Aviation income management method

Abstract

The invention discloses an aviation income management method, which comprises a prediction module and a pricing module, and is characterized in that the prediction module comprises a first step of data preparation, a second step of feature engineering, a third step of model construction, a fourth step of model training, a fifth step of model evaluation and a sixth step of model prediction; the pricing module comprises a step A of establishing an integer programming model by utilizing an optimization engine and a step B of solving a model in the step 1. Compared with the prior art, the optimal pricing solution of the flight per hour in the future can be solved, and the price can be adjusted in time according to the market condition. And a model algorithm is optimized, so that the use effect is better. The main purpose of adopting the deep learning and operation planning optimization method is to reasonably adjust the price during the flight sales period, so that the final income is maximized.

Description

Technical Field

[0001] The invention relates to a revenue management method, in particular to an aviation revenue management method. Background Art

[0002] Aviation revenue management refers to timely and accurate adjustments to revenue management strategies and pricing decisions based on passenger demand forecasts and changes in the market environment, thereby creating opportunities for the company to increase revenue.

[0003] In the original revenue management process, revenue management personnel estimated potential sales based on historical sales data and current sales progress, and then determined ticket prices based on basic operating manuals, evaluation criteria and experience.

[0004] It can be seen that the traditional pricing and revenue management model has shortcomings in the following aspects:

[0005] First, the goal of maximizing revenue cannot be guaranteed: the recommended price is given based on rules manually sorted out according to business scenarios, and no algorithm is used to search for the optimal solution. Second, the ability to estimate the ticket sales volume of each cabin in each hour in the future is insufficient. Third, the decision-making ability is limited: there is no ability to interact with external data (emergencies, competitive changes, etc.), and the decision-making ability is highly dependent on the individual experience of the implementation management.

[0006] To solve the above problems, we made a series of improvements. Summary of the invention

[0007] The purpose of the present invention is to provide an aviation revenue management method to overcome the above-mentioned shortcomings and deficiencies of the prior art.

[0008] An aviation revenue management method includes: a prediction module and a pricing module, which predicts the ticket sales volume of each flight under each pricing condition in each future hour, takes the prediction result as input, constructs an operations optimization model with the goal of maximizing the total flight revenue, and obtains the optimal pricing plan for each future hour after solving the problem. The prediction module includes:

[0009] Step 1 is to prepare data. Prepare the data set for training and testing. The data set contains data of each dimension and the corresponding ticket sales volume.

[0010] Step 2 is feature engineering: processing of missing values, outliers, error values, data formats, and sampling issues in the data.

[0011] Step 3 is to build a model: build a deep neural network model, determine the network structure, number of layers, number of neurons in each layer, and activation function.

[0012] Step 4 is to train the model: by using the labeled training data to adjust the parameters of the model so that it can better fit the data.

[0013] Step 5 is to evaluate the model: After the model training is completed, it is evaluated to understand its performance on unseen data. By evaluating the performance of the model, it is determined whether it has achieved the expected prediction effect, and adjustments and improvements are made.

[0014] Step 6 is model prediction: for new input data, the model will make predictions based on previously learned knowledge and give the corresponding estimated ticket sales results;

[0015] The pricing module includes:

[0016] Step A is to use the optimization engine to establish an integer programming model: by establishing an integer programming model, the optimal solution for the future hourly pricing of the flight is solved.

[0017] Step B is to solve the model of step 1: generate a batch of pricing for each future hour so that the final revenue of each flight is maximized.

[0018] Among them, the integer programming model in step A is:

[0019]

[0020]

[0021] p i ≥p i-1 -a,for i=1,2,...N

[0022] p i ≤p i-1 +a,for i=1,2,...N

[0023] In the integer programming model,

[0024]

[0025] Goal: Maximize revenue and load factor

[0026]

[0027] Constraint 1: The total number of tickets sold from the current time to the departure time does not exceed the remaining number of tickets.

