Information Systems

A decision support system for improved resource planning and truck routing at logistic nodes

Alessandro Hill1 • Jürgen W. Böse1

Published online: 3 October 2016

� Springer Science+Business Media New York 2016

Abstract In this paper, we present an innovative decision

support system that simultaneously provides predictive

analytics to logistic nodes as well as to collaborating truck

companies. Logistic nodes, such as container terminals,

container depots or container loading facilities, face heavy

workloads through a large number of truck arrivals during

peak times. At the same time, truck companies suffer from

augmented waiting times. The proposed system provides

forecasted truck arrival rates to the nodes and predicted

truck gate waiting times at the nodes to the truck compa-

nies based on historical data, economic and environmental

impact factors. Based on the expected workloads, the node

personnel and machinery can be planned more efficiently.

Truck companies can adjust their route planning in order to

minimize waiting times. Consequently, both sides benefit

from reduced truck waiting times while reducing traffic

congestion and air pollution. We suggest a flexible cloud

based service that incorporates an advanced forecasting

engine based on artificial intelligence capable of providing

individual predictions for users on all planning levels. In a

case study we report forecasting results obtained for the

truck waiting times at an empty container depot using

artificial neural networks.

Keywords Decision support systems � Forecasting � Predictive analytics � Truck routing � Resource planning

1 Motivation

Recent numbers on cargo in industrialized countries show

that road based transport dominates the market. More

importantly, it will have a significant stake in the future

since its market share grows faster than for alternative

modes of transport such as for example railroad. Truck

freight exceeded rail freight by a factor of four with a total

of about 1700 billion ton kilometers in 2012 in the Euro-

pean Union [9]. An increase of 50 % leading to about

600 billion ton kilometers in 2030 is estimated only for

Germany [3]. Accordingly, truck deliveries and pick-ups at

logistic nodes [22] such as warehouses, container termi-

nals, freight stations, empty container depots and logistics

centers will further increase.

Truck arrivals at these nodes are typically followed by a

registration procedure at the gate before the subsequent

assignment of the truck to a loading area. Both, the number

of arrivals and the dispatching time can significantly vary

due to various impact factors. The arrivals depend on

highly stochastic business processes of the truck companies

that are associated with the node. Common causes for the

rise of the dispatching time are peak workloads related to

the truck arrivals, insufficient node resources, node-internal

process issues or external factors such as weather. As a

consequence of such delays, the truck waiting times (e.g.,

at the gate) notably increase. This results in a major drop of

the service level provided by the node. At the same time,

the complexity of operations planning at the node increases

in these periods of strongly fluctuating workload which is

likely to decrease efficiency [5, 19]. Thus, truck companies

as well as node operating companies both experience

notable disadvantages.

In order to mitigate the mentioned issues, this paper

describes a concept for a decision support system that is

& Alessandro Hill [email protected]

Jürgen W. Böse

[email protected]

1 Institute of Maritime Logistics, Hamburg University of

Technology, Am Schwarzenberg-Campus 4 (D),

21073 Hamburg, Germany

123

Inf Technol Manag (2017) 18:241–251

DOI 10.1007/s10799-016-0267-3

based on predictive analytics [8]. The presented iLoads 1

system concept essentially aims at supporting operational

planning and control at the truck companies as well as the

nodes by providing forecasted waiting times and truck

arrivals, respectively. This system is innovative since it

extends existing approaches that are currently used in

practice. The most elaborate systems that are used today,

provide either visual real-time gate waiting time informa-

tion through corresponding web cams or simply list trivial

historical information such as yesterday’s waiting time. In

contrast, we incorporate a forecasting engine based on

artificial intelligence to predict waiting times and truck

arrivals. Decision-makers on both sides benefit from real-

time high quality predictions that are tailored to their

individual information needs. The iLoads system is generic

in the sense that it can be implemented at various types of

logistic nodes independent from the precise service it offers

to its customers. The implementation of a system based on

artificial intelligence is motivated by the numerous diverse

dependencies of the highly volatile waiting times in the

described environment. The consideration of additional

external and internal impact factors at the logistic nodes is

crucial for the quality of the forecasts.

The contribution of this paper is twofold. On the one

hand side, we suggest the general concept of providing

forecast information to the actors in this logistic environ-

ment. We propose the application of a standard artificial

neural networks approach to incorporate relevant features.

On the other hand side, we describe the implementation of

such a system based on a real world application and pro-

vide corresponding results.

