Deep learning-based flight dynamic pricing method and system
Through deep learning technology, flight information is preprocessed and feature extraction, and revenue prediction and dynamic pricing are used to use Transformer models, which solves the problem that traditional pricing methods cannot reflect market changes in real time, realizes dynamic adjustment of flight prices, and maximizes airline revenue .
Patent Information
- Application Number
- CN202411211570.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-30
- Publication Date
- 2024-12-27
AI Technical Summary
Traditional flight pricing methods cannot reflect market changes in real time, resulting in airlines being unable to maximize profits.
Deep learning-based dynamic flight pricing method is adopted, and flight information data is preprocessed and feature extraction is performed, revenue prediction is used using the Transformer model, and dynamic pricing is performed based on price sensitivity indicators and pre-set pricing strategies.
It has achieved adjustments to flight prices based on market dynamics, helping airlines maximize profits and improve market competitiveness.
Smart Images
Figure CN119205170A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of flight dynamic pricing, and in particular to a flight dynamic pricing method and system based on deep learning. Background Art
[0002] In the air transport market, flight pricing is a complex and dynamic process that is affected by many factors, including seasonality, cost changes, market demand, etc. Traditional pricing methods often fail to reflect market changes in real time, resulting in airlines being unable to maximize their revenue. Therefore, it is particularly important to develop a dynamic pricing system that can analyze market dynamics in real time and adjust prices accordingly. Summary of the invention
[0003] In order to solve the technical problems existing in the background technology, the present invention proposes a flight dynamic pricing method and system based on deep learning.
[0004] The present invention proposes a flight dynamic pricing method based on deep learning, comprising the following steps:
[0005] S1. Preprocessing the acquired flight information data of a single route to obtain dynamic correlation data;
[0006] S2. Extracting correlation features from dynamic correlation data to obtain pricing correlation features, where the pricing correlation features include price features, time features, flight features, cost features, and seasonality indicators;
[0007] S3: Input pricing-related features and current sales data into the pre-established and trained Transformer model to complete the revenue forecast for the corresponding flight within the preset time period And obtain the price sensitivity index S, the current sales data includes the actual sales volume, the actual ticket price and the sales time point;
[0008] S4. Flight-based revenue forecast The price sensitivity index S and the preset pricing strategy complete the dynamic pricing of the corresponding flights.
[0009] Preferably, the flight information data includes historical flight sales data, price change data, flight schedule, seasonal factors, and flight cost data, and the dynamically associated data includes first-type historical flight sales data, first-type price change data, first-type flight schedule, first-type seasonal factors, and first-type flight cost data.
[0010] Preferably, the data preprocessing specifically includes:
[0011] Clean the flight information data of a single route to remove duplicate and abnormal flight information data;
[0012] Sort the cleaned flight information data in chronological order;
[0013] The sorted flight information data is completed to obtain dynamic associated data.
[0014] Preferably, step S2 specifically includes:
[0015] Extract the departure place, destination and number of flights per day from the first type of flight schedule, and use the extracted departure place, destination and number of flights per day as flight features F = {F1, F2, ..., Fo};
[0016] Extracting a date from the first type of flight schedule and determining whether the corresponding date is a holiday, and when the corresponding date is a holiday, marking the corresponding date as a holiday, and using the extracted date and the holiday mark as a time feature T = {T1, T2, ..., Tm};
[0017] Based on the first type of price change data, the prices are sorted in chronological order to obtain a price sequence P = {P1, P2, ...Pi..., Pn}, and the price sequence P = {P1, P2, ...Pi..., Pn} is used as a price feature, where Pi is the price at the i-th time point;
[0018] Based on the first type of seasonal factors, the dates are marked as factors to obtain the seasonal index S = {S1, S2, ..., Ss} corresponding to each date, where the first type of seasonal factors includes the tourist peak season and the off-season;
[0019] The first type of flight cost data is arranged in chronological order to obtain cost characteristics C = {C1, C2, ..., Cp}.
[0020] Preferably, the method for constructing the pre-established and trained Transformer model includes:
[0021] Standardize the historical flight information data to obtain a feature data set;
[0022] Obtain model parameters θ, and split the feature data set according to a preset strategy to form a training set, a test set, and a validation set;
[0023] Based on the feature data set and model parameters θ, the Transformer network is trained to obtain a Transformer model;
[0024] Adjust the model parameters θ according to the training results until a Transformer model that meets the preset error is obtained.
