Predicting arrival times via machine learning to improve inland container transport planning

(1)

1

Predicting arrival times via machine learning to improve inland container

transport planning

Author: K. Sprenkels (s2386003)

Institution: Rijksuniversiteit Groningen

Program: Technology and operations management (TOM)

Course: Master thesis TOM

1

st

_{assessor: Dr. ir. S. Fazi}

2nd_{assessor: Dr. N. D. Foreest}

(2)

2 Abstract

Purpose: The growth in volume of containers handled by sea port has put increasing pressure on the performance of inland transport. It is acknowledged that this segment can be more efficient by improving transport planning. Hence, predicting container return times at inland terminals can be beneficial to improve container allocation to the available capacitated means of transport. This thesis helps to uncover the factors that influence container return times contributing to efficient inland container transport planning.

Originality/value: Literature of inland terminals within the global container systems is rather new. To our knowledge return times at inland terminals are used in available models, but are typically a known information. No work so far has investigated on the value of such advanced information and how to acquire it.

Design/methodology/approach: This research follows the knowledge discovery in databases methodology on a container flow database of an inland terminal, using machine learning algorithms as data mining tool. Linear regression, neural network, random forest and K-nearest neighbors algorithms have been used to determine both feature (i.e., input variable) importance and predictive performance. In addition, the suitability of a selection of machine learning models for predicting container return times has been benchmarked. This research has used a design science research approach, because it proposes an artefact in order to solve a practical problem. In addition, the findings will be validated using cross-validation and an interview.

(3)

3

Preface

I would first and foremost like to extend a big thanks to my first assessor Stefano Fazi for his clear and constructive feedback. Secondly, I would like to thank Michel van Dijk for his hospitality at the ITV and for making the data used in this thesis available for research purposes.

I would also like to thank Hylke Donker for lending me his expertise on the technicalities of machine learning research and likewise I would like to thank Andrew Ng for tutoring me on the field of machine learning through his free online platform coursera.org.

(4)

4

1. Introduction

The globalisation of supply chains is ever increasing and growing at almost three times of the world GDP growth (UN-ESCAP 2005). This growth is partly due to the containerisation of global trade, which eases the handling and transportation of trade goods. The global container demand in 2017 is an estimated 148 million Twenty-foot equivalent Unit (TEU) (UNCTAD, 2018) and has grown considerably in the past decades. Along with this, the amount of inland container transportation has been growing considerably. High capacity transport has been implemented to alleviate these drawbacks from the growth of container flows in the hinterland (Konings et al., 2013).

The companies providing container transport are the shipping lines. The shipping lines sell transport directly to companies needing transport. Often though, the containers need transportation from the company to the seaport. The shipping lines cannot always provide inland transport so it is contracted to logistic small and medium-sized enterprises (SMEs). Inland transport can be performed using trucks, barges or trains and is often multimodal. Inland container transportation leans on inland terminals to change modality and to reach the sea ports with direct connections. The inland container transport system is thoroughly described by Fazi (2014).

Considering the inland transit of an import container, originating at the sea port, containers preferably travel by high capacity transport (barge or rail) from the seaport to an inland terminal. At the inland terminal, the container is either transported to another inland terminal, or if the terminal is nearby the final destination, the container is trucked to the local companies. After loading/unloading, these containers return to the inland terminal where they depart to the seaport. This research will focus on predicting the container return times using supervised learning algorithms. Container return time is defined in this document as the difference between the gate-out and the gate-in time of the container within the inland terminal. It would be beneficial for future planning if planners know the container return time in advance.

(6)

6

on arrival time prediction using machine learning methodologies with high predictive reliability probabilities (Lin et al., 2013) (Wang et al., 2018) (Barbour et al., 2018). Therefore, predicting container return times using machine learning methodologies could result in enhanced inland terminal planning.

The aim is twofold. Firstly, we will identify factors that influence containers return times, which could provide insights for planners at inland terminals to make more efficient and reliable planning decisions, leading to reduced costs. For example, if it is found certain variables such as time of the day, day of the week, container type and container weight lead to longer container return times, this can be taken into account with planning of the container on (a later) outgoing barge or train. On the other hand, if it is found that a certain company is very fast given certain conditions, planners can decide to let the truck wait at the company and immediately take the container back to the inland terminal. Furthermore, more reliable planning decisions will increase the occupancy of high capacity transport and reduce the need for overbooking, thereby reducing costs and container waiting time and increasing reliability. Secondly, this thesis will benchmark the accuracy of multiple supervised (machine) learning algorithms against each other and the average return times per company.

This will be achieved by using the knowledge discovery in databases (KDD) process (Fayyad et al., 1996). As support for this study, data of processed containers of an inland terminal in the Brabant region are have been used. The four research questions of this paper will be:

1. Can knowledge of container return time be beneficial for container planning? 2. Which methodology can be used to better predict return times?

3. What are the main features that help predicting container return times? 4. What do the results imply for inland terminal management?

(7)

7

2. Theoretical background

This chapter presents a review of literature relevant to the topic of inland terminal operations. The first section is an introduction of the global container supply chain. The second section describes the role of inland container transport within the global container supply chain. The third section summarises the research done on inland terminals, while the fourth section describes the role of time in container systems. Lastly, in the fifth section, the contributions of this thesis to the literature are outlined.

2.1 Global container supply chain

Three main subsystems can be identified within the container logistics chain: ocean transport, sea terminals and inland transport as visualized in figure 1 (Fazi, 2014).

Figure 1: Visual representation of intercontinental shipment of a container. The red flows represent inland transportation which van be over river, road or rail. (Fazi, 2014)

The globalization of supply chains is ever increasing and growing at almost three times of the world GDP growth (UN-ESCAP 2005). The global container demand in 2017 volumes to an estimated 148 million Twenty-foot Equivalent Unit (TEU) (UNCTAD, 2018) and has grown considerably in the past decades as shown in figure 2.

(8)

8

The international container liner shipping industry market structure can best be described as oligopolistic as shown in figure 3, with the top 20 liners sharing over 90% of total US trade in 2006 (Sys, 2009).

Figure 3: Market share of top ten container operators (Source: The Wall Street Journal Dec 1 2016)

The biggest liner, Maersk, has a market share of 17,8% a capacity of 4,175,795 TEU, 698 Ships (Alfaliner, 2019) and approximately 80.220 employees (Maersk annual report, 2018)

(9)

9

The shipping liners sell shipping directly to the companies and subsequently buy pre-carriage from the logistic SMEs running the inland terminals. The size of the shipping liners results in a difficult negotiating position and subsequently results in low profit margins for the logistic SMEs running the inland terminals.

2.2 Inland transport

The amount of inland transportation has been increasing in the last decades. Due to road congestion, environmental concerns, and traffic safety, high capacity transport has been implemented to alleviate these drawbacks from the growth of container flows in the hinterland (Konings et al., 2013).

This high capacity transport comes in the form of container barges on rivers and freight trains. Because most companies cannot receive these ways of high capacity transport, a modality change from high capacity transport to low capacity transport (i.e. road transport) is in place within the hinterland transport chain. This modality change is done within inland terminals.

