, the third type of rule is the structural consistency rule: calculate the total value of the regional sequence after summarizing according to the statistical caliber for each hourly window, and compare it with the overall hourly throughput value. If the difference between the two exceeds the consistency threshold, the relevant records involved in the hourly window are marked as abnormal (the threshold is determined by calculating the average and standard deviation of the absolute value sequence of the difference between the regional total value and the overall value in the historical period, setting the consistency threshold to the average value plus twice the standard deviation, and requiring the threshold to be no less than 1 unit of measurement to avoid misjudgment caused by integer granularity during low-flow periods).

[0029] The judgment results of the above rules are merged in a way that any rule triggers an anomaly. When the anomaly is small, the corresponding value is removed from the sequence and a missing marker is generated at the original position. The missing marker is used for interval identification and method selection for subsequent interpolation correction.

[0030] Interpolation correction is segmented based on the length of the missing interval. First, the start and end positions of consecutive missing segments in each sequence are identified, and the length of the consecutive missing segments is calculated in hours. For intervals with consecutive missing segments not exceeding 2 hours, linear interpolation using adjacent valid points is employed: using the preceding and following valid hour values ​​of the missing interval as endpoints, equal step sizes are allocated according to the relative positions of the missing points within the interval, ensuring the interpolated sequence is continuous at the endpoints and maintains a monotonic trend. The 2-hour threshold is determined using an error-optimal method (threshold determination method: within the historical period, 1-hour, 2-hour, and 3-hour consecutive missing segments are artificially created in the complete sequence; after linear interpolation correction, these are compared with the original values; the mean absolute error is calculated; the missing length with the smallest mean error is selected as the maximum missing length applicable to linear interpolation, resulting in 2 hours). For intervals with consecutive missing lengths exceeding 2 hours, periodic mean imputation is used: for each missing hour, valid samples with the same week and hour attributes from the preceding 8 weeks are taken, and after removing labeled outliers, the arithmetic mean is calculated as the imputation value; the 8-week backtracking length is a fixed parameter (parameter determination method: during the historical period, 4 weeks, 6 weeks, 8 weeks, and 10 weeks are used as candidate backtracking lengths for imputation verification, and the candidate value that best combines the stationarity index of the imputed sequence during the verification period with the subsequent prediction error is used as the final backtracking length, resulting in 8 weeks). When there are fewer than 12 valid samples in the backtracking interval for a missing hour, a downgrade compensation rule is activated: the backtracking span is expanded to 16 weeks to supplement the sample number; if still insufficient, the mean of the same hour for the entire historical period of the sequence is used; the sample number threshold is 12 (threshold determination method: determined by the requirement of covering at least 6 weeks in hourly granular statistics, with 6 corresponding samples for 6 weeks; considering sample loss after outlier removal, the threshold is set to 12 to ensure the stability of the mean estimation). Through the above interpolation correction, a continuous and anomaly-free regional hourly flow sequence and an overall hourly throughput sequence are obtained.

[0031] In this embodiment, standardization is performed after interpolation correction. The goal of standardization is to unify the numerical scales of different regions and business types to a comparable range, thereby avoiding disproportionate impacts of certain channels on the model due to differences in units during subsequent modeling. Standardization parameters include the baseline time period length and the mean and standard deviation of each sequence. The baseline time period length is set to 90 days (parameter determination method: standardization parameters are calculated using 30 days, 60 days, 90 days, and 180 days as candidate baseline lengths and used for subsequent model training and validation; the candidate length with the smallest mean prediction error during the validation period and the smallest drift in the mean of the standardized sequence is used as the final baseline length, resulting in 90 days). The mean is calculated using the arithmetic mean of all hourly values ​​within the baseline time period; the standard deviation is calculated using the dispersion of all hourly values ​​within the baseline time period relative to the mean. The standardization process for any hourly value is as follows: first, subtract the average value of the corresponding sequence within the base time period from the hourly value; then, divide the difference by the standard deviation of the corresponding sequence within the base time period to achieve scale uniformity. When the standard deviation of a sequence is less than 0.001, the sequence is considered an approximately constant sequence within the base time period, and the standardization denominator is 0.001 (threshold determination method: set based on the minimum effective fluctuation of hourly granular data in double-precision storage to avoid distortion of standardized values ​​due to an excessively small denominator).

[0032] Specifically, after standardization, the standardized hourly traffic sequences for each region are organized into a multi-channel traffic time series set using a unified structure of region identifier, service type identifier, and time index. The time index uses consecutive hourly serial numbers from the start to the end of the historical period, ensuring that all channels have consistent lengths and are aligned point-by-point with the overall throughput benchmark sequence. The overall throughput benchmark sequence uses the overall standardized hourly throughput sequence under the same time index. After the multi-channel set is generated, a final check is performed: the missing channel markers are checked hourly to ensure they are cleared, and the consistency rules between the regional total and the overall sequence before standardization are satisfied hourly. After the final check passes, the multi-channel traffic time series set and the overall throughput benchmark sequence are output.

[0033] Furthermore, S2 includes obtaining the traffic time series for each channel, constructing a multi-factor feature vector based on a historical time window, wherein the multi-factor feature vector includes periodic features, peak features, and trend features; using the ridge regression algorithm to independently model the multi-factor feature vector, and suppressing multicollinearity among features in the multi-factor feature vector through L2 regularization constraints to obtain ridge regression model parameters; generating an initial independent prediction value sequence for each channel within a fixed future time window based on the ridge regression model parameters; and outputting the initial independent prediction value sequence, retaining the trend information of the prediction results for each channel within the fixed future time window.

[0034] In this implementation, for the traffic flow time series of each channel, a unified hourly time index is first established. The starting point of the index is the first hour of the historical sequence of the channel, and the ending point is the last hour of the historical sequence. The interval between adjacent indices is fixed at 1 hour, thus ensuring that the traffic flow value of the channel at any given time corresponds to a unique hour position. The historical time window length is set to 48 hours (parameter determination method: on the same airport's historical data, candidate window lengths of 6 hours, 12 hours, 24 hours, 48 ​​hours, and 72 hours are selected. The last 30 days are used as the validation interval, and the rest are used as the training interval. The channel model is trained for each candidate window length and rolling prediction is performed in the validation interval. The average value and error variance of the hourly absolute error are calculated, and the window length with the smallest average value and the smallest error variance is selected, resulting in 48 hours). The fixed future time window length is set to 24 hours (parameter determination method: the daily rolling compilation cycle of the airport operation plan is used as a constraint, the prediction results are used for the next day's resource allocation and time slot capacity pre-allocation, and 24 hours consistent with this cycle is taken as the unified future time window length so that the prediction sequences output by all channels have the same time span and the same hour granularity).

[0035] When constructing a multi-factor feature vector, a standardized flow sequence segment of the channel within a 48-hour historical window is extracted from each prediction start time as a benchmark. Periodic features, peak features, and trend features are extracted from the calendar attributes of the segment and the prediction start time. The three types of features are spliced ​​together in a fixed order to form the same feature vector.

[0036] Specifically, the calculation of periodic features begins with the time attribute of the prediction start time: obtaining the hour number of that time within the day (0 to 23); obtaining the week number corresponding to that time (1 to 7); obtaining the month number corresponding to that time (1 to 12); obtaining the identifier value indicating whether that time is a statutory holiday. The holiday identifier value is determined by the airport's operation calendar; if the date is marked as a holiday in the calendar, the identifier value is 1, otherwise it is 0. Then, the hour number, week number, and month number are converted into discrete indicator codes: 24 positions are reserved for the hour, with the position matching the hour number assigned a value of 1, and the remaining positions assigned a value of 0; 7 positions are reserved for the week number, with the position matching the week number assigned a value of 1, and the remaining positions assigned a value of 0; 12 positions are reserved for the month, with the position matching the month number assigned a value of 1, and the remaining positions assigned a value of 0; the holiday identifier value is a separate position, directly assigned 0 or 1. The dimension and position mapping of the discrete indicator code are generated by the system initialization and remain unchanged over time, thus ensuring that the same time attribute corresponds to the same position on different channels and avoiding drift in the meaning of the time code.

[0037] Furthermore, the calculation of peak characteristics is based on statistics within a 48-hour window. First, the average value of this window is calculated by summing the values ​​for each of the 48 hours and then dividing by 48. Next, the standard deviation is calculated by squaring the difference between each hourly value and the average, summing the results, dividing by 48 to obtain the variance, and finally taking the square root of the variance to obtain the standard deviation. The peak value threshold is set to the average value plus three times the standard deviation (threshold determination method: peak identification is performed on candidate multiples 2, 3, and 4 respectively; within the training interval, the number of false positives and false negatives for each candidate multiple are counted; a false positive is defined as a point identified as a peak but without corresponding upward changes in the following 1 or 2 hours; a false negative is defined as a point not identified as a peak but whose value exceeds twice the window average and then shows a significant decline in the following 1 hour; the number of false positives and false negatives are weighted and summed, with a weight ratio of 1:1, and the multiple with the smallest weighted sum is selected, resulting in 3). After determining the threshold, the system scans hourly along a 48-hour window. When the value of a given hour exceeds the peak value threshold, that hour is marked as a peak point. Peak statistics are generated based on these peak point markings: the maximum peak value is obtained by taking the largest value among all peak points; the number of peak occurrences is obtained by counting the peak point markings; the number of hours from the most recent peak to the prediction starting point is obtained by locating the nearest peak point to the prediction starting point and calculating the hour index difference between the two; if there are no peak points within the 48-hour window, this distance is set to 48; the peak duration is obtained by statistically analyzing the length of consecutive segments formed by consecutive peak points, and the maximum length of these consecutive segments is taken as the peak duration for that window. All of the above peak statistics originate from the defined scanning, counting, locating, and segment statistics processes.

[0038] The trend characteristics were calculated at four levels: multi-scale mean, incremental statistics, net change, and linear fitting direction. The multi-scale mean used three fixed spans: a short span of 3 hours, a medium span of 12 hours, and a long span of 48 hours (parameters determined as follows: candidate short spans were 2, 3, and 4 hours; candidate medium spans were 6, 12, and 18 hours; and candidate long spans were 24, 48, and 72 hours; the model was trained and rolled over for each span combination within the validation interval, and the average absolute error of the validation interval was calculated, with the smallest value selected to obtain the 3-hour, 12-hour, and 48-hour values). The short span mean was obtained by summing the values ​​from the last 3 hours at the end of the window and dividing by 3; the medium span mean was obtained by summing the values ​​from the last 12 hours at the end of the window and dividing by 12; and the long span mean was obtained by summing the values ​​from the entire 48-hour period and dividing by 48. Incremental statistics are constructed using the difference between adjacent hours: for each of the 2nd to 48th hours within the window, the increment is obtained by subtracting the previous hour's value from the current hour's value, resulting in a total of 47 increments; the mean of the increments is the sum of the 47 increments divided by 47; the increment fluctuation amplitude is calculated using the standard deviation of the increments, with the calculation process consistent with the aforementioned standard deviation, but the sample size is replaced with 47. The net change is obtained by subtracting the first hour's value from the last hour's value within the window, representing the overall upward or downward magnitude over the 48 hours.