[0028] p i ≥p i-1 -a,for i=1,2,...N

[0029] Constraint 2: The price reduction at each time point does not exceed the threshold

[0030] p i ≤p i-1 +a,for i=1,2,...N

[0031] Constraint 3: The price increase at each time point does not exceed the threshold

[0032] Among them, p i = price at time point i, d i (p i ) = estimated sales volume from time point i to time point i+1, C i = the number of remaining plans at time point i, a = price adjustment threshold, r = penalty coefficient for the final unsold tickets, r≤0.

[0033] Furthermore, the structure of the deep neural network model of the prediction module includes: a deep neural network, a convolutional neural network, a long short-term memory network and an attention network.

[0034] The prepared data in step 1 includes: extracting features related to ticket sales, flight departure airport, landing airport, take-off time, current sales volume, sales volume in the past N hours, foreign airline prices, foreign airline flights,

[0035] The feature engineering of step 2 includes: processing missing values ​​of data, normalizing data preprocessing, and constructing various deep learning networks based on the characteristics of each dimensional feature, wherein the dimensional features include: constructing a deep neural network to model discrete features such as flight departure airports and landing airports, constructing a long short-term memory network and a convolutional neural network to extract information from sequence features such as sales volume in the past N hours, and constructing an attention network to model the impact of each foreign airline on this flight.

[0036] The model building of step three includes: splicing the information extracted from each network into a deep neural network, and using MSE LOSS to learn model parameters.

[0037] Furthermore, the evaluation indicator for the performance on the evaluation data in step 5 is accuracy.

[0038] Beneficial effects of the present invention:

[0039] Compared with traditional technologies, the present invention can solve the optimal solution for the future hourly pricing of flights, and can adjust prices in time according to market conditions. The model algorithm is optimized to achieve better results. The main purpose of using deep learning and operations optimization methods is to reasonably adjust prices during flight sales to maximize the final profit. DETAILED DESCRIPTION

[0040] The present invention is further described below in conjunction with specific embodiments. It should be understood that the following embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention.

[0041] Example 1

[0042] An aviation revenue management method includes: a prediction module and a pricing module, which predicts the ticket sales of each flight under various pricing conditions in each hour in the future. The prediction results are used as input to construct an operations optimization model with the goal of maximizing the total flight revenue. After solving, the optimal pricing plan for each hour in the future is obtained. The prediction module includes:

[0043] Step 1 is to prepare data. Prepare the data set for training and testing. The data set contains data of each dimension and the corresponding ticket sales volume.

[0044] Step 2 is feature engineering: processing of missing values, outliers, error values, data formats, and sampling issues in the data.

[0045] Step 3 is to build a model: build a deep neural network model, determine the network structure, number of layers, number of neurons in each layer, and activation function.

[0046] Step 4 is to train the model: by using the labeled training data to adjust the parameters of the model so that it can better fit the data.

[0047] Step 5 is to evaluate the model: After the model training is completed, it is evaluated to understand its performance on unseen data. By evaluating the performance of the model, it is determined whether it has achieved the expected prediction effect, and adjustments and improvements are made.

[0048] Step 6 is model prediction: for new input data, the model will make predictions based on previously learned knowledge and give the corresponding estimated ticket sales results;

[0049] Pricing modules include:

[0050] Step A is to use the optimization engine to establish an integer programming model: by establishing an integer programming model, the optimal solution for the future hourly pricing of the flight is solved.

[0051] Step B is to solve the model of step 1: generate a batch of pricing for each future hour so that the final revenue of each flight is maximized.

[0052] Among them, the integer programming model in step A is:

[0053]

[0054]

[0055] p i ≥p i-1 -a,for i=1,2,...N

[0056] pi ≤p i-1 +a,for i=1,2,...N

[0057] In the integer programming model,

[0058]

[0059] Goal: Maximize revenue and load factor

[0060]

[0061] Constraint 1: The total number of tickets sold from the current time to the departure time does not exceed the remaining number of tickets.