The remainder of the paper is organized as follows. In

Sect. 2 we describe the application domain and identify

main user types and associated business processes. The

resulting system requirements are presented in Sect. 3. In

Sect. 4 we give a description of the system architecture and

its interfaces, model components and data components

before concluding this work in Sect. 6.

2 Processes and decision support

The presented iLoads system concept aims at twofold

decision support to simultaneously increase process effi-

ciency at logistic nodes and truck companies. Therefore,

forecasting information is provided to the nodes for

improving internal resource planning and control as well as

to the truck companies and truckers to support their vehicle

routing and scheduling. The requirements regarding future

information are certainly different on both sides. Further-

more, we differentiate between system users according to

their function, such as management and operations, even if

working in the same company.

In Sect. 2.1, we identify the main user types addressed

by the iLoads system and highlight their information needs

for effective decision making. Subsequently, we explain

the relevant business processes at logistic nodes in

Sect. 2.2.

2.1 Basic truck handling process

The focus of the iLoads system concept is particularly on

logistic nodes as integral part of cargo transport networks.

Logistic nodes are handling and storage locations, as

defined in [10]. The offered logistic services include

transport system changeover, load carrier changeover, re-

packaging, short and long term storage. In this context, we

consider logistic nodes at which goods are dropped off or

picked up by trucks. Optionally, the nodes interface with

other transportation means such as ship and railroad.

Figure 1 shows the base process that can be identified at

these nodes. Its two main parts are the administrative truck

handling and the physical truck handling. The administra-

tive handling consists of an eventual waiting time that

occurs before the processing of the documents at the doc-

ument center, also called gate waiting time. This admin-

istrative task might require the driver to register at a desk

but may also be done using an electronic terminal which

typically reduces the time needed. The physical truck

handling process can be divided into intermediate waiting

times and loading or unloading operations. Since multiple

container loading or unloading operations are possible,

several intermediate waiting times might occur.

We define the truck waiting time as the sum of the

administrative waiting time plus the aggregated interme-

diate waiting times before loading and unloading opera-

tions during the physical handling as illustrated in the flow

chart in Fig. 1. We note that periods without physical

activity regarding truck, driver or cargo (e.g., document

processing) are frequently experienced as waiting time for

the truck company. Nevertheless, we follow the waiting

time definition in accordance with the legal situation in

Germany as follows. We consider waiting times as periods

in which the node is unproductive with respect to the

corresponding truck. That is, neither the truck is unloaded

or loaded nor is the corresponding order involved in any

administrative process. This matches the formal definition

of waiting times in major countries of the European Union

(see e.g., [6]). The overall dwell time is the sum of the

truck waiting times and the truck handling time which

corresponds to the total time that the truck spends from

queuing at the gate until its departure from the node’s site. 1 Intelligent logistic order arrival decision support.

242 Inf Technol Manag (2017) 18:241–251

123

2.2 User types and business processes

Regarding the companies which are in the focus of the

iLoads system we can basically distinguish two types of

users according to their information needs. Namely, the

operations planners in charge of the operations manage-

ment [25] on both sides, the logistic nodes and the truck

companies, respectively. We note that depending on the

company size and its organization the responsible per-

sonnel can vary in terms of the number of employees and

the task assignment. Furthermore, executive management

can benefit from forecasted workloads to understand

future trends and trigger strategic initiatives. The pre-

dicted data can also be utilized to feed further analytic

models [20].

2.2.1 Node operations planners

On the node side, the presented decision support system is

most useful to operations planners who benefit from truck

arrival forecasts by increased equipment and personnel

planning accuracy. More detailed, this includes assignment

of employees to shifts, deployment of machinery and usage

of policies on a tactical level. The ability to foresee oper-

ational events allows a more adaptive planning in general.

These planning tasks are typically done for a horizon from

1 day to 1 week.

Additionally, future truck arrival information generates

substantial value for better operations control. Ad hoc

decision support is achieved by responsive short term

forecasts which take into account events (e.g., traffic con-

gestions) or notable changes of influencing factors (e.g.,

weather) which were not present during operational

planning.

In daily business of logistic nodes, the truck arrival rate,

expressed by the number of trucks that arrive during a

specific time period (e.g., 1 hour), is frequently used as an

indicator for the workload. More accurate resource plan-

ning leads to reduced truck waiting times.

Additionally, order specific arrival rates restricted to the

truck type (e.g., light, heavy), the load carrier (e.g., con-

tainer, pallet) or the customer are of interest since they

yield a more detailed estimation of the corresponding

handling effort. Forecasted truck arrival information is

usually not utilized at the nodes. However, corresponding

historical data is frequently considered by the planners.

