Machine Learning Application: Flight Delay Prediction

Yuan Chai

Rosedale Global High School, ZhengZhou, Henan, China

Keywords: Flight Delay Prediction, Machine Learning, Multi-Layer Perceptron, Supervised Learning, Advanced

Transportation Information.

Abstract: As China's civil aviation industry continues to grow, air travel is becoming a popular means of transportation.

However, flight delays have become a significant issue for passengers, leading to various impacts such as

wasted time, financial losses, and emotional stress. For airlines, delays increase operational costs and damage

brand reputation. This paper aims to predict flight delays at Chinese airports using advanced machine learning

techniques, with the goal of improving operational efficiency and providing better service to passengers. The

predictive model presented in this work is designed to foresee flight arrival delays by employing supervised

machine learning algorithmThe paper provides a predictive model that uses supervised machine learning

methods to anticipate flight arrival delays. Flight data from numerous Chinese airports, along with weather

data, were collected and used during the training of the predictive model. A Multi-Layer Perceptron was

applied to build the flight delay prediction model, and extensive data preprocessing was conducted.

Hyperparameter tuning was carried out to optimize performance. The model was evaluated using cross-

validation to ensure its accuracy and generalization ability. Finally, optimization techniques were applied to

address any shortcomings and further enhance the model’s performance. The Validation score, loss and

Accuracy rate of this paper on the data set were 0.958, 0.132 and 92.6%, respectively.

1 INTRODUCTION

As China's civil aviation market develops to expand,

air travel is now a comparatively common form of

transportation for individuals. However, flight delays

have emerged as a significant concern for passengers.

Factors such as typhoons, smog, or aircraft

malfunctions can lead to widespread flight delays.

Flight delays have become a major problem for

both passengers and airlines, as the Civil Aviation

Administration of China (CAAC) has received a large

number of complaints from passengers about the high

rate of flight delays at Chinese airports in the past

decade (Jiang et al., 2020).

For example, Beijing Capital International

Airport (PEK), as China's hub airport, was found to

have an on-time performance rate of 72.74% for

departures in February 2018, significantly lower than

the 87.5% rate of Tokyo Haneda International

Airport, another major airport in Asia (Yu et al.,

2019).

Corresponding author

Flight delays are associated with a range of

impacts, including wasted time, financial losses, and

emotional stress for passengers (Song, Guo, &

Zhuang, 2020). For airlines, delays can increase

operating costs and cause damage to their brand

reputation. Even the delay of one flight can lead to

subsequent delays for multiple other flights.

Therefore, flight delay forecasting is of great

significance in the aviation industry, as it can

significantly improve operational efficiency, help

airlines optimize flight scheduling, and reduce

additional costs caused by delays. For passengers,

being informed of delays in advance can reduce

unnecessary waiting time, increase transparency,

alleviate anxiety, and provide more flexible travel

arrangements (Zhu & Li, 2021).

Additionally, airports and airlines can allocate

resources more efficiently based on the forecast

results, reducing congestion and safety risks, and

enhancing emergency management and decision-

making capabilities. In the end, this improves both the

traveler experience and the air travel system's

efficiency in general.

Chai and Y.

Machine Learning Application: Flight Delay Prediction.

DOI: 10.5220/0013486400004619

In Proceedings of the 2nd International Conference on Data Analysis and Machine Learning (DAML 2024), pages 5-9

ISBN: 978-989-758-754-2

A traditional method for flight delay prediction is

the linear regression model, which makes predictions

by analyzing the linear relationship between flight

delays and certain characteristics. However, this

approach has several drawbacks. First, while the

reality is frequently more complex and significantly

influenced by nonlinear factors, linear regression

makes the assumption that the relationship between

all features and delays is linear.

Secondly, the method assumes that the features

are independent, making it ineffective in dealing with

interactions between features(Shi et al., 2021; Yazdi

et al., 2020). Additionally, linear regression performs

poorly when handling high-dimensional data, is prone

to overfitting, and is sensitive to outliers, leading to

inaccurate prediction results. Therefore, although

linear regression models are simple to use, they have

significant limitations when applied to complex flight

delay forecasts (Kalyan et al., 2020).

In recent years, with the improvement of

computing power and optimization algorithms, multi-

layer perceptron (MLP) models have shown excellent

performance in dealing with complex nonlinear

relationships and high-dimensional data. In addition,

the availability of large-scale data and the

development of deep learning frameworks have made

the training and application of MLP models more

efficient and widespread (Kruse et al., 2022). These

technological advances not only improve the

accuracy of flight delay forecasting, but also drive the

construction of more flexible and intelligent

forecasting systems, opening up new possibilities for

aviation forecasting (Al Bataineh, Kaur, & Jalali,

2022; Sharma, Kim, & Gupta, 2022).

