Multistart Large Neighborhood Search for the liquefied natural gas transportation and trading over long-term time horizons

TL;DR

This work addresses LNG transportation and trading over multi-year horizons, where a heterogeneous vessel fleet, LNG sloshing, and speed- and load-dependent consumption create a highly complex planning problem. The authors propose a three-component Large Neighborhood Search framework: (i) a Big-pairs arc-flow MILP to generate a strong initial plan, (ii) a Small discharges MILP to insert mid-route, lower-volume contracts, and (iii) a tensor-train guided black-box optimizer (TetraOpt) to tune penalty parameters that steer the search toward high-profit regions. The approach yields substantial profit gains over baseline models (e.g., about 35% on production data) with only modest runtime overhead, and it effectively exploits contract flexibility and multi-destination opportunities. The methodology integrates two staged MILPs with a black-box search and is supported by a visualization tool for solution understanding, suggesting practical potential as a decision-support system in LNG trading and shipping.

Abstract

Liquefied Natural Gas (LNG) transportation is a critical component of the energy industry. It enables the efficient and large-scale movement of natural gas across vast distances by converting it into a liquid form, thereby addressing global demand and connecting suppliers with consumers. In this study, we present the Multistart Large Neighborhood Search heuristic for the LNG transportation problem, which involves hundreds of contracts and a planning horizon of two to three years. Our model incorporates several fuel types, LNG sloshing in the tank, and speed- and load-dependent consumption rates. We also consider flexible contracts with LNG volume variability, enabling volume optimizations and multiple discharges. A tensor-train optimizer defines the parameters of Mixed Integer Programming (MIP) models, allowing better solution space exploration. On the historic and artificially generated data, our approach outperforms the baseline linear-programming model by 35% and 44%, respectively, while the time overhead is only several minutes.

Figures (4)

• Figure 1: The Multistart Large Neighborhood Search heuristic for the LNG shipping problem. The penalty optimization (in red) is a black-box optimization problem solved with TetraOpt. Big-pair selection and Small discharge insertion are modeled as ILP and solved with SCIP -- an open-source MILP solver.
• Figure 2: Modeled profit dependency on the data instance for 3 tested approaches. The red line corresponds to the result of the Big-pairs model, the green line -- to one run of the LNS heuristic (taking the baseline as a starting point), and the blue line -- to the best result obtained by Tensor-train guided LNS (with multiple starts).
• Figure 3: A snapshot of the optimized plan for January 10, 2023. Our visualization tool shows the location of all vessels (in ports or on the journey legs). For each vessel, it tracks the volume onboard; we also plot forbidden volume zones (in gray) and allowed volumes (in blue).
• Figure 4: Illustration of the decomposition technique for one vessel. Loading contracts are blue; discharge contracts are red. We mark the contracts taken in the solution with yellow and the contracts from the next subproblem with gray. Note that the vessel may not take the last discharge contracts in the previous subproblem after some more profitable contracts appear in the next subproblem, because we always use the loading contract as an initial one.