Deep learning-based demand forecasting system

Patent metadata

Field	Value
Publication number	US-11948111-B1
Application number	US-202117509812-A
Country	US
Kind code	B1
Filing date	Oct 25, 2021
Priority date	Oct 29, 2020
Publication date	Apr 2, 2024
Grant date	Apr 2, 2024

Abstract

A method of training a neural network to approximate a forecasting error of a passenger-demand forecasting model that includes calculating, using the forecasting model, a historical passenger demand forecast for each key level in a set of key levels and for each departure date in a set of historical departures dates; applying a dropout model to the historical passenger demand forecasts to create a training sample; training, using the historical passenger demand forecasts and the training sample, the neural network, to approximate forecasting errors associated with the forecasting model; calculating, using the forecasting model, a future passenger demand forecast for each key level in the set of key levels and for each departure date in a set of future dates; and approximating, using the trained neural network, the forecasting error associated with the future passenger demand forecasts for the second set of departure dates.

First claim

What is claimed is:

1. A method of training a neural network to approximate a forecasting error of a passenger-demand forecasting model, the method comprising:
calculating, using the forecasting model, a historical passenger demand forecast for each key level in a set of key levels and for each departure date in a first set of departure dates;
wherein the forecasting model is an integrated model that comprises:
a convolutional neural network (“CNN”) model;
a long short-term memory (“LSTM”) model;
wherein the LSTM model is a recurrent neural network;

and
a convolutional LSTM model that integrates the CNN model and the LSTM model together in one architecture;

wherein each key level in the set of key levels is associated with a unique product for which demand is being forecasted; and
wherein the first set of departures dates are historical departures dates;

applying a dropout model to the historical passenger demand forecasts to create a training sample;
training, using the historical passenger demand forecasts and the training sample, the neural network, to approximate forecasting errors associated with the forecasting model;
calculating, using the forecasting model, a future passenger demand forecast for each key level in the set of key levels and for each departure date in a second set of departure dates;
wherein the second set of departures dates are future dates;
wherein each departure date of the second set of departure dates is associated with a plurality of forecast periods;
wherein each forecast period of the plurality of forecast periods is associated with a period of time before departure;
wherein each departure date of the second set of departure dates is associated with a plurality of fare products;
wherein each fare product of the plurality of fare products is a marketable combination of access rights associated with a charging moment;
wherein the CNN model captures, for each departure date of the second set of departure dates and for each forecast period from the plurality of forecasts periods on a given departure date, a booking relationship between the plurality of fare products;
wherein the LSTM model models a time series that is a temporal correlation, across the departure dates of the second set of departure dates, of the booking relationships; and
wherein calculating, using the forecasting model, the future passenger demand forecast for each key level in the set of key levels and for each departure date in the second set of departure dates comprises:
applying a convolutional feature extractor on input data of the CNN model;
wherein the input data comprises historical fractional closures for each fare product of the plurality of fare products;

running the CNN model using the input data of the CNN model to create output(s) of the CNN model; and
running the LSTM model using input data of the LSTM model to create LSTM model output(s);
wherein input data of the LSTM model comprises:
 the output(s) of the CNN model; and
 features used to model seasonality; and

wherein the output(s) of the LSTM model comprise predicted bookings with consideration of cancellation probabilities;

and
approximating, using the trained neural network, the forecasting error associated with the future passenger demand forecasts for the second set of departure dates.

2. The method of claim 1 , further comprising calculating, using the future passenger demand forecasts and the forecasting error associated with the future passenger demand forecasts, a revised future passenger demand forecast for each key level in the set of key levels and for each departure date in the second set of departure dates.

3. The method of claim 1 ,
wherein applying the dropout model occurs offline;
wherein training the neural network occurs offline;
wherein calculating the future passenger demand forecasts for the second set of departures dates occurs in real time; and
wherein approximating the forecasting error associated with the future passenger demand forecasts for the second set of departure dates occurs in real time.

4. The method of claim 1 , wherein the unique product associated with each key level is defined by:
a direction market that indicates a departure location from which a passenger departs and an arrival location at which the passenger arrives;
a day of the week on which the passenger departs from the departure location;
a block of time during the day in which the passenger departs from the departure location;
a class category reflecting a fare paid by the passenger; and
a forecast period reflecting a period of time before departure at which the fare was paid.

5. The method of claim 1 , wherein features used to model seasonality comprise features used to model effects of different days of the week, different weeks of a year, and holidays/non-holidays.

6. A method of training a neural network to approximate a forecasting error of a passenger-demand forecasting model, the method comprising:
calculating, using the forecasting model, a historical passenger demand forecast for each key level in a set of key levels and for each departure date in a first set of departure dates;
wherein the forecasting model is an integrated model that comprises:
a convolutional neural network (“CNN”) model;
a long short-term memory (“LSTM”) model;
wherein the LSTM model is a recurrent neural network; and

a convolutional LSTM model that integrates the CNN model and the LSTM model together in one architecture;

wherein each key level in the set of key levels is associated with a unique product for which demand is being forecasted; and
wherein the first set of departures dates are historical departures dates;

applying a dropout model to the historical passenger demand forecasts to create a training sample;
training, using the historical passenger demand forecasts and the training sample, the neural network, to approximate forecasting errors associated with the forecasting model;
calculating, using the forecasting model, a future passenger demand forecast for each key level in the set of key levels and for each departure date in a second set of departure dates;
wherein the second set of departures dates are future dates;
wherein each departure date of the second set of departure dates is associated with a plurality of forecast periods;
wherein each forecast period of the plurality of forecast periods is associated with a period of time before departure;
wherein each departure date of the second set of departure dates is associated with a plurality of fare products;
wherein each fare product of the plurality of fare products is a marketable combination of access rights associated with a charging moment;
wherein the CNN model captures, for each departure date of the second set of departure dates and for each forecast period from the plurality of forecasts periods on a given departure date, a booking relationship between the plurality of fare products;
wherein the LSTM model models a time series that is a temporal correlation, across the departure dates of the second set of departure dates, of the booking relationships; and
wherein calculating, using the forecasting model, the future passenger demand forecast for each key level in the set of key levels and for each departure date in the second set of departure dates comprises:
applying a convolutional feature extractor on input data of the CNN model;
wherein the input data is a 5D tensor with shape (samples, time, channels, rows, cols);
wherein a sample comprises historical sample size; and
wherein time is a length of each sample;

running the CNN model using the input data of the CNN model to create output(s) of the CNN model; and
running the LSTM model using input data of the LSTM model to create LST

Assignees

• American Airlines Inc

Inventors

• Deng Na
• Segal Benjamin
• Bai Ou
• Thayer Adam
• Pilla Venkata

Classifications

• G06Q10/06315 Primary

  Needs-based resource requirements planning or analysis · CPC title

• G06N3/045

  Combinations of networks · CPC title

• G06Q10/04

  Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem" (market predictions or forecasting for commercial activities G06Q30/0202) · CPC title

• G06Q30/0206

  Price or cost determination based on market factors · CPC title

• G06N3/044

  Recurrent networks, e.g. Hopfield networks · CPC title