[0025] Preferably, the preset strategy is specifically:
[0026] 70% of the feature data set is used as the training set, 20% of the feature data set is used as the test set, and 10% of the feature data set is used as the validation data set.
[0027] Preferably, the preset pricing strategy specifically includes:
[0028] When income forecast When it is greater than the preset revenue target T, the current actual fare is maintained;
[0029] When income forecast When it is less than or equal to the preset income target T, the price adjustment strategy is executed;
[0030] The execution price adjustment strategy specifically includes:
[0031] The price sensitivity index S is compared with the preset threshold S threshold Compare to obtain the price adjustment direction, which includes price increase and price decrease;
[0032] When the price sensitivity index S is greater than the preset threshold S threshold When the price adjustment direction is specifically to reduce the price;
[0033] When the price sensitivity index S is less than the preset threshold S threshold When the price adjustment direction is specifically to increase the price;
[0034] Get the given price constraint P min ≤P≤P max , MaximizeRevenue, and constrain P with a given price min ≤P≤P max , taking MaximizeRevenue as a constraint, adjust the basic fare according to the price adjustment direction to obtain the optimal fare Poptimal.
[0035] Preferably, the calculation process of maximizing revenue MaximizeRevenue is as follows:
[0036] MaximizeRevenue=P×Q(P,S);
[0037] Among them, P is the given ticket price; Q(P,S) is the sales volume under the given ticket price P and price sensitivity index S.
[0038] Preferably, the preset income target T is specifically:
[0039] T=Avg(R hist )×m;
[0040] Among them, Avg(R hist ) is the historical average income; m is the preset percentage value.
[0041] The present invention proposes a flight dynamic pricing system based on deep learning, comprising:
[0042] A data acquisition module is used to perform data preprocessing on the acquired flight information data of a single route to obtain dynamic associated data;
[0043] A feature extraction module is used to extract correlation features from dynamic correlation data to obtain pricing correlation features. Pricing correlation features include price features, time features, flight features, cost features, and seasonal indicators.
[0044] The data processing module is used to input pricing-related features and current sales data into the pre-established and trained Transformer model to complete the revenue forecast of the corresponding flight within the preset time period. And obtain the price sensitivity index S, the current sales data includes the actual sales volume, the actual ticket price and the sales time point;
[0045] Dynamic pricing module for flight-based revenue forecasting The price sensitivity index S and the preset pricing strategy complete the dynamic pricing of the corresponding flights.
[0046] In the present invention, the proposed flight dynamic pricing method and system based on deep learning preprocesses the flight information data of a single route to obtain dynamic correlation data; extracts correlation features from the dynamic correlation data to obtain pricing correlation features; inputs the pricing correlation features and current sales data into a pre-established and trained Transformer model to complete the revenue forecast of the corresponding flight within a preset time period. And obtain the price sensitivity index S; revenue forecast based on flights The price sensitivity index S and the preset pricing strategy complete the dynamic pricing of the corresponding flight. The flight price is dynamically adjusted according to the predicted revenue and price sensitivity index, thus helping airlines maximize their revenue and improve their market competitiveness. BRIEF DESCRIPTION OF THE DRAWINGS
[0047] Figure 1 A schematic diagram of the workflow structure of a flight dynamic pricing method based on deep learning proposed by the present invention;
[0048] Figure 2 A schematic diagram of the system architecture of a flight dynamic pricing system based on deep learning proposed in the present invention. DETAILED DESCRIPTION
[0049] Reference Figure 1 and Figure 2 The present invention proposes a flight dynamic pricing method based on deep learning, comprising the following steps:
[0050] S1. Preprocess the flight information data of a single route to obtain dynamic correlation data.
[0051] In this embodiment, the flight information data includes flight historical sales data, price change data, flight schedule, seasonal factors, and flight cost data, and the dynamically associated data includes first-type flight historical sales data, first-type price change data, first-type flight schedule, first-type seasonal factors, and first-type flight cost data.