[0039] Specifically, the linear fitting directional variable is obtained using a least squares process: First, time numbers 1 to 48 are assigned sequentially to the 48 hours within the window. Then, the mean of the time number and the mean of the flow rate are calculated for each hour. Next, the difference between the time number and its mean, and the difference between the flow rate and its mean are calculated for each hour. These two differences are multiplied and summed to obtain the total coordinated change. Then, the square of the difference between the time number and its mean is calculated for each hour and summed to obtain the total time-discrete change. The total coordinated change is divided by the total time-discrete change to obtain the directional variable. A positive directional variable indicates an overall upward trend, while a negative one indicates an overall downward trend. The larger the absolute value, the steeper the change. Thus, the periodic characteristics, peak characteristics, and trend characteristics are all generated with clear calculation steps, forming a multi-factor feature vector corresponding to the prediction starting point.

[0040] In this embodiment, ridge regression modeling is performed independently for each channel, and training samples are generated hourly by a sliding window. Each available moment within the training interval is selected as a sample moment, requiring at least 48 hours of historical data prior to it and at least 1 hour of actual observations after it for label construction. For each sample moment, a 48-hour window prior to that sample moment is extracted using the aforementioned method, and a feature vector is calculated to form the input; the actual standardized flow value of the first hour after the sample moment is used as the output label. After completing the full interval traversal, the input matrix and output vector are obtained. During model training, the intercept and each feature coefficient are estimated simultaneously. The intercept is used to absorb the baseline offset that still exists after channel standardization, preventing the coefficients from being forced to assume constant terms. The training objective is to minimize the sum of squared prediction errors while incorporating a constraint on the sum of squared coefficients. The constraint strength parameter is set to 10 (parameter determination method: train with candidate strengths of 0.1, 1, 10, 100, and 1000 respectively, and calculate the average absolute error during rolling predictions in the validation interval; simultaneously, perform 5 repeated training sessions within the training interval, divide by time, calculate the standard deviation of each feature coefficient in the 5 training sessions, and take the mean as the coefficient volatility; first select the two candidate strengths with the smallest average absolute error, and then select the one with the smallest coefficient volatility, resulting in 10). This constraint ensures that when there is a correlation between periodic features, peak features, and trend features, the coefficients will not be extremely amplified or cancel each other out due to correlation, thereby improving the stability of the model parameters. After training, the intercept and each feature coefficient, along with the feature position mapping table, are saved together to form the model parameter set for this channel. When saving, the start and end time indices of the training interval and the parameter strength values ​​are included to ensure that subsequent prediction calls use the same parameter version.

[0041] Specifically, the initial independent predicted value sequence is generated hourly within a 24-hour future time window, with the prediction starting point being the current hour. First, an input feature vector for the first hour of the future is constructed: using the observed standardized flow sequence of the most recent 48 hours before the prediction starting point as a window segment, the feature vector is obtained according to the aforementioned periodicity, peak, and trend calculation process. Then, this feature vector is linearly combined with the channel model parameters to obtain the predicted value for the first hour of the future. The calculation process involves multiplying each feature value by its corresponding coefficient and summing the results, then adding the intercept to obtain the prediction result. After obtaining the predicted value for the first hour of the future, this predicted value is used as the value of the channel sequence at the position in the first hour of the future and is used for the next rolling step: when constructing the feature vector for the second hour of the future, the historical window consists of the actual observed values ​​of the 47 hours before the prediction starting point plus the predicted value for the first hour of the future. The feature vector is generated again according to the same feature calculation process, and the predicted value for the second hour of the future is output.

[0042] Furthermore, the above rolling process is repeated hourly until the 24th hour in the future, forming an initial independent predicted value sequence of length 24. Before output, this sequence undergoes three consistency checks: a time index continuity check, which checks whether the hour indices of the 24 predicted times are continuously increasing; a numerical validity check, which checks that each predicted value is finite and does not exhibit undefined states; and a jump check, which compares the absolute value of the difference between adjacent predicted values ​​with a threshold of 6 times the incremental standard deviation within the historical window. If the difference exceeds this threshold, a jump is marked (the threshold is determined by: statistically analyzing the long-term mean of the incremental standard deviation during the training interval, using 4, 6, and 8 times as candidate thresholds to mark jumps in the predicted sequence of the validation interval, and calculating the hit rate of actual observations also showing jumps at the marked jump times, selecting the threshold with the highest hit rate and the fewest false alarms, resulting in 6 times). The checks only produce the marking results and do not perform smoothing or compression processing on the predicted values, thus ensuring that the rises, falls, and peaks / valleys within the future time window maintain their original trend patterns in the output sequence.

[0043] In this embodiment, S3 includes obtaining the initial independent prediction value sequence of each channel, constructing a structural mapping matrix describing the channel affiliation by combining the runway number, terminal area and passenger and cargo identification; reorganizing and accumulating the initial independent prediction value sequence according to the structural mapping matrix to generate the bottom channel vector and the intermediate level aggregate vector; globally summarizing the intermediate level aggregate vector to obtain the top-level overall vector, and concatenating the top-level overall vector, the intermediate level aggregate vector and the bottom channel vector to output a multi-level throughput prediction feature set.

[0044] In this implementation, the generation of the multi-level throughput prediction feature set is based on the same future time window. The length of the future time window is 24 hours (parameter determination method: airport operation plans organize resource allocation and capacity pre-allocation on a daily rolling cycle, and the prediction results serve the hourly planning of the next natural day; the future time window is determined to be 24 hours, consistent with the daily cycle). The time index uses a consecutive 24-hour sequence number starting from the hour of the prediction start point, with the sequence number from 1 to 24. All hierarchical vectors are aligned according to this sequence number to ensure that the bottom layer value, intermediate layer value, and top layer value at any hour position correspond to the same prediction time at the same time position.

[0045] Specifically, the initial independent predicted value sequence for each channel is first obtained. Each channel corresponds to a sequence of length 24 and is bound to a unique channel identifier. The unique channel identifier is formed by concatenating three fields: runway number, terminal area number, and passenger / cargo identifier. The runway number is taken from the airport runway resource coding table, which is generated by the runway direction or runway combination definition in the airport operation management system. Each runway direction or combination in the coding table corresponds to a unique number. The terminal area number is taken from the terminal area division table, which is generated by the airport's planar area management standards to ensure that the same business record belongs to only one terminal area at the same time. The passenger / cargo identifier has two values: 1 for passenger and 2 for cargo (parameter determination method: the throughput statistics divide the business into passenger and cargo categories, which is consistent with the airport statistical report standards. The system configuration table gives the two categories of values ​​in an enumeration manner and is globally unified). After completing the above identifier binding, a channel list is formed. The channel list registers the unique channel identifier and the corresponding 24-hour initial independent predicted value sequence for each channel, which serves as the input source for subsequent structure mapping and vector reassembly.

[0046] It should be noted that the structure mapping matrix is ​​used to describe the membership relationship between channels and hierarchical nodes. Its rows correspond to the bottom-level channels, and its columns correspond to the aggregation nodes. Aggregation nodes are divided into intermediate-level nodes and top-level nodes. Intermediate-level nodes consist of three groups: runway aggregation node group, terminal area aggregation node group, and passenger and cargo aggregation node group; the top-level node is the airport as a whole node. The number of runway aggregation nodes is taken from the total number of runway numbers in the airport runway resource coding table (parameter determination method: based on the number of valid number entries in the runway resource coding table, each change to the coding table generates a new version number and loads a fixed version before the start of the prediction period); the number of terminal area aggregation nodes is taken from the total number of area numbers in the terminal area division table (parameter determination method: based on the number of valid area entries in the area division table, the area entries correspond one-to-one with the airport business affiliation); the number of passenger and cargo aggregation nodes is fixed at 2 (parameter determination method is the same as the aforementioned passenger and cargo identifier value definition). When creating the list of aggregated nodes, first arrange the runway aggregation nodes in ascending order by runway number, then arrange the terminal aggregation nodes in ascending order by terminal area number, then arrange the passenger and cargo aggregation nodes in the order of passenger and cargo identification values, and finally add the overall airport nodes to form a list of aggregated nodes with a fixed column order. This column order remains unchanged within the same forecast period, ensuring that the same node corresponds to the same column position in different batches of calculations.

[0047] In this embodiment, the structural mapping matrix uses 0 and 1 to represent membership relationships. During matrix initialization, all elements are set to 0. Then, the channel list is traversed line by line, performing four positioning and assignment operations for each channel: The first positioning locates the runway aggregation node column corresponding to the channel's runway number, assigning a value of 1 to the element at the intersection of the channel's row and column; the second positioning locates the terminal aggregation node column corresponding to the terminal area number of the channel, assigning a value of 1 to the element at the intersection; the third positioning locates the passenger / cargo aggregation node column corresponding to the passenger / cargo identification of the channel, assigning a value of 1 to the element at the intersection; and the fourth positioning locates the airport overall node column, assigning a value of 1 to the element at the intersection. The positioning process uses a table lookup to column index method: a mapping table from number to column position is established based on the aggregation node list. During channel traversal, the unique column position is directly retrieved from the mapping table using the channel's number, avoiding the same number falling into multiple column positions. After traversal, each row has four elements set to 1, and the remaining elements set to 0.

[0048] Specifically, the generation of the underlying channel vectors is based on the channel list and reorganized. First, the channel list is sorted by unique channel identifier, with the sorting rules being: runway number ascending, terminal area number ascending, and passenger / cargo identifier ascending, ensuring a stable row order in the underlying channel vectors and consistency with the airport structure. Then, a two-dimensional array is created as the underlying channel vector, with the number of rows equal to the total number of channels and a fixed number of 24 columns. The sorted channel list is then filled row by row: the initial 24-hour independent predicted value sequence for each channel is assigned to the corresponding 24 positions in each row according to hourly indices 1 to 24, forming a underlying channel vector with one channel per row and one hour per column. After filling, row-level validation is performed: each row checks that all 24 positions have been assigned values ​​and that the time indices are consecutive. If any missing values ​​are found, the sequence length of the channel list is backtracked to check if it matches the time index binding, until the underlying channel vector satisfies the requirement that each row has a length of 24 and no gaps.