Considering this case, as a container enters the seaport it is transported mostly by container barge to an inland terminal. The capacity of the container barges depends on the waterways used, inland terminal Veghel (ITV) handles barges from 28 to 104 TEU, but the lower Mississippi allows up to 36 barge units per tow carrying up to 36*96=3456 TUE (US Coast guard, 2019). The barge size is not only dependent on the size of the waterway but also on the distance travelled. On longer distances it’s easier to leverage economies of scale and the time constraints imposed by the customers are looser.

Inland terminals might also have a rail connection and use freight trains. Inland terminals store and distribute containers among the local companies, where they get loaded or unloaded and returned. This system in visualized in figure 3.

(10)

10

2.3 The role of inland terminals

Value propositions of inland terminals are its function as a satellite terminal, a load centre and a trans modal centre (Rodrigue et al., 2010). Enabling the triangulation of empty containers and leveraging the economies of scale in intermodal transport (van den Berg et al., 2015). Furthermore, container transport by inland waterways are considerably more environmentally friendly than container transport by roads (Wiercx et al., 2018). These advantages are acknowledged by Gołębiowski (2016) who points out the low energy consumption, low costs and low number of accidents, but acknowledging the slow transport speed as its main weakness. Gołębiowski (2016) also compares other external costs (i.e. costs of air pollution, noise costs and costs of accidents) of the three branches of transport: water, road and rail as presented in table 1.

Table 1: External costs of the tree branches of transport Gołębiowski (2016)

Mathematical modelling seems to be most common methodology in research on inland terminals. Fazi (2014) investigated mode selection, routing and scheduling for inland container transport. Bhattacharya et al. (2013) used an extended version of the California algorithm (Payne and Tignor, 1978) to develop an integrated intermodal freight transport system for intermodal transport cost optimization. They tested the behaviour of the formulated network by simulating different scenarios in Matlab. Mathematical modelling and simulation methods were also used by Wiercx et al. (2019) to determine the optimal configuration and operation of Reach stackers and terminal Yard cranes to contribute to more sustainable terminals and better inland waterway transport.

(11)

11

Wiegmans et al. (2014) used regression analysis to determine inland port performance and in 2017 Stochastic Frontier Analysis as introduced by Aigner et al. (1977) was used by Wiegmans and Witte (2017) for analysing the efficiency of inland terminals.

2.4 The value of time in inland transport

Containers circulating in the global supply chain are generally owned by the shipping lines, and are in itself not a very valuable item. A new 40ft container may cost around $2500-$2800 (Veenstra, 2015), but a shipping line may own several million of them. No records are kept for the total amount of containers globally, but According to Drewry Maritime Research, the global container fleet accounted for about 32.9 million TEU in 2012 (World Shipping Council, 2019).

To optimize the utilization of the containers ocean carriers have devised a penalty system to stimulate short return times. If the container stays at the seaport too long it is said to be in “demurrage”. If the container is not returned to the seaport within time it is said to be in “detention”. The number of free days and the penalties are negotiable (Veenstra, 2015). According to Storm (2011) Maersk handles a free demurrage time of 3 days and a free detention time of 5 days when transported by barge. A daily demurrage penalty of €45 for the first 7 days and € 75 for the remaining days and a daily detention penalty of €35. These costs seem steep considering that renting a refurbished 40ft container is possible for as little as €2 per day (bdcontainers.com 2019).

A demurrage and detention costs minimization model is provided by Fazi and Roodbergen (2018) using a generalized bin packing formulation. Similarly to the model proposed by Fazi et. al. (2015), This study models container return times based on average truck travel distance. Not considering other container features.

Optimizing inland terminal operations has therefore, not only the opportunity to increase customer satisfaction trough faster delivery times, but also directly influences the logistic SMEs bottom line through reducing demurrage and detention (D&D) costs and by optimizing barge loading.

(12)

12

Machine learning methodologies work well for predictive analytics, multiple studies have been conducted aiming to predict arrival times using machine learning methodologies. Lin et al. (2013) used artificial neural network models to predict real-time bus arrivals based on GPS data and fare collection system data. They identify the key factors that affect bus arrival times and create a model with a predictive reliability probability of more than 85%. Wang et al (2018) introduce a hybrid model to address short-term trajectory prediction in terminal manoeuvring by application of machine learning methods, classifying estimated time of arrival (ETA) prediction with more than 95% accuracy. Lastly, Barbour et al. (2018) used support vector analysis for ETA prediction of freight traffic on US railroads and observe improvements exceeding 20% over baseline methods. They also include an analysis of feature importance and conclude that the relative importance of predictive factors is location specific.

Shen et al. (2019) conducted a study to optimally reposition empty containers, considering free detention time and liner carrier cooperation. They used CPLEX solver to get the optimal results. CPLEX is a software package solving linear integer programming problems, being similar to linear regression algorithms.

2.5 The contribution of this thesis

This research takes the perspective of planners within logistic SMEs operating inland terminals. Therefore, only the data available for these SMEs will be used in this research. This research will only consider the part of the hinterland container transport chain between the inland terminal and the local companies as visualized in figure 5.

Figure 5: Visualisation of the system boundary: The red rectangle is the system boundary of this research

Planners at the inland terminal plan the high capacity transport between the seaport and the inland terminal and have complete control of the times and routes. Next to their own trucking service for the transport between the inland terminal and the local companies, the ITV uses external trucking companies. They have limited control on the times and routes of these trucks. Furthermore, they lack control of the time the containers are at the local companies.

(13)

13

the suitability of a selection of these methods for predicting the container return time. This would give the SMEs running inland terminals a reasonable estimation of the variability of the parts of this system where they lack control.

Learning algorithms are known for their ability to predict based on past data. Predicting arrival times can help planners make scheduling decisions, thus optimizing inland terminal operations. This research aims to help logistic SMEs improve scheduling decisions by predicting the return times of the containers to inland terminals based on past data using supervised learning algorithms.

(14)

14

3. Methodology

This chapter will elaborate on the research design, KDD process, the data requirements and an overview and motivation for the chosen algorithms.

3.1 Research design

It is challenging to categorise the subject of this research as it draws from both container logistics literature as data science literature, thereby revealing a research gap. The main research objective is to create an artefact solving the problem of predicting container return times and benchmarking the methods used. Therefore, a design science research (DSR) approach is applicable (Wieringa, 2009; Hevner, 2007). In order to create the artefact, data has to be drawn from a case. Therefore, a case study is embedded within this design science research.

The general structure of a problem-solving process using design science research is the regulative cycle (Wieringa, 2009), which is summarised as follows: It stars by investigation of a practical problem, it specifies specific solution designs, it validates these designs and lastly implements the solutions, thereafter the cycle repeats. This cycle is visually represented in figure 6. The regulative cycle consists of four steps: (1) system description, (2) system analysis, (3) proposed design, (4) testing and validating proposed designs. This research broadly follows these steps.

Figure 6: The regulative cycle (Wieringa 2009)

(15)

15

period 01-09-2016 and 31-10-2016. The qualitative data was gathered by the means of two semi-structured interviews. To make a selection of algorithms a machine learning specialist was interviewed on which algorithms are most likely to be suitable. To validate the findings cross-validation (Goodfellow, 2016) was used and the director logistics of the ITV was interviewed on what features are likely to have predictive value for the algorithms.