2.2.2 Truck company operations planners

The second main user type of the iLoads system is oper-

ations planning personnel in truck companies which is

responsible for planning and control of the vehicle fleet.

The forecasted waiting times can be used for truck routing,

truck scheduling and related tasks [4, 11]. Furthermore,

related inter-terminal traffic coordination [13] can benefit

from the information provided by the system. So-called

dispatchers can use forecasted truck waiting times to

improve the operational fleet management. This includes

the daily or weekly order assignment to trucks and drivers

followed by the route planning. In practice, it is of major

importance to plan the tours in accordance with existing

time window restrictions. Therefore, the periods in which a

truck has to visit a logistic node can depend on scheduled

appointment times of preceding and subsequent jobs. For

instance, the pick-up of an empty container has to happen

an appropriate time before the packing date agreed with the

customer, or, a truck might be unavailable during certain

periods due to previously planned trips.

Fig. 1 Base process at logistic nodes

Inf Technol Manag (2017) 18:241–251 243

123

As at logistic nodes, the operational planning at truck

companies is done from 1 day to 1 weak in advance. This

includes the assignment of the orders to the available

trucks, followed by the determination of the individual

tours. Ultimately, this induces the number of resources

(i.e., trucks and drivers) required to manage the workload.

Besides efficient resource planning, tour delays due to

truck waiting times can be reduced which increases punc-

tuality. Furthermore, the logistic nodes benefit from

smoothed peak workloads since the tour planner will try to

schedule truck arrivals within periods of low waiting time

(see Fig. 2).

Currently, only a few logistic nodes provide information

about truck waiting times to their customers [1]. One rea-

son might be the lack of digital information on actual truck

waiting times. In some cases, nodes offer a webcam service

to show the current situation at the gate [12]. Such visual

information can be used by dispatchers to get a rough idea

about current waiting time. Against this backdrop, it is not

surprising that most dispatchers do not anticipate truck

waiting times at all. However, today’s economic and eco-

logic damage caused by truck waiting times is consider-

able. Several approaches were undertaken to clarify the

general waiting time situation in major ports [15, 17, 21]).

A survey among more than 550 German logistics services

providers in 2012 revealed that in 50 % of the cases the

waiting times at warehouses exceed 1 h (Bundesminis-

terium für Verkehr und digitale Infrastruktur [7]).

In contrast to truck arrival rates, the calculation of

waiting times requires basic statistical compilation. Typi-

cally this is achieved by considering average hourly wait-

ing times which have to be derived from the individual

waiting times. In addition, auxiliary measures such as the

maximal, or minimal, hourly waiting time could be useful

in practice.

We note that the information needs of independent

truckers basically correspond to those of dispatchers.

Nevertheless, differences between both user groups exist

on the soft- and hardware level since the former truckers

are permanently on the road in contrast to dispatchers being

located in an office on site at the company. Independent

truckers, organized as one-man companies, basically use

the system as the truck companies. Since they are contin-

uously on the road, they particularly benefit from short

waiting times.

3 System requirements

In this section we describe the iLoads system requirements

in detail. Based on the embedding of the system into the

relevant business processes in Sect. 2.2 we provide the

essential functional requirements and define the necessary

data sources.

3.1 User requirements

The iLoads system has to efficiently provide different

views on the truck arrival data and likewise for the truck

waiting time. In the following we assume that historical

information as well as predicted information is provided by

the system.

3.1.1 Display and forecast horizon

As mentioned in Sect. 2, the use of the iLoads within the

different processes implies individual user needs. To pre-

sent the corresponding information in a meaningful way,

we suggest the inclusion of minimal and maximal truck

waiting times in addition to the average waiting times. The

estimation of relevant key performance indicators is left to

the user but could represent a practical extension. Another

main feature that has to be addressed is the option of

customizing the forecast horizons as well as the overall

timespan displayed which includes historical data. In this

regard, the user has to be able to clearly distinguish

between past and future data and adjust the corresponding

horizons individually. Naturally, this implies real-time

Fig. 2 Benefits of forecast information on truck waiting

times and arrival rates at logistic

nodes

244 Inf Technol Manag (2017) 18:241–251

123

reporting functionality. The forecasting horizon should

range from 30 min for operations control up to multiple

months for resource planning.

Furthermore, the operations manager has to be able to

adjust the forecast granularity. That is, constant time

periods, or buckets, in which the underlying arrivals are

summarized and the waiting times are averaged, respec-

tively. In practice, these buckets should comprise from

10 min to 1 month subject to the chosen forecast horizon.