[0049] Furthermore, the intermediate-level aggregation vector is obtained by cumulatively calculating the underlying channel vector node by node and hour by hour using the structure mapping matrix. For each intermediate-level aggregation node, the corresponding column of the node is first selected in the structure mapping matrix, and all channel rows with elements equal to 1 in that column are selected to form the channel set of that node. Then, the hourly accumulation is performed from 1 to 24: at a certain hourly number, the cumulative value is first initialized to 0, and then the value of the underlying channel vector at that hourly number is taken out one by one according to the row number order in the channel set. This value is added to the cumulative value and the cumulative value is updated until the channel set is traversed. The cumulative value is the aggregation prediction value of the aggregation node for that hour. The aggregation prediction values ​​of the aggregation node at the 24-hour number are sequentially arranged to form an aggregation sequence of length 24. The above selection and hourly accumulation process is repeated for all nodes in the runway aggregation node group to form the runway aggregation vector group; the same process is repeated for all nodes in the terminal aggregation node group to form the terminal aggregation vector group; the same process is repeated for the two nodes in the passenger and cargo aggregation node group to form the passenger and cargo aggregation vector group. Arrange the three aggregate vectors in the order of the nodes in the column list to form an intermediate-level aggregate vector with 24 columns and the number of rows equal to the total number of intermediate-level nodes.

[0050] Specifically, the top-level overall vector is obtained by globally summarizing the intermediate-level aggregated vectors, while introducing consistency checks to avoid duplicate accumulation. The top-level overall vector uses passenger and freight aggregation node groups as the sole source of aggregation (parameter determination method: the passenger and freight aggregation node groups are consistent with the overall throughput caliber in business statistics, and the number of nodes is fixed at 2, ensuring a unique aggregation path and avoiding duplicate superposition caused by runway aggregation and terminal aggregation participating simultaneously). At each hourly sequence number, the aggregated predicted value of the passenger aggregation node and the aggregated predicted value of the freight aggregation node for that hour are taken, and the two are added together to obtain the overall predicted value for that hour. The top-level overall vector of length 24 is calculated sequentially for the 24 hourly sequences. After completing the top-level overall vector, a consistency check is performed: For each hourly index, the values ​​of all channels in the bottom-level channel vector at that hourly index are summed row by row to obtain the total value of all bottom-level channels. This total value is then subtracted from the overall predicted value of the top-level overall vector at the same hourly index to obtain the difference. The difference threshold is set to 0.000001 (threshold determination method: the values ​​are represented using double-precision floating-point in calculation and storage, and there is a rounding error in the accumulation process. The threshold is set to 10 to the power of -6 to cover the worst-case rounding accumulation error in the 24-hour accumulation, while ensuring that any difference caused by structural omissions or duplications is significantly greater than this threshold). When the absolute value of the difference is greater than 0.000001, the source of the anomaly is located: first, check whether there are any missing or duplicate row numbers in the channel set filtering results at that hourly index, then check whether there are any missing values ​​in the four assignments of the channel unique identifier and the structure mapping matrix, until the absolute value of the difference for all hourly indexes is no greater than 0.000001. After the consistency check passes, the top-level overall vector is confirmed to be valid.

[0051] In practical implementation, the multi-level throughput prediction feature set is concatenated to form a unified output structure. The concatenation order prioritizes hierarchical semantics: the first row places the top-level overall vector; followed by the intermediate-level aggregated vectors, with the row order strictly following the order of the aggregated node list; finally, the bottom-level channel vectors are placed, with the row order strictly following the sorted order of the channel list. During the concatenation process, each row undergoes the same column width validation, ensuring the number of columns is 24. If a row's column count is not 24, the time index for that row is considered inconsistent, and the generation process of its source sequence is traced back. To ensure that subsequent steps can accurately locate the meaning of each row, a row index table is generated: the top-level overall row is registered with a hierarchical type number of 1 and the node name is "Airport Overall"; the intermediate-level runway aggregated row is registered with a hierarchical type number of 2 and the corresponding runway number; the intermediate-level terminal aggregated row is registered with a hierarchical type number of 3 and the corresponding terminal area number; the intermediate-level passenger and cargo aggregated row is registered with a hierarchical type number of 4 and the passenger and cargo identification values; and the bottom-level channel row is registered with a hierarchical type number of 5 and the three field values ​​of the channel's unique identifier. The row index table and the concatenated set of predicted features are provided synchronously in the same output object, so that any predicted value can be traced directly to its hierarchical affiliation and channel source through the row number.

[0052] In this embodiment, S4 includes obtaining the bottom channel vector sequence in the multi-level throughput prediction feature set, constructing a cross-correlation matrix for the bottom channel vector sequence; calculating the traffic proportion value based on the cross-correlation matrix and the intermediate level aggregated vector to obtain the proportion weight distribution sequence of the intermediate level channels; performing regression analysis on the proportion weight distribution sequence and the top-level overall vector to determine the overall influence intensity value between each intermediate level of the top level; and constructing a feature importance mapping table by integrating the overall influence intensity value, the proportion weight distribution sequence and the cross-correlation matrix, and extracting a set of weight coefficients to characterize the difference in the influence of different channels on the overall throughput prediction result.

[0053] In this implementation, the bottom-level channel vector sequence is first located in the feature set. This sequence consists of a total of 24 predicted sequences for each channel, arranged in a fixed order according to their unique identifiers. The unique identifiers are composed of the runway number, terminal area number, and passenger / cargo identification, ensuring consistency in position for the same channel across all batches of calculations. Subsequently, the intermediate-level aggregated vector and the top-level overall vector are located within the same data object. The intermediate-level aggregated vector is arranged in a fixed order according to the node list, while the top-level overall vector is a 24-length overall predicted sequence. All three types of vectors are strictly aligned on hourly indices 1 to 24.

[0054] The cross-correlation matrix is ​​constructed pairwise for each underlying channel vector sequence, with rows and columns corresponding to the underlying channel indices. For any channel sequence, the 24-hour average is calculated by summing the values ​​at each of the 24-hour positions and dividing by 24. Then, the average is subtracted from the value at each hour position to form a mean-reduced sequence, used to remove the influence of the baseline term on the correlation. For any pair of channel sequences, two mean-reduced sequences are obtained, and then the co-variance is calculated hourly: at hour index 1, the two mean-reduced values ​​are multiplied to obtain product 1; at hour index 2, the two mean-reduced values ​​are multiplied to obtain product 2, and so on up to hour index 24. Finally, the 24 products are summed to obtain the total co-variance. Subsequently, the scaling factor for each sequence is calculated: for each mean-reduced sequence, the mean-reduced values ​​at each of the 24-hour positions are squared and summed to obtain a sum of squares, and then the square root of the sum of squares is taken to obtain the scaling factor. The lower limit threshold for the scale is set to 0.000001 (the threshold is determined by rounding errors in double-precision floating-point operations due to the square of the 24th power and the summation; the lower limit is set to 10 to the power of -6 to mask the problem of excessively small divisors caused by near-constant sequences, while ensuring that the scale of normally fluctuating sequences is much larger than this lower limit). When the scale of any sequence is less than 0.000001, the correlation coefficient of the channel pair is set to 0, and the channel is marked as a low-fluctuation channel to avoid amplification of correlation caused by extremely small scales. When the scale meets the lower limit condition, the total amount of coordinated change is divided by the product of the two scales to obtain the correlation coefficient of the channel pair; the value of the correlation coefficient increases as the two sequences change in the same direction and decreases as the two sequences change in opposite directions. The correlation coefficients are calculated for each channel pair according to the above rules, and then filled into the corresponding row and column positions of the cross-correlation matrix. The matrix symmetry is naturally satisfied by changing the order of the channel pairs to obtain the same correlation coefficient. The diagonal position of the matrix corresponds to the correlation between the same channel and itself. The scale quantity is set to 1 when the lower limit condition is met, and set to 0 when the lower limit condition is triggered, while retaining the low fluctuation mark to ensure that the credibility of the correlation index of the channel can be identified in the subsequent fusion stage.

[0055] The proportion weight distribution sequence is calculated hourly based on the intermediate-level aggregated vector and the top-level overall vector. First, an aggregated prediction sequence of length 24 is extracted node by node according to the intermediate-level node list, and simultaneously, a 24-length overall prediction sequence of the top-level overall vector is extracted. For each hourly index, the overall prediction value for that hour is used as the denominator, and the node aggregated value for that hour is used as the numerator for each intermediate-level node. The numerator is divided by the denominator to obtain the flow proportion value of that node in that hour. The lower limit threshold of the denominator is set to 0.000001 (the threshold is determined using the same numerical caliber as the lower limit of the cross-correlation matrix scale, used to suppress abnormal amplification of ratios caused by extremely small overall values). When the overall prediction value for a certain hour is less than 0.000001, the proportion values ​​of all nodes in that hour are set to 0, and that hour is marked as a low-flow hour to ensure that the proportion sequence does not contain non-physically meaningful extreme values. When the overall prediction value meets the lower limit condition, the proportion calculation is completed for each intermediate-level node to obtain the proportion distribution for that hour. Then, hourly normalization correction is performed: at the same hour index, the proportions of all nodes are added together to obtain the total proportion; the threshold for the difference between the total proportion and 1 is set to 0.000001 (threshold determination method: there is a rounding error in the accumulation of multiple node proportions, and this threshold covers the upper bound of this error and is used to identify significant deviations caused by structural omissions or duplications). When the absolute value of the difference between the total proportion and 1 is greater than 0.000001, the total proportion is used as the normalization denominator, and the proportions of all nodes in that hour are divided by the normalization denominator one by one, so that the normalized total proportion of that hour is 1; when the absolute value of the difference is not greater than 0.000001, the original proportion value remains unchanged. The above calculation and correction are repeated for hour indices 1 to 24 to obtain the proportion weight distribution sequence of the intermediate level channel. The proportion weight distribution sequence corresponds to a set of node proportion values ​​for each hour, and after normalization correction, it satisfies the constraint that the sum is 1.