3.2 The knowledge discovery in databases process

Data mining is the analysis step of the knowledge discovery in databases process (Fayyad 1996). The process of extracting information and knowledge hidden in data with approaches such as modelling and rule generation using computer technologies, especially artificial intelligence methods such as machine learning algorithms (Shao et al., 2008). The purpose of data mining is to use the discovered knowledge to assist decision-making (Shao et al., 2008). Waller et al. (2013) predict that data science, predictive analytics and big data (DPB) will revolutionize the field of Supply Chain Management (SCM) and state that data scientists also need deep domain knowledge and analytical skills.

The field of datamining has been in a remarkable and ongoing growth which has been fuelled by a confluence of factors: the explosive growth of data collection, the storing of data, the access to data, competitive pressure in a globalized economy, the development of commercial data mining software suites and, the tremendous growth in computing power and storage capacity(Larose & Larose, 2015). The limiting factor not being the absence of data but the skills to translate this data into knowledge (Larose & Larose, 2015), summarised by the now famous quote from John Naisbitt from his 1984 book Megatrends: “We are drowning in information but starved for knowledge”.

Knowledge discovery in databases is achieved by using methods, algorithms and other techniques to extract useful information from data. This process can be divided into nine steps as illustrated in figure 7 (Fayyad et al., 1996).

1. Develop an understanding of the application domain and identify the goal of the KDD process. 2. Select data on which discovery is to be performed

3. Clean and pre-process data

4. Data reduction and projection for finding the useful features to represent the data depending on the goal of the task

5. Matching the goal of the KDD process to a particular data-mining method. 6. Exploratory analysis and model hypothesis selection.

7. Data mining: searching for patterns of interest. 8. Interpreting mined patterns.

(16)

16

Figure 7: The steps that compose the knowledge discovery process (Fayyad et al. 1996)

Note that KDD is an iterative process and can contain loops between any two steps.

The goal of the KDD process in this setting is to predict the container return times. The data on which this discovery is to be performed is the “container statistics MCA for ITV in period 01-09-2016 to 31-10-2016” dataset. This database is from Multimodaal Coördinatie- en adviescentrum (MCA) Brabant containing the container statistics of inland terminal Veghel (ITV). In this context these are the first two steps of the KDD process.

3.3 Data requirements

Due to advances in computer technology, we can currently store and process large amounts of data. Most data acquisition devices are digital and record reliable data. Nevertheless, stored data becomes only useful if it is analysed and turned into useful information to, for example, make predictions (Alppaydin, 2010).

3.3.1 Database size

(17)

17

3.3.2 Making features numerical

Data contains examples (the records in the dataset) composed of features. A feature is an individual measurable property or characteristic of a phenomenon being observed, and an example is a single instance of the logged features (Bishop, 2006). Features are usually numeric, but structural features such as strings and graphs are also found in data. As machine learning algorithms cannot process these structural features they should be converted into numerical data. This can be done using dummy variables where the string is converted to multiple binary features (often called one-hot-encoding).

3.3.3 Missing data

Many datasets contain missing or erroneously entered data fields. Two techniques for dealing with missing data are proposed by Lakshminarayan et al. (1996). The first, Autoclass, is a program developed by Cheeseman et al. (1988) to automatically discover clusters in data and is based on Bayesian classification theory (Hanson et al., 1990). The second, C4.5, is a supervised learning algorithm for decision tree induction developed by Quinlan (1993). The performance of these techniques varies and will not be used as this is out of the scope of this research. Instead, the features and examples containing missing data will be removed from the dataset.

3.3.4 Outliers

Outliers can be difficult to handle as some of them can be caused by measurement errors, while others may represent phenomena of interest (Liu et al., 2002). Outliers can most easily be detected using data visualisation such as graphs and boxplots. Comprehensive articles on outlier detection are provided by Su and Tsai (2011) and Rousseeuw and Hubert (2011).

Outliers can have a substantial negative effect on the performance of machine learning algorithms, because machine learning algorithms search for generalizations within the data. If the outliers are a result of erroneous data they should be removed. Depending on the algorithm, for example in linear regression, it may be better to remove outliers from the dataset and explain them separately.

3.3.5 Feature scaling

(18)

18

3.3.6 Splitting the dataset

For testing of the accuracy of machine learning algorithms the dataset should be split in two different distinct datasets. The training/validation set and the test set. This ensures an unbiased indication of the models’ performance; this is called cross-validation (Goodfellow, 2016). The training set is used to train the algorithm, and the validation set to optimize the model’s parameters. When using cross-validation, the part that is used for training and the part that is used for validation are permuted. After training and tuning of the parameters the algorithms accuracy is obtained using the test set (Cichosz, 2015). This split is visualized in figure 7.

Figure 7: Visualisation of the splits of the dataset

The most straightforward way of splitting the dataset is randomly, using a random (but pre-determined) seed. The conventional ratio between the splits is given by Goodfellow (2016) and is approximately 80% and 20% for the training/validation set and test set respectively.

3.4 Choice of algorithms

Within the field of machine learning, two main tasks coexist: supervised and unsupervised learning. The difference between them is that supervised learning uses a ground truth, meaning that knowledge of the output values is available within the training and validation data set. Therefore, the algorithm approximates the relation between the input and output observable in the data (Mohri et al., 2018). Within this research setting output values are available within the training data; therefore, supervised learning algorithms are used. This research will use regression algorithms opposed to classification algorithms as the results of regression algorithms are more easily interpretable.

(19)

19

The choice is made to use a linear regression analysis as the results are relatively well interpretable. The coefficients of a normal linear regression algorithms where too large to be interpretable, therefore a ridge regression algorithm was used. This algorithm penalizes large coefficients, thus making the coefficients better interpretable and preventing overfitting. Secondly, an Elasticnet linear regression algorithm was used. This algorithm sets small coefficients to 0. Preventing overfitting on insignificant variables, thereby increasing the prediction accuracy.

In order to choose which algorithms to benchmark machine learning specialist dr. H. C. Donker was interviewed. He suggested to use a neural network and random forest algorithms as these algorithms are broadly applicable and generally give good results. For bus arrival time prediction Lin et al. (2013) similarly proposes neural network models, which they propose as one of the most efficient methods for bus arrival time predictions. Wang (2018) predicts arrival times using a multi-cell neural network. For this research a MLP regressor will be used as it is the simplest type of neural network.

Random forest performs very well compared to many other algorithms (Liaw & Wiener, 2002) and is more robust with respect to noise (Pal, 2005). Lastly, the K-nearest neighbors method will be used as this algorithm followed from a flowchart proposed by Q. Lanners (2019) on towardsdatascience.com.

3.5 Algorithm description and chosen hyperparameters

To get a valid comparison between the algorithms used all algorithms have the same split between the train/validation set and the test set. Therefore, all the algorithms are trained on the same data, and the predictive performance is well comparable. The test set is 20% of the complete dataset containing 842 individual instances on which the models are tested. Hyperparameters are the parameters whose value is set before the learning process begins.

In total five algorithms were tested.

- A ridge regressor, which is a linear regression which penalizes high coefficients to prevent overfitting and to make the coefficients better interpretable.

(20)

20

- A Multi-layered perceptron (MLP) regressor, which works as a neural network outputting one specific value. Through trial and error, it was found that 5 hidden layers with each 500 nodes gave good results.