[0062] p i ≥p i-1 -a,for i=1,2,...N

[0063] Constraint 2: The price reduction at each time point does not exceed the threshold

[0064] p i ≤p i-1 +a,for i=1,2,...N

[0065] Constraint 3: The price increase at each time point does not exceed the threshold

[0066] Among them, p i = price at time point i, d i (p i ) = estimated sales volume from time point i to time point i+1, C i = the number of remaining plans at time point i, a = price adjustment threshold, r = penalty coefficient for the final unsold tickets, r≤0.

[0067] The structure of the deep neural network model of the prediction module includes: deep neural network, convolutional neural network, long short-term memory network and attention network.

[0068] The data prepared in step 1 include: extracting features related to ticket sales, flight departure airport, landing airport, take-off time, current sales volume, sales volume in the past N hours, foreign airline prices, foreign airline flights,

[0069] The feature engineering of step 2 includes: processing missing values ​​of data, normalizing data preprocessing, and constructing deep learning networks based on the characteristics of each dimensional feature. The dimensional features include: constructing a deep neural network to model discrete features such as flight departure airports and landing airports, constructing a long short-term memory network and a convolutional neural network to extract information from sequence features such as sales volume in the past N hours, and constructing an attention network to model the impact of each foreign airline on this flight.

[0070] The model building in step three includes: splicing the information extracted from each network into a deep neural network, and using MSE LOSS to learn the model parameters.

[0071] The evaluation metric for the performance on the evaluation data in step 5 is accuracy.

[0072] The present invention is designed based on the deep learning theory and the operational research profit maximization theory, and mainly adopts the deep neural network model and the integer programming model respectively. Its characteristics are that it can solve the optimal solution for the future hourly pricing of flights, and can adjust the price in time according to market conditions. The main difference from the greedy allocation algorithm of current domestic revenue management is that the greedy allocation algorithm is an algorithm based on intuition or experience, and cannot achieve the goal of maximizing revenue. In terms of specific effects, the algorithm currently used in revenue management has a higher degree of optimization than the greedy allocation algorithm and has better use effect.

[0073] This solution is a new model built based on practical problems. No one has made a similar one using the same principle model or algorithm.

[0074] This algorithm first constructs a deep learning model to predict the estimated ticket sales volume of each flight at each future hour and each pricing situation. Through rich feature mining and extraction and advanced model structure, a high estimation accuracy is achieved. Compared with the traditional greedy allocation algorithm, the application of deep learning models can automatically learn and extract features from a large amount of data, thereby achieving complex pattern recognition and prediction tasks. The main application of deep learning models in airline revenue management is demand forecasting, that is, predicting the passenger demand for each fare sub-class of each flight in the future based on historical data and market information. The advantage of deep learning models is that they can process high-dimensional, nonlinear, and non-stationary data, capture the dynamic changes and influencing factors of demand, and improve the accuracy and robustness of predictions.

[0075] Then, an integer programming model with the goal of maximizing revenue is established. Compared with the instability and uncertainty of the empirical summary of the traditional greedy algorithm, it can make the linear or nonlinear function of one or more integer variables reach the maximum or minimum value while satisfying some constraints. In airline revenue management, the integer programming model is used to obtain the optimal pricing for each hour in the future by solving it. The process of searching for the optimal solution with an algorithm is realized. The advantage of the integer programming model is that it can accurately describe the constraints and objective functions of the problem, and use an effective algorithm to solve the optimal solution or approximate optimal solution. In addition, this algorithm fully models the competition of foreign airlines in the prediction module and realizes the ability to interact with external data, thereby making better decisions.

[0076] In this embodiment, there are three price levels to choose from, namely 100 yuan, 150 yuan and 180 yuan. There are still 3 hours before the flight takes off, and the remaining number of tickets is 60. The price adjustment threshold is set to 50 yuan.