The initial step in this paper was data pre-

processing, which included cleaning, addressing

missing values, and normalizing the data. These

actions were performed to guarantee the accuracy of

the data and to reduce the noise's influence on the

model. Next, feature engineering was performed,

with the importance of data features being analyzed

and new features being created to enhance the model's

predictive capabilities. On the basis of this

framework, the model was developed, selecting

suitable machine learning algorithms and fine-tuning

the model's hyperparameters to maximize efficiency.

Subsequently, the model was evaluated using

cross-validation and other techniques to assess its

accuracy and generalization ability, ensuring stable

performance across different datasets. When the

problem is finally solved using optimization

technology, the model's performance is further

enhanced, and the test set's ultimate accuracy rate is

92.6%.

2 METHODS

In this paper, a comprehensive approach to data prep

aration and model development was undertaken to cr

eate a flight delay prediction model. The procedure s

tarted with extensive data pre-processing, which incl

uded vital operations including data normalization, d

ata cleansing, and management of missing variables.

These steps were essential to ensure the integrity and

quality of the data, as well as to minimize the impac

t of noise and inconsistencies on the model's perform

ance. Following this, an extensive feature engineerin

g phase was carried out. This phase included the anal

ysis of feature importance and the creation of new, m

eaningful features that could capture the underlying

patterns in the data more effectively. The enhanced f

eature set significantly improved the model's predicti

ve capability by providing richer and more relevant i

nformation. The procedures of paper is shown in Fig

ure 1.

Figure 1: Paper Procedures.(Picture credit : Original)

2.1 Data Overview

Table 1: Data Overview.

Name Type

Departure Airport String

Fli

ht Numbe

Strin

Scheduled De

arture Time DateTime

Scheduled Fli

ht Duration Float

Arrival Special Circumstance

String

Arrival Weathe

Strin

…… ……

Flight Delay

Prediction

Data collection and

rocessin

Weathe

Flight

informatio

Airpor

Other

Feature en

ineerin

Modelin

Model evaluation Model

DAML 2024 - International Conference on Data Analysis and Machine Learning

The collected data are shown in Table 1. These d

ata above encompass various flight-related informati

on, including airport details, flight numbers, schedul

ed times, aircraft information, and weather condition

s. This information is critical for predicting flight del

ays. Factors such as scheduled departure time, flight

month, departure weather, and arrival weather can he

lp identify potential causes of delays and predict pos

sible delay scenarios. This enables the implementatio

n of appropriate measures to mitigate the impact of fl

ight delays on operations.

2.2 Data Processing

During the data processing phase, data related to flig

hts, weather, city-airport mapping and special cases

needs to be processed. The main processing steps inc

luded filling in missing values to ensure data integrit

y; grouping and sorting the data to accurately calcula

te prior delays and takeoff intervals, which are crucia

l for subsequent delay analysis; refining the timing o

f special circumstances data to precisely match it wit

h flight data, thereby enhancing the model's ability to

identify causes of delays; and formatting and catego

rizing weather data, including labeling extreme weat

her conditions, to help the model better understand th

e impact of weather on flight delays. These steps not

only cleaned and optimized the data but also contribu

ted to building an efficient and accurate predictive m

odel. Aggregate data are shown in Table 2.

2.3 Feature Engineering

In this paper, feature engineering is a crucial step tha

t significantly impacts the effectiveness and perform

ance of the model's training. This step of encoding ca

tegorical data allows the model to efficiently process

non-numerical information, such as airport codes an

d weather conditions. By converting these classificat

ion labels into values that can be interpreted by the m

odel, the integrity and consistency of the input data t

o the model are guaranteed. This process is essential

for maintaining a balanced weight among different fe

atures in the model, especially when dealing with dis

tance-based machine learning algorithms. By guaran

teeing that every feature has the same weight, norma

lization keeps any one feature from unduly impactin

g the result of model training.

2.4 Model Construction

The MLP is a supervised learning algorithm designe

d to predict outcomes or classify data by emulating t

he structure and functioning of the human brain's neu

ral network. An input layer is taken the data, one or

more hidden layers process it, and an output layer pr

oduces the final predictions. The MLP is a sort of fee

dforward artificial neural network made up of numer

ous layers of nodes (Heidari et al., 2019). Each node,

or neuron, in the network is connected to the neuron

s in adjacent layers by weighted connections, which

determine the strength and influence of the signals p

assing through them.

The principles are as follows:

 Forward Propagation: Data in MLP moves from t

he input layer to the output layer via the hidden l

ayers. After multiplying the input by the appropri

ate weight and adding a bias term, each neuron p

rocesses the outcome through an activation funct

ion to provide an output.

 Activation Function: The role of the activation fu

nction is to introduce non-linear characteristics to

the network. Without them, the neural network

would be unable to capture complex data pattern

s. Activation functions like ReLU, Sigmoid, and

Tanh are frequently utilized (Oostwal, Straat, &

Biehl, 2021).

