Civil aviation promotion scene-oriented in-event load prediction method and system

Abstract

The invention discloses an in-event load prediction method and system for a civil aviation promotion scene, and belongs to the technical field of artificial intelligence and big data analysis, and the method constructs a closed-loop technical scheme comprising a data enhancement module, a time sequence prediction model and an online fine tuning module. The method comprises the following steps: firstly, learning daily and promotion flow modes through a conditional variation auto-encoder, generating synthetic enhancement data with promotion features, and processing the synthetic enhancement data through an overlapped average algorithm; then, a hierarchical encoder-decoder structure is combined with dimension segmentation embedding and a double-stage attention mechanism, and cross-variable and cross-time-period multi-scale feature fusion is achieved; and finally, dynamically constructing a candidate set in a deployment stage, calculating a priority score, screening key samples, and carrying out rapid parameter updating on a model output layer to cope with concept drift. According to the scheme, technical innovation is realized in three aspects of data generation, feature modeling and online optimization, and the accuracy and robustness of load prediction in a promotion scene are remarkably improved.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and big data analysis technology, specifically to a method and system for predicting in-process load in civil aviation promotion scenarios. Background Technology

[0002] In the process of digital transformation in the civil aviation industry, airlines and online travel platforms frequently launch promotional activities to enhance their market competitiveness. These activities often trigger sudden, non-linear surges in traffic, with peak traffic reaching several times the daily average. However, traditional load forecasting techniques are primarily geared towards scenarios with stable data distribution and struggle to effectively address the drastic changes in data distribution during promotional periods, leading to significant delays in resource allocation. Furthermore, complex business dependencies exist between various interfaces of the civil aviation information system; failure to accurately model these cross-interface dependencies will further limit forecast accuracy. More critically, the scarcity of promotion-related samples in historical data makes it difficult for models to fully learn their unique distribution patterns, ultimately weakening their ability to predict sudden traffic spikes.

[0003] Current mainstream load forecasting methods can be divided into two categories: statistical model-based and deep learning-based. Statistical model methods construct mathematical relationship models by analyzing historical data. Their core idea is to allow the model to learn that load changes have inherent statistical regularities, establishing mathematical mapping relationships through the temporal correlation and periodicity of historical data. However, when facing nonlinear and multi-scale fluctuations caused by promotional scenarios, statistical model-based methods struggle to construct accurate mathematical mapping relationships. Deep learning models demonstrate advantages in handling high-dimensional nonlinear problems. However, single models are prone to underfitting due to their simple structure, while hybrid models, although adaptable to multivariate data, are limited by the scarcity of promotional data and struggle to fully learn specific distribution patterns. Furthermore, the concept drift phenomenon that occurs during promotional activities drastically reduces the predictive performance of existing models.

[0004] Existing methods suffer from three main drawbacks: First, at the data level, the scarcity of promotional samples prevents the model from capturing sudden traffic spikes, weakening its generalization ability to abnormal scenarios. Second, at the feature modeling level, traditional methods lack explicit modeling mechanisms for cross-interface dependencies and cannot effectively integrate multi-scale temporal features, resulting in insufficient capture of complex patterns. Finally, at the adaptive level, static models lack online optimization mechanisms to address concept drift, failing to dynamically adjust parameters during promotions, leading to significant deviations between predicted results and actual load.

[0005] To overcome the aforementioned bottlenecks, a comprehensive solution covering data augmentation, feature fusion, and online optimization is needed. Specifically, generative models should be used to alleviate the data scarcity problem, a hierarchical architecture should be used to explicitly model spatiotemporal dependencies, and a lightweight fine-tuning mechanism should be introduced to achieve real-time model adaptation. Only by forming a closed-loop system of "data generation - model training - online adaptation" can the accuracy and robustness of load prediction in promotional scenarios be systematically improved. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method and system for predicting load during civil aviation promotion scenarios. By constructing a closed-loop system of data augmentation, feature fusion and online optimization, accurate load prediction can be achieved during promotion periods.

[0007] This invention is achieved using the following technical solution: On the one hand, the present invention provides a method for real-time load forecasting in civil aviation promotion scenarios, the method comprising: S1: Construct a data augmentation module to learn the traffic patterns of daily traffic data and promotional traffic data in historical traffic data. Generate a large amount of synthetic augmented data with promotional characteristics based on the promotional traffic data. Process the obtained synthetic augmented data using an overlapping average algorithm to obtain promotional augmented traffic data. S2: Construct a time-series prediction model, using daily traffic data from historical traffic data and promotional enhanced traffic data generated by the data augmentation module as training data to jointly train the time-series prediction model offline, thereby improving the basic generalization ability of the time-series prediction model for the distribution of daily and promotional traffic data. S3: Build an online fine-tuning module. After the time series prediction model is trained offline, it enters the deployment stage. The online fine-tuning module automatically filters key samples based on real-time traffic and triggers rapid parameter updates of the time series prediction model to cope with concept drift and improve the accuracy of traffic prediction in promotional scenarios.

[0008] Furthermore, in S1: The data augmentation module is built upon the Conditional Variational Autoencoder (CVAE) model, retaining the encoder-decoder structure of CVAE. The encoder compresses time-series data into a latent variable distribution, and the decoder reconstructs the data based on conditional variables, resulting in TimeVAE, which serves as the core foundation model for the data augmentation module. Specifically, it includes: The encoder performs feature extraction and latent variable generation, including input window data. and time phase , window data and time phase The concatenated data is used as a joint input. A convolutional neural network is then used to extract features from the concatenated data, learn the correlation between local patterns in the time series data and temporal conditions, and finally output latent variables. ; The decoder performs multi-component reconstruction and conditional injection, including: processing the input's implicit variables. and time phase The data is then stitched together to integrate the stochastic style of the latent variables with the deterministic information of the time conditions. This integrated deterministic information is then decomposed using trend, seasonality, and residual sub-modules. The outputs of these three sub-modules are summed to reconstruct the complete window data. .

