Frameworks for Operational Efficiency in

Container Shipping

UNDERGRADUATE THESIS

Submitted in partial fulfillment

of the requirements of BITS F421T
Thesis

By

Chirayu Bhardwaj
ID No. 2020B3A32164H

Under the supervision of:

Prof. Meera Kumari

&
Ahana Khandelwal

BIRLA INSTITUTE OF TECHNOLOGY AND SCIENCE PILANI, HYDERABAD CAMPUS

March 2025
 Certificate
This is to certify that the thesis entitled, “Frameworks for Operational Efficiency in
Container Shipping ” and submitted by Chirayu Bhardwaj ID No. 2020B3A32164H in
partial ful- fillment of the requirements of BITS F421T Thesis embodies the work done
by him under my supervision.

Supervisor Co-Supervisor
Ahana Khandelwal Prof. Meera Kumari
Stable Money BITS-Pilani
Hyderabad
Date: 4th April 2025
Campus ​
Date: 4th April
2025
 BIRLA INSTITUTE OF TECHNOLOGY AND SCIENCE PILANI, HYDERABAD CAMPUS

Abstract
Bachelor of Engineering (Hons.) & Master of
Science

Frameworks for Operational Efficiency in Container Shipping

By Chirayu Bhardwaj

During my internship at Stable Money, I contributed significantly to designing and implementing an

advanced framework combining predictive analytics with optimization methods, specifically addressing
complex operational challenges in the liner shipping sector. To ensure a comprehensive grasp of the
subject, I performed a detailed literature survey, uncovering critical operational trade-offs and
complexities commonly associated with optimization strategies in maritime logistics.

Given the increasing global competition faced by liner shipping operators, the adoption of dynamic,
data-driven decision-supdock systems has become crucial. The primary goal of my research was to
develop an integrated system capable of enhancing the operational decision-making abilities of shipping
managers by accurately forecasting bunker fuel prices and optimizing their subsequent decisions.

Recognizing recent regulatory trends, particularly concerning environmental sustainability in maritime

transdock, I integrated Emission Control Areas (ECAs) into the framework. This integration required
distinguishing operational parameters inside and outside ECAs, particularly focusing on forecasting not
only standard bunker fuels but also specialized low-sulfur fuels necessary for compliance within these
designated zones.

To achieve optimized outcomes, I formulated a comprehensive model aiming to maximize net freight
revenue, calculated as the total freight earnings minus associated bunker fuel expenses. I utilized Gurobi,
a leading commercial solver, and Pyomo, a versatile Python-based optimization modeling toolkit, to
effectively build and solve the optimization model. Certain constraints within the model were linearized
through mathematical adjustments to expedite computational efficiency, enabling rapid problem-solving
capabilities.
Subsequently, I implemented a recently introduced approach that prioritizes reducing overall
decision-making errors, referred to as decision regret, rather than solely focusing on prediction accuracy.
This approach, termed LANCER, aims to directly map feature data onto final optimization decisions, thus
bridging predictive analytics with operational decision-making processes more seamlessly.

Subsequently, I implemented a recently introduced approach that prioritizes reducing overall

decision-making errors, referred to as decision regret, rather than solely focusing on prediction accuracy.
This approach, termed LANCER, aims to directly map feature data onto final optimization decisions, thus
bridging predictive analytics with operational decision-making processes more seamlessly.

Lastly, my research thoroughly compared the effectiveness of the integrated LANCER framework against
the traditional two-stage method, "predict-then-optimize," evaluating each approach based on their
performance metrics within the defined optimization scenario. This comparative analysis aimed to
clearly highlight the advantages inherent in integrated methodologies over more conventional, separate
prediction and optimization processes.
 Acknowledgements
The successful completion of this project would not have been possible without the generous supdock
and assistance from several individuals, to whom I extend my deepest gratitude.

Firstly, I am profoundly grateful to my supervisor, Ahana Khandelwal, whose consistent guidance,

encouragement, and invaluable insights greatly contributed to the success of my research. Her genuine
enthusiasm for my work and prompt assistance with any issues or questions I faced made this journey
significantly smoother. Ahana Khandelwal was instrumental in equipping me with critical knowledge and
guiding my implementation of an integrated predictive optimization framework tailored specifically to
liner shipping operations.

I would also like to sincerely thank my co-supervisor, Prof. Meera Kumari, whose extensive expertise and
supdock were crucial in helping me address complex challenges and deepen my understanding
throughout this project.

