You are on page 1of 7

Optimization of the rail freight car fleet sizing problem

Miloš Milenković Nebojša Bojović Nataša Glišović Libor Švadlenka


Division for Management in Division for Management in Dept. of Mathematical Dept. of Transport
Railway, Rolling stock and Railway, Rolling stock and Sciences, State University of Management,
Traction, The Faculty of Traction, The Faculty of Novi Pazar, Marketing & Logistics,
Transport and Traffic Transport and Traffic Vuka Karadžića bb, 36300 Jan Perner Transport
Engineering, University of Engineering, University of Novi Pazar, Serbia Faculty, University of
Belgrade, Belgrade, Pardubice, Studentská
Vojvode Stepe 305, 11000 Vojvode Stepe 305, 11000 95, 532 10 Pardubice,
Belgrade, Serbia Belgrade, Serbia Czech Republic

Abstract
An efficient rail freight car planning and distribution system is of great importance for most railway companies. Efficiency
can be represented with respect to the total number of cars needed in the system, and how well they are utilized to satisfy
customer demand. In this paper, a rolling horizon (model predictive) control approach to the problem of simultaneous
optimization of the rail freight car fleet size and allocation is developed. Developed dynamic model considers
heterogeneous fleet with substitution possibilities between freight car types. Numerical example is given to illustrate the
model and solution methodology.

KEYWORDS
freight cars, fleet size, allocation, model predictive control

1. INTRODUCTION
In the area of freight transport, the railroads of almost all countries face with strong competition and a
prominent trend of market reduction. It has become imperative for rail systems to develop better planned
instruments for more rational and efficient utilization of freight cars that represent a great amount of total
investments.
Rail transport frequently contains fleets of rail-cars which circulate on networks, carrying people or goods
(Sayarshad and Ghoseiri, 2009). The capacity of rail transport is directly related to the number of available rail-
cars. Owners and operators of rail transport invest in rail-cars in order to provide the capacity needed to meet
demands. Determining the optimal number of rail-cars for a particular system requires a tradeoff between the
cost of owning rail-cars and the potential costs or penalties associated with not meeting some demands as a
result of not using enough rail-cars. Serving demands results in the relocation of rail-cars. The consequent
movement of rail-cars between various locations is often imbalanced, and this implies the need for optimal
allocation of empty rail-cars over the network. Thus, the fleet size of rail-cars which are available for service at
any given time (and their locations) depends upon the rail-car allocation strategy.
Rail freight cars are expensive capital items, and fleet sizing is an important issue for both researchers and rail
freight transport service providers. Many models for fleet management have been proposed in literature. We
refer the reader to (Milenković and Bojović, 2013) for a detailed and comprehensive review of alternative
approaches to rail freight car fleet sizing and allocation. However, existing approaches are mainly concentrated
on homogeneous rail freight car fleet and/or empty flows. There are only few papers which consider both,
heterogeneous flows and simultaneous movement of loaded and empty freight cars. Also, almost all
approaches are based on multiperiod ILP or MILP formulations that maximize profit or minimize cost of rail
freight car fleet system over a given time horizon. Solutions of these approaches are valid only if the demands,
forecasts, disturbances and parameters remained the same. This is never the case, and therefore, in order to
develop an efficient approach for solving rail freight car fleet sizing and allocation problem to find an optimal
policy in a dynamic environment a rolling horizon approach is used in this paper.
The remainder of this paper is organized as follows. After introduction, Section 2 presents the model for rail
freight car fleet sizing and allocation. In Section 3, the MPC solution approach is described. Numerical example
is provided in Section 4. Concluding remarks are given in the last section.