[0009] Furthermore, after constructing the data augmentation module, the process also includes model training of the constructed data augmentation module. The specific process includes: During the training phase of the CVAE model, the CVAE model aims to maximize the lower bound of log-likelihood evidence (ELBO) under certain conditions. By optimizing this objective, the CVAE model can learn from window data. Its corresponding one-hot encoded time phase vector Conditional probability distribution Its objective function can be expressed as,

[0010] In the formula, Indicates the parameters of the encoder, Indicates the parameters of the decoder; This represents a time series window of data; Indicates and The corresponding one-hot encoded time phase vector; This represents the low-dimensional latent variables learned by the CVAE model from the data; This represents the posterior probability distribution defined by the encoder of the CVAE model; This represents the likelihood probability distribution defined by the CVAE model decoder; express Divergence is used to regularize the latent space; Indicates that in the given data and time phase vector Under the condition of encoder Defined posterior probability distribution Below, for the log-likelihood function The expected value.

[0011] Furthermore, the data augmentation module learns the traffic patterns of daily and promotional traffic data from historical traffic data, generates a large amount of synthetic augmented data with promotional characteristics based on the promotional traffic data, and processes the obtained synthetic augmented data using an overlapping average algorithm to obtain promotional augmented traffic data. This also includes: S101: The historical traffic data is preprocessed using a deeply modified CVAE model: the CVAE model is used to learn the normal traffic pattern of daily traffic data and the promotional characteristics of promotional traffic data. The promotional characteristics include sudden sharp increases in traffic, duration, and subsequent slow decline. S102: The CVAE model calculates the corresponding time phase index based on the timestamp of each time point in the promotional traffic data. The time phase index is then strictly bound to the promotional traffic data extracted from the historical traffic data and stored together as time phase data information. S103: The CVAE model combines temporal phase data information to generate a large amount of independent augmented data with promotional characteristics that are discontinuous in time; S104: Using the overlapping average algorithm, the generated independent enhanced data is smoothly "stitched" into a continuous time series, and given column names and timestamps consistent with the format of historical traffic data, to obtain promotional enhanced traffic data that can be used directly in the future.

[0012] Furthermore, the calculated time phase index is a category label obtained by discretizing the 24 hours of a day according to a preset time granularity.

[0013] Furthermore, in S2: The construction of the time-series prediction model involves building a hierarchical feature extraction and fusion module, including: S201: In multivariate time series prediction, the training data for each dimension is divided into segments of length [missing information]. The fragments are embedded using Dimension-wise Segment Embedding (DSW) to construct a hierarchical encoder; S202: Based on the Two-Stage Attention (TSA) mechanism, it explicitly captures dependencies across variables and time periods, and finally constructs a hierarchical encoder-decoder structure and frequency domain-time domain dual-path processing to achieve multi-scale information fusion and accurate prediction.

[0014] Furthermore, in S201, the multivariate time series prediction is a specific application branch of the time series prediction model, focusing on processing time series data containing multiple related variables; the step of dividing the training data for each dimension into segments of length [missing information]. The fragment, and its embedding method using dimensional segmentation, is expressed as follows:

[0015] In the formula, Represents historical observation values. express The set of real numbers for all elements in a dimension matrix; Indicates the prediction of future sequence values. express The set of real numbers for all elements in a dimension matrix; and These represent the maximum time steps in the past and future, respectively; The largest variable dimension; This indicates the length of the training data segment for each dimension; Indicates the dimension of the variable; Indicates the index of the segment; Indicates the first The fragment index corresponds to the first... Observations in each variable dimension; Indicates the first The first time step Observations in each variable dimension; The expression for the dimension segmentation embedding method is as follows:

[0016] In the formula, This indicates the final input to the model corresponding to the [number]th [item]. The fragment index corresponds to the first... Input of one variable dimension; Indicates the first The fragment index corresponds to the first... Embedding of a sequence of variable dimensions; Represents position embedding, for each Each vector provides a unique representation of its location information.

[0017] Furthermore, in S202, the two-stage attention mechanism is used to process a two-dimensional vector array representing a multivariate time series, wherein each vector corresponds to a segment of the original sequence; the entire vector array is sequentially passed through a cross-time stage and a cross-dimensional stage to obtain the corresponding temporal dependencies and dimensional dependencies. The construction of the hierarchical encoder-decoder structure includes: S202-1: The encoder achieves multi-scale feature extraction through temporal vector merging and two-stage attention; S202-2: The decoder fuses features through a temporal attention module and a multi-head attention mechanism.

[0018] The encoder achieves multi-scale feature extraction through temporal vector merging and two-stage attention, including: When the number of layers At that time, the output of the DSW embedding is directly used as the initial representation; When the number of layers When adjacent time vectors are merged through a gating mechanism, a coarse-grained representation is formed.

[0019] Furthermore, S3 also includes: S301: Dynamically construct a candidate set based on the current prediction sample data; S302: Filter and fine-tune sample data based on candidate data in the candidate set; S303: Update the parameters of the last output layer of the time series prediction model using fine-tuned samples, fine-tune the model, and then make predictions to obtain the final prediction results.

[0020] Furthermore, in S301, the dynamic construction of the candidate set based on the current predicted sample data involves selecting the sample closest to the predicted sample in the time dimension, along with the augmented data generated by the data augmentation module, as the candidate set based on the proximity principle. This serves as the basis for subsequent selection and fine-tuning of samples. The two parts of the candidate set maintain a 1:1 ratio in terms of the number of rock samples. The in-phase principle is applied for screening, retaining the percentage of samples that are in phase with the test sample and the percentage of samples to the left and right of the corresponding phase within each period. Using % of the data, we obtain a candidate set of predicted samples.