3.1.2 Specific data display

Besides showing the general arrival data for all trucks more

specific views on the arrival rates and waiting times are

required for effective order and customer oriented decision

support. The mentioned information has to be given for

various predefined order types which are commonly cate-

gorized by the truck type, cargo type and the customer.

3.1.3 Threshold visualization

A meaningful visual presentation of the relevant informa-

tion should include clear signals to indicate action required

by the user. This can be accomplished by the use of tem-

perature schemes and traffic light systems which qualify

the recent or future situation. Corresponding thresholds

have to be defined by the planning personnel based on their

experience. Moreover, the forecast information can be

translated directly into best practices such as required

resource quantities or operational plans and strategies. In

both cases such an interpretation should be parameterizable

by the user and, therewith, flexible with respect to opera-

tional or strategic changes.

3.1.4 Accessibility

The different user types described in Sect. 2.2 access the

system for different purposes and in particular from their

individual workplaces. Certainly, they are all equipped

with a devices that have access to the information system.

Thus, the user interface has to be accessible to multiple

device types such as desktop computers, tablets and smart

phones using their corresponding operation systems.

3.1.5 Response time

Users with operative duties expect information in real time

within their commonly fast paced environment. In other

words the forecast has to be presented within seconds to

support ad hoc decision making. Even on a tactical level,

the information has to be provided continuously, whereas

the management has rather low requirement regarding the

systems response time.

3.1.6 Forecast accuracy

Regarding helpfulness, reliability and user acceptance of

the system, we require a certain minimum forecast quality.

Depending on the business, an accuracy within 25 % can

be acceptable from a practitioner’s point of view. We note

that one idea of the presented system is to outperform

straightforward approaches such as simple hourly or daily

performance averages through the incorporation of

sophisticated forecasting methods.

3.2 Data and sources

The presented system relies on sufficient input data to

produce satisfactory forecasts. Corresponding base infor-

mation is given by historical data which is typically hosted

by the logistic node. The suggested intelligent forecasting

methodology will furthermore utilize external information

to increase the forecasting quality. In both cases the system

has to adapt to the interfaces provided by the host systems

for the sources described below.

3.2.1 Historical data

The historical data which is relevant for the forecasts is

collected by the logistic node and commonly organized in

its Terminal Operating System (TOS). This typically con-

sists of work schedules, customer and order information

whereat we are mainly interested in operational data store

[16]. As a minimum requirement it should comprise actual

waiting times and arrivals. The following check points are

sufficient to collect this data (see Fig. 1). In the following

we describe the necessary time stamps.

• Truck arrival time: The point in time when the truck arrives at the site of the logistic node. Optical character

recognition (OCR) device are widely used to capture

these arrivals.

• Administrative wait start and end time: The truck arrival time followed by the time at which the driver

hands out information to the node’s administration,

either electronically or paper based. The order infor-

mation is commonly entered in the node’s TOS at the

time of registration and, therefore, available in an

internal database.

• Intermediate wait start and end times: The pairs of start and end times at which the truck arrives at a loading or

unloading area and when the actual physical process

starts. In practice, these waiting time can be negligible but

may also exceed the administrative waiting times notably.

Optionally, truck departure times, i.e., points in time

when the truck leaves the site, are useful to calculate the

overall dwell times. The actual waiting time corresponds to

Inf Technol Manag (2017) 18:241–251 245

123

the sum of the administrative waiting time plus the cumu-

lative intermediate waiting times as illustrated in Fig. 1. The

arrival rates are calculated by counting the number of arri-

vals during a specific period (e.g., each hour).

3.2.2 Forecasting parameters

A forecasting parameter corresponds to information that

has an impact on the time series that will be predicted.

Such a forecast supporting parameter is given as a time

series and is used to improve the forecast quality by

exploiting the presumed correlation. Therefore, we assume

that historical data for these predictor time series, or pre-

dictors, is available as well as some forecasted information.

Typical predictors with a trivial forecast are weekdays,

holidays and shift schedules whereas for example weather

conditions, traffic and economic indices need to be fore-

casted themselves first in order to be available. We cate-

gorize predictor series according to the place of collection

of the corresponding data as follows.

• Node-specific forecasting parameters This information is collected by the logistic node and encompasses

observations that originate from business and opera-

tions at the node. Examples are personnel and machin-

ery schedules and operational strategies (e.g.,

dispatching modes, storage policies).

• External forecasting parameters Third party data that describes the economic, environmental and traffic

related situation during the forecasting horizon that

has an influence on the forecasted time series; for

instance weather information (e.g., rain, thunderstorms,

frost, heatwaves) and traffic information (e.g., conges-

tions, travel speed) from urban traffic control systems.