- A random forest regressor, which is a collection of decision trees containing random values to ensure different splits for each decision tree (Rebala, 2019). The average value of all decision trees is outputted as the answer. In this case 10000 decision trees were used, and the dataset was split down to a binsize of 1. The computational time was quite long and the result was only marginally better than the result for 1000 decision trees. It is advised to increase the binsize and decrease the number of decision trees when computating for larger datasets. - A K-nearest neighbors regressor, which finds the neirest neighbors of a random instance within

the multidimensional feature space, thereby categorizing the instances. In this case the algorithm searches for 20 nearest neighbors.

The code used for making the predictions can be found in appendix B. For all other hyperparameters the default values of the sci-kit modules were used. These can be found on scikit.org.

(21)

21

4. Results

This chapter will describe the results of the study. Firstly, the features, outliers and the algorithms with subsequent hyperparameters will be described. Secondly, to further understand the specifics of this case, a visual analysis is done on the container return times. Thirdly, the coefficients of the linear regression algorithms will be presented. Lastly, a benchmark of prediction accuracy of multiple algorithms is given.

4.1 Features and database structuring

This research will consider 7 features: - Day of the week

- Container departure time - Container type

- Container weight - Loading or unloading - Trucking company - Destination company

These features will be split into subsequent variables as presented in table 1. Most variables are binary (dummy variables) with the exception of container weight. For the container weight variable feature scaling is applied, meaning that the container weight of each instance is divided by the container weight of the heaviest container of the dataset (30.980 kg).

These features are used as they are available within the database. More features can be added if available, which will be addressed in the discussion. For privacy purposes the trucking companies and destination companies will not be named, and the estimated travel time to the companies could not be included as company addresses were unavailable.

(22)

22

of variables will also have a positive effect on computation time, as computation time scales exponentially with the number of variables.

4.2 Expectations based on interview

An interview was conducted in order to be able to validate the results. When the management of the ITV was asked what were the most important factors determining container return time the most important variable would be the will of the customer and the container load, for example: if the container is loaded with boxes or pallets could make a profound difference in the loading/unloading speed. Regrettably these variables are not available within the database.

The day of the week was expected to have not much influence on the container return time with the exception of Friday, as many containers will not return till Monday. It was also mentioned that for individual destination companies the day of the week might have an influence on the container return time due to the availability of labour within the destination companies. The container departure time was expected to be very important as containers leaving in the late afternoon and evening would be likely to return the following day. The container type and container weight were expected to have a small influence, as these variables could give an indication on the type of load. It was expected that if the container was loaded or unloaded and the type of trucking company would have little influence on the container return time. Lastly, the destination company was expected to have a mayor influence on the container return time due to the differences in distance to the inland terminal, differences of the type of load for each company and the differences in loading/unloading speed of the individual destination companies.

The destination company addresses where requested, but could not be given due to privacy concerns. Therefore, the distances to the inland terminal could not be included in this research. Nevertheless, most destination companies are located close to the inland terminal as this case is subject to Dutch infrastructure.

4.3 Visual analysis

(23)

23

distribution per day (appendix A) clearly shows that most containers return on the same day as when they departed. Some containers arrive the day after, except if the container departed on Friday, then these containers arrive 3 days after, because containers only sparsely travel within the weekends.

Graph 1: Container return times of the complete dataset.

4.4 Linear regression coefficients

The coefficients of the variables are given in table 1. These coefficients should be interpreted as follows: The algorithm predicts the container return times based on the variables. The predicted container return time is the intercept value + the coefficient of each binary variable + the weight coeffient*the relative weight. For example: If a full 40HC container weighing 50% of the heaviest container in the dataset departs on Monday between 6:00 and 9:00 to company 1 with trucking company 1 the predicted return time of the Ridge regression with removed outliers is 0,712 - 0,205 - 0,356 + 0,122 – 0,223 – 0,5*0,348 + 0,036 + 0,176 = 0,088 days = 2,11 hours.

Note that due to the linearity of these algorithms the return time prediction can be negative and these predictions are not that accurate. Nevertheless, these coefficients are well interpretable. Negative coefficients suggest that containers subject to that variable are faster and positive coefficients suggest that containers subject to that variable are slower.

0 0,2 0,4 0,6 0,8 1 1,2 0 100 200 300 400 500 600 700 800 900 1 6 ₁₁ ₁₆ ₂₁ ₂₆ ₃₁ ₃₆ ₄₁ ₄₆ ₅₁ ₅₆ ₆₁ ₆₆ ₇₁ ₇₆ ₈₁ ₈₆ ₉₁ ₉₆ 101 106 111 116 121 Am o u n t

Return time (Hours)

Containter return time

(24)

24

Table 1: Overview of the variable properties and subsequent linear regression coefficients. The first 4 columns outline the structure of the dataset. The instances column represents the number of instances subject to the variable are found within the dataset. The average return time column represents the average return time of these instances. The last three columns show the coefficients found in the three experiments. Negative coefficients suggest that containers subject to that variable are faster and positive coefficients suggest that containers subject to that variable are slower. The averages are in days and the coefficients can be interpreted as such.

Features

Variable

names Variable meaning Variable type Instances

Average return time (days) Ridge regression coefficients (outliers not removed) Ridge regression coefficients (Outliers removed) ElasticNet Regression coefficients (Outliers removed) Total amount = 4896 Average return time complete dataset = 0,5804813 Intercept value = 0,997935 Intercept value = 0,7122235 Intercept value = 0,5583418

day of the week D1 Monday binary 1087 0,405 -0,358 -0,205 -0,101

D2 Tuesday binary 982 0,412 -0,334 -0,151 -0,054

D3 Wednesday binary 948 0,466 -0,271 -0,094 0,000

D4 Thursday binary 932 0,491 -0,301 -0,091 0,000

D5 Friday binary 925 1,167 0,284 0,540 0,600

D6 Saturday binary 22 1,710 0,980 N/A N/A

D7 Sunday binary 0 N/A N/A N/A N/A

Container

departure time T1 0:00-6:00 binary 308 0,370 -0,359 -0,370 -0,271

T2 6:00-9:00 binary 1561 0,304 -0,348 -0,356 -0,310

T3 9:00-12:00 binary 1285 0,448 -0,194 -0,166 -0,136

T4 12:00-15:00 binary 1028 0,836 0,138 0,188 0,189

T5 15:00 - 18:00 binary 612 1,199 0,380 0,425 0,415

T6 18:00 - 24:00 binary 69 1,254 0,384 0,278 0,114

Container type CT1 40HC binary 2616 0,545 0,031 0,122 0,000

CT2 20DV binary 684 0,539 -0,038 0,097 0,000

CT3 40RH binary 570 0,438 -0,381 -0,273 -0,059

CT4 40DV binary 415 0,587 -0,017 0,104 0,064

CT5 45RH binary 315 _1,488 _-0,162 _-0,060 _0,230

CT6 45HC binary 260 0,218 -0,017 0,011 -0,201

CT7 20RF binary 28 0,174 -0,532 N/A N/A

CT8 20TK binary 8 3,134 1,116 N/A N/A

Loading/Unloading L1 Loading binary 1931 0,590 0,143 0,223 0,157

L2 Unloading binary 2965 _0,582 _-0,143 _-0,223 _-0,156

container weight W container weight

Continious

between 0 and 1 2965 0,582 -0,160 -0,348 -0,122

Trucking company TC1 Largest volume binary 3493 0,613 -0,051 0,036 0,074

TC2 binary 307 0,428 0,066 0,307 -0,056 TC3 binary 279 0,454 -0,172 -0,081 0,000 TC4 binary 239 0,147 -0,456 -0,379 -0,369 TC5 binary 147 0,890 -0,137 -0,040 0,037 TC6 binary 87 0,452 -0,215 -0,215 -0,053 TC7 binary 79 0,572 -0,015 -0,016 0,000 TC8 binary 60 0,484 -0,092 0,021 0,000 TC9 binary 58 0,644 -0,026 0,063 0,000 TC10 binary 40 0,940 -0,031 0,124 0,000 TC11 binary 37 0,550 0,114 0,181 0,067