[0021] Furthermore, in S302, the step of filtering and fine-tuning sample data based on candidate data in the candidate set includes: calculating and sorting priority scores for all samples in the candidate set, and filtering out the samples with the highest priority scores as fine-tuning samples for the current prediction sample. Furthermore, the specific calculation expression for the priority score is as follows:

[0022] In the formula, This represents a function that gradually decreases as the time interval increases, used to evaluate the timeliness of a sample. It is a coefficient. For time intervals; The similarity function used is the cosine similarity function. For the prediction sample, Samples within the candidate set; This represents the model's prediction residuals for samples in the candidate set, used to evaluate the amount of information that the sample can provide for model updates; Indicates time fractions; Represents the similarity score; This represents the residual fraction.

[0023] Furthermore, in S303, the fine-tuning samples are the 10 samples with the highest priority scores; the output layer is a linear layer during execution.

[0024] On the other hand, the present invention provides a real-time load prediction system for civil aviation promotion scenarios, comprising: Data Augmentation Module: Used to learn the traffic patterns of daily traffic data and promotional traffic data in historical traffic data, generate synthetic augmented data with promotional characteristics based on promotional traffic data, and obtain promotional augmented traffic data through overlapping average algorithm; Time-series prediction model: Based on daily traffic data in historical traffic data and promotional enhanced traffic data generated by the data augmentation module, the model is trained offline to improve its basic generalization ability to the distribution of daily and promotional traffic data. Online fine-tuning module: During the deployment phase of the time series prediction model, key samples are automatically selected based on real-time traffic, triggering rapid parameter updates of the model to cope with concept drift, thereby improving the accuracy of traffic prediction in promotional scenarios.

[0025] Furthermore, the data augmentation module is built based on a conditional variational autoencoder (CVAE) and includes an encoder and a decoder, wherein: The encoder is used to concatenate window data and time phase into a joint input, extract features and generate latent variables through a convolutional neural network; The decoder is used to concatenate latent variables with time phase, and decompose and reconstruct window data through trend submodule, seasonality submodule and residual submodule.

[0026] Furthermore, the time series prediction model divides multivariate time series into segments and embeds them through Dimensional Segmentation Embedding (DSW), and explicitly captures cross-variable and cross-time period dependencies by combining a two-stage attention (TSA) mechanism. It constructs a hierarchical encoder-decoder structure and frequency domain-time domain dual-path processing to achieve multi-scale information fusion and prediction.

[0027] Furthermore, the online fine-tuning module dynamically constructs a candidate set, calculates sample priority scores to filter fine-tuning samples, and updates the parameters of the last output layer of the time series prediction model, ultimately outputting the prediction results.

[0028] Furthermore, the data augmentation module, time-series prediction model, and online fine-tuning module work together to form a closed-loop system from data generation and model training to online optimization, in order to solve the problem of insufficient prediction accuracy caused by data scarcity, concept drift, and cross-interface dependencies in promotional scenarios.

[0029] Compared with existing technologies, this invention demonstrates significant technological breakthroughs and innovative advantages in three aspects: 1. Innovative Advantages of the Data Augmentation Module: Addressing the pain point of scarce promotional data in traditional methods, this solution develops an innovative data augmentation technique based on Conditional Variational Autoencoders (CVAEs). This module, through an asymmetric conditional injection architecture, injects only time phase conditions and latent variables into the seasonality submodule, achieving precise control over periodic patterns. By using an overlapping averaging algorithm, independent window samples are smoothly "stitched" into a continuous time series. The generated high-quality simulation data possesses an interpretable decomposition structure (baseline level, trend, seasonality, residuals), alleviating the data scarcity problem and providing the model with rich training samples for promotional scenarios, greatly improving the model's generalization ability.

[0030] 2. Technological Breakthrough in the Hierarchical Feature Extraction and Fusion Module: This module employs Dimensional Segmentation Embedding (DSW) combined with a two-stage attention (TSA) mechanism to explicitly capture cross-variable and cross-time-period dependencies. A gating mechanism enables coarse-grained merging of adjacent time vectors, and combined with frequency-time domain dual-path processing, frequency domain representation modulation is performed using Discrete Fourier Transform to construct a hierarchical encoder-decoder structure. This design can accurately capture complex and changing features, explicitly model cross-interface dependencies, and significantly improve the model's feature extraction capabilities for complex temporal patterns and cross-interface dependencies.

[0031] 3. Real-time Adaptability of the Online Fine-tuning Module: Addressing the challenge of "concept drift" in data distribution during promotional scenarios, this solution employs an online fine-tuning mechanism based on priority score selection. A candidate set is constructed using the nearest neighbor principle and the in-phase principle. A comprehensive score is then calculated by combining a timeliness function, a similarity function, and a prediction residual function. Only the last output layer of the model needs fine-tuning to achieve rapid adaptation. This lightweight strategy maintains a significant improvement in prediction accuracy while achieving an optimized balance between computational efficiency and performance, enabling the model to maintain high accuracy throughout the promotional period.

[0032] These advantages work together to form a complete technical loop from data preparation to model deployment and online optimization, solving the core bottleneck problem faced by existing technologies in handling civil aviation promotion scenarios.

[0033] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures pointed out in the description, claims, and drawings. Attached Figure Description

[0034] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0035] Figure 1 This is a schematic diagram of a real-time load prediction method for civil aviation promotion scenarios according to the present invention; Figure 2 This is a schematic diagram of the core basic module based on the CVAE model in the data augmentation module of this invention; Figure 3 This is a flowchart illustrating the encoder in the hierarchical encoder-decoder structure of this invention. Figure 4 This is a flowchart illustrating the process of constructing the online fine-tuning module of this invention. Detailed Implementation

[0036] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, 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.