Another way to categorize predictor data is to differ-

entiate between deterministic predictors and stochastic

predictors. These are on the one hand parameters that are

known even for future periods, such as the day of the week

or the holiday schedule. On the other hand information on

the future weather or personnel sick leaves is not known in

advance and could at most be integrated using forecasted

data itself. For more detailed information about forecasting

with multiple predictors we refer to [24].

4 System architecture and technologies

In this section we describe the iLoads system based on the

requirements defined in Sect. 3 such that it can be

embedded seamlessly into the business processes summa-

rized in Sect. 2. As typical for a decision support system

we introduce the corresponding model components con-

sisting of the user interfaces, the intelligent forecasting

module and the data components [23]. First, we present the

system architecture given in Fig. 3, followed by the dif-

ferent components.

4.1 User interfaces

The interface of the decision support system to the user is

web based and its functionality is twofold. User informa-

tion is transferred to the system at the time of the forecast

request. Conversely, the retrieved forecast is presented to

the user. To provide decision support, the latter incorpo-

rates statistical evaluation of the forecast data (e.g., average

values) and customizable time series that are derived from

the raw forecasting series (e.g., truck arrivals per time

unit). Furthermore, the user gets access to the complete

historical data. This is accomplished by a dynamic html-

based interface which can be accessed by every web-

browsing capable device. For this purpose, web application

frameworks are widely available which additionally facil-

itate database connectivity and provide security features as

well as developing tools. Applications that are tailored to

different operating systems (e.g., Android, iOS, Windows

Mobile) could optionally enhance the usability. Following

the requirements from Sect. 3, we differentiate between the

users at the logistic node and the users at the truck com-

panies, where the latter could either be multi-truck orga-

nizations or self-employed truck drivers (see Fig. 3).

4.1.1 Truck company

To support the dispatcher’s truck routing and scheduling,

two forecast horizons are offered. An intra-day setting

supports the dispatcher’s short term decisions that affect the

truck planning and control until the end of the day. Addi-

tionally, the planning of weekly operations is supported by a

second scenario. Each outlook optionally uses a specified

order type argument which is supported by the node (e.g.,

container drop-off, pick-up) to increase the forecast accu-

racy. In Fig. 4 two representations are given for an intra-day

waiting time forecast. The estimated general minimum,

average and maximum waiting times are depicted in a line

plot (left). A more intuitive temperature scheme shows the

expected waiting time evolution for the remaining hours of

the day based on a node dependent threshold setting (right).

As Fig. 4 shows historical data can be considered for dis-

playing and is presented to the user in the same fashion as the

forecasted data. A special case of a truck company with a

single employee is a self-employed truck driver.

4.1.2 Logistic node

The display of the truck arrival rates, which are related to

the expected workload for the node, is done similar to the

246 Inf Technol Manag (2017) 18:241–251

123

waiting times. At each node, a useful extension could take

care of the calculation of various resource demands based

on the truck arrivals to support the decision maker with

resource planning. For example a proposal for the number

of required container stackers or the needed interchange

staff could be presented to the planner based on the hourly

orders. Moreover, the deduction of key performance indi-

cators such as for instance average wait truck time and

daily order volumes can easily be added for reporting.

4.2 Forecasting module

The forecasting module is the cloud based heart of the

iLoads system and is designed as a hosted service. It

consists of the following components which are responsible

for calculating the forecasts as well as handling the input

data from the different sources, organizing the forecasted

information and preparing the raw data for its adequate

presentation to the user. Even though the presented iLoads

architecture is designed for a single node and its collabo-

rating trucking companies, the integration of multiple

nodes neither affects the forecasting module nor the gen-

eric structure of the system. Solely the corresponding

databases have to be adequately extended. For more

information on cloud-related literature we refer to [14].

4.2.1 Data handler

The data handler retrieves the input data that is needed to

perform the forecast. In addition to reading raw data from

external sources (e.g., historical data, forecasting parame-

ters) it writes the information to the forecast input database.

The update of the latter could be done based on the push or

pull principle. Most importantly, the latest forecast inputs

Fig. 3 The system architecture with its components and their

relationships

Fig. 4 Waiting time forecast representations. Minimal,

average and maximal waiting

times against actual (left) and

temperature scheme for point-

wise prediction (right)

Inf Technol Manag (2017) 18:241–251 247

123

are available just-in-time when a new forecast is calculated

by the forecasting engine. The integration of additional

data sources requires the modification of the retrieval

routines used within this component.