[0056] The overall impact strength value was determined through regression analysis of the top-level overall vector and the proportion weight distribution sequence. The regression samples used hourly indices as sample points, totaling 24 sample points. Sample selection was based on low-flow hours: low-flow hours were first removed, and the remaining hours were retained as valid samples; the lower limit threshold for the number of valid samples was set at 12 (threshold determination method: hourly fitting needs to cover at least half a day's span to ensure representativeness across peak and valley periods, the lower limit is half of 24, determined as 12). When the number of valid samples was less than 12, the entire 24 hours were used for fitting, and low-flow hours were assigned a smaller weight, with the smaller weight set at 0.1 (parameter determination method: in historical rolling forecasts, candidate weights of 0.05, 0.1, and 0.2 were fitted respectively, and the average value and residual fluctuation of the overall fitting residuals were compared. The one with the smallest average value and the smallest fluctuation was selected as 0.1). The regression fitting uses the proportion of intermediate-level nodes as input and the overall top-level prediction value as the fitting target. It solves for the influence intensity value corresponding to each intermediate-level node, ensuring that the sequence formed by weighting the proportions of each node according to its influence intensity closely matches the change pattern of the overall top-level vector within a 24-hour period. The solution process employs a standard linear regression numerical solution procedure: first, an input table is constructed, where each row corresponds to an hourly sample point, and each column corresponds to an intermediate-level node, with the element of that row and column representing the proportion of that node for that hour; then, a target column is constructed, where each row represents the overall top-level prediction value for that hour; subsequently, the influence intensity value is solved based on the minimum residual criterion, and a quadratic regularization constraint is introduced with a strength parameter of 10 (the parameter is determined by running rolling fitting with candidate parameters 0.1, 1, 10, and 100 respectively, calculating the average change of the influence intensity value of each node in adjacent prediction batches, and comparing the average of the fitting residuals; the parameter with the smallest change and no increase in the average residual is selected as 10). This suppresses the intensity value fluctuations caused by the correlation between node proportion inputs, ensuring the continuity of the intensity value during rolling updates over time. After fitting, a set of influence intensity values ​​are obtained. Each intermediate level node corresponds to a specific value, which is used to characterize the sensitivity of the change in the proportion of that node to the overall change.

[0057] The feature importance mapping table is generated by fusing the cross-correlation matrix, the weight distribution sequence, and the overall influence intensity value, and a set of weight coefficients is extracted accordingly. For each bottom-level channel, its structural affiliation is first determined: the runway number, terminal area number, and passenger / cargo identification are parsed from the channel's unique identifier, and the runway node, terminal node, and passenger / cargo node matching the above numbers are located in the intermediate-level node list; then, the influence intensity values ​​of these three nodes are extracted from the overall influence intensity value, and the average of the three values ​​is calculated to obtain the structural strength index of the channel. The averaging process involves adding the three values ​​one by one and then dividing by 3, so that the structural strength index is directly related to the channel's structural affiliation. Then, the scale contribution index is calculated: The passenger and cargo node proportion sequence and the terminal node proportion sequence corresponding to the channel are extracted from the proportion weight distribution sequence. First, the 24-hour average proportion is calculated by adding each of the 24 items in the proportion sequence and then dividing by 24. Next, the peak hour index is determined. The peak hour index is the hour position with the largest value in the top-level overall vector within 24 hours (parameter determination method: the overall peak period represents the period of strongest resource competition; the position of the overall maximum value is used as the unique peak hour index, making the peak definition deterministic). The node proportion at this peak hour index is taken as the peak proportion. The average proportion and the peak proportion are averaged at a 1:1 ratio to obtain the scale contribution index, ensuring that the scale contribution reflects both the contribution during the entire time period and the contribution during the peak period. The correlation strength index is then calculated: the row corresponding to the channel is extracted from the cross-correlation matrix, and a set of similar channels is filtered. The set of similar channels is defined as all bottom-level channels with the same terminal area number and the same passenger and cargo identification. The correlation coefficient of each of the filtered similar channels with this channel is read one by one. The absolute value of the correlation coefficient is taken first, and then the absolute values ​​are added one by one and divided by the number of similar channels to obtain the average absolute correlation, which is used as the correlation strength index. When this channel is marked as a low-fluctuation channel, the correlation strength index is directly set to 0 to avoid the correlation interpretation bias caused by low fluctuation.

[0058] In practice, to integrate the three types of indicators under the same dimension, full-channel normalization is performed on the structural strength indicator, scale contribution indicator, and correlation strength indicator. For any indicator, the minimum and maximum values ​​are first scanned across the entire channel range, and then the range is calculated as the maximum value minus the minimum value; the lower limit threshold of the range is set to 0.000001 (the threshold is determined by the fact that the range across the entire channel range has rounding errors under floating-point operations, and the lower limit is set to 10 to the power of negative 6 to avoid the normalization denominator being too small due to the range being close to 0).

[0059] Specifically, when the range is less than 0.000001, the normalized result of the index across all channels is uniformly set to 0.5 to ensure that the index makes a neutral contribution in the fusion; when the range meets the lower limit condition, the normalized index in the range of 0 to 1 is obtained by subtracting the minimum value from the channel index and then dividing by the range for each channel. After completing the three types of normalization, the channel importance score is calculated, and the fusion weights are set to 0.5, 0.3, and 0.2 (the parameter determination method is as follows: in the historical rolling forecast, the candidate weight combinations 0.6, 0.2, 0.2 with 0.5, 0.3, 0.2 with 0.4, 0.4, 0.2 are respectively formed into weight coefficient sets and entered into the subsequent weighting adjustment process. The average value of the final overall prediction error is compared, and the combination with the smallest average value is selected as 0.5, 0.3, and 0.2). Among them, 0.5 corresponds to the structural strength normalization index, 0.3 corresponds to the scale contribution normalization index, and 0.2 corresponds to the correlation strength normalization index. The channel importance score is obtained by multiplying the three indices by their corresponding fusion weights and then adding them one by one. Then, a set of weight coefficients is extracted from the channel importance scores: First, the importance scores of all channels are added together to obtain the total score. The lower limit threshold of the total score is set to 0.000001 (the threshold is determined by ensuring that the upper bound of the cumulative rounding error is still less than the threshold when the number of channels is large; this threshold is used to identify abnormal situations where the total score of all channels is close to 0). When the total score is less than 0.000001, the average score of all channel weight coefficients is taken, and the average score is 1 divided by the total number of channels to ensure that the total weight coefficient is 1. When the total score meets the lower limit condition, the importance score of each channel is divided by the total score to obtain the weight coefficient of that channel. The resulting set of weight coefficients satisfies the condition that the total score is 1, and directly represents the difference in the impact of different channels on the overall prediction in the subsequent weighted adjustment of channel prediction values.

[0060] S5 includes acquiring historical traffic load data and independent predicted values ​​for each physical channel, generating a traffic peak period percentage sequence; constructing a comprehensive competition interaction matrix based on the traffic peak period percentage sequence, the capacity allocation bias value determined by the channel type difference, and the channel dynamic influence coefficient; applying the comprehensive competition interaction matrix to the independent predicted values ​​to extract the channel competition correction set; and combining the channel competition correction set with the independent predicted values ​​for weighted adjustment to obtain a channel intermediate prediction value sequence considering the competition relationship between channels.

[0061] In this implementation, two types of input data are first collected and time-aligned: the first is the historical traffic load data of each physical channel, including the actual traffic value counted hourly and the corresponding capacity baseline value for that hour; the second is the independent predicted value sequence of each physical channel within a future time window, with the future time window length set at 24 hours (parameter determination method: the prediction results serve the hourly operation organization and resource allocation for the next natural day, the operation plan is rolled over daily, and the time span is consistent with one natural day, determined to be 24 hours). The alignment process uses hourly indexes as the main thread, mapping both historical load data and future prediction sequences to a unified hourly numbering system, ensuring that the historical load characteristics, capacity baseline, and future prediction values ​​of the same channel at the same hour position have the same temporal semantics.

[0062] It should be noted that the peak traffic period percentage sequence is generated channel by channel. For a given channel, the load ratio is first calculated hourly within the historical period. The calculation order for the load ratio is as follows: take the actual traffic value of that hour as the numerator, take the capacity baseline value of that hour as the denominator, and perform division to obtain the load ratio for that hour. The peak judgment threshold is set to 0.85 (threshold determination method: during the historical period, peak hours are marked using 0.75, 0.80, 0.85, and 0.90 respectively. The marking results are compared hourly with the airport congestion period records, and the threshold with the highest hit rate and the fewest false alarms is selected, resulting in 0.85). When the load ratio is not less than 0.85, the hour is marked as a peak hour. Subsequently, a peak frequency table of weekday and hour number combinations is constructed: for each combination, the total number of times the combination appears in the historical period is counted, and then the number of times it is marked as a peak hour is counted. The peak frequency of the combination is obtained by dividing the peak frequency by the total number of times. The peak frequency falls within the range of 0 to 1. After completing the frequency table, a peak period percentage sequence for the next 24 hours is generated: For each future hour, the weekday number and hour number are obtained, and the corresponding peak frequency is retrieved from the peak frequency table as the peak period percentage value for that hour. To ensure the percentage sequence is consistent with the future prediction intensity, prediction intensity correction is also required: First, the historical mean is calculated for the same hour of the same week within the historical period. The historical mean is obtained by summing up the samples of the same type and then dividing by the number of samples. Then, the independent prediction value for the future hour is compared with the historical mean. When the independent prediction value is higher than the historical mean, the excess ratio is calculated. The excess ratio is calculated by subtracting the historical mean from the independent prediction value and then dividing by the historical mean. Subsequently, the peak period percentage for that hour is adjusted upwards towards 1, with the adjustment amount being the excess ratio, and the maximum percentage after adjustment is 1, ensuring that the percentage value has a clear boundary. When the independent prediction value is not higher than the historical mean, the percentage value remains unchanged. This yields a peak period percentage sequence with a channel length of 24.

[0063] It should be noted that the capacity allocation bias value is used to reflect the capacity tilt caused by differences in channel type. The bias calculation is completed in two steps: generating the original bias value and normalizing within the same domain. The original bias value is obtained by multiplying the capacity baseline by the type priority coefficient: the capacity baseline is the average value of the corresponding capacity baseline value of the channel over the 24 hours of the day. The average value is calculated by adding the 24-hour capacity baseline values ​​one by one and then dividing by 24; the type priority coefficient is 1.20 for passenger channels, 0.90 for cargo channels, and 1.00 for runway direction channels (parameter determination method: a numerical priority is formed based on the airport operation support strategy, with passenger resources taking priority over cargo resources with a coefficient greater than 1, and runway direction as the basic resource with a coefficient of 1.00; this set of coefficients is determined to be the above values ​​through the stability test of historical operation allocation records and maintains a consistent caliber under version management). After the initial bias values ​​are generated, they are normalized within the same domain: First, the competition domain is determined. The competition domain is defined as a set of channels that share the same physical resource unit. The physical resource unit is a set with the same runway number, terminal area number, or cargo area number. Then, the initial bias values ​​of all channels within the competition domain are summed. Finally, the initial bias value of each channel is divided by the sum of the values ​​in the domain to obtain the capacity allocation bias value of that channel. This ensures that the sum of the bias values ​​within the competition domain is 1, thereby implementing the capacity tilt constraint into the specific competition relationship.