[0037] In one embodiment, please refer to Figure 1 A method for real-time load forecasting in civil aviation promotion scenarios, the method comprising the following steps: Step S1: Construct a data augmentation module, use the data augmentation module to learn the traffic patterns of daily traffic data and promotional traffic data in historical traffic data, generate a large amount of synthetic augmented data with promotional characteristics based on promotional traffic data, and process the obtained synthetic augmented data using an overlapping average algorithm to obtain promotional augmented traffic data; In this embodiment, since historical traffic data often contains very little (scarce) promotional traffic data, and the distribution of traffic data during promotional periods often differs significantly from that of daily traffic data, the data augmentation module in this embodiment aims to improve the adaptability of the time-series prediction model during promotional periods. It provides unified augmented data with temporal phase to meet the needs of both the initial pre-training and subsequent online fine-tuning of the time-series prediction model. This allows the data augmentation module to generate high-quality, temporally diverse synthetic data, providing sufficient and effective augmented samples for the initial training and fine-tuning processes. This effectively combats the deterioration in model prediction performance caused by data pattern drift, thereby ensuring the long-term stability of the entire prediction system with high accuracy.

[0038] Furthermore, in this embodiment, the data augmentation module is constructed by retaining the encoder-decoder structure of the Conditional Variational Autoencoder (CVAE) model. The encoder compresses the time-series data into a latent variable distribution, and the decoder reconstructs the data based on the condition variables, resulting in TimeVAE as the core foundation model of the data augmentation module. Its core technical architecture is as follows: Figure 2 As shown, it includes: The encoder performs feature extraction and latent variable generation, including input window data. and time phase , window data and time phase The concatenated data is used as a joint input. A convolutional neural network is then used to extract features from the concatenated data, learn the correlation between local patterns in the time series data and temporal conditions, and finally output latent variables. ; The decoder performs multi-component reconstruction and conditional injection, including: processing the input's implicit variables. and time phase The data is then stitched together to integrate the stochastic style of the latent variables with the deterministic information of the time conditions. This integrated deterministic information is then decomposed using trend, seasonality, and residual sub-modules. The outputs of these three sub-modules are summed to reconstruct the complete window data. .

[0039] In this embodiment, CVAE maps time-series data and time conditions to the latent variable space through an encoder, and then achieves hierarchical control over the "trend, periodicity, and randomness" of the time-series data through the multi-component structure of the decoder. This architecture retains CVAE's condition generation capabilities and improves the controllability and interpretability of time-series data generation through "asymmetric injection" (only the seasonal submodule receives time conditions) and "multi-component decomposition" (independent modeling of trend, seasonality, and residuals). The above steps clearly present the core process of CVAE in time-series data processing, namely "encoding-decoding-reconstruction," and the precise control mechanism of time conditions on periodic patterns.

[0040] Furthermore, in this embodiment, after the data augmentation module is constructed, the model training process for the constructed data augmentation module is also included. The specific process includes: During the training phase of the CVAE model, the CVAE model aims to maximize the lower bound of log-likelihood evidence (ELBO) under certain conditions. By optimizing this objective, the CVAE model can learn from window data. Its corresponding one-hot encoded time phase vector Conditional probability distribution Its objective function can be expressed as,

[0041] In the formula, Indicates the parameters of the encoder, Indicates the parameters of the decoder; This represents a time series window of data; Indicates and The corresponding one-hot encoded time phase vector; This represents the low-dimensional latent variables learned by the CVAE model from the data; This represents the posterior probability distribution defined by the encoder of the CVAE model; This represents the likelihood probability distribution defined by the CVAE model decoder; express Divergence is used to regularize the latent space; Indicates that in the given data and time phase vector Under the condition of encoder Defined posterior probability distribution Below, for the log-likelihood function The expected value, the output of the 0th layer of the encoder, and H is the feature representation of the input sequence encoded by the encoder.

[0042] The above embodiments, through conditional modifications to the TimeVAE model and specific injection strategies, combined with corresponding data processing procedures, successfully implemented a controllable and high-fidelity time-series data augmentation scheme. The generated synthetic data contains rich temporal contextual information, which can significantly improve the training performance of downstream time-sensitive prediction models, and has significant practical application value.

[0043] Furthermore, in this embodiment, the construction of the data augmentation module, which learns the traffic patterns of daily traffic data and promotional traffic data in historical traffic data, generates a large amount of synthetic augmented data with promotional characteristics based on the promotional traffic data, and processes the obtained synthetic augmented data using an overlapping average algorithm to obtain promotional augmented traffic data, further includes: Step S101: Use a deeply modified CVAE model to preprocess historical traffic data: Use the CVAE model to learn the normal traffic pattern of daily traffic data and the promotional characteristics of promotional traffic data. The promotional characteristics include sudden sharp increases in traffic, duration, and subsequent slow decline. Step S102: The CVAE model calculates the corresponding time phase index based on the timestamp of each time point in the promotional traffic data. The time phase index is then strictly bound to the promotional traffic data extracted from the historical traffic data and stored together as time phase data information. Step S103: The CVAE model combines temporal phase data information to generate a large amount of independent augmented data with promotional characteristics that are discontinuous in time; Step S104: Using the overlapping average algorithm, the generated independent enhanced data are smoothly "stitched" into a continuous time series, and given column names and timestamps consistent with the format of historical traffic data, to obtain promotional enhanced traffic data that can be used directly in the future.

[0044] Furthermore, the calculation of its "time phase" index is a category label obtained by discretizing the 24 hours of a day according to a preset time granularity (e.g., 5 minutes).

[0045] In one embodiment, promotional enhanced traffic data is generated for the constructed data enhancement module. The specific implementation data is used to further illustrate this: "The CVAE model combines "time phase" information (e.g., the time phase corresponding to the start of the flash sale at 10:00 AM) to generate a large amount of synthetic enhanced data with these "limited-time flash sale" characteristics." For example, when generating simulated daily flight booking data, the system calculates the "time phase" for each time point (e.g., 8:00 AM, 12:00 PM, 6:00 PM). When generating data, the CVAE model precisely injects the phase information representing "8:00 AM" into the decoder's "seasonal submodule," which is responsible for generating daily periodic changes, ensuring that the synthetic data shows the expected morning peak at 8:00 AM. Meanwhile, the "trend submodule," responsible for generating the overall booking growth trend, is not affected by this "8:00 AM" phase information, thus preventing phase information from erroneously influencing the generation of the overall trend and ensuring that the generated data has both accurate periodicity and a reasonable long-term trend.