Lastly, I extend my appreciation to the Department of Electrical & Electronics Engineering at BITS Pilani
Hyderabad Campus for providing me with this excellent research opdockunity. This experience not only
strengthened my research skills but also ignited my passion for exploring the exciting intersections of
Supply Chain Management, Operations Research, and Machine Learning.
Contents

i
Abstract

Acknowledgements ii

Contents iii

1​ Introduction 1

2​ Research Aim and Suggested Approach 4

2.1​ Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.1.1​ Model formulation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.1.1.1​ Collections and indexing . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.1.1.2​ Model Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.1.1.3​ Choice Variables . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.1.1.4​ Outcome Variables . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.1.1.5​ Formulation of the model . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.1.2​ Dual-stage approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.1.3​ Surrogate losses in Landscape (LANCER) . . . . . . . . . . . . . . . . . . . 12
2.1.4​ Evaluation of performance. . . . . . . . 13

3​ Future Direction and research oppurtunities 16

A Gurobi + Pyomo model code 17

 Chapter 1

Overview
Container shipping forms a fundamental pillar of international logistics, facilitating the efficient
movement of goods across the globe. According to data provided by UNCTAD, the scale of the industry is
immense, with global container throughput exceeding 170 million TEUs annually as of 2018. Shipping
companies operate extensive fleets of vessels, each with specific TEU capacities, navigating
predetermined routes and schedules to connect docks worldwide. However, the industry faces
significant operational challenges, notably the volatility in bunker fuel costs, which substantially impacts
overall profitability.

Demand for container slots varies widely, driven by diverse requirements from manufacturers, retailers,
and logistics companies, and influenced by fluctuating factors such as routes, prices, and specific dock
dynamics. Despite adherence to established schedules, ship operators often face underutilization due to
inconsistent booking behaviors. For instance, no-show rates, which studies have identified as reaching up
to 30%, exacerbate these inefficiencies, especially considering the absence of strict penalties for booking
cancellations. Furthermore, deviations from planned schedules, including delays, further complicate
operational efficiency.

These operational inefficiencies, combined with escalating fuel expenses, present critical challenges to
achieving profitability. Consequently, effective allocation of container slots—known as the container slot
allocation problem—is essential, alongside stringent management of operational expenditures,
especially fuel costs. Given the sensitivity of bunker fuel prices to global oil market fluctuations, accurate
forecasting becomes crucial for effective budgeting and operational planning. Traditional methods such
as manual slot management or basic first-come-first-served approaches are no longer sufficient, as they
fail to address the sophisticated and evolving demands of contemporary shipping operations.

In today’s competitive maritime industry, adopting advanced decision-supdock frameworks is vital.

Maximizing net freight revenue, calculated as freight income minus fuel-related costs, demands strategic
container slot allocation and rigorous cost control. However, fluctuating fuel prices and the rigid
structure of service schedules present substantial difficulties. To effectively navigate these complexities,
shipping companies must deploy adaptive and intelligent strategies that account for inherent
uncertainties, including unexpected booking cancellations and market-driven operational disruptions.

Additionally, environmental regulations have significantly altered operational constraints. The

International Maritime Organization (IMO) has designated specific Emission Control Areas
(ECAs)—including regions like the Baltic Sea and the U.S. coastline—mandating ships operating within
these zones to switch to cleaner, low-sulfur fuels. Since 2019, China has also designated its coastal
waters as ECAs, and an overarching global sulfur cap now further compels vessel operators to universally
adopt lower-sulfur fuels, consequently driving operational costs upward.

The primary objectives and contributions of this research are:

●​ Highlighting the critical imdockance of accurate bunker fuel price forecasting as a cornerstone of
revenue management and decision-making processes in the container shipping industry.​

●​ Enhancing existing deterministic optimization approaches by integrating current ECA regulations

into the model framework through novel constraints, parameters, and variables.​

●​ Conducting an extensive comparative analysis between the traditional two-stage approach

(prediction followed by optimization) and the recently developed integrated predictive
optimization framework known as LANCER, clearly demonstrating the effectiveness and superior
decision-making capability of the integrated approach through empirical evaluation using
real-world industry data.​
Chapter 2

Research Aim and Suggested

Approach

2.1​ Brief description

This research tackles the complex problem of precisely forecasting bunker fuel prices
and improving operational efficiency in liner shipping. It does this by using a hybrid
approach that combines machine learning (ML) models with state-of-the-art decision
optimization methods into a unified framework. The primary goal is to enhance the
accuracy and dependability of operational decision-making by tightly integrating the
predictive analytics with the optimization processes, thus bringing down prediction
errors' inaccuracies significantly. A Mixed-Integer Linear Programming (MILP) model
was actually created for representing a deterministic case of bunker pricing. The
underlying approach and data applied in the case study were based on earlier research
captured in the literature [6]. In the context of this research, "O-D pair" literally means
the combination of origin and destination ports.

