You are on page 1of 9

Proceedings of 7th Transport Research Arena TRA 2018, April 16-19, 2018, Vienna, Austria

Optimizing production schedules in classification yards

Henning Preis*, Stefan Frank, Sebastian Bäcker, Rainer König


Center for Rail Logistics (CRL), Technische Universität Dresden, 01062 Dresden

Abstract

Classification yards play an important role in railroad networks for operating single wagonload transportation.
Here, inbound trains are disassembled, the rail cars are sorted and then assembled to form outbound trains. The
efficient and resource-conserving operation of shunting yards has a profound impact on the quality and
profitability of single wagonload rail freight services. Preparing production schedules for classification yards that
ensure punctuality and efficiency is a challenging task that becomes even harder with the specification of
customer-oriented, digital supported transportation plans for each wagon.
Against this background we introduce a decision support approach that helps yard managers to find optimal
schedules for the resource allocation in the yard. The main component is an optimization approach that involves
time constraints and capacity restrictions as well as an objective function that represents goals depending on
planning level. At the strategic level of planning, these goals are determined by the utilization of resources (e.g.
engines, staff, tracks). However, during the operation, including ad-hoc disturbances, it is useful to minimize the
loss of quality instead (e.g. tardiness, missing connections). Besides the mathematical formulation of the
problem we discuss solution methods, introduce an optimization framework and outline computational results
that show the economical potential.
Furthermore, we discuss the opportunities of real-time control and outline the path towards fully automated
classification yards. With this work we support the transforming process of the railway freight transport system
to meet the prospective challenges in the digital era.

Keywords: Rail Freight; Classification Yard; Production Schedule; Optimization; Digital Yard Management

*
Corresponding author. Tel. +49 351 46336712
E-mail-address: henning.preis@tu-dresden.de
Preis / TRA2018, Vienna, Austria, April 16-19, 2018

1. Introduction

Providers of rail freight transport are forced to permanently optimize their production schemes in the context of
the strong intermodal and intramodal competition. The single wagonload transportation system in particular
requires availability and control of the necessary resources to a great extent. Classification yards, in which trains
are disassembled, sorted, controlled and reassembled according to customer requirements, represent crucial
points in the production network, whose efficient operation decisively determines the quality and cost of the
transport services provided.

The production scheduling in classification yards is by far not trivial, since a variety of dependencies have to be
considered. The high complexity results from the combination of time constraints, requirements concerning car
transitions between the inbound and outbound trains, different limitations of the resource availabilities and other
operational specifications for staff members and engines. Accordingly, this work is carried out by experts with
many years of experience, who rely on well-established scenarios, which are adapted to suit a specific situation.
Against the background of increasing performance requirements, in particular the reduction of the car dwell
times, transparency to meet customer expectations and fading expert knowledge, this manual approach is no
longer suitable. Furthermore, changing framework conditions in terms of the availability of digitalized process
information are a basis for computer-aided support for managers.

A decision support system involving an optimization model for scheduling activities and resource allocation is
beneficial in order to create optimized results by means of quantitative methods. The presented approach
required for this purpose takes up general concepts of the optimized sequence planning and adapts these to the
requirements of the application in train formation. It can be applied to different stages of production planning
and scheduling with corresponding model set ups, long term decisions with aggregated data as well as
operational arrangements in real time.

The remainder of this article is organized as follows. Section 2 defines challenges appearing in classification
yards in context with digitalisation and provides a literature review on related optimization approaches. Section 3
introduces an optimization framework with the focus on data preparation. Section 4 provides a mixed-integer
programming model which captures the general requirements of scheduling and resource allocation in
classification yards. Section 5 discusses solution strategies and shows the results for some numerical examples.
Section 6 discusses the potential of this approach for strategic decisions concerning the performance of
classification yards on the one hand and operational planning and controlling on the other hand.