Proceedings of the 4th International Symposium & 26th National Conference on Operational Research 155
June 4-6, 2015, Chania, Greece
2. MODEL FOR RAIL FREIGHT CAR FLEET SIZING AND ALLOCATION
This paper proposes a new formulation and a solution procedure for optimizing the rail freight car fleet size
and allocation problem wherein car demands were assumed to be deterministic, heterogeneous fleet of cars
assumed and partial substitutability between freight car types included. Developed model considers state,
control and capacity constraints. Then, a model predictive control approach is presented to find an optimal
policy of considered problem.
The model proposed in this paper includes substitution possibilities between freight car types. Namely, on
some occasions, a transport can be performed by several car types. In case when demand for a car can be
fulfilled by more than one car type, the demand may be specified for an aggregate car type. Demand for an
original car type will be fulfilled both from available cars of that type as well as from decision to substitute the
type demanded by another type. Demand for an aggregate car type can be fulfilled by only by original car
type. The aggregation of car types is pre-specified, i.e. which original car types can be used to fulfill a demand
for aggregate car type (Joborn, 2001).
Consider the following notation for model formulation:
• N : number of stations in a railway network;
• P : number of periods over a planning horizon;
• T σ : set of all original and aggregate car types
• T : set of all original car types, T ⊆ T σ
• T a : set of all aggregate car types, T a ⊆ T σ
• T t : set of car types, original and aggregate, that can be substituted by the type t . if the type t is an
original type, then the type itself is included in T t .
• T t : set of all car types that can substitute for the type k. If the type k is an original type, then the type itself
is included in T t . Thus, T t are all car types that can be used to fulfill a demand for cars specified for type
t
• Dijt ( n) : demand for a rail freight car of type t between station i and station j over period n ,
i , j ∈ N ; t ∈T σ ; n ∈ P , i ≠ j
• Sit ( n) : number of empty and loaded cars of type t at station i at the end of period n , i , j ∈T ; n ∈ P
• Vijt ( n) : unmet demand for freight car loads of type t between stations i and j over period n ,
i , j ∈ N ; t ∈T σ ; n ∈ P , i ≠ j
• Eijt ( n) : empty car flows of type t between station i and station j over period n , i , j ∈ N ; t ∈T ; n ∈ P , i ≠ j
• Fijt ( n) : loaded car flows of type t between station i and station j over period n , i , j ∈ N ; t ∈T ; n ∈ P , i ≠ j
• Eijtt ' ( n) : empty car flows of type t used as a substitution for car type t ' between station i and station j
over period n , i , j ∈ N ; t ∈T ; t ' ∈T t ; n ∈ P , i ≠ j
• Fijtt ' ( n) : loaded car flows of type t used as a substitution for car type t ' between station i and station j
over period n , i , j ∈ N ; t ∈T ; t ' ∈T t ; n ∈ P , i ≠ j
• hit : unit holding cost for a rail freight car of type t at station i per period, t ∈T
• v ijt : unit shortage cost for a rail freight car of type t between stations i and j per period,
i , j ∈ N ; t ∈T σ ; i ≠ j ;
• eijt : unit cost of empty movement of rail freight cars of type t from station i to station j ,
=i 1,..., = =
N ; j 1,..., N ; t 1,...,T ; i ≠ j
• lijt : unit cost of loaded movement of rail freight cars of type t from station i to station j ,
i , j ∈ N ; t ∈T ; n ∈ P , i ≠ j
• rijtt ' : cost of substitution a car of type t ' by a car of type t , i , j ∈ N ; t ∈T ; t ' ∈T t ; n ∈ P ; i ≠ j
• Qt : unit ownership cost for a car of type t travelling between stations per period, t ∈T
• θijt ( m, n) : portion of loaded cars of type t arrived during period n , and dispatched during period m
from station i to station j , n > m , i , j ∈ N ; t ∈T ; n ∈ P ; i ≠ j
• αijt ( m, n) : portion of empty cars of type t arrived during period n , and dispatched during period m from
station i to station j , i , j ∈ N ; t ∈T ; n ∈ P ; n > m; i ≠ j
• x ijtm ( n) : number of loaded cars of type t dispatched from station i during the period m and arrived at
station j during the period n , i , j ∈ N ; t ∈T ; n ∈ P ; n > m; i ≠ j .
• y ijtm ( n) : number of empty cars dispatched from station i during the period m and arrived at station j
during the period n , i , j ∈ N ; t ∈T ; n ∈ P ; n > m; i ≠ j .
• ϑt : ratio of the number of empty cars of type t to the total number of freight cars.
• X ( n) : state vector with components Sit ( n) , Vit ( n) , Dijt ( n) , x ijtm ( n) , y ijtm ( n)
• U( n) : control vector with components Fijt ( n) , Eijt ( n) , Eijtt ' ( n) , Fijtt ' ( n)