4.2.2 Forecasting engine

For the information system, we suggest a forecasting

engine that implements time series forecasting algorithms

based on artificial intelligence. The relevant time series are

the average waiting times and the number of truck arrivals

with respect to the chosen bucket (e.g., 1 h buckets).

Interval based forecasting information, i.e., claiming

that the future value is within a certain range, is favorable

in terms of user friendliness compared to point forecasts for

larger waiting times. One way to produce an interval based

forecast using a series of pointwise forecasts is the

extrapolation of the time series containing the minimal and

maximal waiting times (see Fig. 4). By adding the point-

wise prediction of the average waiting time a meaningful

outlook can be provided to the user.

Numerous successful applications prove that artificial

intelligence based methods are capable of producing high

quality forecasts. For a survey of related techniques for

load forecasting we refer to [2]. However, any time series

forecasting methodology may be suitable. An overview of

relevant alternative methods can be found in [18]. For the

presented decision support system we suggest the imple-

mentation of artificial neural networks (ANNs). For an

overview of artificial neural network based methods we

refer to [26] and will not go into detail in this paper.

Several commercial and non-commercial software

packages allow the seamless integration of efficient imple-

mentations of ANNs within the system. As a part of the

forecasting module, the forecasting engine is called peri-

odically to generate predictions (e.g., waiting times, arrival

rates). The parameterization of the method used (e.g., ANN

structure and training strategy) have to be tailored to the

node-specific data and to the forecast type (see Sect. 2.1).

As a quality control measure, the forecast information

can be stored along with the actual observations, which

yields the forecast errors. Moreover, this information can

be used to fine-tune the forecasting engine or it could serve

as an alarm system to monitor the analytic performance of

the iLoads system.

4.2.3 Forecasting databases

The forecasting module utilizes two databases to store its

direct input and output (see Fig. 3). A forecast input

database is used to store the consolidated historical data

and the forecasting parameters which are collected from

different sources in different formats. In a preprocessing

step this database is updated prior to the computation of the

forecast. Herewith, changes concerning the input data

sources or formats only require an adaptation of this pre-

ceding update process. A forecast output database holds the

actual forecasting information which is updated periodi-

cally. Both databases can be integrated into a database

management system (DBMS) located in the cloud based

environment. The latter can be accessed by the forecast

engine and the request handler and each request from a

truck company or a logistic node will trigger the trans-

mission of the latest forecast according to the type of user

and the desired forecast information.

4.2.4 Request handler

To meet the requirements of the user groups described in

Sect. 2.1, a dedicated unit handles the forecast requests and

the preparation of the raw forecast data. Besides the cal-

culation of auxiliary data representations used by the user

interface, it is responsible for the extraction and aggrega-

tion of specific key figures that correspond to the request of

the individual user. This processing step includes the cal-

culation of derived time series and is performed on request

based on the latest forecast data.

As a minimum requirement, the system provides the

average waiting times to the truck company (see Fig. 4),

although minimal and worst case waiting times are desir-

able. At the same time, a request from the node is answered

by providing the predicted workloads. Depending on the

forecasting methodology, more advanced interval based

measures such as quantiles could be integrated as well.

4.3 Internal processes

In the following, we describe the workflows foreseen in the

decision support system. Two main interleaved processes

take place to provide proper forecast information to the

users, as illustrated in Fig. 5.

4.3.1 Periodic forecast calculation

The large number of requests from truck companies in a real

time application and the computational forecasting effort

make it practically impossible to run the forecasting engine

on demand. Therefore, we suggest a periodic update of the

base forecasts (waiting times, arrival rates) which are then

submitted to the user at the time of request. Moreover, the

actual forecasting is driven by the availability of new input

data. In theory, the prediction needs to be updated once new

historical data and forecasting parameters are available. In

practice, this dynamic recalculation is unneeded since the

advantages of real time computations are negligible if the

recalculation period is chosen reasonably small.

248 Inf Technol Manag (2017) 18:241–251

123

The implementation of this static forecasting schedule

requires the computation of multiple forecasts and their

storage in the forecast output database to serve individual

requests from nodes or truck companies. This process is

illustrated in the flow chart in Fig. 5.

4.3.2 Forecast retrieval

Forecast information is retrieved from two different sides,

the logistic node and the truck companies. Each request is

user dependent and requires an individual forecast, as

explained above. However, a shared process, the forecast

retrieval, takes place when a user requests forecast infor-

mation. After the request is triggered via the web interface,

the corresponding base forecast (waiting times or arrival

rates) is read from the forecast output database by the

request handler. After its preparation the individual answer

is provided to the user via the web interface.