[0077] Therefore, according to the model of the present invention, the ticket sales volume for the first hour, second hour and third hour in the future for each price level is predicted by the deep learning model as follows:

[0078]

[0079]

[0080] Then establish the following integer programming model

[0081]

[0082] x 11 *20+x 12 *40+x 13 *30+x 21 *11+x 22 *24+x 23 *16+x 31 *5+x 32 *15+x 33 *10≤60

[0083] x 11 +x 21 +x 31 =1

[0084] x 12 +x 22 +x 32 =1

[0085] x 13 +x 23 +x 33 =1

[0086] |x 11 *100+x 21 *150+x 31 *180-x 12 *100-x 22 *150-x 32 *180|≤50

[0087] |x 13 *100+x 23 *150+x 33 *180-x 12 *100-x 22 *150-x 32 *180|≤50

[0088] The optimization engine can be used to find the optimal pricing for each hour in the next three hours.

[0089] The specific implementation modes of the present invention are described above, but the present invention is not limited thereto. The present invention can also have various changes without departing from the purpose of the present invention.

Claims

1. An aviation revenue management method, comprising: The prediction module and the pricing module predict the ticket sales volume of each flight under each pricing condition in each hour in the future. The prediction module is characterized in that the prediction result is used as input to construct an operations optimization model with the goal of maximizing the total revenue of the flight, and the optimal pricing plan for each hour in the future is obtained after solving the problem. The prediction module includes: Step 1 is to prepare data. Prepare the data set for training and testing. The data set contains data of each dimension and the corresponding ticket sales volume. Step 2 is feature engineering: processing of missing values, outliers, error values, data formats, and sampling issues in the data. Step 3 is to build a model: build a deep neural network model, determine the network structure, number of layers, number of neurons in each layer, and activation function. Step 4 is to train the model: by using the labeled training data to adjust the parameters of the model so that it can better fit the data. Step 5 is to evaluate the model: After the model training is completed, it is evaluated to understand its performance on unseen data. By evaluating the performance of the model, it is determined whether it has achieved the expected prediction effect, and adjustments and improvements are made. Step 6 is model prediction: for new input data, the model will make predictions based on previously learned knowledge and give the corresponding estimated ticket sales results; The pricing module includes: Step A is to use the optimization engine to establish an integer programming model: by establishing an integer programming model, the optimal solution for the future hourly pricing of the flight is solved. Step B is to solve the model of step 1: generate a batch of pricing for each future hour so that the final revenue of each flight is maximized. Among them, the integer programming model in step A is: p i ≥p i-1 -a,for i=1,2,...N p i ≤p i-1 +a,for i=1,2,...N In the integer programming model, Goal: Maximize revenue and load factor Constraint 1: The total number of tickets sold from the current time to the departure time does not exceed the remaining number of tickets. p i ≥p i-1 -a,for i=1,2,...N Constraint 2: The price reduction at each time point does not exceed the threshold p i ≤p i-1 +a,for i=1,2,...N Constraint 3: The price increase at each time point does not exceed the threshold Among them, p i = price at time point i, d i (p i ) = estimated sales volume from time point i to time point i+1, C i = the number of remaining plans at time point i, a = price adjustment threshold, r = penalty coefficient for the final unsold tickets, r≤0.

2. The aviation revenue management method according to claim 1, characterized in that: The structure of the deep neural network model of the prediction module includes: a deep neural network, a convolutional neural network, a long short-term memory network and an attention network. The prepared data in step 1 includes: extracting features related to ticket sales, flight departure airport, landing airport, take-off time, current sales volume, sales volume in the past N hours, foreign airline prices, foreign airline flights, The feature engineering of step 2 includes: processing missing values ​​of data, normalizing data preprocessing, and constructing various deep learning networks based on the characteristics of each dimensional feature, wherein the dimensional features include: constructing a deep neural network to model discrete features such as flight departure airports and landing airports, constructing a long short-term memory network and a convolutional neural network to extract information from sequence features such as sales volume in the past N hours, and constructing an attention network to model the impact of each foreign airline on this flight. The model building of step three includes: splicing the information extracted from each network into a deep neural network, and using MSE LOSS to learn model parameters.

3. The aviation revenue management method according to claim 1, characterized in that: The evaluation indicator for the performance on the evaluation data in step 5 is accuracy.

Images