Deplano, I (2020) Operational Research and Machine Learning Applied to Transport Systems. Doctoral thesis, Liverpool John Moores University.

	Text 2020deplanophd_chapter3internal.pdf - Submitted Version Restricted to Repository staff only Download (52kB)
	Text 2020deplanophdLicenseCopy_ITOR12740_2019-11-06internal.pdf - Supplemental Material Restricted to Repository staff only Download (52kB)
Preview	Text 2020deplanophd.pdf - Published Version Download (4MB) | Preview

Abstract

The New Economy, environmental sustainability and global competitiveness drive inno- vations in supply chain management and transport systems. The New Economy increases the amount and types of products that can be delivered directly to homes, challenging the organisation of last-mile delivery companies. To keep up with the challenges, deliv- ery companies are continuously seeking new innovations to allow them to pack goods faster and more efficiently. Thus, the packing problem has become a crucial factor and solving this problem effectively is essential for the success of good deliveries and logistics. On land, rail transportation is known to be the most eco-friendly transport system in terms of emissions, energy consumption, land use, noise levels, and quantities of people and goods that can be moved. It is difficult to apply innovations to the rail industry due to a number of reasons: the risk aversion nature, the high level of regulations, the very high cost of infrastructure upgrades, and the natural monopoly of resources in many countries. In the UK, however, in 2018 the Department for Transport published the Joint Rail Data Action Plan, opening some rail industry datasets for researching purposes. In line with the above developments, this thesis focuses on the research of machine learning and operational research techniques in two main areas: improving packing operations for logistics and improving various operations for passenger rail. In total, the research in this thesis will make six contributions as detailed below. The first contribution is a new mathematical model and a new heuristic to solve the Multiple Heterogeneous Knapsack Problem, giving priority to smaller bins and consid- ering some important container loading constraints. This problem is interesting because many companies prefer to deal with smaller bins as they are less expensive. Moreover, giving priority to filling small bins (rather than large bins) is very important in some industries, e.g. fast-moving consumer goods. The second contribution is a novel strategy to hybridize operational research with ma- chine learning to estimate if a particular packing solution is feasible in a constant O(1) computational time. Given that traditional feasibility checking for packing solutions is an NP-Hard problem, it is expected that this strategy will significantly save time and computational effort. The third contribution is an extended mathematical model and an algorithm to apply the packing problem to improving the seat reservation system in passenger rail. The problem is formulated as the Group Seat Reservation Knapsack Problem with Price on Seat. It is an extension of the Offline Group Seat Reservation Knapsack Problem. This extension introduces a profit evaluation dependent on not only the space occupied, but also on the individual profit brought by each reserved seat. The fourth contribution is a data-driven method to infer the feasible train routing strate- gies from open data in the United Kingdom rail network. Briefly, most of the UK network is divided into sections called berths, and the transition point from one berth to another is called a berth step. There are sensors at berth steps that can detect the movement when a train passes by. The result of the method is a directed graph, the berth graph, where each node represents a berth and each arc represents a berth-step. The arcs rep- resent the feasible routing strategies, i.e. where a train can move from one berth. A connected path between two berths represents a connected section of the network. The fifth contribution is a novel method to estimate the amount of time that a train is going to spend on a berth. This chapter compares two different approaches, AutoRe- gressive Moving Average with Recurrent Neural Networks, and analyse the pros and cons of each choice with statistical analyses. The method is tested on a real-world case study, one berth that represent a busy junction in the Merseyside region. The sixth contribution is an adaptive method to forecast the running time of a train journey using the Gated Recurrent Units method. The method exploits the TD’s berth information and the berth graph. The case-study adopted in the experimental tests is the train network in the Merseyside region.

Item Type:	Thesis (Doctoral)
Uncontrolled Keywords:	packing problems; operational research; forecasting train running time; machine learning; container loading problem
Subjects:	T Technology > T Technology (General) T Technology > TF Railroad engineering and operation
Divisions:	Maritime & Mechanical Engineering (merged with Engineering 10 Aug 20)
Date Deposited:	10 Jul 2020 18:28
Last Modified:	07 Sep 2022 16:01
DOI or ID number:	10.24377/LJMU.t.00013211
Supervisors:	Nguyen, TT, Zhang, Q and Hussain, A
URI:	https://researchonline.ljmu.ac.uk/id/eprint/13211

View Item