2. Challenges and approaches of operation scheduling in classification yards

The function of classification yards is to disassemble inbound trains, to sort the car groups according to
destinations, to assemble outbound trains and to ensure their technical operability before the departure. For this
purpose, various sub-tasks have to be dealt with, which are taken over by various resources in the yard. These
include personnel and engines as well as infrastructure facilities, in particular tracks of different function groups.
In the interplay of operational requirements and practical conditions, operational schedules for the involved
resources have to be found on the premise of meeting a previously defined goal (e.g. efficient resource use and
timeliness of all processes).

A large number of classification yards are hump yards using artificial hills to separate and classify rail cars by
the means of gravity. Since there is a strong limited number of humps (usually one), humps represent bottlenecks
of classification yards. For that reason several scientific approaches address the problem of finding optimal
hump sequences of inbound trains, a simplified scheduling problem with only one resource (see e.g. Yagar et al.
(1983), Kraft (2000), Wiehl (2016)). But also other resources, especially the restricted number of available tracks
and engines, may influence the efficiency of the yard. The number of available classification tracks may be
influenced by the sorting procedure due to different block to track assignments (see e.g. Daganzo et al. (1982),
Dahlhaus et al. (2000) and Hansmann and Zimmermann (2008)). Also the number of track pulls can be
optimized (see Gatto et al. (2009)), which results in an optimized utilization of engine and shunting staff. A
detailed overview of related optimization problems is given by Boysen et al. (2012).
Preis / TRA2018, Vienna, Austria, April 16-19, 2018

Besides the introduction of optimization approaches for classification yards there is also a great progress of
digital technology. Information and communication systems delivers huge volumes of dynamic data and enable
communication between the stakeholders of the transportation processes. In context of managing railroad
systems and components such as classification yards there is a high potential to generate noticeable value (see
Zarembski (2014)).

3. Optimization Framework

Shunting operations in classification yards include recurring types of tasks which are executed in varying order
at designated places in the yard. Following this characteristics, it is reasonable to consider the task of production
scheduling in this yard as a generalized Job-Shop-Scheduling Problem (see Pinedo (2012)). The preparation of
data is aligned in this direction (see figure 1).

On the one hand we define the yard characteristics including standard shunting procedures and classification
schemes (single stage sorting, multi-stage sorting, with/without humping etc.). This provides us with a list of all
operations along with their duration, resource requirements and preceding operations. Furthermore, we capture
the track properties including length and connectivity as well as a list of all available resources (engines, staff).
The data describing the yard characteristics are static and filed in specific yard models.

On the other hand we prepare the transport volume at the yard. This includes the inbound and outbound trains
with their scheduled arrival and departure times, the characteristics of rail cars (axles, weight, type of load) and
the scheduled transitions between inbound and outbound relations. This data is updated on a daily basis,
regarding individual transportation plans for each rail car.

Data Preparation, Data Preparation,


(yard characteristics) (quantity structure)

Shunting Timetables
Procedures Inbound/Outbound

Infrastructure Rail car transitions

Resource Technical
availability parameters

Interpretation
Optimization

CY-PSP
(Classification Yard Production Visualization
Schedulig Problem )
Optimization Data

Solution Procedure Evaluation

Fig. 1: Optimization framework for the operation scheduling in classification yards

Based on this information we can formulate the Classification Yard Production Scheduling Problem (CY-PSP)
based on a Generalized Job-Shop-Scheduling-formulation adapted to this specific scenario. This is a Mixed
Integer Linear Program (MILP) for which known solution approaches are available. Once the specific instance
of the CY-PSP is solved, we can visualize the operation schedule in form of a Gantt-Chart (see figure 2), where
the horizontal axis shows day and time while the vertical axis shows different resources, for example two
engines AE1 and AE2. Additionally, the results of the optimization allow us to calculate the minimum number
and utilization of all resources in the yard and evaluate the performance.
Preis / TRA2018, Vienna, Austria, April 16-19, 2018