156
• Λ( n) : matrix in the state vector. May be represented in terms of unit matrices, zero matrices and matrices
Λ1 , Λ 2 ( n) , Λ 3 ( n) and Λ 4 .
to πits N1−1 ∑ j =1(ϑt ⋅ e jit + (1− ϑt )l jit , t ∈T
N
• πits : supplying costs that are equivalent =
• πitu : costs of unmet demand for each station, t ∈T σ

• πith : freight car holding costs in every station, t ∈T


• G : system control matrix
• Φ : system disturbance matrix
• d( n) : vector of disturbances
• J : system performance measure which represents the optimality criterion for dispatching empty and
loaded cars. In this paper it is also denoted as cost functional.
• Γ( P ) : positive, semidefinite matrix in quadratic term of the state vector in the last period of the planning
horizon
At ( n) : positive definite matrix with components πits and qt ∑ s = m saijts ( n + s − m) and qt ∑ s = m sbijts ( n + s − m)
P P

where a jit , n +1− m ( n=) θ jit ( m, n + 1), m < n, n ∈ P and b jit , n +1− m ( n=) α jit ( m, n + 1), m < n, n ∈ P
• Bt : positive definite matrix with components πitu and rijtt '
• Lt : matrix with components πith
• O2 , O3 : zero matrices of appropriate dimensions

Basic relations between the defined quantities can be now defined as follows:
N
1) Sit ( n) + ∑
Sit ( n += ∑ (F jit ( m) ⋅ θ jit ( m, n + 1) +E jit ( m) ⋅ α jit ( m, n + 1))
=j 1 m < n +1
N
(1)
− ∑ ∑(Fijtt ' ( n + 1) + Eijtt ' ( n + 1)), t =
1,..., T
t '∈T t j =1

1) Vijt ' ( n) + Dijt ' ( n + 1) − ∑ Fijtt ' ( n + 1),=


Vijt ' ( n += t 1,..., T , ∀t ' ∈T σ (2)
t ∈K t '

Fijt ( n), Eijt ( n), Fijtt ' ( n), Eijtt ' ( n),Vijt ' ( n), Sit ( n) ≥ 0 (3)

Relations (1) are conservation of flow constraints for freight cars at each station in each time period which
include the effects of travel times for car movements through the θ and α terms. In other words, the number
of freight cars of type t in next period is equivalent to the number from the last period plus loaded and empty
freight car flows which will arrive during the n +1period minus the dispatched loaded and empty freight car
flows of type t and types t ' which can be substituted by cars of type t . Relations (2) ensure that all demand is
accounted for. Unmet demand in period n +1 must equal unmet demand from previous period plus new
demand minus the cars that are used to fulfill demand in time period n +1. The last relation reflects the natural
assumption of non negativity of freight car flows, states of cars in every station and unmet demands (Beajon
and Turnquist (1991), Sayarshad and Ghoseiri (2009)). Therefore, all state variables and control actions are
always nonnegative values.

2.1. Objective function


In this approach, the system performance measure represents the cost minimization objective function which
includes costs of owning and distributing empty and loaded cars of considered types on a considered rail
network. The goal is to derive an optimal policy which maps any value of the state to the control which best
fulfil the given objective. As the system objective function incorporates the cost of moving and holding empty
and loaded cars, there are no practical constraints on its being represented in quadratic form. Let the system
performance be measured by the following cost:
1 1 P −1  A( n) L   X ( n)
=J X ( P )T Γ( P ) X ( P ) + ∑[ X ( n), U( n)]T  T (4)
2 2 n=0  L B  U( n) 

The first term is a cost of rail freight car fleet system in the last period over the planning horizon which is the
function of the state vector as a result of control actions during the previous periods.
The second term in objective function represents a cost of rail freight car fleet system in P −1 periods over the
planning horizon.