[0046] In one embodiment, an overlapping average algorithm is used to smoothly "stitch" the generated independent enhanced data into a continuous time series, and assigns it column names and timestamps consistent with historical traffic data format, resulting in promotional enhanced traffic data that can be directly used subsequently. Further explanation is provided using implementation data: Suppose the CVAE model generates two adjacent 1-hour flow data windows: the flow rate for the last 10 minutes of the first window is [100, 110, ..., 190], and the flow rate for the first 10 minutes of the second window is [80, 90, ..., 170]. If the windows are directly concatenated, a sudden jump in flow rate may occur at the boundary between the two windows (e.g., a jump from 190 to 80). Using an overlap averaging algorithm, the system calculates the average of the overlapping portions of the two windows (e.g., a weighted average of the last 5 minutes of the first window and the first 5 minutes of the second window), thus smoothly connecting the two windows to form a continuous flow rate curve, avoiding unnatural abrupt changes.

[0047] Step S2: Construct a time-series prediction model. Use daily traffic data from historical traffic data and promotional traffic data generated by the data augmentation module as training data to train the time-series prediction model offline, thereby improving the basic generalization ability of the time-series prediction model for the distribution of daily and promotional traffic data. Furthermore, in this embodiment, the core of constructing the time-series prediction model is a hierarchical feature extraction and fusion module, including: Step S201: In multivariate time series prediction, the training data for each dimension is divided into segments of length [length missing]. The fragments are embedded using Dimension-wise Segment Embedding (DSW) to construct a hierarchical encoder; Furthermore, in multivariate time series forecasting, the goal is to base forecasting on historical observations. Predicting future sequences Value, of which and These represent the maximum time steps in the past and future, respectively. The largest variable dimension; a basic assumption in this prediction process is that there are inherent relationships between multiple dimensions (such as different variables of the query interface and the ticketing interface), and effectively utilizing this cross-dimensional dependency helps to improve prediction accuracy; Furthermore, the training data for each dimension is divided into segments of length [length missing]. The fragment, and its embedding method using dimensional segmentation, is expressed as follows:

[0048] In the formula, Represents historical observation values. express The set of real numbers for all elements in a dimension matrix; Indicates the prediction of future sequence values. express The set of real numbers for all elements in a dimension matrix; and These represent the maximum time steps in the past and future, respectively; The largest variable dimension; This indicates the length of the training data segment for each dimension; Indicates the dimension of the variable; Indicates the index of the segment; Indicates the first The fragment index corresponds to the first... Observations in each variable dimension; Indicates the first The first time step Observations in each variable dimension; The expression for the dimension segmentation embedding method is as follows:

[0049] In the formula, This indicates the final input to the model corresponding to the [number]th [item]. The fragment index corresponds to the first... Input of one variable dimension; Indicates the first The fragment index corresponds to the first... Embedding of a sequence of variable dimensions; Represents position embedding, for each Each vector provides a unique representation of its location information.

[0050] Step S202: Based on the Two-Stage Attention (TSA) mechanism, the dependencies across variables and time periods are captured, and finally a hierarchical encoder-decoder structure and frequency domain-time domain dual-path processing are constructed to achieve multi-scale information fusion and accurate prediction.

[0051] Furthermore, the two-stage attention mechanism is used to process a two-dimensional vector array representing a multivariate time series, where each vector corresponds to a segment of the original sequence; the entire vector array is sequentially passed through a cross-time stage and a cross-dimensional stage to obtain the corresponding temporal and dimensional dependencies.

[0052] Furthermore, such as Figure 3 As shown, the construction of the hierarchical encoder-decoder structure includes: Step S202-1: The encoder achieves multi-scale feature extraction through temporal vector merging and two-stage attention, including, (1). When the number of layers At that time, the output of the DSW embedding is directly used as the initial representation, and the expression is:

[0053] In the formula, Indicates the first In the layer, the corresponding first The fragment index corresponds to the first... The hidden state of each variable; (2). When the number of layers At that time, adjacent time vectors are merged through a gating mechanism to form a coarse-grained representation, expressed as follows:

[0054] In the formula, , represents a learnable merge matrix, where The shape represents the number of rows = Number of columns = The set of real numbers that are all elements in a matrix; This indicates vector concatenation; It represents the Hadamardi (or Hadama) stack; Represents the gate vector; This represents the learnable weight matrix, used to calculate the gate vector; Indicates the hidden state of the previous layer; This represents the sigmoid activation function; where, when the number of fragments... When the value is odd, it is automatically padded with zeros for alignment. The merged representation is then processed through a dual-path approach in the frequency and time domains to obtain the result.

[0055] In the formula, , indicating the first Layers and dimensions are The input tensor, where, The shape represents the "length dimension". × Variable Dimensions ×Feature Dimension "The set of real numbers of all elements in a matrix; Discrete Fourier Transform (DFT) along the time dimension" Converted to frequency domain representation,

[0056] In the formula, Indicates the number of frequency points to retain, the default is 0. Learnable frequency domain weights After modulation, the inverse IDFT recovers the time-domain signal, and its expression is:

[0057] In the formula, Indicates the number of frequency points retained. , A natural number, representing the sum of the retained frequency points; Representing a dimension as A complex matrix.

[0058] Step S202-2: The decoder fuses features through a temporal attention module and a multi-head attention mechanism.