Fig. 2: Visualization of the scheduling results for all resources

4. Model formulation of the Classification Yard Production Scheduling Problem (CY-PSP)

The task of production scheduling in classification yards is to capture all the activities involved and allocate the
necessary resources according to an objective (e.g. minimal use of resources). Individual activities that are
handled by a particular resource in a known time period are called operations. The set of all operations is then
determined by the operation list . The operation represents the idle state at the beginning
and at the end of the work sequences. For each operation the following parameters are known:

...category of operation (e.g. "ABD" for humping)


...set of necessary resource types (e.g. one engine and one staff member)
...referenced train number
...duration
...release time (earliest starting time calculated by adding up the durations of preceding
operations beginning at the scheduled arrival times of related inbound trains)
...due time (latest starting time calculated by subtracting the durations of succeeding operations
and of the operation itself from the scheduled departure times of related outbound trains)

The sequence of operations required for the complete processing of inbound and outbound trains is given by a
precedence parameter which takes the value of 1, if operation is a direct successor of operation and
otherwise the value of 0. This definition allows different processing patterns for trains which are chosen
according to the technology of the yard. For example the car inspection can be carried out after the entrance of
an inbound train or after the coupling of the outbound train. One example of an precedence graph, which is a
representation of the activity-on-node-network of all operations completed by arcs showing their operational
sequence, is given in figure 3.

Fig. 3: Precedence graph of operations in classification yards


Preis / TRA2018, Vienna, Austria, April 16-19, 2018

The set of resources of the classification yard are separated in the categories active resources and passive
resources . Active resources are all mobile devices or personnel who move to the position of the train and
perform the work there. On the other hand, passive resources are occupied or released by the movement of the
trains or car groups. These are all tracks in the yards, separated into groups of inbound tracks, hump,
classification tracks and outbound tracks. For all resource types , the costs are known, which are
incurred for the utilization of a resource of this type per working day. Additionally, a limit is given for the
maximum number of available resources of the type . For all resource types transition times are
given between all relevant operations and , which represent the time requirement for reaching the start position
for the execution of the subsequent operation. Passive resources do not have such transition times.

The model described below aims to create feasible operation sequences for all participating resources by
assigning the corresponding operations with temporal and sequence-dependent restrictions. Therefore, we define
subsets with , that means with all operations that have to be performed by resource
type . For this subsets we apply constraints for feasible sequences. Decisions are represented by variables ,
which are assigned 1 when a resource of the type changes directly from the operation to the operation ,
otherwise with 0. Furthermore, variables are assigned to each , which represent the starting times. The
operation scheduling problem for a classification yard (CY-PSP) is:

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

(9)

(10)

The objective function (1) minimizes the costs of the resources usage in the classification yard. All job
sequences starting from the idle state represent the use of a resource and are included in the objective value
with the respective usage costs. The constraints (2) ensure the observance temporal feasibility of successive
operations, where represents a sufficiently large number. For the punctual processing of all operations
according to the timetable of arrivals and departures of trains, all operations are forced to start within the defined
time window (3). Constraints (4) - (7) define feasible operation sequences for all resources. This means that all
operations are serviced by the requested resources (4), flow conservation must be ensured (5), and that feasible
temporal setups for the successive operations handled by a resource are applied (6). In addition, all successive
operations take place on a train are forced to be assigned to the same track in a strict consecutive order (7).
Compliance with the upper limit of resources of type is ensured by restrictions (8). Constraints (9) and (10)
restrict the domains of the variables.
Preis / TRA2018, Vienna, Austria, April 16-19, 2018

5. Solution approaches and computational results

The CY-PSP is modelled as a mixed-integer linear program in order to use common practice solution methods. It
is common practise to use commercial branch-and cut-solvers. We implemented the CY-PSP in C# using an
interface to IBM CPLEX 12.5.1. With some problem-specific parameter settings of the optimization procedure
we were able to solve a large number of medium-sized instances (about 350 operations) to optimality within a
reasonable time frame (see table 1). For larger instances and shorter processing times we propose a heuristic
approach based on sequencing rules on the one hand and a exact specialized solver approach on the other hand.