157
2.2. Constraints
Constraints that restrict the feasibility region of variables defined in the previous section are following.
System dynamics: The model under consideration is described by a time-discrete control system whose states
vary according to the following difference equation:
X ( n=
+ 1) Λ[ n, a( n), b( n)]X ( n) + GU( n) + Φd( n),=n 0,1,..., P − 1 (5)
These linear first order dynamic constraints reflect the basic relations (1)-(2) of considered problem. The state
in the next period is a function of state, the control and the vector of demand rates d( n) that are considered
as disturbances into the system. The initial state is X (0) known and given.
Non-negative control constraints: For each period
= n 0,1,..., P − 1and each car type t ∈T number of loaded
and empty freight car flows is a non-negative number. Thus, we have constraints:
U( n) ≥ O2 (6)

Where O2 represent a 2NT ( N − 1) zero matrix.

Non-negative state constraints: For each period


= n 0,1,..., P − 1 the current state in rail freight car system
must be a non-negative number. Thus, we have constraints:
X ( n) ≥ O3 (7)

Where O3 represent a T ( N 2 + 2N( N − 1)( P − 1)) zero matrix.

Capacity constraints: Additionally, we also impose the constraints in which the number of rail freight cars in
each station will be limited by the capacities of stations:
DX ( n) ≤ K (8)

3. MPC CONTROLLER
This section describes the MPC approach in detail. The controller for rail freight car fleet system is
parameterized by a set of discrete decision periods 0,1,..., m,..., n,..., P , a positive definite matrix H of
dimensions and a vector F .
At time n , the controller uses the optimal solution of a QP problem in which the decision variables are the
controls over the time horizon n, n + 1,..., n + P − 1 . Given the current state X ( n) and the controls Ue (i ) ,
i= n, n + 1,..., n + P − 1, the prediction of the state over the time horizon, is generated (details below) and appears
in the objective and constraints of the QP (Le et al. (2013)). Online MPC optimization for rail freight car fleet
sizing problem therefore comprises the minimization over Ue ( n) of a quadratic objective subject to linear
constraints:
1
minimize Ue ( n)T HUe ( n) + FUe ( n) (9)
Ue ( n ) 2
subject to Ae ( n)Ue ( n) ≤ Be ( n) (10)

where
 −I   O2 
O3 + Ni X ( n) + Ei d( n) , i = 0,..., P -1
Ae ( n)=  −C i  , Be ( n) = (11)
   
DC i   −DRi − DSi + K 
This optimization problem is known as a quadratic programming (QP) problem. H matrix is a positive definite
matrix and constraints are linear. Therefore, this problem represents a convex problem which is solved by
Matlab’s QP solver (quadprog) for each period n . Only the first element of the optimal predicted input
sequence will be selected as input to the rail freight car system and therefore for updating the state X ( n) and
right side of inequalities, Be ( n) in each time period within the planning horizon.

158
4. NUMERICAL EXAMPLE
This section demonstrates the developed MPC approach on a hypothetical rail network composed from two
stations ( N = 2 ) due to the space restrictions. Rail freight operations are performed with six car types (
T σ = 1,2,3,4,5,6 ), four original ( T = 1,2,3,4. ) and two aggregate types of rail freight cars ( T a = 5,6. ). This
assumption is based on existence of four basic types (open, closed, flat, other) and common substitutability
between open and flat as well as closed and other car types. Substitution possibilities for this case are given
on Figure 1. The planning horizon is P = 4 where each day represents a decision period.

Figure 1. Substitution possibilities between car types

Table 1. contains the input data on: unit costs of empty trips ( e jit ) and unit costs of loaded trips ( l jit ) of all
original car types as well as unit car shortage costs ( pijt ) for all original and aggregate car types. Cost terms are
given in monetary units per a car for all routes. Proportions of arrivals for loaded rail freight car flows together
with the quadratic term in the unit ownership cost as well as proportions of arrivals for empty freight car flows
of different types are given in Table 2. The coefficient of empty trips is 0.35 for all freight car types.