TC12 binary 21 0,304 -0,253 N/A N/A

TC13 binary 14 0,300 -0,108 N/A N/A

TC14 binary 12 0,661 0,015 N/A N/A

TC15 binary 10 1,083 -0,118 N/A N/A

TC16 binary 7 1,918 0,681 N/A N/A

TC17 binary 5 3,962 0,835 N/A N/A

TC18 Smallest volume binary 1 0,097 -0,036 N/A N/A

Destination

Company C1 Largest volume binary

465 0,918 0,174 0,176 0,207 C2 binary 426 0,730 0,009 -0,064 0,000 C3 binary 309 0,770 0,250 0,183 0,172 C4 binary 301 0,868 0,306 0,256 0,000 C5 binary 270 _0,048 _-0,370 _-0,397 _-0,295 C6 binary 243 0,399 -0,327 -0,465 -0,005 C7 binary 171 0,097 -0,306 -0,265 -0,062 C8 binary 155 1,967 1,208 1,075 0,582 C9 binary 145 0,304 -0,409 -0,503 -0,253 C10 binary 144 _1,168 _0,818 _0,797 _0,393

(25)

25

4.5 Prediction accuracy benchmark

The common way for benchmarking machine learning algorithm accuracy is by comparing the mean absolute error, the mean squared error and the root mean squared error. These key figures are presented in table 2. As a benchmark the average over the complete dataset is also presented. In this case the average return time of the complete dataset (0,58 days) is used as a prediction on every instance in the test set. For comparability purposes, all of the algorithms have the same split in the train/validation set and the test set. Therefore, the algorithms are trained and tested on the same instances. The size of the test set is 20% of the total dataset which is 842 instances for the database where the outliers are removed and 979 instances for the database where the outliers are not removed.

Table 2: key figures on predication accuracy for each algorithm type.

Table 2 shows how removing the outliers has limited influence on the prediction accuracy for linear regression algorithms.

From table 2 follows that the random forest regressor is the most accurate algorithm within this setting. Nevertheless, this comparison gives an overly simplistic view on algorithm accuracy, as it only considers the mean errors, and fails to give information about the accuracy of the individual predictions. Therefore, another presentation is provided in graph 2 and table 3.

Algorithm type

Mean absolute error (days)

Mean squared error

Root mean square error Average over complete

dataset 0,597 0,814 0,902

Ridge regressor (outliers

not removed) 0,434 0,559 0,747

Ridge regressor (outliers

removed) 0,440 0,548 0,740 ElasticNet Regressor (Outliers removed) 0,438 0,573 0,757 MLP regressor (outliers removed) 500,500,500,500,500 0,291 0,567 0,753

Random Forest Regressor 0,280 0,481 0,693

K-nearest neighbors

(26)

26

Graph 2 shows the percentage of predictions within a certain error. From this graph it can be concluded that the Multi-layered perceptron (MLP) regressor has the highest accuracy, although when the error is large it is on average larger than the error of the random forest regressor.

Graph 2 also clearly outlines the differences in prediction accuracy of linear regression algorithm versus more complex algorithms, suggesting that within this case the interrelationships between the variables are very important.

(27)

27

Table 3 is added as the graph is badly readable for the part where the error is very large. Note that some predictions exceed an error of 36 hours, these are mostly instances where the actual return time is very large, which no of the algorithms could predict.

(28)

28

5. Discussion

This chapter will discuss the results of the study. Firstly, actions that can be taken to improve predictive performance will be outlined. Next, the implications for inland terminal management will be discussed. Lastly, the limitations of this research and suggestions for future research are discussed.

5.1 Limitations of this research

Even though this research uses about 5000 datapoints, this data is actually quite limited as it only includes 2 months of data. Previous researchers using similar methods use bigger datasets (Lin et al., 2013; Wang et al., 2018; Barbour et al., 2018; Shen et al., 2019). It is likely that predictive performance will increase if more relevant data is inputted into the models. It is known that at least 3 years of quality data is available within the ITV. Secondly, the number of input variables could be increased, suggestions for possible relevant input variables are:

- If 3 years of data is available, a variable for seasonality could be introduced. For example: a variable for departure month could be added.

- A variable for which planner planned the container could be added. This data was available within the used dataset, but was unused due to privacy concerns. The limited number of planners employed by the ITV meant the data could not be anonymized.

- A variable for expected truck travel time could be added. This data was not available for this research, because due to privacy concerns the addresses of all of the destination companies could not be provided.

- A variable for which truck transported the container could be added.

- Another opportunity is to add local weather data. This data can be found on the site of the KNMI and could be added into the model. A limitation though is that in practice this weather data will have to be compared to weather predictions, as the weather data is not yet available while making a planning. Therefore, adding this variable might be more interesting for research purposes than for practical purposes.

(29)

29

in this model are multiplied this would create 6*6= 36 variables instead of 6+6=12 variables. This could also be done with destination company and container type.

This research makes no use of hyperparameter optimization tools (Bengio, 2000; Hutter, 2015) as this was beyond the scope of this research. Furthermore, due to the limited amount of data there was no need to optimize the calculation time of the algorithms as reflected by the choice of hyperparameters. If the amount of data is increased, and the models become more complex and the calculation time takes longer, hyperparameter optimization tools might prove important. Choosing the right/optimal hyperparameters will have an overall positive effect on algorithm performance and calculation time.

In this research the choice is made to remove all the variables containing more than 30 instances due them being outliers. Removing these outliers improved performance slightly, nevertheless the threshold of 30 instances was chosen arbitrarily, more complex ways to determine outliers are presented by Rebala (2019). Among the outliers removed were 9 instances with negative return time and 26 instances with a return time of 0. These instances are likely due to erroneous data and were easily spotted. However, it is unknown how many erroneous instances where not noticed and removed. Erroneous data has a large negative effect on predictive performance; therefore, it is paramount that the data used is of high quality.

The coefficients found using the linear regression analysis may be a strong indicator but cannot be totally conclusive due the possibility of unexpected correlative effects (e.g. the coefficients found for container weight may be incorrectly interpreted that heavy containers are transported faster, whereas it is more likely that full containers are transported faster and full containers are always heavier than empty containers). These effects become more prevalent as the features become more interdependent, and can never totally be negated. Therefore, these coefficients should not be interpreted too rashly.

(30)

30

purposes.”(1981) (e.g. The planners might cheat the system to improve their perceived performance by disproportionally valuating container return times)

The most important limitation of the complex algorithms is that due to their complexity it is virtually uncontrollable how these algorithms find their solutions. Therefore, it is difficult to make conclusive statements on the actual patters these algorithm, nevertheless recent developments in explainable AI (Artificial intelligence) have made it possible to quantify the feature contributions for individual predictions (Lundberg et al., 2020; Štrumbelj, 2014).