5.1. Heuristic approach

In the heuristic algorithm a solution is derived by determining the operation sequences for all available resources
and setting start times for all operations based on these sequences. The Operation Sequencing Algorithm (OSA)
(see figure 4) is the core of the heuristic. OSA tries to create a partial solution by iterating all resource types
(Lines 11-14) and determines operation sequences for all available resources (Line 12) by following the basic
sequencing rule “Earliest Due Date” (see Baker (1984)). Afterwards it updates the time windows of all
operations in regards to the found sequences (Line 13). This ensures that there are no conflicts between already
found and newly determined sequences. Once all available resources of all resource types have an operation
sequence, it is saved as a partial solution .
In order to use the OSA efficiently, it is embedded in the Resource Reduction Loop (RRL). Here the resource
types are first ordered according to their utilization (Line 1), which is calculated based on the workload
proportional to the available time interval. This order is passed to the OSA (Line 2) together with the
corresponding maximum number of available resources . This improves the changes for the OSA to find
an initial partial solution. Afterwards, the RRL tries to find better solutions by reducing the number of available
resources for each resource type one by one (Lines 6-8). Starting with the most costly resource types (Line
5), the reduced number of resources is passed to the OSA in order to find a partial solution. Finally, the RRL
takes the partial solution with the lowest costs and determines the start times for all operations (Line 9). Since
the OSA has already reduced the time windows of all operations according to the found operation sequences, the
start times are set to the lower bound of the time window ( ) and we obtain a solution .

Resource Reduction Loop (RRL)


1: Order resource types ( ) by utilization descending
2: if ( 𝐴 𝑎 = 𝑎𝑙 𝑒 ) then
3: return 𝑙𝑙
4: else
5: Order resource types ( ) by cost descending
6: for = 1, . . . , do
7: while ( 𝐴( 1) = 𝑒 ) do
8: ∶= 1
9: ∶= 𝑒 𝑎 𝑒 ( )
10: return

Operation Sequencing Algorithm (OSA)


11: for = 1, . . . , do
12: if ( 𝐺𝑒 𝑒𝑞 𝑒 𝑒( )= 𝑒 ) then
13: 𝑈 𝑎𝑒 𝑒𝑊 𝑤 ()
14: else return 𝑎𝑙 𝑒
15: ⟸ Determined Sequences of
16: return 𝑒

Fig. 4: Heuristic approach including recursive subroutines RRL (Resource Reduction Loop)
and OSA (Operation Sequencing Algorithm)

5.2. Specialized solver approach

In order to improve the solving behavior a separate solver specialized to the problem at hand was implemented.
The goals of this solver are the fast feasibility testing of given scenarios and the verification of high-quality,
respectively optimal, solutions. We developed a four-stage approach to reach these goals. As a first step of the
Preis / TRA2018, Vienna, Austria, April 16-19, 2018

algorithm several pre-processing techniques are performed to narrow the solution space of a given scenario
instance by strengthening the bounds of time windows and resource utilization. Afterwards in step 2 a feasibility
test is performed assuming full resource availability in order to determine if the input data is viable and therefore
a feasible solution can be generated. To detect those feasibilities, respectively infeasibilities, we encoded the
model from the previous section by means of Boolean Satisfiability (SAT) (see Biere et al. (2009)). Therefore,
high-specialized algorithms following the principles of tree search with unit propagation and conflict driven
clause learning (see Davis and Putnam (1960), Davis et al. (1962), Marques-Silva and Sakallah (1996)) are used
to achieve the given goals. In the third part of the approach minimal and maximal requirements of individual and
also interdependent resources are determined to further tighten the area of feasible (and also optimal) solutions in
the search space. Hereby, the SAT-solving approach is used again under the varying resource requirements in
combination with traditional techniques from the field of Operations Research and Constraint Programming (see
Nemhauser and Wolsey (1988), Kumar (1992)). In the last step of the algorithm, a bounding search based on
Demeulemester (1995) is used to close the gap between lower bound and upper bound. The lower bound is given
by the theoretic minimal cost of resource utilization while the upper bound is determined by the least-cost
feasible solution found so far. The convergence of this overall optimization approach is summarized in figure 5.
Note that, instead of the feasibility test in step two of the algorithm, the previously presented heuristic or any
other can be used, but because a heuristic may fail to find a feasible solution (if one exists) and also is not able to
detect infeasibilities from the theoretic point of view, the presented exact approach overcomes these weakness.