5 Case study

In this section we demonstrate the capability of the system

by conducting experiments within a real world application.

As explained in Sect. 1, we decided to apply a machine

learning approach based on artificial neural networks to be

able to integrate additional features. To this end, we

implemented a forecasting module based on commercial

software. We report results that we achieve through the

latter for a data set that contains historical data from an

empty container depot.

5.1 Real world application and data

In our experiments we use actual truck waiting time data

from a large maritime empty container depot in Northern

Germany. The data set spans 133 days and the number of

truck arrivals range between 500 and 800 per workday. The

used time series is derived from data based on two given

time stamps for each truck arrival. The time point TIN specifies the time of arrival of the truck on the site and TOUT denotes the time when the driver delivered the documents at

the interchange before heading to the assigned unloading or

loading point within the depot. Hence, the truck waiting time

is given by TOUT - TIN. In this scenario we did not make a

difference between loading and unloading and did not take

into account additional order properties such as the number

of empty containers involved or the container type. The time

series of waiting times contains one value for each hour

within the horizon, resulting in 3192 data points. Each of

these values is computed as the average of the waiting times

for trucks with an arrival time TIN in the corresponding hour

interval. Note that the company works in a two-shift mode

(6 a.m.–8 p.m.) resulting in zero values during night time

and weekends.

5.2 Forecasting implementation

It is known that several commercial vendors provide

powerful and flexible frameworks to model and optimize

artificial neural networks. We decided to use the MATLAB

Neural Networks Toolbox 2 in version 8.5 (R2015a) which

is widely used in the industry. The net is parametrized as a

two layer architecture with a closed loop for additional

feedback. In an autoregressive fashion we use default lags

of size two, i.e. the two past realizations that contribute to

the value at the next time-step, for our multi-step predic-

tion. A preprocessing procedure was implemented to

eliminate the overnight non-working periods. We add

predictor information in form of the weekday, the daytime

and public holidays. Furthermore, we experiment with the

number of lags.

Fig. 5 The workflow of retrieving forecast information

together with the periodic

forecasting mechanism

2 The MathWorks, Inc. (http://www.mathworks.com/help/nnet).

Inf Technol Manag (2017) 18:241–251 249

123

5.3 Results

We forecast the truck waiting time for the last week within

our historical data horizon. This corresponds to the last 168

data points of the time series. The evaluation is based on

the actual (average) waiting times in this period using the

mean squared error (MSE) and the mean average per-

centage error (MAPE).

The forecast accuracy could be increased by the inte-

gration of the weekday, daytime and holiday predictors.

Additionally, we experimented with different values for the

number of lags. More detailed, we compared the forecast

quality when using 2, 3, 4 and 5 lags. The best results are

achieved when increasing the number of lags to 5. More-

over, we eliminated the night periods (8 p.m.–6 a.m.) in a

preprocessing step which further improved the results. The

obtained MAPE is 29.9 and the MSE is 26.6. For an overall

average waiting time of 30 min this corresponds to a

±5 min accuracy which is sufficient in many instances

considering the requirements for decision support in

operational practice of truck companies. Figure 6 shows

the forecast for the waiting time in minutes during the

evaluation week compared to the actual historical data.

6 Conclusion

In this work we present a concept for an innovative deci-

sion support system that provides forecasting based deci-

sion support to two types of users. On the one hand it

predicts truck arrivals to logistic nodes such as empty

container depots, packing facilities or terminals. These

facilities can utilize this estimation of future workloads to

improve their resource planning. On the other hand the

estimated waiting times at these nodes are made available

for truck companies that do business at these nodes. These

are able to adjust their route planning in order to reduce the

waiting times at the nodes. The expected result of this

optimization-driven interplay is smoothed peak workloads

at the nodes due to adaptive truck routing and reduced

truck waiting times because of more accurate resource

deployment at the nodes.

Another main strength of the developed system is its

flexibility and, in particular, the implementation at

relatively little effort. Moreover, it can be embedded into

the analytics landscape of the involved companies to

enhance business intelligence.

Acknowledgments The research Project 17694 N, entitled ‘‘Truck Waiting Time Forecasting at Logistic Nodes’’ (Lkw-Wartezeitprog-

nose für logistische Knoten) at the Institute of Maritime Logistics at

Hamburg University of Technology was funded by the German

Federal Ministry for Economic Affairs and Energy (Vorhaben der

Industriellen Gemeinschaftsförderung, IGF). We thank Sabine Wer-

ner for the fruitful discussion.