Table 1. Cost parameters for rail freight car fleet sizing and allocation problem
Freight car type
Route 1 2 3 4 5 6
ejit Ijit pijt ejit Ijit pijt ejit Ijit pijt ejit Ijit pijt pijt pijt
A-B 300 100 300 310 120 200 280 130 320 290 140 500 390 270
B-A 320 120 400 330 140 300 300 150 420 280 170 400 400 330

Table 2. Proportion of loaded and empty cars arrived during period n and unit car ownership costs Qt

Freight car type


1 2 3 4
n
A-B B-A A-B B-A A-B B-A A-B B-A
Loaded cars
1 0.25 0.40 0.23 0.20 0.20 0.20 0.15 0.20
2 0.35 0.30 0.42 0.35 0.30 0.30 0.45 0.35
3 0.40 0.30 0.35 0.45 0.50 0.50 0.40 0.45
Empty cars
1 0.22 0.15 0.30 0.25 0.27 0.23 0.34 0.30
2 0.32 0.35 0.30 0.30 0.33 0.23 0.30 0.46
3 0.46 0.50 0.40 0.45 0.40 0.54 0.36 0.24
Qt 100 150 130 110

The demand for transportation is considered as deterministic. For all days over the planning horizon, freight
car types and all origin-destination combinations demand is presented in Table 4.

Table 4. Daily demand between of stations on considered part


Daily demand Daily demand
Route Freight car type Route
1 2 3 4 1 2 3 4
1 30 25 35 33 27 18 29 24
2 29 24 20 31 12 14 17 16
3 22 26 28 35 17 19 21 15
A-B B-A
4 27 29 32 36 19 21 23 27
5 25 19 12 11 19 21 13 10
6 17 19 14 13 26 11 23 8

159
Table 5 contains the values of unit holding costs ( hit ) per car per period and the estimate of the initial number
of cars ( Sit (0) ) at all stations as well as unmet demands ( Vijt ) from preceding period for origin-destination pairs.

Table 5. Unit holding cost, initial number of cars and unmet demand
Unit holding cost and initial Unmet demand from
number of cars in stations preceding period
Freight
Station Route
car type
A B A-B B-A
hit S0 hit S0 V12t V21t
1 50 20 40 17 4 11
2 40 15 30 16 6 14
3 25 22 20 18 8 12
4 35 18 30 22 9 8
5 5 7
6 9 4

Table 6. presents the numbers of loaded and empty cars, that were initially in transit at the beginning of the
horizon.

Table 6. Number of loaded and empty cars dispatched before the beginning of the cycle
Loaded cars Empty cars
Freight car
Route One Two One Two
type
period periods period periods
1 20 13 5 3
2 30 16 7 5
A-B
3 13 14 4 5
4 18 17 5 6
1 15 10 6 4
2 14 19 5 6
B-A
3 11 19 3 4
4 12 14 4 7

Now, by using developed solution procedure the optimal control law for this rail freight car fleet and allocation
system composed from two stations on planning period of four days, can be determined. Computer procedure
for performing all calculations is developed using MATLAB and the obtained results are summarized and
shown in Figure 1. For proper functioning of the described system optimal size ( FS ) of 243 freight cars is
needed. The optimal value of total costs corresponding to this required rail freight car fleet size is=J 5.23 ⋅ 106
monetary units.

Figure 2. Computational results

5. CONCLUSIONS
In this paper, an approach based on MPC is proposed for rail freight car fleet sizing and allocation problem.
Rail freight car demands and travelling times are assumed to be deterministic and unmet demands are
backordered. Presence of various types of rail freight cars as well as substitutability between car types is
covered by the approach. Developed approach has been tested on a numerical example. However, in real rail

160
freight transport operation, there is a great uncertainty in model parameters and there is a need to handle
uncertainty through stochastic or fuzzy stochastic approach.

REFERENCES
1. Beaujon G.J., Turnquist, M.A., 1991. A model for fleet sizing and vehicle allocation. Transportation Science, Vol. 25, No.
1, pp. 19-45.
2. Joborn M., 2001. Optimization of empty freight car distribution in scheduled railways. PhD Dissertation, Linkoping
University, Linkoping, Sweden.
3. Le T., Vu H.L., Nazarthy Y., Vo Q.B., Hoogendoorn S., 2013. Linear-quadratic model predictive control for urban traffic
networks. Transportation Research Part C, Vol. 36, pp. 498-512.
4. Milenkovic M., Bojovic N., 2013. A fuzzy random model for rail freight car fleet sizing problem. Transportation
Research Part C Vol. 33, pp. 107-133.
5. Sayarshad H.R., Ghoseiri K., 2009. A simulated annealing approach for the multi periodic rail-car fleet sizing problem.
Computers and Operations Research, Vol. 36, No. 6, pp. 1789-1799.

161

You might also like