Lastly, because this research draws data from a case that is subject to Dutch infrastructure, the actual container return times are generally fast. Predicting container return times can potentially be even more valuable in cases where the container return times are longer, such as in rural areas. It is unclear how much the predictive performance can be increased, but it is likely that there will be an upper limit where the predictions cannot be further improved due to the natural noise within the system.

5.2 Future research

To gain understanding of the potential value for inland terminal planning and the upper limit of predictive performance a research is required using as much relevant data as possible. The accuracy of these predictions should be compared with the predictions of experienced planners.

(31)

31

learning, this research is the tip of the iceberg. Among classification algorithms, many different modelling methodologies are available.

Currently, the hinterland container supply chain is modelled in the literature mostly using mathematical models. These models use assumptions and generalisations for container travel- and arrival times (Fazi, 2015). These models can be extended using machine learning research based on cases to further account for the complexities of the hinterland container supply chain.

This research predicts container return times as it was assumed that this would be the most useful variable to predict for inland terminal management as this time is an unknown for inland terminal management and sufficient data was available. However, other labels can be chosen given sufficient available data. For example: Using the same methodology container waiting time, container arrival time, barge travel time, barge utilization, expected detention costs or the size of the container flow within a certain timeframe can be predicted. Predicting the other metrics could be valuable for both practical and research purposes.

Lastly, this research further confirms that machine learning methodologies are an invaluable tool in the toolbox of an operations researcher and supply chain manager, as Waller et al. predicted in 2013, and can be applied across disciplines.

5.3 Prediction in practice

As the hinterland supply chain system is constantly changing, updating the algorithm to include recent instances is very important if implemented in practice. In this research the dataset is pre-processed manually. However, if standardized, this could be done automatically by the use of Python scripts, Excel macros or a range of dedicated tools such as TensorFlow and PyTorch. Automation makes it possible to update the model daily. It is also advised that an extra weight is added to more recent instances, making recent instances more important for the algorithm.

(32)

32

discernible from the used dataset. Nevertheless, for the logistic SMEs running the inland terminals reliably predicting which containers would qualify could improve the bottom line.

(33)

33

6. Conclusion

In this research, multiple models are proposed to predict container return times within the hinterland container supply chain. Linear regression algorithms can be used for system analysis purposes due to their relative simplicity. The more complicated models could find fruition within practice as their predictive performance is good considering the limited amount of data used.

Most findings from the results of the linear regression analyses where in line with expectations, such as a high penalty for departing on Friday or in the evening, which is consistent among the different experiments. Furthermore, the predictive performance was tested using cross-validation (Goodfellow, 2016) therefore, these results will be perceived as valid. More remarkable model inclinations are the overall negative penalty on reefer (i.e. refrigerated) containers (40RH, 45RH, 20 RF), suggesting these containers get priority within this system. It is hypothesized that this is due to both the likelihood of these containers containing fresh fast-moving products and the containers themselves being more valuable.

Most remarkable are the models’ inclinations towards containers that depart while full. This difference might be explained by full containers getting priority and containers generally getting unloaded faster than they are getting loaded, However, the coefficients found by the algorithm are very large. The difference in predictions between full and empty departing containers is consistently more than 0,25 days = 6 hours, not considering container weight. These findings are in conflict with the expectations based on the interview, therefore it is likely that an unaccounted-for effect is observed. Moreover, due to this reason the found coefficient for container weight might not be actually representative of container weight.

The found coefficients on trucking company can generally give an indication of performance. The coefficients found for the destination company mostly involves the container handling speed of the destination company. The actual distance to the company is likely only a small contributing factor as most of the companies are located near to the inland terminal.

(34)

34

7. References

1. Aigner, D., Lovell, C., & Schmidt, P. (1977). Formulation and estimation of stochastic frontier production function models. Journal of Econometrics, 6(1), 21-37. doi:10.1016/0304-4076(77)90052-5

2. Alpaydin, E. (2009). Introduction to machine learning. MIT press.

3. Alphaliner "Alphaliner – Top 100 Operated Fleets As Per 20 october 2019". Retrieved 20 October 2019 from: https://alphaliner.axsmarine.com/PublicTop100/index.php

4. B. Rambali, L. Baert, D. Thoné & D. L. Massart (2001) Using Experimental Design to Optimize the Process Parameters in Fluidized Bed Granulation, Drug Development and Industrial Pharmacy, 27:1, 47-55, DOI: 10.1081/DDC-100000127

5. Barbour, W., Martinez Mori, J. C., Kuppa, S. and Work, D. B. (2018) “Prediction of Arrival Times of Freight Traffic on Us Railroads Using Support Vector Regression,” Transportation Research Part C, 93, pp. 211– 227. doi: 10.1016/j.trc.2018.05.019.

6. Bengio Y (2000) “Gradient-Based Optimization of Hyperparameters,” Neural computation, 12(8), pp. 1889–900.

7. Bishop, C. M. (2006). Pattern recognition and machine learning. springer.

8. Cai J, Shen L, Jin Z and 2019 IEEE International Conference on Systems, Man and Cybernetics, SMC 2019 2019 10 06 - 2019 10 09 (2019) “Empty Container Repositioning with Consideration of Free Detention Time and Liner Carrier Cooperation,” Conference Proceedings - IEEE International Conference on

Systems, Man and Cybernetics, 2019-october, pp. 2097–2102. doi: 10.1109/SMC.2019.8914419.

9. Cheeseman, P., Kelly, J., Self, M., Stutz, J., Taylor, W., and Freeman, D. 1988. Bayesian Classification. In Proceedings of American Association of Artificial Intelligence(AAAI), 607-611, San Mateo, California:Morgan Kaufmann Publishers Inc.

10. Cichosz, P. (2015). Data mining algorithms : Explained using r. Chichester, West Sussex: John Wiley & Sons. (2015). Retrieved October 8, 2019

11. Cost container renting, retrieved on 20-10-2019 from: https://bdcontainers.com/diensten/zeecontainer-huren/

12. Craighead, C., Hanna, J., Gibson, B., & Meredith, J. (2007). Research approaches in logistics: Trends and alternative future directions. The International Journal of Logistics Management, 18(1), 22-40.

13. Dengiz, O., Smith, A. E. and Nettleship, I. (2006) “Two-Stage Data Mining for Flaw Identification in Ceramics Manufacture,” International Journal of Production Research, 44(14), pp. 2839–2851. doi: 10.1080/00207540500534454.

14. Dietvorst BJ, Simmons JP and Massey C (2015) “Algorithm Aversion: People Erroneously Avoid Algorithms After Seeing Them Err,” Journal of experimental psychology. General, 144(1), pp. 114–26. doi: 10.1037/xge0000033.

(35)

35

16. Fazi S (2014) Mode selection, routing and scheduling for inland container transport. PhD thesis, TU Eindhoven, Beta Research School, 2014

17. Fazi, S., & Roodbergen, K. (2018). Effects of demurrage and detention regimes on dry-port-based inland container transport. Transportation Research. Part C: Emerging Technologies, 89, 1-18.

18. Fazi, S., Fransoo, J., & Van Woensel, T. (2015). A decision support system tool for the transportation by barge of import containers: A case study. Decision Support Systems, 79, 33-45.