[0052] In this embodiment, data preprocessing specifically includes:
[0053] Clean the flight information data of a single route to remove duplicate and abnormal flight information data;
[0054] Sort the cleaned flight information data in chronological order;
[0055] The sorted flight information data is completed to obtain dynamic associated data.
[0056] In this example, all input features, including not only the current sales data but also all other features used to predict flight prices and revenues, are normalized. Standardization is performed to eliminate the impact of dimensions and optimize model performance. Among them, Z is the standardized data point; x is the original data point; μ is the mean of the original data; δ is the standard deviation of the original data. After processing, all data points in each feature will be converted to a standard normal distribution with a mean of 0 and a standard deviation of 1. This helps the model better handle data of different dimensions and ranges. Convert time series data (such as historical prices and sales volumes) into serialized form for model processing. Add position encoding to the sequence data so that the model can understand the position information of each element in the sequence. Set the final feature vector expression to x = [P1, T1, F1, C1, S1, Q1, Pr1, Tsale1,…, Pn
[0057] ,Tn,Fn,Cn,Sn,Qn,Prn,Tsale n].
[0058] S2. Extract correlation features from dynamic correlation data to obtain pricing correlation features. Pricing correlation features include price features, time features, flight features, cost features, and seasonal indicators.
[0059] In this embodiment, step S2 specifically includes:
[0060] Extract the departure place, destination and number of flights per day from the first type of flight schedule, and use the extracted departure place, destination and number of flights per day as flight features F = {F1, F2, ..., Fo};
[0061] Extracting a date from the first type of flight schedule and determining whether the corresponding date is a holiday, and when the corresponding date is a holiday, marking the corresponding date as a holiday, and using the extracted date and the holiday mark as a time feature T = {T1, T2, ..., Tm};
[0062] Based on the first type of price change data, the prices are sorted in chronological order to obtain a price sequence P = {P1, P2, ...Pi..., Pn}, and the price sequence P = {P1, P2, ...Pi..., Pn} is used as a price feature, where Pi is the price at the i-th time point;
[0063] The dates are marked based on the first type of seasonal factors to obtain seasonal indexes S = {S1, S2, ..., Ss} corresponding to each date, where the first type of seasonal factors include the tourist peak season and the off-season;
[0064] The first type of flight cost data is arranged in chronological order to obtain cost characteristics C = {C1, C2, ..., Cp}.
[0065] S3: Input pricing-related features and current sales data into the pre-established and trained Transformer model to complete the revenue forecast for the corresponding flight within the preset time period And obtain the price sensitivity index S, the current sales data includes the actual sales volume, the actual ticket price and the sales time point.
[0066] In this embodiment, the method for constructing a pre-established and trained Transformer model includes:
[0067] Standardize the historical flight information data to obtain a feature data set;
[0068] Obtain model parameters θ, and split the feature data set according to a preset strategy to form a training set, a test set, and a validation set;
[0069] Based on the feature data set and model parameters θ, the Transformer network is trained to obtain a Transformer model;
[0070] Adjust the model parameters θ according to the training results until a Transformer model that meets the preset error is obtained.
[0071] In this embodiment, the preset strategy is specifically:
[0072] 70% of the feature data set is used as the training set, 20% of the feature data set is used as the test set, and 10% of the feature data set is used as the validation data set.
[0073] In the Transformer model operation stage: Self-attention mechanism: Calculate the query (Q), key (K), and value (V) representation of the input data. The input feature vector x passes through three different linear layers ω Q ,ω k ,ω v , mapped to query (Q), key (K) and value (V), where Q = xω Q , K = xω k 、V=xω v ;ω Q ,ω k ,ω v It is a weight matrix that has been trained in advance.
[0074] Calculate the attention score by dot product For each day i and j, calculate the attention score:
[0075]
[0076] This score reflects the similarity between the feature value labels such as ticket price, passenger flow and holidays on the i-th day and those on the j-th day.
[0077] In this embodiment, the Softmax function is applied to normalize the attention score of each i using the softmax function to obtain the attention weight: wij=softmax(AttentionScoreij); this will ensure that the total attention on the i-th day is 1.