[0064] In this embodiment, the channel dynamic impact coefficient is used to reflect the channel's sensitivity to changes in overall throughput and its structural position. The dynamic impact coefficient is generated based on the channel weight coefficient, which is taken from the set of preceding weight coefficients. To reflect the spillover effect of strongly correlated channels, a correlation correction is introduced: First, the correlation coefficients between the channel and other channels in the same domain are taken from the cross-correlation matrix. The absolute values ​​of each item are then averaged to obtain the average absolute correlation. The trigger threshold is set at 0.60 (the threshold is determined by statistically analyzing the mean and standard deviation of the average absolute correlation of all channels over the historical period, taking the mean plus one standard deviation as the strong correlation threshold, with a final value of 0.60). When the average absolute correlation is higher than 0.60, the base weight coefficient is adjusted upwards. The adjustment ratio is the portion of the average absolute correlation exceeding 0.60, and the maximum adjustment result is limited to within twice the base weight coefficient (the parameter is determined by comparing the impact of different adjustment upper limits on the stability of subsequent competitive corrections in the historical rolling forecast, selecting the upper limit that minimizes the overall prediction error and prevents large-scale compression of the channel sequence, resulting in twice the value). When the average absolute correlation is not higher than 0.60, the base weight coefficient remains unchanged. This results in the dynamic influence coefficients of each channel, which remain unchanged within the same day. The update cycle is 1 day (parameter determination method: competitive correction serves the daily rolling plan, and daily updates can suppress hourly noise and maintain strategy consistency).

[0065] It should be noted that the comprehensive competitive interaction matrix is ​​constructed hourly over the next 24 hours. The matrix is ​​organized according to the directed relationships between channels, and the matrix elements represent the competitive pressure intensity of the source channel on the target channel in that hour. For any pair of channels, resource sharing is first determined: if the two channels have the same runway number, terminal area number, or cargo area number, the resource sharing flag is set to 1; otherwise, it is set to 0. The resource sharing determination is obtained directly by retrieving the channel's unique identifier from the resource mapping table. Then, the peak overlap value is calculated: the peak overlap value is obtained by multiplying the peak period percentages of the two channels in that hour. The larger the peak overlap value, the more likely competition is to occur in that hour. Next, the bias difference value is calculated: the absolute value of the difference between the capacity allocation bias values ​​of the two channels is taken as the bias difference value. The larger the bias difference value, the more obvious the resource tilt. Finally, the dynamic influence coefficient of the target channel is taken as the influence amplification value, so that the target channel, which is more sensitive to the overall situation, exhibits more significant fluctuations under competition. The final competition intensity is synthesized in the following order: Base competition value = Resource sharing marker × Peak summation value; Tilt correction value = Base competition value × (1 + Bias difference value); Competition intensity = Tilt correction value × (1 + Impact amplification value). To prevent excessive competition intensity during extreme peak periods, the upper limit threshold for competition intensity is set to 1.50 (threshold determination method: statistically analyze the intraday maximum value distribution of competition intensity in historical rolling forecasts, select the upper limit that minimizes the overall forecast error and prevents large-scale zeroing of the corrected channel forecast, resulting in 1.50). When the competition intensity exceeds 1.50, it is truncated to 1.50. The above calculation is repeated for each hour of the next 24 hours to form a comprehensive competition interaction matrix set sliced ​​by hour.

[0066] Specifically, the channel competition correction set is obtained by applying the comprehensive competition interaction matrix to independent predicted values. For a specific target channel in a future hour, the competition pressure value is first calculated: the competition intensity of all source channels pointing to the target channel in that hour is extracted item by item, and weighted and accumulated using the independent predicted value of the corresponding source channel in that hour as the weight. The accumulation process starts from 0 and adds item by item until the entire set of source channels is traversed, thus obtaining the competition pressure value of the target channel in that hour. Then, the competition pressure value is converted into a competition correction amount: first, the bias suppression factor is calculated, which is 1 minus the target channel capacity allocation bias value, so that channels with lower biases can withstand higher correction magnitudes; then, the competition pressure value is multiplied by the bias suppression factor to obtain the bias correction amount; finally, the bias correction amount is multiplied by the target channel dynamic influence coefficient to obtain the final competition correction amount, so that channels that are more sensitive to the overall situation show a more obvious decline under competition. To avoid numerical distortion caused by corrections exceeding independent forecast values, the upper limit of the correction amount is set to 0.80 (parameter determined by comparing the impact of 0.60, 0.70, 0.80, and 0.90 on the channel sequence shape and overall error in historical rolling forecasts, and selecting a proportion that reflects competitive suppression without triggering large-scale zeroing, resulting in 0.80). When the competitive correction amount exceeds 0.80 times the independent forecast value for that hour, the competitive correction amount is limited to 0.80 times the independent forecast value for that hour. The competitive correction amount is calculated hourly for the target channel over the next 24 hours, forming a competitive correction amount sequence of length 24 for that channel; this calculation is repeated for all channels to obtain a set of channel competitive correction amounts.

[0067] In this implementation, the intermediate prediction value sequence for a channel is obtained by weighting and adjusting the competitive correction amount and the independent prediction value. For a future hour and a specific channel, the initial intermediate prediction value is obtained by subtracting the competitive correction amount for that hour from the independent prediction value. When the initial intermediate prediction value is less than 0, the intermediate prediction value for that hour is set to 0 to maintain consistency in physical meaning. Subsequently, the variation range constraint between adjacent hours is implemented to prevent non-business-related linear jumps caused by the competitive correction: First, the hourly increment sequence of the channel is constructed within the historical period. The hourly increment sequence is calculated hourly by the difference in actual flow between adjacent hours. Then, the standard deviation of the hourly increment is calculated. The standard deviation is calculated by first calculating the mean of the hourly increment, then squaring and summing the difference between each hourly increment and the mean, dividing the sum by the sample size to obtain the variance, and taking the square root of the variance to obtain the standard deviation. The variation range threshold is set to 6 times the standard deviation of the hourly increment (the threshold is determined by statistically analyzing the long-term mean and fluctuation range of the standard deviation of the hourly increment within the historical period, selecting an upper bound multiple that can cover sudden fluctuations while suppressing abnormal linear jumps, and determining it to be 6 times under the condition of minimizing rolling prediction error). For the intermediate forecast values ​​of two consecutive hours in the future, calculate the absolute value of the difference between them. If this absolute value exceeds the aforementioned threshold, adjust the intermediate forecast value of the following hour to the threshold boundary, keeping the adjustment direction consistent with the original difference direction to ensure the sequence is continuous and does not disrupt the peak-valley trend. Apply this constraint hourly in sequence for the next 24 hours to obtain the final channel intermediate forecast value sequence.

[0068] Furthermore, S6 includes obtaining the intermediate predicted value sequence of the channel, merging and summing the intermediate predicted value sequence of the channel according to the preset channel hierarchical structure mapping relationship to obtain a preliminary overall predicted throughput sequence; comparing the preliminary overall predicted throughput sequence with the airport capacity hard constraint threshold, and generating a throughput overflow difference sequence if overflow exists; constructing a multi-dimensional capacity constraint projection vector based on the throughput overflow difference sequence, and inversely decomposing the multi-dimensional capacity constraint projection vector to obtain a channel projection correction set; applying the channel projection correction set to adjust the intermediate predicted value sequence of the channel, and outputting the final overall predicted throughput sequence that meets the airport capacity hard constraint threshold.

[0069] In this implementation, the intermediate predicted value sequence for each channel is first obtained and hierarchical merge summation is performed. The intermediate predicted value sequence is organized by channel unique identifier, which consists of runway number, terminal area number, and passenger / cargo identification. The total number of channels is determined by the airport channel configuration table and remains unchanged within the same prediction period. The hierarchical structure mapping relationship is expressed using a pre-set mapping table, which contains at least three items: channel unique identifier, the identifier of the intermediate level node to which the channel belongs, and the identifier of the channel participating in the overall aggregation. The intermediate level node identifier covers runway nodes, terminal nodes, and passenger / cargo nodes. The node list is derived from the airport infrastructure configuration and includes a version number. A fixed version is loaded at the beginning of the prediction period to avoid duplicate or missed entries caused by structural changes within the same period. Merge and summation is performed hourly by hourly index: When the hourly index is a fixed value, the overall accumulator is first set to 0. Then, all channels are traversed one by one, and the intermediate predicted value of the channel at the hourly index is read. If the identifier of the channel participating in the overall aggregation is 1, the value is added to the overall accumulator and the overall accumulator is updated. The overall accumulator after the traversal is the preliminary overall predicted throughput value for that hour. Intermediate-level aggregation values ​​are generated synchronously: A node accumulator is created for each intermediate-level node and set to 0. When traversing the channels one by one, the mapping table is used to determine whether the channel belongs to the node. If it does, the intermediate predicted value of the channel for that hour is added to the node accumulator. The node accumulator after the traversal is the node aggregation value for that hour. The above overall accumulation and node accumulation are repeated for hourly indices 1 to 24 to obtain the preliminary overall predicted throughput sequence and the intermediate-level node aggregation sequence set.

[0070] It should be noted that after the merge and summation are completed, hourly consistency checks are performed to avoid distortion of the overall values ​​due to structural duplication or omissions. The check uses a single-caliber intermediate-level node set for comparison, with the caliber being the passenger and freight node set. This is because the passenger and freight caliber is consistent with the overall throughput caliber, and the number of nodes is fixed at 2, ensuring a unique aggregation path. For each hourly index, the aggregated values ​​of passenger and freight nodes are added to obtain the node total value. Then, the preliminary overall predicted throughput value is subtracted from the node total value to obtain the difference. The difference judgment threshold is set to 0.000001 (threshold determination method: channel values ​​are accumulated multiple times in double-precision floating-point calculations, resulting in rounding errors. The threshold is set to 10 to the power of -6 to cover the upper bound of the error of multi-channel accumulation within 24 hours, while also being able to distinguish significant deviations caused by structural omissions or duplications). When the absolute value of the difference is greater than 0.000001, check in a fixed order: first check if the channel participation overall summary identifier is misplaced, then check if the passenger and freight node affiliation is missing or duplicated, then check if the channel unique identifier has duplicate records in the mapping table. After the check is completed, re-execute the merge summation and verification for that hour until all hour indices 1 to 24 satisfy the condition that the absolute value of the difference is not greater than 0.000001.

[0071] Subsequently, a capacity hard constraint comparison is performed, generating a throughput overflow difference sequence. Airport capacity hard constraint thresholds are given hourly, and the sequence is derived from the capacity configuration table. This table, created by the operations support department, establishes a unified overall hourly upper limit based on the hourly processing caps for runways, terminals, and cargo areas. The overall hourly upper limit is conservatively calculated: the minimum value of each capacity upper limit constraining the overall throughput within the same hour is converted to the unified throughput standard and used as the hourly hard constraint threshold, ensuring that the hard constraint threshold does not exceed any individual resource constraint. The capacity configuration table is version-managed, using the same version within the forecast period. The comparison calculation is performed hourly by hourly index: at a given hourly index, if the preliminary overall predicted throughput value is greater than the hourly capacity hard constraint threshold, the throughput overflow difference is the difference between the two; otherwise, it is not greater than the hourly capacity hard constraint threshold, and the throughput overflow difference is 0. This judgment and difference calculation are repeated for hourly indices 1 to 24, forming a throughput overflow difference sequence. Each item in the sequence is a non-negative number, representing the total amount that needs to be reduced for that hour.