Fig. 5: Convergence of the special solver approach

5.3. Computational results

Experiments were performed on 50 academic instances and 200 realistic instances based on real-world scenarios
adapted from german classification yards. These instances contain 40 to 100 inbound and outbound trains, which
imply 300 to 900 operations, 4 to 10 homogeneous resource types with up to 12 units of each type. In
preliminary tests all 250 instances were proofed to be feasible. Table 1 shows the mean results of the three
different solution approaches which are obtained within a time limit of two hours (CPLEX and Special Solver),
respectively 10 minutes (heuristic approach).

Table 1: Experimental results


Class Solution approach Initial solutions Optimal solution Time [s] Gap [%]
found found
Academic CPLEX 12.5.1 50 31 2749.3 8.6
(50) Heuristic 41 1 108.6 51.5
Special Solver 50 50 24.6 0.0

Realistic CPLEX 12.5.1 200 24 6581.8 21.3


(200) Heuristic 135 0 195.9 97.6
Special Solver 200 182 890.7 0.9
Preis / TRA2018, Vienna, Austria, April 16-19, 2018

For the academic instances both solvers were able to find initial solutions within two hours while the heuristic
achieved this goal for about 80 percent of the instances. With the help of several acceleration techniques (e.g.
cutting planes, branching rules) the standard solver was able to close the gap between lower bound and upper
bound for 31 instances. The special solver approach solved all instances to optimality within a mean time of less
than 30 seconds. For the real-world instances the solving behaviour is as expected. The standard solver performs
quite well in terms of getting an initial solution but bad in closing the gap. In contrast, the special solver found
90 percent of the optimal solutions and closed the mean gap to less than one percent.

6. Benefits and further investigations

The processes in classification yards are planned in several stages. Goal of the digitalisation is to provide all
these stages with consistent information and to make coordinated decisions. Economic efficiency is just as
important as a high-quality service. The proposed optimization framework for production scheduling can be used
here with great benefit.

6.1. Strategic level

At the strategic level fundamental decisions on the performance of the yard depending on the availability of
resources are made. With the use of the optimization framework it is possible to assess the classification yards
with regard to their performance which leads to statements for optimized overall network planning. Different
scenarios of network flows may be processed to find best results in terms of resource utilization and robustness.
The determination of maximum dwell times can be used to offer reliable time chains to the customers of single
wagonload transportation. On the other hand the optimization framework provides information on the necessity
to reserve resources. Especially forward-looking decisions on infrastructure measures, e.g. the decommissioning
of tracks, can be evaluated. Simulations using our approach also allow an assessment of technology changes by
using digital supported, automated technology like coupling robots etc.

6.2. Tactical level

At the tactical level concrete shift plans for personnel are created to cope with the requirements of the weekly
timetable. The automated generation of workable operation schedules provides better results in terms of resource
utilization and balanced workload. This will help to ensure that uniform network-wide requirements are met. A
great advantage also lies in the time-saving procedure for regional planners. With the help of our optimization
framework they can reduce the amount of routine work and can concentrate on specific features. Unused
capacities and available expert knowledge can be used for further service-oriented decisions regarding the
overall system. In special situations it is possible to react quickly to intermittent situations and changes (e.g.
blocking of tracks, building conditions etc.).