[0078] The Softmax function is an activation function commonly used in machine learning and deep learning, especially when dealing with multi-classification problems. It converts a vector or a set of real-number inputs into another vector of the same length as output, where each element of the output vector has a value between 0 and 1, and the sum of all elements is 1. This makes the Softmax function very suitable for use as an activation function for the output layer of a multi-classification model because it can provide a probability distribution.
[0079] In the self-attention mechanism, the Softmax function is used to calculate the attention weights, which reflect the importance of each element in the input sequence to the current element. By processing the original attention score (i.e. the result of the dot product) through the Softmax function, we can get a normalized weight distribution, so that the model can take into account the relevance of elements at different positions when processing sequence data.
[0080] For example, in flight dynamic pricing, the Softmax function can help the model determine which day's ticket price in the past week has the greatest impact on today's ticket price when predicting future ticket prices. Through the attention weights processed by the Softmax function, the model can more reasonably allocate the degree of attention to ticket prices on different days.
[0081] In this embodiment, the attention weights are used to weight the sum V, and the obtained attention weights are used to weight the sum of the values V to obtain the self-attention output:
[0082] Output i =∑ j w ij ·V j ;
[0083] This will generate a new feature representation for each day that incorporates information from the entire series. This representation is like a snapshot of the entire series, highlighting the most important information for predicting fares while ignoring less relevant details. In this way, each day's representation contains useful signals from the entire series, helping us better understand the data and make more accurate predictions.
[0084] In this embodiment, the input data is divided into multiple heads, and each head calculates self-attention independently.
[0085] The input is the feature representation after the self-attention mechanism, that is, the self-attention output obtained by processing the query (Q), key (K), and value (V) and weighted summation. i .
[0086] The split feature representation is that the output of the self-attention is logically split into h parts, corresponding to h different "heads". Each head will process a subset of the input data.
[0087] The number of heads h is determined based on experimental verification and theoretical analysis, aiming to balance the learning ability and computational efficiency of the model. When selecting h, factors such as the model's ability to capture flight data features, prediction accuracy, and resource consumption are considered. The specific number of heads is determined based on a series of preliminary experiments by comparing the performance of the model on the validation set under different head number configurations, ensuring that the selected number of heads can optimally meet the needs of the flight dynamic pricing task.
[0088] We recommend choosing a moderate number of heads, such as h = 8, as a starting point, as this is a common choice that has been shown to work well across a variety of tasks.
[0089] The linear transformation process is that for each head h, the input data is transformed independently (i.e., different weight matrices ) Generate the corresponding query Q h , key K h Sum value V h .
[0090] The specific calculation of self-attention is to use Q in each head h and key K h Calculate the attention score AttentionScore h , the formula is
[0091] Softmax normalization is to apply the Softmax function to the attention scores in each head to obtain the normalized attention weights and weighted sum V: Use the normalized attention weights to value V h Perform weighted summation to get the output of each head
[0092] Output Head of each head h It captures the feature representation of the input data in different subspaces. By capturing the feature representation of different subspaces in each head, the model is able to build a comprehensive feature set that can more comprehensively reflect the intrinsic characteristics and patterns of the input data. In the flight dynamic pricing model, this helps to predict fares more accurately because it takes into account multiple aspects of the data, such as seasonal fluctuations, price trends, passenger flow changes, etc.
[0093] In this embodiment, the Transformer model feedforward network processing process is as follows:
[0094] A feedforward network is a component in a sequence processing model that consists of two linear layers and a nonlinear activation function in between. 1 The output of self-attention or multi-head attention is transformed from dimension d model Mapping to a higher dimension d ff , usually d ff Yes model A nonlinear activation function (such as ReLU) is applied to the output of the first linear layer to introduce nonlinear characteristics to the model and enhance its expressive power. 2 The output of self-attention or multi-head attention is transformed from dimension d ff Mapping back to the original dimension d model The feedforward network performs nonlinear transformations on the output of self-attention, which helps the model capture more complex features and patterns.