[0072] In this implementation, when the throughput overflow difference is greater than 0 at a certain hour index, a multi-dimensional capacity constraint projection vector for that hour is constructed, and a set of channel projection corrections is generated accordingly. The multi-dimensional dimension is composed of the intermediate-level node dimension and the overall dimension. The overall dimension is used to carry the total reduction for that hour, and the intermediate-level dimension is used to carry the structural amortization reduction. The projection target value of the overall dimension for that hour is directly taken as the throughput overflow difference for that hour. The projection target value of the intermediate-level dimension is generated by jointly amortizing the structural proportion and the structural influence intensity. The amortization process is strictly executed hourly and has normalization constraints: First, the node proportion is calculated by dividing the node aggregation value for that hour by the preliminary overall predicted throughput value for that hour; when the preliminary overall predicted throughput value for that hour is less than 0.000001, the proportion of all nodes for that hour is set to 0 and the amortization calculation for that hour is terminated (threshold determination method: the same error caliber as the consistency verification threshold is used to avoid the proportion being abnormally amplified due to an excessively small denominator). Second, the node influence intensity is obtained. The node influence intensity is taken as the overall influence intensity value formed in the previous steps, and the node influence intensity remains fixed within the same prediction day. Then, the node allocation weight is calculated: the node's proportion is multiplied by its influence intensity to obtain the node's allocation weight. The allocation weights of all nodes are summed to obtain the total weight, with a lower limit threshold of 0.000001 (threshold determined to cover floating-point operation errors and avoid instability caused by the total weight approaching 0). When the total weight is less than 0.000001, the hourly overflow difference is evenly distributed to each node according to the number of nodes; when the total weight is not less than 0.000001, the allocation weight of each node is divided by the total weight to obtain a normalized allocation coefficient, making the sum of the normalized allocation coefficients equal to 1. Finally, the hourly overflow difference is multiplied by the normalized allocation coefficient of each node to obtain the projected target value for each node dimension. At this point, the hourly multidimensional capacity constraint projection vector is determined, with the overall dimension being the overflow difference and the intermediate-level dimensions being the node projected target values.

[0073] It should be noted that the inverse decomposition of the multidimensional capacity constraint projection vector is used to allocate node reduction amounts to specific channels, forming a channel projection correction set. The inverse decomposition is performed hourly by hourly index, and then node-by-node within each hour. For a given hourly index, the channel projection correction amount for that hour is first initialized to 0 for all channels. Then, each intermediate-level node is processed: the node's projection target value for that hour is read as the node's reduction amount; the channel set to which the node belongs is retrieved through a mapping table, with the channel set retrieval using the channel's unique identifier as the key to ensure no duplicate channels within the set. A reduction allocation coefficient is calculated for each channel within this set, determined jointly by the channel's dynamic influence coefficient and capacity allocation bias: first, the channel's dynamic influence coefficient is taken, derived from the preceding weight coefficient and corrected for correlation, remaining unchanged within the same prediction day; then, the channel's capacity allocation bias value is taken, derived from the channel type priority and capacity benchmark normalized; the reduction allocation coefficient is calculated by multiplying the channel's dynamic influence coefficient by 1 and subtracting the capacity allocation bias value, ensuring that channels with lower capacity biases and greater sensitivity to the overall situation bear a higher reduction share. The total allocation coefficients are summed item by item within the node's channel set, with a lower limit of 0.000001 (the threshold is determined in the same way as before). When the total allocation coefficients are less than 0.000001, the node reduction amount is evenly distributed to each channel according to the number of channels for that node. When the total allocation coefficients are not less than 0.000001, the reduction allocation coefficient of each channel is divided by the total allocation coefficients to obtain a normalized allocation coefficient, making the sum of the normalized allocation coefficients equal to 1. Then, the node reduction amount is multiplied by the normalized allocation coefficient of each channel to obtain the intra-node reduction amount for that channel. The intra-node reduction amount is accumulated into the channel projection correction amount for that channel in that hour, until the channel set for that node is traversed. The above allocation and accumulation are repeated for all nodes to obtain the initial channel projection correction amount for each channel in that hour.

[0074] In this implementation, after the initial channel projection correction is generated, channel reduction space constraints are applied to prevent a channel from being allocated a reduction amount exceeding its own value. For a channel at a certain hour index, the reduction space is the difference between the channel's median predicted value for that hour and 0, which is numerically equal to the channel's median predicted value for that hour. If the initial channel projection correction is greater than the reduction space, the channel projection correction is truncated to the reduction space, and the excess is included in the unallocated reduction. The unallocated reduction is redistributed according to the remaining channel reduction space for the same hour. The redistribution strictly follows the remaining reduction space ratio rule: the remaining reduction space is calculated for each channel in the set that has not been truncated, and the remaining reduction space is the channel's median predicted value for that hour minus the currently allocated channel projection correction; the remaining reduction space is summed for all remaining reduction spaces to obtain the total remaining space; the unallocated reduction is allocated according to the proportion of each channel's remaining reduction space to the total remaining space and accumulated to the corresponding channel projection correction. The maximum number of iterations for reallocation is set to 3 (parameter determination method: within the historical channel scale range, hours that still show truncation after one reallocation are mainly concentrated in peak periods; 3 iterations can absorb the vast majority of unallocated reductions within acceptable computational latency and prevent abnormal oscillations in the channel sequence). If unallocated reductions still exist after 3 iterations, it indicates that the reduction space of all channels in that hour is insufficient to absorb all overflow differences. In this case, the achievable reduction for that hour is taken as the sum of the intermediate predicted values ​​of all channels minus the sum of the minimum values ​​after the capacity constraints of all channels, which is numerically equal to the sum of the intermediate predicted values ​​of the channels in that hour. Then, the final overall predicted throughput value for that hour is directly set as the capacity hard constraint threshold to ensure that the overall output does not exceed the limit. At the same time, the channel predicted values ​​are reduced proportionally according to the proportion of the intermediate predicted values ​​of the channels. After reduction, a constraint of not less than 0 is applied to each channel to ensure that the physical meaning of the channel layer values ​​is consistent.

[0075] It should be noted that after obtaining the set of channel projection corrections, this set is used to adjust the intermediate prediction value sequence of the channels and output the final overall predicted throughput sequence. The application process is performed hourly, channel by channel: at a certain hour index, for a certain channel, the intermediate prediction value of that hour is subtracted from the channel projection correction value of that hour to obtain the channel prediction value after capacity constraints. If the result is less than 0, the channel prediction value for that hour is set to 0. After the subtraction is completed for all channels for that hour, the prediction values ​​of all channels for that hour are summed item by item according to the hierarchical mapping relationship to obtain the final overall predicted throughput value for that hour. The above subtraction and summation are repeated for hour indices 1 to 24 to form the final overall predicted throughput sequence. Before output, an hourly hard constraint verification is performed: For each hour, the final overall predicted throughput value is compared with the hourly capacity hard constraint threshold. If the final overall predicted throughput value is greater than the capacity hard constraint threshold, the overall value for that hour is reduced to the capacity hard constraint threshold, and all channel prediction values ​​for that hour are proportionally reduced by the same reduction ratio. The reduction ratio is the sum of the capacity hard constraint threshold divided by the hourly channel prediction values. After reduction, a not-less-than-zero constraint is applied again for each channel. The verification difference threshold is set to 0.000001 (threshold determination method: absorbing rounding errors caused by proportional reduction and re-summing). The verification is considered successful when the difference between the final overall predicted throughput value and the capacity hard constraint threshold is not greater than 0.000001. At this point, the final overall predicted throughput sequence that meets the airport capacity hard constraint threshold is output, while the predicted sequence after channel-level capacity constraints is simultaneously retained.

[0076] S7 includes obtaining the corrected channel prediction value sequence, which is obtained by superimposing the channel projection correction set onto the original prediction value; decomposing the corrected channel prediction value sequence according to the regression coefficient matrix of the ridge regression model to obtain the channel prediction contribution value set; performing operations on the preset weight coefficient set and the channel prediction contribution value set to generate a normalized weighted factor vector; and using the normalized weighted factor vector to perform weighted fusion on the corrected channel prediction value sequence to output the final airport future throughput trend prediction result that meets the airport capacity constraint conditions.

[0077] In this implementation, a sequence of corrected channel prediction values ​​is first generated. For each channel and each hourly index, the intermediate channel prediction value for that hour is read, along with the corresponding channel projection correction amount. The channel projection correction amount is expressed using a signed metric, with negative values ​​corresponding to capacity reduction and positive values ​​corresponding to capacity replenishment. The signed metric is used to satisfy the calculation structure of superimposing the correction amount onto the original prediction value. The superposition calculation is performed hourly: the channel projection correction amount is added to the intermediate channel prediction value to obtain the corrected channel prediction value for that hour. After superposition, a non-negativity constraint is applied: when the corrected channel prediction value is less than 0, the corrected channel prediction value for that hour is set to 0, ensuring that the throughput value is consistent with the measurement metric. After superposition and non-negativity constraint are applied to all channels and hourly indices, a set of corrected channel prediction value sequences is formed. The set structure is consistent with the intermediate channel prediction value sequence, both being hourly sequences multiplied by the number of channels by 24.

[0078] Furthermore, a set of channel prediction contribution values ​​is generated based on the regression coefficient matrix of the ridge regression model. The regression coefficient matrix is ​​formed by arranging the regression coefficients of each channel's ridge regression model in channel order, consistent with the order of the channel's unique identifiers, and the order of feature terms is consistent with the feature definition order in step 2, ensuring that coefficients at the same position correspond to the same semantic feature term. To transform the coefficient matrix into the contribution intensity of the channel in the fusion stage, the channel coefficient intensity is first calculated for each channel: all regression coefficients of the channel are read item by item, and the absolute value of each coefficient is taken and accumulated item by item. The accumulated result is used as the channel coefficient intensity of the channel. The absolute value is used to eliminate the offsetting effect of the difference in positive and negative signs on the intensity characterization, so that the intensity only reflects the sensitivity. After obtaining the channel coefficient intensity, a channel prediction contribution sequence is generated: for each channel and each hour index, the corrected channel prediction value for that hour is multiplied by the channel coefficient intensity of that channel, and the product is used as the contribution value of that channel for that hour. The multiplication is used to express the combined effect of the prediction scale of the channel for that hour and the sensitivity of the channel model on the strength of the contribution. The contribution sequence of a channel is calculated hourly for each of the 24-hour indexes. This calculation is repeated for all channels to obtain a set of predicted contribution values. The lower limit threshold for the sum of contributions is set to 0.000001 (the threshold is determined by using the contribution value as the normalized denominator in subsequent division operations; the threshold is used to shield the case where the denominator is close to 0 due to floating-point rounding errors; this threshold is set to 10 to the power of -6, making it significantly smaller than the normal fluctuation range of percentage-type indicators, while being sufficient to cover the rounding accumulation error caused by multiplication and accumulation of multiple channels over 24 hours). When the sum of the contribution values ​​of all channels at a certain hour index is less than 0.000001, the normalized denominator of the contribution for that hour is considered unstable, and subsequent normalization enters the equal division branch.