[0059] In this embodiment, as Figure 3 As shown, assuming the encoder has N layers, using... , ,…, (in , indicating the output of layer 0 of the encoder. (This refers to the feature representation of the input sequence encoded by the encoder) represents the N+1 outputs of the encoder, and the computational complexity of each encoder layer is O(n). Based on a hierarchical encoder-decoder architecture, the decoder progressively fuses the encoder output features with its own position embedding through a temporal attention module and a multi-head attention mechanism. The layer calculation process can be described as: when At this time, the position embedding matrix can be learned. As input, the initial features are obtained through the TSA layer. ;when Then output from the previous layer. As input to the TSA layer; then passed through a multi-head attention module to... For querying, output by encoder Cross-encoder-decoder feature interaction is performed for keys and values, followed by two residual connections and layer normalization, and finally output layer by layer. to The computational complexity of each decoder layer is O(n). .

[0060] In one embodiment, the construction of a hierarchical encoder-decoder structure and frequency-time dual-path processing are used to achieve multi-scale information fusion and accurate prediction. Specific implementation data will be used to further illustrate this: Suppose we need to predict flight bookings for multiple routes (multivariate). First, the DSW embeds booking data for each route every 5 minutes into a vector. Then, the hierarchical encoder uses a gating mechanism to merge adjacent 5-minute vectors into 10-minute vectors, then into 20-minute vectors, progressively extracting features at different time granularities. At each merged layer, the data simultaneously enters two paths: one path directly processes the time series data through an MLP to capture local changes such as sudden booking peaks; the other path transforms the data to the frequency domain using a DFT to identify periodic patterns such as weekly or daily bookings, adjusts these frequency components using learnable weights, and then transforms it back to the time domain using an IDFT. Finally, the results from the two paths are fused, enabling the model to both perceive sudden events and predict periodic patterns.

[0061] In one embodiment, the two-stage attention (TSA) mechanism is further illustrated using specific implementation data: When predicting flight bookings, the TSA mechanism first analyzes booking trends across different time windows, such as the past 24 hours and the past 7 days, at the "cross-time phase" level. Simultaneously, at the "cross-variable" level, it considers other variables related to flight bookings, such as flight search volume and payment success rate. TSA intelligently identifies, for example, whether a current surge in flight bookings is highly correlated with a significant increase in flight search volume for a specific route an hour ago, or negatively correlated with a sudden drop in payment success rate 30 minutes ago. Through this two-stage attention mechanism, the model can integrate this complex cross-variable and cross-time-phase information to make more accurate predictions.

[0062] Step S3: Build an online fine-tuning module. After the time series prediction model is trained offline, it enters the deployment stage. The online fine-tuning module automatically filters key samples based on real-time traffic and triggers rapid parameter updates of the time series prediction model to cope with concept drift, thereby improving the accuracy of traffic prediction in promotional scenarios.

[0063] Furthermore, such as Figure 4As shown, the online fine-tuning module, after the time-series prediction model has been trained offline, enters the deployment phase. The online fine-tuning module automatically filters key samples based on real-time traffic and triggers rapid parameter updates to address concept drift, thereby improving the accuracy of traffic prediction in promotional scenarios. Specifically, this includes: Step S301: Dynamically construct a candidate set based on the current prediction sample data; Furthermore, the dynamic construction of the candidate set based on the current predicted sample data involves selecting the sample closest to the predicted sample in the time dimension, along with the augmented data generated by the data augmentation module, as the candidate set based on the proximity principle. This serves as the basis for subsequent selection and fine-tuning of samples. The two parts of the candidate set maintain a 1:1 ratio in terms of the number of rock samples. The in-phase principle is applied for screening, retaining the percentage of samples that are in phase with the test sample and the percentage of samples to the left and right of the corresponding phase within each period. Using % of the data, we obtain a candidate set of predicted samples.

[0064] Furthermore, in one embodiment, when the "limited-time flash sale" event starts at 10:00 AM, the model needs to be fine-tuned online. The system dynamically constructs a candidate set, which includes: 1. Real traffic data for the last few minutes (e.g., the past 15 minutes).

[0065] 2. An equal amount of synthetic augmented data generated by the data augmentation module, simulating a "flash sale" scenario around 10 a.m.

[0066] The system then filters samples based on the "in-phase principle." For example, if the current prediction time is 10:05 AM, the system will retain candidate samples whose time phase falls within the "morning peak" (e.g., 9:30-10:30) range, while excluding samples whose time phase belongs to the "early morning trough," even if they may be close in time. This ensures that the fine-tuned samples are highly matched to the time context of the current promotional activity.

[0067] Step S302: Filter and fine-tune sample data based on candidate data in the candidate set; Furthermore, the step of filtering and fine-tuning sample data based on candidate data in the candidate set includes: calculating and sorting priority scores for all samples in the candidate set, and selecting the samples with the highest priority scores as fine-tuning samples for the current prediction sample. Furthermore, the specific calculation expression for the priority score is as follows:

[0068] In the formula, This represents a function that gradually decreases as the time interval increases, used to evaluate the timeliness of a sample. It is a coefficient. For time intervals; The similarity function used is the cosine similarity function. For the prediction sample, Samples within the candidate set; This represents the model's prediction residuals for samples in the candidate set, used to evaluate the amount of information that the sample can provide for model updates; Indicates time fractions; Represents the similarity score; This represents the residual fraction.

[0069] Furthermore, in one embodiment, when an airline conducts a "flash sale" event, the model needs to predict real-time traffic and perform online fine-tuning. Assume the current prediction time is... The online fine-tuning module filters samples from the candidate set. A candidate sample A is selected if it occurs 5 minutes before time T. Small in size and highly time-sensitive, its traffic pattern is highly similar to the current real-time traffic pattern. (High), and the model produced a large error when predicting sample A ( If the flow pattern is large (e.g., large), then sample A will have a very high priority score. Conversely, a sample occurring one hour before time T, even if its flow pattern is similar, will have a relatively low priority score due to its low timeliness. In this way, the system can prioritize the samples that best help the model adapt to the current sudden flow changes for fine-tuning.