 Back-propagation and Gradient Descent: When t

raining an MLP, the back-propagation algorithm

is used to update the weights and biases in the ne

twork. This process involves calculating the grad

ient of the loss function with respect to each weig

ht and using gradient descent to iteratively adjust

the weights to minimize the error.

Table 2: Display of integrated data.

Depcode Arrcode Flightno Planned departu

re time

Planned time of

arrival

Planned fli

ght time

PlannedDe

time

PlannedAr

rtime

KWE KHN GY7113 1496271600 1496277900 1.750000 23 0

HGH URC 3U8953 1496273500 1496299800 7.583333 23 6

TAO WNZ SC4731 1496273400 1496280300 1.916667 23 1

LXA BPX TV9849 1496273400 1496277900 1.250000 23 0

TAO SZX SC4731 1496273400 1496289600 4.500000 23 4

Machine Learning Application: Flight Delay Prediction

3 EXPERIMENTAL RESULT

3.1 Data Splitting

To fairly assess the model's performance on unidenti

fied data, the data set is initially split into training an

d test sets. 25% of the data is utilized as the test set t

o gauge the model's capacity for generalization, whil

e the remaining 75% is randomly allocated to the trai

ning set for model training and parameter tuning. Thi

s department contributes to the impartiality and depe

ndability of test findings.

3.2 Model Training

The MLP was selected as the classification model du

e to its suitability for handling data with complex no

nlinear relationships. The MLP configuration include

s two hidden layers, each with 50 neurons, and uses t

he ReLU activation function, which helps avoid the

vanishing gradient problem, thereby making the train

ing process more stable. The model uses Stochastic

Gradient Descent as the optimizer and applies L2 reg

ularization (alpha=1e-4) to prevent overfitting.

3.3 Model Assessment

A detailed evaluation of the MLP model's performan

ce was conducted. The accuracy of the model on the

test set was calculated, with the MLP method measur

ing the proportion of correctly predicted samples, ref

lecting the prediction validity of the model. Furtherm

ore, to gain a comprehensive understanding of the m

odel's training process and structural details, several

key metrics were analyzed: the model's total number

of layers, the number of iterations, the ultimate loss,

and the output layer's activation function. This infor

mation provided insights into the model's internal me

chanisms, allowed for the assessment of training effi

ciency, and offered crucial data to support further mo

del optimization.

During the training of MLP model, a significant i

ssue was that the loss value remained consistently hi

gh, even after multiple iterations, without showing a

notable decrease. This situation may indicate that the

model struggled to fit the training data effectively. T

he persistently high loss value could be due to the m

odel's structure being unsuitable for handling specifi

c data features or an improper learning rate setting le

ading to inefficient optimization. Additionally, altho

ugh the accuracy reached 92.6%, the high loss value

may suggest that the model's predictive performance

is unstable on certain samples. The optimization resu

lts are shown in Table 3.

3.4 Output Result

The MLP model developed for flight delay predictio

n achieved a significant level of accuracy, with a fina

l accuracy rate of 92.6% on the test set. Despite this

high accuracy, the model encountered issues during t

raining, particularly with the loss value. The loss re

mained persistently high across multiple iterations, i

ndicating that the model struggled to fit the training

data effectively. This issue suggests that either the m

odel's structure was not well-suited to handling certai

n data features, or that the learning rate was not opti

mally configured, leading to inefficient optimization.

The high loss value also implies that the model’s pre

dictive performance might be unstable for certain sa

mples, raising concerns about its robustness.

In response to these challenges, the model under

went a detailed optimization process. The ReLU acti

vation function was employed to address the vanishi

ng gradient problem, which is common in deep neura

l networks and can hinder learning efficiency. The o

ptimization also included the use of Stochastic Gradi

ent Descent (SGD) with a learning rate of 0.01, striki

ng a balance between learning speed and stability. Fu

rthermore, an early halting mechanism was added to

avoid overfitting and improve the model's capacity to

generalize to new data.

Table 3: Optimization Outcome.

Fli

htno Fli

htDe

code Fli

htArrcode PlannedDe

time PlannedArrtime

rob

GY7113 KWE KHN 23 0 1

3U8953 HGH URC 23 6 1

SC4731 TAO WNZ 23 1 1

TV9849 LXA BPX 23 0 1

SC4731 TAO SZX 23 4 1

CZ6591A SZX NGB 23 1 0

GX8873 NNG TAO 23 4 1

GX8873 NNG XFN 23 1 1

3U8946 INC CTU 23 1 1

MU2392 INC XIY 23 0 1

DAML 2024 - International Conference on Data Analysis and Machine Learning

After these optimizations, the model showed imp

roved performance. The loss value decreased signific

antly, reaching an average of 0.13 after 63 iterations,

and the validation accuracy also increased, reaching

a score of 95.79%. This indicates that the optimizatio

ns were successful in stabilizing the training process

and improving the model’s predictive accuracy. The

final model, with its reduced loss and enhanced valid

ation accuracy, is more robust and better suited to pr

edict flight delays accurately.