[0079] Next, a normalized weighted factor vector is generated. The preset weight coefficient set is taken from the weight coefficient set output in step 4 and arranged in order of channel unique identifier. The preset weight coefficient set satisfies the non-negativity constraint and its sum is 1 (parameter determination method: the weight coefficient set is generated by fusing structural strength, scale contribution, and correlation strength and then normalizing. The normalization process is completed by dividing each channel score by the sum of the scores, so that the sum is 1 and each term is non-negative). At each hourly index, the original weighting factor is first generated: for each channel, the preset weight coefficient of the channel is multiplied by the contribution value of the channel in that hour, and the product is the original weighting factor of the channel in that hour. Then, the original weighting factors of all channels are accumulated one by one to obtain the original factor sum. If the sum of the original factors is not less than 0.000001 (threshold determination method: using the same numerical caliber as the lower limit threshold of the total contribution, both used for floating-point operation stability protection to avoid abnormal amplification of the normalization result due to an excessively small denominator), then the original weighted factor of each channel is divided by the sum of the original factors to obtain the normalized weighted factor vector for that hour, so that the sum of the normalized weighted factors of all channels for that hour is 1; if the sum of the original factors is less than 0.000001, then the normalized weighted factor vector for that hour is evenly divided according to the number of channels, and the average value is 1 divided by the total number of channels (parameter determination method: this branch is used for deterministic backoff when the denominator is unavailable, and the even division makes the vector sum 1 and keeps the factors of each channel non-negative). Repeat the above calculation process of multiplication generation, summation, normalization division or even division backoff for hour indices 1 to 24 to obtain a normalized weighted factor vector sequence covering 24 hours.

[0080] Finally, the corrected channel prediction value sequence is weighted and fused using a normalized weighted factor vector, and the final airport future throughput trend prediction result is output. Fusion is performed hourly by hourly index: at a certain hourly index, the fusion accumulator is set to 0; all channels are traversed one by one, reading the corrected channel prediction value and the normalized weighted factor for that hour; the two are multiplied to obtain the weighted value for that hour; then the weighted value is added to the fusion accumulator, and the fusion accumulator is updated; the fusion accumulator after the traversal is complete is the fused predicted throughput value for that hour. The above channel-by-channel multiplication and item-by-item accumulation are repeated for hourly indices 1 to 24 to form the fused predicted throughput sequence. After fusion, capacity constraint verification is performed: the airport capacity hard constraint threshold sequence is read, and the fused predicted throughput value for each hour is compared with the capacity hard constraint threshold for that hour at each hour index; when the fused predicted throughput value is greater than the capacity hard constraint threshold, the fused predicted throughput value for that hour is reduced to the capacity hard constraint threshold for that hour, and the normalized weighted factor vector for that hour is simultaneously scaled proportionally so that the fused predicted throughput value obtained after fusing and accumulating the scaled factor vector with the corrected channel prediction value sequence is consistent with the capacity hard constraint threshold; the scaling ratio is the capacity hard constraint threshold divided by the calculated result of the fused predicted throughput value before the reduction, and after scaling, all channel factors are accumulated item by item and normalized division is performed so that the sum of the scaled factor vector is 1 and each component is non-negative. The verification difference threshold is set to 0.000001 (threshold determination method: the fusion process involves multiple multiplications and accumulations, and scaling and re-normalization involve division operations; the threshold is used to absorb the rounding errors introduced by the above floating-point operations to ensure that the final output does not exceed the limit). When the difference between the fused predicted throughput value and the capacity hard constraint threshold is not greater than 0.000001, the verification for that hour is considered passed. After performing fusion, verification, downsizing, and factor calibration on the 24-hour index hourly, the final airport future throughput trend prediction result sequence that meets the airport capacity constraint conditions is output.

[0081] To verify the prediction stability, interpretability, and capacity constraint adaptability of the airport throughput trend prediction method described in this invention under complex operating scenarios, this embodiment constructs an experimental sample based on historical operating data of a certain airport.

[0082]

[0083] As shown in the table above, Scheme 1 directly predicts the total airport throughput as a whole, without distinguishing between different business channels such as runway direction, terminal, passenger and cargo transport. Therefore, it cannot fully reflect the differentiated changes between different channels. Its MAPE is 8.74%, reaching 12.96% during peak hours, and it experienced 21 instances of capacity exceeding limits, indicating that this scheme is prone to prediction bias during peak hours. Scheme 2, based on Scheme 1, introduces channel decomposition, breaking down the airport throughput into multiple channel flows with clear business meanings and modeling them separately. This reduces the MAPE from 8.74% to 7.91%, and the peak-hour MAPE from 12.96% to 11.38%. This result shows that channel decomposition can improve the model's ability to characterize changes in different business flows.

[0084] Scheme 3 further replaces ordinary least squares regression with a ridge regression model, reducing MAPE to 6.82% and RMSE to 493.5. This result indicates that in airport throughput forecasting, different input features are often correlated; for example, planned flight volume, historical traffic, and peak period proportions may simultaneously affect the forecast results. Ridge regression, through L2 regularization, can suppress abnormal fluctuations in model parameters, thereby improving forecast stability. Scheme 4, based on Scheme 3, introduces multi-level weighted fusion, further reducing MAPE to 6.24%. This result shows that weighted fusion of channel forecast results according to runway direction, terminal, and passenger / cargo type can more accurately reflect the differences in the contribution of different channels to the airport's total throughput.

[0085] Scheme 5 introduces a comprehensive competition and interaction matrix based on Scheme 4, reducing the peak-hour MAPE from 8.36% to 7.42%. This result indicates that during peak hours, competition may arise between channels due to sharing runway, terminal area, or cargo area resources. Correcting channel prediction results using the comprehensive competition and interaction matrix can reduce peak-hour prediction distortion caused by the superposition of independent channel predictions.

[0086] Scheme 6, the embodiment of this invention, further introduces a capacity hard constraint projection mechanism based on Scheme 5, reducing MAPE to 5.31%, and during peak hours to 6.88%, while reducing the number of capacity overruns from 8 to 0. This result demonstrates that this invention not only improves the accuracy of airport throughput trend prediction but also ensures that the prediction results comply with overall airport capacity limits and channel capacity constraints, thereby enhancing the executability and reliability of the prediction results in actual airport operation and management.

[0087] To ensure the comparability of ablation experiment results, all schemes 1 to 6 in the table were tested using the same dataset, the same data partitioning method, and the same prediction evaluation metrics. The experimental data came from historical airport operation records, with hourly granularity. The sample included fields such as historical airport throughput, flow rate in each runway direction, passenger flow rate in each terminal, cargo flow rate, planned flight volume, peak hour percentage, capacity allocation bias, channel dynamic impact coefficient, and airport hourly capacity limit. The experiment aimed to predict hourly throughput for the next 24 hours, comparing the model output with the actual operational data for the corresponding hours to calculate the prediction error under different schemes.

[0088] Specifically, the raw airport operation data is first preprocessed to remove severely missing records, abnormal timestamps, and records that clearly do not conform to the actual operation patterns of the airport. For a small number of missing values, interpolation between adjacent time periods or imputation using the mean of the same period is used. Then, the samples are divided into training, validation, and test sets according to time sequence. The training set is used for model parameter learning, the validation set is used to determine the ridge regression regularization parameters, the upper limit threshold of competition intensity, and the capacity constraint parameters, and the test set is used only for final performance evaluation. All schemes are predicted on the same test set to avoid incomparable evaluation results due to sample differences.

[0089] Scheme 1 obtains data by directly building an airport total throughput prediction model. This scheme does not perform channel splitting of airport throughput; instead, it takes historical total throughput, planned flight volume, holiday markers, and historical lag characteristics as inputs and directly outputs the predicted airport total throughput for the next 24 hours. By comparing this predicted value with the actual total throughput for the corresponding hour in the test set, the MAPE, peak-hour MAPE, RMSE, and number of capacity overruns for Scheme 1 can be obtained.

[0090] Scheme 2 obtains data through a combination of channel decomposition and ordinary least squares regression. This scheme first breaks down airport throughput into multiple channel flows based on runway direction, terminal area, and passenger / cargo type. Then, it establishes an ordinary least squares regression prediction model for each channel to obtain the predicted value for the next 24 hours. Finally, the predicted values ​​for all channels are aggregated to form the total airport throughput prediction value, and the error is calculated by comparing it with the actual total throughput of the test set. Because this scheme incorporates channel decomposition, its error is lower than that of Scheme 1. However, ordinary least squares regression is susceptible to multicollinearity, and its prediction stability remains limited.

[0091] The data for Scheme 3 is obtained by replacing ordinary least squares regression with ridge regression, based on Scheme 2. Specifically, a ridge regression model is built for each channel, and an L2 regularization term is added to the loss function to limit excessive fluctuations in the regression coefficients. After the ridge regression model is trained, the prediction results for each channel over the next 24 hours are output, and the prediction values ​​for each channel are then summarized to obtain the predicted total airport throughput. The only difference between Scheme 3 and Scheme 2 is the channel prediction model; therefore, the error change of Scheme 3 compared to Scheme 2 reflects the contribution of ridge regression and L2 regularization to the prediction stability.

[0092] Scheme 4 obtains data by adding multi-level weighted fusion to Scheme 3. Specifically, after obtaining the ridge regression prediction results for each channel, the contribution weights of different channels are further calculated according to the runway direction layer, terminal level, and passenger / cargo type layer, and the channel prediction results are fused and corrected based on these weights. The corrected channel prediction values ​​are then aggregated into the airport's total throughput prediction value. The difference between Scheme 4 and Scheme 3 lies in whether multi-level weighted fusion is introduced; therefore, the difference in their evaluation results reflects the improvement in prediction accuracy brought about by multi-level structural modeling.