References

1. APM Terminals (2015) APM Terminals. [Online] Available at:

https://tos.portgot.se/truckservicetime. Accessed 13 Oct 2015

2. Bansal R, Pandey J (2005) Load forecasting using artificial

intelligence techniques: a literature survey. Int J Comput Appl

Technol 22(2–3):109–119

3. BMVI (2015) Prognostizierte Verkehrsleistung im deutschen

Güterverkehr im Jahr 2030 nach Verkehrsträgern (in Milliarden

Tonnenkilometer). [Online] Available at: http://de.statista.com/

statistik/daten/studie/7143/umfrage/prognose-zur-verkehrsleistung-

im-gueterverkehr

4. Böse JW (ed) (2011) Handbook of terminal planning. Springer,

Berlin

5. Bretzke W-R, Barkawi K (2013) Sustainable logistics: responses

to a global challenge. Springer, Berlin

6. Bundesamt für Güterverkehr (2011) Marktbeobachtung Güter-

verkehr—Sonderbericht zur Situation an der Laderampe. Bun-

desamt für Güterverkehr, Köln

7. Bundesministerium für Verkehr und digitale Infrastruktur (2014)

Handbuch: Schnittstelle Laderampe—Gute Beispiele. Bun-

desministerium für Verkehr und digitale Infrastruktur, Berlin

8. Davenport TH, Harris JG (2007) Competing on analytics: the

new science of winning. Harvard Business Press, Boston

9. Comission European (2014) EU transport in figures: statistical

pocket book 2014. European Union, Luxembourg (Belgium)

10. European Commission (2010) Integration in the intermodal goods

transport of non EU states: rail, inland/coastal waterway modes:

Project Report WP 2.4 (INTERREG IIIb CADSES). Brussels:

European Commission

11. Golden BL, Raghavan S, Wasil E (2008) The vehicle routing

problem: latest advances and new challenges: latest advances and

new challenges. Springer, Berlin

12. HCS Hamburg Container Service GmbH (2015) HCS WebPortal -

2.10.0.0. [Online] Available at: http://www.hcs-depot.de. Acces-

sed 13 Oct 2015

13. Heilig L, Lalla-Ruiz E, Voss S, (2016) port-IO: A mobile cloud

platform supporting context-aware inter-terminal truck routing.

In: Istanbul: proceedings of the 24th european conference on

information systems (ECIS)

Fig. 6 Forecasted hourly average truck waiting times

versus historical data for the last

week of the horizon

250 Inf Technol Manag (2017) 18:241–251

123

14. Heilig L, Voss S (2014) A scientometric analysis of cloud

computing literature. IEEE Trans Cloud Comput 2(3):266–278

15. Huynh N et al (2011) Truck delays at seaports: assessment using

terminal webcams. Transp Res Rec J Transp Res Board Iss

2222:54–62

16. Inmon WH, Imhoff C, Sousa R (2001) Corporate information

factory. Wiley, New Jersey

17. Lam FS, Park J, Pruitt C (2007) An accurate monitoring of truck

waiting and flow times at a terminal in the Los Angeles/Long

Beach Ports. METRANS, Long Beach

18. Makridakis S, Wheelwright S, Chang Y (1998) Forecasting

methods and applications, 3rd edn. Wiley, New York

19. Merk O, Busquet B, Aronieti RA (eds) (2015) The impact of

mega-ships. Int Transp Forum OECD, Paris

20. Nelson GB (2004) Real time decision support: creating a flexible

architecture for real time analytics. SAS Institute, Montreal

21. Phillips EE (2015) The Wall Street Journal. [Online] Available

at: http://www.wsj.com/articles/oakland-port-installing-sensors-on-

nearby-streets-to-measure-truck-wait-times-1435085331. Accessed

13 Oct 2015

22. Rimien_e K, Grundey D (2007) Logistics centre concept through evolution and definition. Eng Econ 54(4):87–95

23. Sauter VL (2010) Decision support systems for business intelli-

gence. Wiley, New Jersey

24. Stock JH, Watson MW (2006) Forecasting with Many Predictors.

In: Elliott G, Granger C, Timmermann A (eds) Handbook of

economic forecasting. Elsevier, Amsterdam, pp 515–554

25. Swink M, Melnyk SA, Cooper MB, Hartley JL (2014) Managing

operations. McGraw-Hill/Irwin, New York

26. Zhang G, Patuwo BE, Hu MY (1998) Forecasting with artificial

neural networks: the state of the art. Int J Forecast 14(1):35–62

Inf Technol Manag (2017) 18:241–251 251

123

Reproduced with permission of copyright owner. Further reproduction prohibited without permission.