[0070] Step S303: Update the parameters of the last output layer of the time series prediction model using fine-tuned samples, fine-tune the model, and then make predictions to obtain the final prediction results.

[0071] Furthermore, the fine-tuning samples are selected from the 10 samples with the highest priority scores; the output layer is a linear layer during execution.

[0072] In this application, the term "concept drift" is defined as: an irreversible and fundamental change in the potential dependency between the target variable and historical data over time.

[0073] In one embodiment, a real-time load prediction system for civil aviation promotion scenarios is provided, comprising: Data Augmentation Module: Used to learn the traffic patterns of daily traffic data and promotional traffic data in historical traffic data, generate synthetic augmented data with promotional characteristics based on promotional traffic data, and obtain promotional augmented traffic data through overlapping average algorithm; Time-series prediction model: Based on daily traffic data in historical traffic data and promotional enhanced traffic data generated by the data augmentation module, the model is trained offline to improve its basic generalization ability to the distribution of daily and promotional traffic data. Online fine-tuning module: During the deployment phase of the time series prediction model, key samples are automatically selected based on real-time traffic, triggering rapid parameter updates of the model to cope with concept drift, thereby improving the accuracy of traffic prediction in promotional scenarios.

[0074] Furthermore, the data augmentation module is built based on a conditional variational autoencoder (CVAE) and includes an encoder and a decoder, wherein: The encoder is used to concatenate window data and time phase into a joint input, extract features and generate latent variables through a convolutional neural network; The decoder is used to concatenate latent variables with time phase, and decompose and reconstruct window data through trend submodule, seasonality submodule and residual submodule.

[0075] Furthermore, the time series prediction model divides multivariate time series into segments and embeds them through Dimensional Segmentation Embedding (DSW), and explicitly captures cross-variable and cross-time period dependencies by combining a two-stage attention (TSA) mechanism. It constructs a hierarchical encoder-decoder structure and frequency domain-time domain dual-path processing to achieve multi-scale information fusion and prediction.

[0076] Furthermore, the online fine-tuning module dynamically constructs a candidate set, calculates sample priority scores to filter fine-tuning samples, and updates the parameters of the last output layer of the time series prediction model, ultimately outputting the prediction results.

[0077] Furthermore, the data augmentation module, time-series prediction model, and online fine-tuning module work together to form a closed-loop system from data generation and model training to online optimization, in order to solve the problem of insufficient prediction accuracy caused by data scarcity, concept drift, and cross-interface dependencies in promotional scenarios.

[0078] Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for real-time load forecasting in civil aviation promotion scenarios, characterized in that: A data augmentation module is constructed to learn the traffic patterns of daily traffic data and promotional traffic data in historical traffic data. Based on the promotional traffic data, a large amount of synthetic augmented data with promotional characteristics is generated. The obtained synthetic augmented data is processed using an overlapping average algorithm to obtain promotional augmented traffic data. A time-series prediction model is constructed, using daily traffic data from historical traffic data and promotional enhanced traffic data generated by the data augmentation module as training data to jointly train the time-series prediction model offline, thereby improving the basic generalization ability of the time-series prediction model to the distribution of daily and promotional traffic data. An online fine-tuning module is built. After the time series prediction model is trained offline, it enters the deployment stage. The online fine-tuning module automatically selects key samples based on real-time traffic and triggers rapid parameter updates of the time series prediction model to cope with concept drift and improve the accuracy of traffic prediction in promotional scenarios.

2. The method for real-time load forecasting for civil aviation promotion scenarios as described in claim 1, characterized in that: The data augmentation module, based on the CVAE model, retains the encoder-decoder structure of CVAE. The encoder compresses time-series data into a latent variable distribution, and the decoder reconstructs the data based on condition variables, resulting in TimeVAE, which serves as the core foundation model for the data augmentation module. Specifically, it includes... The encoder performs feature extraction and latent variable generation, including: input window data. and time phase , window data and time phase The concatenated data is used as a joint input. A convolutional neural network is then used to extract features from the concatenated data, learn the correlation between local patterns in the time series data and temporal conditions, and finally output latent variables. ; The decoder performs multi-component reconstruction and conditional injection, including: processing the input latent variables. and time phase The data is then stitched together to integrate the stochastic style of the latent variables with the deterministic information of the time conditions. This integrated deterministic information is then decomposed using trend, seasonality, and residual sub-modules. The outputs of these three sub-modules are summed to reconstruct the complete window data. .

3. The method for real-time load forecasting for civil aviation promotion scenarios as described in claim 2, characterized in that, After constructing the data augmentation module, the process also includes training a model for the constructed data augmentation module, specifically: During the training phase of the CVAE model, the CVAE model aims to maximize the lower bound of log-likelihood evidence (ELBO) under certain conditions. By optimizing this objective, the CVAE model can learn from window data. Its corresponding one-hot encoded time phase vector Conditional probability distribution Its objective function is expressed as follows: In the formula, Indicates the parameters of the encoder, Indicates the parameters of the decoder; This represents a time series window of data; Indicates and The corresponding one-hot encoded time phase vector; This represents the low-dimensional latent variables learned by the CVAE model from the data; This represents the posterior probability distribution defined by the encoder of the CVAE model; This represents the likelihood probability distribution defined by the CVAE model decoder; express Divergence is used to regularize the latent space; This represents the expected value.

4. The method for real-time load forecasting for civil aviation promotion scenarios as described in claim 1, characterized in that, The method of using a data augmentation module to learn the traffic patterns of daily and promotional traffic data in historical traffic data, generating a large amount of synthetic augmented data with promotional characteristics based on the promotional traffic data, and processing the obtained synthetic augmented data using an overlapping average algorithm to obtain promotional augmented traffic data, also includes: A deeply modified CVAE model is used to preprocess historical traffic data: the CVAE model is used to learn the normal traffic pattern of daily traffic data and the promotional characteristics of promotional traffic data. The promotional characteristics include the sudden sharp rise in traffic, the duration, and the subsequent slow decline. The CVAE model calculates the corresponding time phase index based on the timestamp of each time point in the promotional traffic data. The time phase index is then strictly bound to the promotional traffic data extracted from historical traffic data and stored together as time phase data information. The CVAE model combines temporal phase data information to generate a large amount of independent augmented data with promotional characteristics that are discontinuous in time. The overlapping average algorithm is used to smoothly "stitch" the generated independent enhanced data into a continuous time series, and assign it column names and timestamps with the same format as historical traffic data, so as to obtain promotional enhanced traffic data that can be used directly later.