[0095] In this embodiment, the residual connection and layer normalization processing process of the Transformer model are specifically as follows: the residual connection adds the input of the feedforward network (the output of self-attention or multi-head attention) to the output of the feedforward network to form a residual connection Residual = Input + FFNOutput; the output of the residual connection is layer normalized, and the result of the layer normalization of the residual connection is normalized by the normalization operation LayerNorm (Residual) to reduce the internal covariate shift and accelerate the model training. Residual connections help solve the gradient vanishing problem in deep networks and allow gradients to propagate more effectively in the network. Encoder stacking: Encoder layer structure, each encoder layer contains a self-attention mechanism, multi-head attention, a feedforward network, and residual connections and layer normalization steps. The purpose of stacking: By stacking multiple encoder layers, the model can abstract and extract features layer by layer, and each layer further learns the high-level representation of the data based on the previous layer. Implementation of stacking: The output of each layer is used as the input of the next layer, forming a deep network that can capture more complex data features and patterns. Selection of stacking layers: The number of encoder layers is selected based on experimental verification and task requirements to achieve the best model performance and generalization ability.
[0096] The above steps together constitute the core of the Transformer model, providing powerful sequence processing capabilities for the flight dynamic pricing task.
[0097] After completing step S3, we get the following key outputs, which will serve as the basis for the dynamic pricing decision in step S4: Self-attention output: The weighted value Output obtained through the self-attention mechanism i , reflecting the importance of features at different time points. Multi-head attention output: The feature representation generated by each head h is Head h , these representations capture the characteristics of the input data in different subspaces. Feedforward network output: data processed by the feedforward network, including feature representations after nonlinear transformation. Residual connection and layer normalization processing: the results of residual connection Residual and layer normalization LayerNorm (Residual) help stabilize deep network training. Encoder stack output: comprehensive feature representations after processing by multiple encoder layers, which contain high-level features abstracted from the original data. Final prediction output: the model's final predicted flight revenue R predicted And price sensitivity index S. Among them, the model outputs the predicted value of flight revenue R predicted , which is a comprehensive result based on all processed features and model parameters. The price sensitivity index S takes into account the integration of multi-dimensional features, including seasonal factors, cost changes, etc., to provide a comprehensive sensitivity measure; it can be interpreted as the expected impact of price changes on flight demand under current market conditions.
[0098] S4. Flight-based revenue forecast The price sensitivity index S and the preset pricing strategy complete the dynamic pricing of the corresponding flights.
[0099] In this embodiment, the preset pricing strategy specifically includes:
[0100] When income forecast When it is greater than the preset revenue target T, the current actual fare is maintained;
[0101] When income forecast When it is less than or equal to the preset income target T, the price adjustment strategy is executed;
[0102] The execution price adjustment strategy specifically includes:
[0103] The price sensitivity index S is compared with the preset threshold S threshold Compare to obtain the price adjustment direction, which includes price increase and price decrease;
[0104] When the price sensitivity index S is greater than the preset threshold S threshold When the price adjustment direction is specifically to reduce the price;
[0105] When the price sensitivity index S is less than the preset threshold S threshold When the price adjustment direction is specifically to increase the price;
[0106] Get the given price constraint P min ≤P≤P max , MaximizeRevenue, and constrain P with a given price min ≤P≤P max , taking MaximizeRevenue as a constraint, adjust the basic fare according to the price adjustment direction to obtain the optimal fare Poptimal.
[0107] In this embodiment, the calculation process of maximizing revenue MaximizeRevenue is as follows:
[0108] MaximizeRevenue=P×Q(P,S);
[0109] Among them, P is the given ticket price; Q(P,S) is the sales volume under the given ticket price P and price sensitivity index S.
[0110] In this embodiment, the preset income target T is specifically:
[0111] T=Avg(R hist )×m;
[0112] Among them, Avg(R hist ) is the historical average income; m is the preset percentage value.
[0113] In this embodiment, using the output of step S3, step S4 will perform the following dynamic pricing process: Revenue forecast evaluation: Evaluate R predicted Whether the preset income target R is met or exceeded target . The preset income target R target It is formulated by the airline's business leaders, and the basis for formulation includes but is not limited to expert judgment, historical data reference, etc.