[0093] Scheme 5 obtains data by adding a comprehensive competitive interaction matrix to Scheme 4. Specifically, for each hour of the next 24 hours, resource sharing markers, peak overlap values, capacity allocation bias differences, and target channel dynamic impact coefficients are calculated between channels, forming a comprehensive competitive interaction matrix sliced ​​by hour. This matrix is ​​then used to competitively correct the predicted values ​​for each channel, reducing prediction bias caused by multiple channels simultaneously competing for runway, terminal, or cargo area resources during peak hours. The difference between Scheme 5 and Scheme 4 lies in whether or not the competitive relationship between channels is considered; therefore, the reduction in MAPE during peak hours reflects the corrective effect of the competitive interaction matrix on peak prediction distortion.

[0094] Scheme 6 obtains data by further incorporating capacity hard constraint projection into Scheme 5. Specifically, after competitive interactive correction, the predicted results for each channel are compared with the airport's hourly total capacity limit and the corresponding channel's capacity limit. When the predicted result exceeds the capacity limit, the predicted value for each channel is adjusted according to the capacity constraint projection rules to ensure the final predicted result falls within the feasible capacity range. After adjustment, the error is calculated between the final predicted value and the actual value in the test set. The difference between Scheme 6 and Scheme 5 lies in whether or not capacity hard constraint projection is performed; therefore, the number of capacity overruns is reduced from 8 to 0, demonstrating the effectiveness of the capacity hard constraint projection mechanism in improving the feasibility of the predicted results.

[0095] The MAPE in the table is used to represent the average percentage error of the predicted value relative to the actual value. It is calculated as follows: MAPE = absolute value of the difference between the predicted value and the actual value / actual value × 100%, and then the average value is taken for all prediction periods in the test set.

[0096] Peak-hour MAPE is the average percentage error calculated in the same way after selecting hourly samples from the test set where airport throughput is at its peak. Peak hours can be determined based on the airport's historical throughput distribution, for example, selecting the hours with the highest throughput in the day as peak hours, or determining the morning and evening peak hours based on the airport's actual operation and management rules.

[0097] RMSE is used to represent the root mean square level of prediction error. It is calculated as: RMSE = square root of the average of the squares of the differences between the predicted and actual values. RMSE is more sensitive to larger prediction deviations, and therefore can be used to reflect the model's error control capability under abnormal fluctuations or peak periods.

[0098] The number of capacity overruns refers to the number of times, within the test set, the hourly predicted throughput output by the model exceeds the airport's hourly capacity limit, or the predicted value for any channel exceeds the corresponding channel's capacity limit. This metric is used to evaluate whether the prediction results conform to the airport's actual operational capacity boundaries. Unlike simple error metrics, the number of capacity overruns reflects the feasibility of the prediction results in actual airport operation and management.

[0099] Using the above method, schemes 1 to 6 progressively add technical modules under the same testing conditions, with each scheme adding only one major technical feature compared to the previous one. Therefore, the error variations between schemes can reflect the independent contribution of the corresponding technical features. Experimental results show that channel decomposition can improve the model's ability to characterize changes in traffic flow for different services; ridge regression can suppress model fluctuations caused by the correlation of multiple factors; multi-level weight fusion can enhance the correlation expression between different service levels; the comprehensive competitive interaction matrix can reduce prediction distortion caused by resource competition during peak hours; and capacity hard constraint projection can ensure that the final prediction results meet airport capacity limits.

[0100] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for rapidly predicting future airport throughput trends based on ridge regression algorithm, characterized in that, include: S1. Obtain the multi-channel traffic time series set and the overall throughput benchmark series; S2. For the traffic time series of each channel, each channel is independently modeled, and an initial independent prediction value sequence of each channel is generated within a fixed future time window to suppress multicollinearity between features and retain the trend information of each channel's prediction results. S3. Organize the independent prediction values ​​of all channels into a multi-level throughput prediction feature set that includes the bottom channel vector, the intermediate level aggregate vector, and the top level overall vector. S4. Based on the multi-level throughput prediction feature set, calculate the correlation coefficient between the bottom-level channels, the weight distribution of the intermediate-level channels, and the overall influence intensity between the top-level intermediate levels to obtain the weight coefficient set. S5. Based on the set of weighting coefficients, combined with the priority ranking of the proportion of peak traffic periods in the channels and the capacity allocation bias caused by the differences in channel types, the independent predicted values ​​of each channel are weighted and adjusted to obtain the intermediate predicted value sequence of the channels. S6. Accumulate and summarize all intermediate predicted values ​​of all channels layer by layer upwards according to the hierarchical structure to obtain a preliminary overall predicted throughput sequence. Compare the preliminary overall predicted throughput sequence with the airport capacity hard constraint threshold. If the preliminary overall predicted throughput sequence exceeds the capacity hard constraint threshold, trigger the capacity constraint projection calculation to adjust the prediction result; otherwise, proceed to the next step. S7. After projection of the capacity constraint, the channel prediction results are normalized and weighted and fused based on the set of weight coefficients and the channel prediction contribution output by the ridge regression model to obtain the final airport future throughput trend prediction result that meets the airport capacity constraint conditions.

2. The method for rapidly predicting the future throughput trend of an airport based on the ridge regression algorithm according to claim 1, characterized in that: S1 includes: The system acquires historical airport operation records, aggregates them by hourly window, and generates hourly traffic sequences for each region and overall airport hourly throughput sequences. Specifically, it acquires hourly departure and arrival international and domestic passenger and cargo traffic sequence data for each runway direction, terminal, and cargo area of ​​the airport during historical periods, and acquires overall airport hourly throughput sequence data for the corresponding time period. Outliers are detected in the hourly flow sequences of each region and the overall hourly throughput sequence of the airport. The gaps after removing outliers are corrected by interpolation to obtain the corrected flow sequences without outliers. The corrected abnormal flow sequence is standardized to generate a standardized flow value sequence. A multi-channel flow time series set and an overall throughput benchmark sequence are obtained based on the standardized flow numerical sequence.

3. The method for rapidly predicting the future throughput trend of an airport based on the ridge regression algorithm according to claim 1, characterized in that: S2 includes: Obtain the traffic time series for each channel, and construct a multi-factor feature vector based on the historical time window. The multi-factor feature vector includes periodic features, peak features, and trend features. The ridge regression algorithm is used to independently model the multi-factor feature vectors. The L2 regularization constraint is used to suppress multicollinearity among features in the multi-factor feature vectors to obtain the ridge regression model parameters. Based on the parameters of the ridge regression model, an initial independent prediction sequence for each channel is generated within a fixed future time window; Output the initial independent prediction value sequence, and retain the trend information of each channel prediction result within a fixed future time window.

4. The method for rapidly predicting the future throughput trend of an airport based on the ridge regression algorithm according to claim 1, characterized in that: S3 includes: Obtain the initial independent prediction sequence for each channel, and construct a structural mapping matrix describing the channel affiliation by combining the runway number, terminal area, and passenger and cargo identification. Based on the structure mapping matrix, the initial independent predicted value sequence is reorganized and accumulated to generate the bottom-level channel vector and the intermediate-level aggregate vector; The top-level overall vector is obtained by globally summarizing the intermediate-level aggregated vectors. The top-level overall vector, the intermediate-level aggregated vectors, and the bottom-level channel vectors are then concatenated to output a multi-level throughput prediction feature set.

5. The method for rapidly predicting the future throughput trend of an airport based on the ridge regression algorithm according to claim 1, characterized in that: S4 includes: Obtain the bottom channel vector sequence from the multi-level throughput prediction feature set, and construct a cross-correlation matrix for the bottom channel vector sequence; Based on the cross-correlation matrix and the intermediate-level aggregation vector, the traffic proportion value is calculated, and the proportion weight distribution sequence of the intermediate-level channels is obtained. By performing regression analysis between the weight distribution sequence and the overall top-level vector, the overall influence intensity value among the intermediate levels of the top-level is determined. A feature importance mapping table is constructed by integrating the overall impact intensity value, the proportion weight distribution sequence, and the cross-correlation matrix. The set of weight coefficients is extracted to characterize the differences in the impact of different channels on the overall throughput prediction results.

6. The method for rapidly predicting the future throughput trend of an airport based on ridge regression algorithm according to claim 1, characterized in that, S5 includes: Obtain historical traffic load data and independent predicted values ​​for each physical channel, and generate a sequence of traffic peak period percentages. Based on the peak traffic period percentage sequence, the capacity allocation bias value determined by the channel type difference, and the channel dynamic influence coefficient, a comprehensive competitive interaction matrix is ​​constructed. The matrix is ​​organized according to the directed relationship between channels, and the matrix elements represent the competitive pressure intensity of the source channel on the target channel in that hour. The comprehensive competition interaction matrix is ​​applied to the independent predicted values ​​to extract the channel competition correction set, which is obtained by applying the comprehensive competition interaction matrix to the independent predicted values. By combining the set of channel competition corrections with the independent prediction values ​​and making a weighted adjustment, a sequence of intermediate prediction values ​​for the channels that takes into account the competition relationship between channels is obtained.

7. The method for rapidly predicting the future throughput trend of an airport based on the ridge regression algorithm according to claim 1, characterized in that, S6 includes: Obtain the intermediate predicted value sequence of the channel, and merge and sum the intermediate predicted value sequence of the channel according to the preset channel hierarchy structure mapping relationship to obtain the preliminary overall predicted throughput sequence; The preliminary overall predicted throughput sequence is compared with the airport capacity hard constraint threshold. If there is an overflow, a throughput overflow difference sequence is generated.

8. The method for rapidly predicting the future throughput trend of an airport based on the ridge regression algorithm according to claim 7, characterized in that: S6 further includes: A multidimensional capacity-constrained projection vector is constructed based on the throughput overflow difference sequence, and the multidimensional capacity-constrained projection vector is decomposed inversely to obtain the channel projection correction set; The channel projection correction set is applied to adjust the intermediate prediction value sequence of the channel, and the final overall predicted throughput sequence that meets the airport capacity hard constraint threshold is output.

9. The method for rapidly predicting the future throughput trend of an airport based on the ridge regression algorithm according to claim 1, characterized in that, S7 includes: The corrected channel prediction value sequence is obtained by superimposing the channel projection correction value set onto the initial independent prediction value. The modified channel prediction value sequence is decomposed based on the regression coefficient matrix of the ridge regression model to obtain a set of channel prediction contribution values.

10. The method for rapidly predicting the future throughput trend of an airport based on the ridge regression algorithm according to claim 9, characterized in that: The S7 also includes: The preset set of weight coefficients and the set of channel prediction contribution values ​​are used to generate a normalized weighted factor vector. The corrected channel prediction value sequence is weighted and fused using a normalized weighted factor vector to output the final airport future throughput trend prediction result that meets the airport capacity constraint.

Images