19. Feng J.C.-X & Kusiak, A. (2006) “Data Mining Applications in Engineering Design, Manufacturing and Logistics,” International Journal of Production Research, 44(14), pp. 2689–2694.

20. Friedman, J. H. (1998). Data mining and statistics: What's the connection?. Computing Science and

Statistics, 29(1), 3-9.

21. Gołębiowski, C. (2016). Inland water transport in poland. Transportation Research Procedia, 14, 223-232. doi:10.1016/j.trpro.2016.05.058

22. Goodfellow, I., Bengio, Y., & Courville, A. (2016). Deep learning. MIT press.

23. Goodhart, Charles (1981). "Problems of Monetary Management: The U.K. Experience". In Courakis, Anthony S. (ed.). Inflation, Depression, and Economic Policy in the West. pp. 111–146. ISBN 978-0-389-20144-1.

24. Hanson, R., Stutz, J. and Cheeseman, P. (1990). Bayesian Classification Theory, Technical Report, (FIA-90-12-7-01), NASA, Ames.

25. Hastie, T., Tibshirani, R., & Friedman, J. (2009). The elements of statistical learning : Data mining,

inference, and prediction (Second ed., Springer series in statistics). New York: Springer.

26. Hevner, A. R. (2007). A three cycle view of design science research. Scandinavian journal of information

systems, 19(2), 4.

27. https://scikit-learn.org/stable/tutorial/machine_learning_map/index.html

28. Hutter, F., Lücke Jörg and Schmidt-Thieme, L. (2015) “Beyond Manual Tuning of Hyperparameters,” KI -

Künstliche Intelligenz : German Journal on Artificial Intelligence - Organ des Fachbereichs "Künstliche Intelligenz" der Gesellschaft für Informatik e.V, 29(4), pp. 329–337. doi: 10.1007/s13218-015-0381-0.

29. Lakshminarayan, K., Harp, S. A., Goldman, R. P., & Samad, T. (1996, August). Imputation of Missing Data Using Machine Learning Techniques. In KDD (pp. 140-145).

30. Lanners Q. (2019). Choosing a Scikit-learn Linear Regression Algorithm. Retrieved from (11-1-2020): https://towardsdatascience.com/choosing-a-scikit-learn-linear-regression-algorithm-dd96b48105f5 31. Larose, D. T. and Larose, C. D. (2015) Data mining and predictive analytics. Second edn. Hoboken, New

Jersey: Wiley (Wiley series on methods and applications in data mining)

32. Li, L., Negenborn, R., & De Schutter, B. (2015). Intermodal freight transport planning - a receding horizon control approach. Transportation Research Part C, 60, 77-95. doi:10.1016/j.trc.2015.08.002

33. Liaw, A., & Wiener, M. (2002). Classification and regression by randomForest. R news, 2(3), 18-22. 34. Lin Y, Zou N, Jia L and Yang X (2013) “Real-Time Bus Arrival Time Prediction: Case Study for Jinan,

(36)

36

35. Liu, X., Cheng, G., & Wu, J. (2002). Analyzing outliers cautiously. Ieee Transactions on Knowledge and

Data Engineering, 14, 432-437.

36. Lundberg, S. M., Erion, G., Chen, H., DeGrave, A., Prutkin, J. M., Nair, B., ... & Lee, S. I. From local explanations to global understanding with explainable AI for trees.

37. Malik, M. M., Abdallah, S. and Ala’raj, M. (2018) “Data Mining and Predictive Analytics Applications for the Delivery of Healthcare Services: A Systematic Literature Review,” Annals of Operations Research, 270(1-2), pp. 287–312. doi: 10.1007/s10479-016-2393-z.

38. Matthew Steward, March 24 2019, The actual difference between statistics and machine learning, , Retrieved (26-9-2019) from: https://towardsdatascience.com/the-actual-difference-between-statistics-and-machine-learning-64b49f07ea3

39. Mohri, M., Rostamizadeh, A., & Talwalkar, A. (2018). Foundations of machine learning (Second ed., Adaptive computation and machine learning) [Second edition.]. Cambridge, Massachusetts: MIT Press. (2018). Retrieved September 4, 2019

40. Møller - Mærsk "Group Annual Report 2018" (PDF). A.P.. Retrieved on 20-10-2019 from: https://investor.maersk.com/static-files/9295e4f7-97f8-4ea7-a503-fe716bf99bef

41. Murphy, K. (2012). Machine learning : A probabilistic perspective (Adaptive computation and machine learning series). Cambridge, Mass.: MIT Press. (2012). Retrieved September 26, 2019,

42. Naisbitt, J. (1984) Megatrends : ten new directions transforming our lives. 1St [updated] paperback edn. New York: Warner Books.

43. Nielsen, M. A. (2015). Neural networks and deep learning (Vol. 2018). San Francisco, CA, USA:: Determination press.

44. Pal, M. (2005). Random forest classifier for remote sensing classification. International Journal of

Remote Sensing, 26(1), 217-222.

45. Pyne, H.J. Tignor, S.C., 1978. Freeway incident detection algorithms based on decision tree with states. Transportation Research Record 682, 378-382

46. Quinlan, J.R., 1993. C4.5 Programs For Machine Learning, San Mateo, California: Morgan Kaufmann Publishers Inc.

47. R. Konings, E. Kreutzberger, and V. Maraˇs. Major considerations in developing a huband-spoke network to improve the cost performance of container barge transport in the hinterland: the case of the port of rotterdam. Journal of Transport Geography, 29:63–73, 2013

48. Rebala, G., Ravi, A., & Churiwala, S. (2019). An introduction to machine learning. Cham: Springer. doi:10.1007/978-3-030-15729-6

49. Rodrigue, J., & Notteboom, T. (2009). The terminalization of supply chains: Reassessing the role of terminals in port/hinterland logistical relationships. Maritime Policy & Management, 36(2), 165-183. doi:10.1080/03088830902861086

(37)

37

51. Rousseeuw, P., & Hubert, M. (2011). Robust statistics for outlier detection: Robust statistics for outlier detection. Wiley Interdisciplinary Reviews: Data Mining and Knowledge Discovery, 1(1), 73-79. doi:10.1002/widm.2

52. Shao, Q., Rowe, R. C. and York, P. (2008) “Data Mining of Fractured Experimental Data Using Neurofuzzy Logic-Discovering and Integrating Knowledge Hidden in Multiple Formulation Databases for a Fluid-Ded Granulation Process,” 97(6), pp. 2091–2101. doi: 10.1002/jps.21098.

53. Storm, R. (2011). Controlling container demurrage and detention through information sharing (thesis). Technical report, Rotterdam School of Management, Rotterdam.

54. Štrumbelj, E., & Kononenko, I. (2014). Explaining prediction models and individual predictions with

feature contributions. Knowledge and information systems, 41(3), 647-665.

55. Su, X., & Tsai, C. (2011). Outlier detection. Wiley Interdisciplinary Reviews: Data Mining and Knowledge

Discovery, 1(3), 261-268. doi:10.1002/widm.19

56. Sys, C., & The International Forum on Shipping Ports and Airports (IFSPA) 2008. (2009). Is the container liner shipping industry an oligopoly? Transport Policy, 16(5), 259-270. doi:10.1016/j.tranpol.2009.08.003

57. Taylor, E. L. (2017) “Making Sense of ‘algorithm Aversion,’” Research World, 2017(64), pp. 57–57. doi: 10.1002/rwm3.20528.