[0114] Price sensitivity analysis: Determine the price adjustment strategy based on S. If S<threshold S threshold If S>threshold S threshold The price may need to be adjusted downward. The threshold S used in price sensitivity analysis threshold The threshold is initially set by comprehensively considering market research, historical sales data, consumer behavior analysis and industry standards, and is automatically adjusted according to real-time sales feedback and market dynamics during the model operation. The threshold uses machine learning algorithms to predict the relationship between price and sales volume, combines risk assessment to determine the adjustment strategy, and refers to expert system recommendations for fine-tuning to ensure that the pricing strategy is in line with market trends and meets compliance requirements. Through the established feedback loop mechanism, S threshold Continuously optimize to adapt to changing market conditions and achieve precise execution of dynamic pricing strategies.
[0115] Pricing strategy decisions: Combine revenue forecasts with price sensitivity indicators to determine if ticket prices need to be adjusted.
[0116] Optimization algorithm pricing: Use optimization algorithms, such as linear programming, to determine the optimal fare P optimal The goal is to maximize revenue MaximizeRevenue=P×Q(P,S) under the given price constraint Pmin≤P≤Pmax. Here Q(P,S) represents the sales volume under the given ticket price P and price sensitivity index S. Price sensitivity index S is a key parameter that reflects the degree of consumer response to price changes and directly affects the sensitivity of the sales function Q. When the ticket price P increases, if the S value is larger, it means that demand is more sensitive to price and sales may drop faster; on the contrary, if the S value is smaller, demand is not sensitive to price and sales are less affected by price changes. When determining the optimal ticket price, we also need to consider the price constraint, that is, P min ≤P≤P max , which ensures that the fare will not be lower than the cost limit or higher than the upper limit accepted by the market. Through linear programming, we can find the optimal solution P under these constraints optimal, it can maximize revenue while satisfying consumers’ price sensitivity S.
[0117] Real-time pricing implementation: The calculated optimal fare P optimal Implement into flight sales and monitor sales data and market feedback.
[0118] Dynamic pricing feedback: Collect real-time sales data after implementation, including new sales volumes, ticket prices, and timings, and use this data to update model parameters.
[0119] Model parameter update: Update the model parameters θ based on the new sales data and revenue forecasts, using the learning rate α and gradient Adjustment
[0120] Iteration of pricing strategy: Repeat the process of step S4 according to the updated model parameters and market conditions to achieve continuous optimization of dynamic pricing.
[0121] Reference Figure 1 and Figure 2 The present invention proposes a flight dynamic pricing system based on deep learning, comprising:
[0122] A data acquisition module is used to perform data preprocessing on the acquired flight information data of a single route to obtain dynamic associated data;
[0123] A feature extraction module is used to extract correlation features from dynamic correlation data to obtain pricing correlation features. Pricing correlation features include price features, time features, flight features, cost features, and seasonal indicators.
[0124] The data processing module is used to input pricing-related features and current sales data into the pre-established and trained Transformer model to complete the revenue forecast of the corresponding flight within the preset time period. And obtain the price sensitivity index S, the current sales data includes the actual sales volume, the actual ticket price and the sales time point;
[0125] Dynamic pricing module for flight-based revenue forecasting The price sensitivity index S and the preset pricing strategy complete the dynamic pricing of the corresponding flights.
[0126] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.
Claims
1. A flight dynamic pricing method based on deep learning, characterized in that: The following steps are involved: S1. Preprocessing the acquired flight information data of a single route to obtain dynamic correlation data; S2. Extracting correlation features from dynamic correlation data to obtain pricing correlation features, where the pricing correlation features include price features, time features, flight features, cost features, and seasonality indicators; S3: Input pricing-related features and current sales data into the pre-established and trained Transformer model to complete the revenue forecast for the corresponding flight within the preset time period And obtain the price sensitivity index S, the current sales data includes the actual sales volume, the actual ticket price and the sales time point; S4. Flight-based revenue forecast The price sensitivity index S and the preset pricing strategy complete the dynamic pricing of the corresponding flights.
2. The flight dynamic pricing method based on deep learning according to claim 1, characterized in that: The flight information data includes historical flight sales data, price change data, flight schedules, seasonal factors, and flight cost data; the dynamically associated data includes first-type historical flight sales data, first-type price change data, first-type flight schedules, first-type seasonal factors, and first-type flight cost data.
