| dc.description.abstract | Unlike other airline operations planning problems, optimization in flight dispatching is
not common in literature. Flight dispatchers are centrally located and monitor multiple
flights in different places simultaneously. Their work involves planning fuel requirements,
routing, and weather monitoring, both before and during a flight. An area of opportunity
exists in the assignment of work amongst dispatchers.
A desk contains a series of flights, and is served by a dispatcher or a series of dispatchers
working consecutive shifts. In this work, we do not consider shifts and instead focus on
assigning flights amongst a set number of desks. Our goal is to balance the workload of
each desk, which is measured by the sum of each desk’s maximum workload throughout
the day.
Two formulations are presented that model the assignment of flights to desks, which we
call the Flight Dispatching Problem. The Flight Dispatcher Schedule Formulation (FDSF)
assigns flights amongst a set number of desks. The Set Covering Formulation (SCF) selects
from known schedules (the assignment of flights to a single desk) to cover all flights with the
specified number of desks (i.e., schedules). The base implementation solves the SCF using
a column generation approach that creates new schedules with each iteration. Additional
variants are also modelled where we limit which flights are assigned to the same desk.
Testing is performed on European Airline Data and American Airlines data. The
instances range in size from 46 to 297 flights in one day. We find that the FDSF solves to
optimality quickly for small instances but not for the larger ones. The base implementation
converges within two hours for the small and mid-size instances. Gaps are reduced using
an improvement heuristic in some cases. For the larger instances, neither implementation
solves within two hours and the gaps after that time are very large. Constraining the
flight assignments provides trade-offs between computation time (which is typically faster)
and solution quality (which is typically worse). We also tested the case where load varies
throughout the flight.
For the base implementation, most of the computation time for larger instances is spent
in the pricing problem. In some cases, this is improved by generating multiple columns in
each iteration instead of just one. The solution of the pricing problem is an area where
future work could be focused to improve the computational performance. Other areas for
future work include modelling dispatch zones instead of decomposing the problem by zones,
changing the balance metric in the objective, incorporating uncertainty, and including the
shift component of the dispatching problem. | en |