5. The method for real-time load forecasting for civil aviation promotion scenarios as described in claim 1, characterized in that, The construction of the time series prediction model includes: In multivariate time series prediction, the training data for each dimension is divided into segments of length [length missing]. The fragments were embedded using a dimension segmentation embedding method to construct a hierarchical encoder; Based on a two-stage attention mechanism, the system explicitly captures dependencies across variables and time periods, and ultimately constructs a hierarchical encoder-decoder structure and frequency-time dual-path processing to achieve multi-scale information fusion and accurate prediction.

6. The method for real-time load forecasting in civil aviation promotion scenarios as described in claim 5, characterized in that, The training data for each dimension is divided into segments of length [length missing]. The fragments, and the way they are embedded using a dimension-segmented embedding method, are expressed as follows: In the formula, Represents historical observation values. express The set of real numbers for all elements in a dimension matrix; Indicates the prediction of future sequence values. express The set of real numbers for all elements in a dimension matrix; and These represent the maximum time steps in the past and future, respectively; The largest variable dimension; This indicates the length of the training data segment for each dimension; Indicates the dimension of the variable; Indicates the index of the segment; Indicates the first The fragment index corresponds to the first... Observations in each variable dimension; Indicates the first The first time step Observations in each variable dimension; The expression for the dimension segmentation embedding method is as follows: In the formula, This indicates the final input to the model corresponding to the [number]th [item]. The fragment index corresponds to the first... Input of one variable dimension; Indicates the first The fragment index corresponds to the first... Embedding of a sequence of variable dimensions; Represents position embedding, for each Each vector provides a unique representation of its location information.

7. The in-process load forecasting method for civil aviation promotion scenarios as described in claim 5, characterized in that: The two-stage attention mechanism is used to process a two-dimensional vector array representing a multivariate time series, where each vector corresponds to a segment of the original sequence; the entire vector array is passed through a cross-time stage and a cross-dimensional stage in sequence to obtain the corresponding temporal and dimensional dependencies. The construction of the hierarchical encoder-decoder structure includes: The encoder achieves multi-scale feature extraction through temporal vector merging and two-stage attention; The decoder fuses features through a temporal attention module and a multi-head attention mechanism.

8. The method for real-time load forecasting for civil aviation promotion scenarios as described in claim 7, characterized in that, The encoder achieves multi-scale feature extraction through temporal vector merging and two-stage attention, including: When the number of layers At that time, the output of the DSW embedding is directly used as the initial representation; When the number of layers When adjacent time vectors are merged through a gating mechanism, a coarse-grained representation is formed.

9. The method for real-time load forecasting for civil aviation promotion scenarios as described in claim 1, characterized in that, The online fine-tuning module, after the time-series prediction model has been trained offline, enters the deployment phase. This module automatically filters key samples based on real-time traffic and triggers rapid parameter updates to the time-series prediction model to address concept drift, thereby improving traffic prediction accuracy in promotional scenarios. It also includes: A candidate set is dynamically constructed based on the current prediction sample data; Fine-tuning sample data based on candidate data in the candidate set; The parameters of the last output layer of the time series prediction model are updated using fine-tuned samples. After fine-tuning the model, predictions are made to obtain the final prediction results.

10. The in-process load forecasting method for civil aviation promotion scenarios as described in claim 9, characterized in that: The dynamic construction of a candidate set based on the current predicted sample data involves selecting the sample closest to the predicted sample in the time dimension, along with the augmented data generated by the data augmentation module, as the candidate set based on the proximity principle. This serves as the basis for subsequent selection and fine-tuning of samples. The two parts of the candidate set maintain a 1:1 ratio in terms of the number of rock samples. The in-phase principle is applied for screening, retaining the in-phase samples and their corresponding phase percentages within each period. Using % of the data, we obtain a candidate set of predicted samples.

11. The method for real-time load forecasting for civil aviation promotion scenarios as described in claim 9, characterized in that: The process of filtering and fine-tuning sample data based on candidate data in the candidate set includes: calculating and sorting priority scores for all samples in the candidate set, and selecting the samples with the highest priority scores as fine-tuning samples for the current prediction sample. The specific expression for calculating the priority score is as follows: In the formula, This represents a function that gradually decreases as the time interval increases, used to evaluate the timeliness of a sample. It is a coefficient. For time intervals; The similarity function used is the cosine similarity function. For the prediction sample, Samples within the candidate set; This represents the model's prediction residuals for samples in the candidate set, used to evaluate the amount of information that the sample can provide for model updates; Indicates time fractions; Represents the similarity score; This represents the residual fraction.

12. A real-time load forecasting system for civil aviation promotion scenarios, characterized in that, include: Data Augmentation Module: Used to learn the traffic patterns of daily traffic data and promotional traffic data in historical traffic data, generate synthetic augmented data with promotional characteristics based on promotional traffic data, and obtain promotional augmented traffic data through overlapping average algorithm; Time-series prediction model: Based on daily traffic data in historical traffic data and promotional enhanced traffic data generated by the data augmentation module, the model is trained offline to improve its basic generalization ability to the distribution of daily and promotional traffic data. Online fine-tuning module: During the deployment phase of the time series prediction model, key samples are automatically selected based on real-time traffic, triggering rapid parameter updates of the model to cope with concept drift, thereby improving the accuracy of traffic prediction in promotional scenarios.

Images