3. The flight dynamic pricing method based on deep learning according to claim 2, characterized in that: The data preprocessing specifically includes: Clean the flight information data of a single route to remove duplicate and abnormal flight information data; Sort the cleaned flight information data in chronological order; The sorted flight information data is completed to obtain dynamic associated data.
4. The flight dynamic pricing method based on deep learning according to claim 2, characterized in that: Step S2 specifically includes: Extract the departure place, destination and number of flights per day from the first type of flight schedule, and use the extracted departure place, destination and number of flights per day as flight features F = {F1, F2, ..., Fo}; Extracting a date from the first type of flight schedule and determining whether the corresponding date is a holiday, and when the corresponding date is a holiday, marking the corresponding date as a holiday, and using the extracted date and the holiday mark as a time feature T = {T1, T2, ..., Tm}; Based on the first type of price change data, the prices are sorted in chronological order to obtain a price sequence P = {P1, P2, ...Pi..., Pn}, and the price sequence P = {P1, P2, ...Pi..., Pn} is used as a price feature, where Pi is the price at the i-th time point; Based on the first type of seasonal factors, the dates are marked as factors to obtain the seasonal index S = {S1, S2, ..., Ss} corresponding to each date, where the first type of seasonal factors includes the tourist peak season and the off-season; The first type of flight cost data is arranged in chronological order to obtain cost characteristics C = {C1, C2, ..., Cp}.
5. The flight dynamic pricing method based on deep learning according to claim 1, characterized in that: The method for constructing the pre-established and trained Transformer model includes: Standardize the historical flight information data to obtain a feature data set; Obtain model parameters θ, and split the feature data set according to a preset strategy to form a training set, a test set, and a validation set; Based on the feature data set and model parameters θ, the Transformer network is trained to obtain a Transformer model; Adjust the model parameters θ according to the training results until a Transformer model that meets the preset error is obtained.
6. The flight dynamic pricing method based on deep learning according to claim 5, characterized in that: The preset strategy is specifically: 70% of the feature data set is used as the training set, 20% of the feature data set is used as the test set, and 10% of the feature data set is used as the validation data set.
7. The flight dynamic pricing method based on deep learning according to claim 1, characterized in that: The preset pricing strategy specifically includes: When income forecast When it is greater than the preset revenue target T, the current actual fare is maintained; When income forecast When it is less than or equal to the preset income target T, the price adjustment strategy is executed; The execution price adjustment strategy specifically includes: The price sensitivity index S is compared with the preset threshold S threshold比 Yes, to obtain the price adjustment direction, which includes price increase and price decrease; When the price sensitivity index S is greater than the preset threshold S threshold When the price adjustment direction is specifically to reduce the price; When the price sensitivity index S is less than the preset threshold S threshold When the price adjustment direction is specifically to increase the price; Get the given price constraint P min ≤P≤P max , MaximizeRevenue, and constrain P with a given price min ≤P≤P max , taking MaximizeRevenue as a constraint, adjust the basic fare according to the price adjustment direction to obtain the optimal fare Poptimal.
8. The flight dynamic pricing method based on deep learning according to claim 7, characterized in that: The calculation process of maximizing revenue MaximizeRevenue is as follows: MaximizeRevenue=P×Q(P,S); Among them, P is the given ticket price; Q(P,S) is the sales volume under the given ticket price P and price sensitivity index S.
9. The flight dynamic pricing method based on deep learning according to claim 7, characterized in that: The preset income target T is specifically: T=Avg(R hist )×m; Among them, Avg(R hist ) is the historical average income; m is the preset percentage value.
10. A flight dynamic pricing system based on deep learning, characterized in that: include: A data acquisition module is used to perform data preprocessing on the acquired flight information data of a single route to obtain dynamic associated data; A feature extraction module is used to extract correlation features from dynamic correlation data to obtain pricing correlation features. Pricing correlation features include price features, time features, flight features, cost features, and seasonal indicators. The data processing module is used to input pricing-related features and current sales data into the pre-established and trained Transformer model to complete the revenue forecast of the corresponding flight within the preset time period. And obtain the price sensitivity index S, the current sales data includes the actual sales volume, the actual ticket price and the sales time point; Dynamic pricing module for flight-based revenue forecasting The price sensitivity index S and the preset pricing strategy complete the dynamic pricing of the corresponding flights.