58. The wall street journal Dec 1 2016, Maersk Line to Buy German Shipping Line Hamburg Süd in $4 Billion Deal, retrieved from (20-10-2019): https://www.wsj.com/articles/maersk-line-buys-german-container-shipper-hamburg-sud-1480590315

59. Tseng, T.-L., Huang, C.-C., Jiang, F. and Ho, J. C. (2006) “Applying a Hybrid Data-Mining Approach to Prediction Problems: A Case of Preferred Suppliers Prediction,” International Journal of Production

Research, 44(14), pp. 2935–2954. doi: 10.1080/00207540600654525.

60. Tukey, J. (1962). The future of data analysis. The Annals of Mathematical Statistics, 33(1), 1-67. 61. UNCTAD (United Nations Conference on Trade and Development). 2018. Review of Maritime Transport

2018. United Nations, New York and Geneva.

62. UN-ESCAP (United Nations Economic and Social Commission for Asia and the Pacific). 2005. Regional Shipping and Port Development Strategies (Container Traffic Forecast), Report ST/ESCAP/ 2398. United Nations, New York

63. United States Coast Guard, Local notice to mariners Mississippi river system, light list volume V, notice number 11-19, March 13, 2019

64. Van den Berg, R., & De Langen, P. (2015). Towards an 'inland terminal centred' value proposition. Maritime Policy and Management, 42(5), 499-515.

65. Veenstra A. Ocean transport and the facilitation of trade C.Y. Lee, Q. Meng (Eds.), Handbook of Ocean Container Transport Logistics: Making Global Supply Chains Effective, Springer (2015), pp. 429-450 66. Voss, C., Johnson, M. & Godsell, J., 2016. Case Research. In C. Karlsson, ed. Research Methods for

(38)

38

67. Waller, M. A. and Fawcett, S. E. (2013) “Data Science, Predictive Analytics, and Big Data: A Revolution That Will Transform Supply Chain Design and Management,” Journal of Business Logistics, 34(2), pp. 77– 84. doi: 10.1111/jbl.12010.

68. Wang, Z., Liang, M. and Delahaye, D. (2018) “A Hybrid Machine Learning Model for Short-Term Estimated Time of Arrival Prediction in Terminal Manoeuvring Area,” Transportation Research Part C, 95, pp. 280–294. doi: 10.1016/j.trc.2018.07.019.

69. Wiegmans, B., & Witte, P. (2017). Efficiency of inland waterway container terminals: Stochastic frontier and data envelopment analysis to analyze the capacity design- and throughput efficiency. Transportation Research Part A, 106, 12-21. doi:10.1016/j.tra.2017.09.007

70. Wiegmans, B., Witte, P., & Spit, T. (2015). Inland port performance: A statistical analysis of dutch inland ports. Transportation Research Procedia, 8, 145-154. doi:10.1016/j.trpro.2015.06.050

71. Wiercx, M., Van Kalmthout, M., & Wiegmans, B. (2019). Inland waterway terminal yard configuration contributing to sustainability: Modeling yard operations. Research in Transportation Economics, 73. 72. Wieringa, R. (2009). Design Science as Nested Problem Solving. International Conference on Design

Science Research in Information Systems and Technology, 1–12. https://doi.org/0.1145/1555619.1555630

73. World Shipping Council. (2019). Global container fleet. Retrieved from http://www.worldshipping.org/ about-the-industry/containers/global-container-fleet

74. Zijm, W., Klumpp, M., Clausen, U., & Ten Hompel, M. (Eds.). (2016). Logistics and supply chain innovation

: Bridging the gap between theory and practice (Lecture notes in logistics). Cham: Springer.

(39)

39

Appendix A: Graphs of return time distribution per variable

Day of the week

Graph 3: Return time distribution per day Departure time window

Graph 4: Return time distribution per departure time window 0 0,05 0,1 0,15 0,2 0,25 1 6 ₁₁ ₁₆ ₂₁ ₂₆ ₃₁ ₃₆ ₄₁ ₄₆ ₅₁ ₅₆ ₆₁ ₆₆ ₇₁ ₇₆ ₈₁ ₈₆ ₉₁ ₉₆ 101 106 111 116 121 Re lat iv e amou n t Hours

Return time distribution per day

Saturday Monday Tuesday Wednesday Thursday Friday

0 0,05 0,1 0,15 0,2 0,25 1 5 9 ₁₃ ₁₇ ₂₁ ₂₅ ₂₉ ₃₃ ₃₇ ₄₁ ₄₅ ₄₉ ₅₃ ₅₇ ₆₁ ₆₅ ₆₉ ₇₃ ₇₇ ₈₁ ₈₅ ₈₉ ₉₃ ₉₇ 101 105 109 113 117 121 Re lat iv e amou n t Hours

Distribution return times per departure time window

(40)

40 Container type

Graph 5: Return time distribution per container type Loading/unloading

Graph 6: Return time distribution for loading/unloading 0 0,05 0,1 0,15 0,2 0,25 0,3 0,35 0,4 0,45 1 5 9 ₁₃ ₁₇ ₂₁ ₂₅ ₂₉ ₃₃ ₃₇ ₄₁ ₄₅ ₄₉ ₅₃ ₅₇ ₆₁ ₆₅ ₆₉ ₇₃ ₇₇ ₈₁ ₈₅ ₈₉ ₉₃ ₉₇ 101 105 109 113 117 121 Re lat iv e amou n t Hours

Return time distribution per container type

40HC 20DV 40RH 40DV 45RH 45HC 20RF 20TK 0 0,02 0,04 0,06 0,08 0,1 0,12 0,14 0,16 0,18 0,2 1 5 9 ₁₃ ₁₇ ₂₁ ₂₅ ₂₉ ₃₃ ₃₇ ₄₁ ₄₅ ₄₉ ₅₃ ₅₇ ₆₁ ₆₅ ₆₉ ₇₃ ₇₇ ₈₁ ₈₅ ₈₉ ₉₃ ₉₇ 101 105 109 113 117 121 Re lat iv e amou n t hours

Return time distribution loading/unloading

(41)

41 Trucking company

Graph 7: Return time distribution for the 5 trucking companies with the largest volume

Destination company

Graph 8: Container volume per company 0 50 100 150 200 250 300 350 400 450 500 1 7 ₁₃ ₁₉ ₂₅ ₃₁ ₃₇ ₄₃ ₄₉ ₅₅ ₆₁ ₆₇ ₇₃ ₇₉ ₈₅ ₉₁ ₉₇ 103 109 115 121 127 133 139 145 151 amou n t o f co n ta in ers Company number

volume per company

(42)

42

Graph 9: Return time distribution for the 5 companies with the largest container volume 0 0,1 0,2 0,3 0,4 0,5 0,6 0,7 1 5 9 ₁₃ ₁₇ ₂₁ ₂₅ ₂₉ ₃₃ ₃₇ ₄₁ ₄₅ ₄₉ ₅₃ ₅₇ ₆₁ ₆₅ ₆₉ ₇₃ ₇₇ ₈₁ ₈₅ ₈₉ ₉₃ ₉₇ 101 105 109 113 117 121 Re lat iv e amou n t hours

Predicting arrival times via machine learning to improve inland container transport planning