6.3. Operational level

At the operation level ad-hoc corrections and replanning activities are processed to reduce the effects of
disturbances.The Optimization framework can be used as the core of a yard control model which, in the sense of
an online optimization approach, provides improved reactivity to the system. It considers the actual arrival times
of the trains and the current situation of resource availability and responds adequately to changes. Here, the
objective function changes from the minimization of resource usage to maximizing customer service, which can
be obtained for example by reducing outbound delays assuming the adherence of all scheduled car transitions. In
conjunction with corresponding feedback on the system state, the transparency about the process status and
reliable forecasts of the completion time imply great advances for customer service.

Acknowledgements

The research is located at the Center for Rail Logistics (CRL), a joint work between the TU Dresden and the DB
Cargo AG. The authors would like to thank the colleagues of DB Cargo for their great support and their valuable
input.
Preis / TRA2018, Vienna, Austria, April 16-19, 2018

References

Baker, K. R., 1984. Sequencing rules and due date assignments in a job shop. Management Science 30, 1093–1104.
Biere, A., Heule, M. J. H., van Maaren, H., Walsh, T., 2009. Handbook of Satisfiability. IOS Press, Netherlands.
Boysen, N., Fliedner, M., Jaehn, F., Pesch, E., 2012. Shunting yard operations: Theoretical aspects and applications. European Journal of
Operational Research 220, 1-14.
Daganzo, C. F., Dowling, R. G., Hall, R. W., 1983. Railroad classification yard throughput: The case of multistage triangular sorting.
Transportation Research Part A (17), 95–106.
Dahlhaus, E., Horak, P., Miller, M., Ryan, J. F., 2000. The train marshalling problem. Discrete Applied Mathematics 103, 41–54.
Davis, M., Putnam, H., 1960. A Computing Procedure for Quantification Theory, Journal of the ACM, 201–215.
Davis, M., Logemann, G., Loveland, D., 1962. A machine program for theorem proving. Communications of the ACM 5(7).
Demeulemeester, E.,1995. Minimizing Resource Availability Costs in Time-Limited Project Networks. Management Science, INFORMS,
41, 1590-1598.
Gatto, M., Maue, J., Mihalák, M., Widmayer, P., 2009. Shunting for dummies: An introductory algorithmic survey. In: Ahuja, R., Möhring,
R., Zaroliagis, C. (Eds.), Robust and online large-scale optimization, Lecture Notes in Computer Science, Vol. 5868. Springer,
Berlin/Heidelberg, 310–337.
Hansmann, R. S., Zimmermann, U. T., 2008. Optimal sorting of rolling stock at hump yards. In: Krebs, H.-J., Mathematics – Key
Technology for the Future. Springer, Berlin/Heidelberg, 189–203.
Kraft, E. R., 2000. A Hump Sequencing Algorithm for Real Time Management of Train Connection Reliability. Transportation Quarterly
Vol. 54 (4), 95-115.
Kumar, V., 1992. Algorithms for Constraint-satisfaction Problems: A Survey. Artificial Intelligence Magazine (13), 32-44.
Marques-Silva, J. P., Sakallah, K. A., 1996. GRASP—a new search algorithm for satisfiability. Proceedings of the 1996 IEEE/ACM -
International Conference on Computer-aided Design, IEEE Computer Society.
Nemhauser, G. L., Wolsey, L. A., 1988. Integer and combinatorial optimization. Wiley-Interscience series in discrete mathematics and
optimization, Wiley.
Pinedo, M., 2012. Scheduling: Theory, Algorithms and Systems, Springer, Berlin/Heidelberg.
Wiehl, A., 2016. Adapting Exact and Heuristic Procedures in Solving an NP-Hard Sequencing Problem. In: Operations Research
Proceedings 2014, Springer International Publishing Switzerland.
Yagar S, Saccomanno F.F., Shi Q., 1983. An efficient sequencing model for humping in a rail yard. Transportation Research Part A (17),
251–262.
Zarembski, A. M., 2014. Some examples of big data in railroad engineering. in: IEEE International Conference on Big Data. IEEE, 96–102.

You might also like