Machine Learning Approach To Predict Flight Delays

Abstract

Air transport provides efficient, well organized, and time effective services. Even though flights are the fastest way to transport, its delay leads to customer dissatisfaction. Many factors effect flight delays, some of them are weather, operational imperfection, baggage loading etc. In this paper, we are developing a predictive system which predicts flight delays based on weather data. Flightdata set taken from US DEPARTMENT OF TRANSPORTATION and weather data set from HOURLY LAND-BASED WEATHER OBSERVATIONS FROM NOAA. We have implemented Ensemble method, Decision tree and Random forest on the balanced data set. For balancing the data set we are using sampling techniques. The algorithms are applied on the combined flight and weather data set to predict the flight delays.

Introduction

Airlines are playing crucial role in fastest mode of transport. But when flights get delayed, it not only effects customer convenience but also it certainly effects Flight Company’s reputation. Flight companies have to pay huge compensation to the customers as they have to obey certain rules. In order to avoid all those losses, it is a must to predict flight delays. There have been introduced many predictive models to predict flight delays. The Federal Aviation System (FAA) decides a flight is delayed only when the flight is late by 15 minutes than its programmed time. Supervised Machine Learning algorithms are used to predict arrival flight delays. By using Machine learning it is capable to improve models with bulk amounts of datasets like flight dataset and weather dataset. Random Forest, Ada Boost, K-N-Neighbours, Decision Tree are applied to construct models to predict whether the flight will be delayed to not. Flight data and weather data are merged and fed in the model. By using this data, constructed model accomplished a binary classification to predict flight delays.

Gradient Boosted Decision Tree is applied to predict whether the flight will be delayed or not. This model attained the great coefficient of determination of above 92% in case of flight arrivals and above 94% in case of flight departures. Artificial neural networks techniques are used for the benefit of application where the prediction model is built to prognosticate the flight delays. A type of ANN structure DMP-ANN is found which is worthy enough to predict delays. This is used to predict the flight delays with less mean square root error. A two-stage predictive model is evolved making use of supervised machine learning algorithms to speculate the flight delays. In 1st stage binary classification is performed to predict the flight delays and in the 2nd stage regression is done to predict the value of delay in minutes. By using Gradient Boosting technique, accuracy of departure delay prediction is 84% and the accuracy of arrival delay prediction is 94%. Early warning grading standards formed by the combination of flight operation and the data of flight delays. First of all we get the number of flights delayed on regular intervals of time from 7am to 10 pm. The section which contains the number of flights delayed in regular intervals of time as large emergence probability is treated as common condition, while the number of flight delayed in regular intervals of time as minimum emergence probability is treated as danger condition and then it is necessary to do early warning.

A type of Bayesian Network Structure Learning algorithm called Target-fixed Stochastic-ordered which is used to build a predictive model to predict the flight delay. After surveying the composite and undetermined association between the flight delays and practicable influence components, Bayesian network change to replicate and determine about flight delay in busy hub-airport. Two learning methods were implemented with three models. 1st model is an approximation for arrival delay based on framework learning Bayesian network with Expectation-Maximization algorithm. 2nd model is an approximation for arrival delay based on construction learning of Bayesian network with K2 algorithm. The model well read by K2 is demonstrated more acceptable for modeling the estimation of flight delay, with a better approximation rate than 1st model. A model established on the genetic algorithm to get the most out of the flight delays to lower significant air traffic flight delays. It holds two instances. First instance is to convey the delays to a number of delayed flights to stay away from delay, which can expand the flight well timed rate.

Another instance is to pass around the delay losses to heterogeneous airlines to protect the civility and equilibrium the delay loss of the air service and travelers. A prediction model of the flight delay generation is introduced in which the disapproving flight resources and the unfavorable airfield resources are examined to furnish a more productive technique for the prediction of flight delay generation. Simulation exhibit that the model and algorithm supply a constructive method for measuring prediction of flight delay transmission. A new technique is introduced which is based on content based recommendation system. According to the transmission of the delay, this new technique vigilance the target airport by observing the status of corresponding airports. The discovered status is balanced with the previous data in order to predict the solemnity of the delay. In this paper, Ensemble method is implemented to predict flight delays with better accuracy. Also Random Forest and Decision Tree algorithms are implemented with improved accuracy when compared to previous papers.

Proposed methodology

1. Sampling Technique: Oversampling

Oversampling is the technique used in this model to modify the class diffusion of a data set. The dataset is not balanced when the classification groupings are not equally constituted. To balance the dataset we have applied oversampling. Performance of classifiers is just made better by using this oversampling. Minority and majority class are balanced using this sampling technique.

2. Algorithms Implemented

Ensemble method, Random forest algorithm and Decision tree algorithms are utilized to advance the predictive model for flight delay survey. Classification based approach is applied in this model.

1. Ensemble Method

Data split samples are drawn in selecting variables from the complete training set alternative of bootstrap sample of instruction set. From range of values, splits are chosen absolutely at random. Extra tree classifier is normally economical to train from a measurement point of view but can improve much larger. Extra tree classifier can some time conclude superior to Random forest. In this paper, this method is taken to show good accuracy in predicting flight delays.

2. Random Forest Algorithm

Random Forest Algorithm is a supervised classification algorithm. It can be used for both classification and regression problems. It is used for handling the missing values. It is used to model the categorical values. We have implemented this algorithm on balanced dataset to get good accuracy.

3. Decision Tree

Decision tree is a popular tool in machine learning. It is one of the types of Supervised Machine Learning where the data is constant split to a definite framework. Decision nodes and leaves are existed. These leaves are the final outcomes and the decision nodes are the place where data is spitted.

Data analysis

1. Data Set Description

Flight on time performance Transportation Statistics collection of data from U. S. Department of Transportation And weather data collected from Hourly land based weather observations taken from NOAA. And these two are combined together and processed using predictive models to predict flight delays. This flight data set contains data of 70 airports in United States. It is taken that flight is considered to be delayed only if it is delayed by more than fifteen minutes only. Also the flights which are diverted are eliminated from the flight data set. In the flight data set we have taken column headers like month, year, day of week, carrier, day of month, origin and destination airport id, departure delays, and arrival delays and finally cancelled. In weather data set we have taken columns like year, adjusted month, adjusted delay, adjusted hour, time zone, visibility, dry bulb Fahrenheit, dry bulb Celsius, dew point Fahrenheit, dew point Celsius, relative humidity, wind speed. After taking these two data sets unnecessary columns are dropped from both flight data and weather data. In weather data, date column is spitted into year, month and day columns.

2. Data Preprocessing

In preprocessing redundant attributes are removed from data sets and column names are renamed as the data sets are having some column headers as same names. Numerical values like origin airport id, destination airport id are converted to categorical values as they are not actually numerical values built they represent some identity. And later the columns which are not necessary in this prediction process are removed. Later the two tables are merged which is necessary for the prediction model. Extra tree classifier is applied on the obtained data set. Random forest and decision tree algorithms are also applied to improve the performance when compared to previous papers. Later the categorical values are gained converted into numerical as machine learning algorithms disclose good performance by with numerical variables only.

Conclusion

In this paper, a prediction model is permitted to classify flight delays influenced by bleak weather order. A model is constructed on the data sets of both flight delays and weather data sets and sampling technique is applied to balance the data. Extra tree classifier, Decision tree and Random forest algorithms are applied on the balanced data to predict flight delays with better accuracy. We can further improve accuracy using deep neural networks.