However, despite these improvements, the initial

issue with high loss values highlights the importance

of careful model design and parameter tuning, partic

ularly when dealing with complex datasets. The succ

ess of the optimization process also underscores the

need for ongoing refinement and testing to ensure th

e model’s stability and reliability in various operatio

nal conditions. The final results are shown in Table

Table 4: Results of testing dataset

Validation score 0.957987

Loss 0.131940

Accurac

rate 0.926278

4 CONCLUSIONS

This study utilized an optimized Multi-Layer Percept

ron model to effectively predict flight delays, achievi

ng a 92.6% accuracy by handling complex nonlinear

data relationships. The model's success was attribute

d to techniques such as learning rate adjustment and

early stopping mechanisms. Looking ahead, integrati

ng more complex architectures like ResNet and lever

aging larger, more diverse datasets are expected to fu

rther enhance prediction reliability, improving operat

ional efficiency and passenger experience.

REFERENCES

Jiang, Y., Miao, J., Zhang, X., & Le, N. (2020, October 1).

A multi-index prediction method for flight delay base

d on long short-term memory network model. IEEE X

plore. https://doi.org/10.1109/ICCASIT50869.2020.93

68554

Yu, B., Guo, Z., Asian, S., Wang, H., & Chen, G. (2019).

Flight delay prediction for commercial air transport: A

deep learning approach. Transportation Research Part

E: Logistics and Transportation Review, 125, 203–221.

https://doi.org/10.1016/j.tre.2019.03.013

Song, C., Guo, J., & Zhuang, J. (2020). Analyzing passeng

ers’ emotions following flight delays- a 2011–2019 cas

e study on SKYTRAX comments. Journal of Air Trans

port Management, 89, 101903. https://doi.org/10.1016/

j.jairtraman.2020.101903

Zhu, X., & Li, L. (2021). Flight time prediction for fuel loa

ding decisions with a deep learning approach. Transpo

rtation Research Part C: Emerging Technologies, 128,

103179. https://doi.org/10.1016/j.trc.2021.103179

Shi, T., Lai, J., Gu, R., & Wei, Z. (2021). An Improved Ar

tificial Neural Network Model for Flights Delay Predic

tion. International Journal of Pattern Recognition and

Artificial Intelligence, 35(08), 2159027. https://doi.org

/10.1142/s0218001421590278

Yazdi, M. F., Kamel, S. R., Chabok, S. J. M., & Kheirabad

i, M. (2020). Flight delay prediction based on deep lear

ning and Levenberg-Marquart algorithm. Journal of Bi

g Data, 7(1). https://doi.org/10.1186/s40537-020-0038

0-z

Kalyani, N. L., Jeshmitha, G., Sai U., B. S., Samanvitha,

M., Mahesh, J., & Kiranmayee, B. V. (2020, October

1). Machine Learning Model - based Prediction of Flig

ht Delay. IEEE Xplore. https://doi.org/10.1109/I-SMA

C49090.2020.9243339

Kruse, R., Sanaz Mostaghim, Borgelt, C., Braune, C., & St

einbrecher, M. (2022). Multi-layer Perceptrons. Texts i

n Computer Science, 53–124. https://doi.org/10.1007/9

78-3-030-42227-1_5

Al Bataineh, A., Kaur, D., & Jalali, S. M. J. (2022). Multi-

Layer Perceptron Training Optimization Using Nature

Inspired Computing. IEEE Access, 10, 36963–36977.

https://doi.org/10.1109/access.2022.3164669

Sharma, R., Kim, M., & Gupta, A. (2022). Motor imagery

classification in brain-machine interface with machine

learning algorithms: Classical approach to multi-layer

perceptron model. Biomedical Signal Processing and C

ontrol, 71, 103101. https://doi.org/10.1016/j.bspc.2021.

103101

Heidari, A. A., Faris, H., Mirjalili, S., Aljarah, I., & Mafarj

a, M. (2019). Ant Lion Optimizer: Theory, Literature R

eview, and Application in Multi-layer Perceptron Neur

al Networks. Nature-Inspired Optimizers, 23–46. https:

//doi.org/10.1007/978-3-030-12127-3_3

Oostwal, E., Straat, M., & Biehl, M. (2021). Hidden unit s

pecialization in layered neural networks: ReLU vs. sig

moidal activation. Physica A: Statistical Mechanics an

d Its Applications, 564, 125517. https://doi.org/10.101

6/j.physa.2020.125517

Machine Learning Application: Flight Delay Prediction