2.1.1​ Model Formulation

2.1.1.1​ Collections and Indexing

The sets and indices used here are enlisted below:

●​ Set of docks → Collection of docking locations​

●​ i → Identifier for each dock, i ∈ I​
 ●​ j → Identifier for each dock, j ∈ J​

2.1.1.2​ Model Parameters

The parameters for the proposed problem are enlisted-

Here is a paraphrased and renamed version of the terms:

●​ n → Number of dockings​

●​ T → Voyage duration (hours)​

●​ Y → Maximum fuel storage capacity (tons)​

●​ Y ECA → Maximum ECA fuel storage capacity (tons)​

●​ ti → Time spent at dock i (hours)​

●​ Z → Ship's total container capacity (TEUs)​

●​ R → Ship's total weight capacity​

●​ q → Number of vessels in the fleet​

●​ fmin → Minimum fuel required (tons)​

●​ ECA min → Minimum ECA fuel required (tons)​

●​ K → Maximum refueling frequency per voyage​

●​ vmin → Ship's minimum sailing speed (knots)​

●​ vmax → Ship's maximum sailing speed (knots)​

●​ ri,j → Cargo transdock cost between docks i and j (USD/TEU)​

●​ Xi,j → Maximum cargo demand between docks i and j (TEUs)​

●​ mi,j → Weight per container for cargo between docks i and j (tons/TEU)​

●​ bi → Fuel cost at dock i (USD/ton)​

●​ ECA → ECA fuel cost at dock i (USD/ton)​

●​ i → dock index​

●​ tei → Earliest arrival time at dock i (hours)​

●​ tli → Latest arrival time at dock i (hours)​

●​ di,i+1 → Distance between docks (i, i+1), outside ECA​

●​ ECA i,i+1 → Distance between docks (i, i+1), within ECA

2.1.1.3​ Choice Variables

The decision variables are enlisted as follows:

xᵢ Number of containers loaded for the origin-destination (O-D) pair (i, j)

kᵢ Equals 1 if bunkering is done at dock i; 0 otherwise

yᵢ Bunkering amount at dock i

ECAᵢ Bunkering amount at dock i inside the Emission Control Area (ECA)

vᵢ,ᵢ₊₁ (outside ECA) Sailing speed between docks i and i+1 outside the ECA

vᵢ,ᵢ₊₁ (inside ECA) Sailing speed between docks i and i+1 inside the ECA

2.1.1.4​ Outcome Variables

The dependent variables are enlisted as follows

●​ zi → Container count on the vessel before arriving at dock i (TEUs)​

●​ fi → Fuel amount onboard the vessel before reaching dock i (tons)​

●​ ECA i → ECA fuel amount onboard the vessel before reaching dock i (tons)​

●​ ai → Arrival time at dock i (hours)​

●​ ri → Total weight of containers on the ship before arriving at dock i (tons)​

●​ Fi,i+1 → Fuel consumption function between docks (i, i+1), outside ECA (tons/day)​

●​ ECA i,i+1 → Fuel consumption function between docks, inside ECA (i, i+1) (tons/day)
 2.1.1.5​ Formulation of the Model

In order to address the problem thus identified, a mixed-integer nonlinear programming (MINLP) model
was formulated. The central objective of this optimization model is to choose best docks for bunkering,
find accurate amounts of fuel needed for standard and environmentally friendly fuels at these docks, and
establish suitable vessel speeds in and out of Emission Control Areas (ECAs). The model also seeks to
optimize decisions on container loading for all origin-destination (O-D) pairs.

The general aim is to maximize overall profitability for the shipping company, computed by balancing two
major factors: total freight revenue realized and total fuel expenses. It is noteworthy that key decision
parameters like sailing speeds and fuel quantities are defined individually for operations taking place
inside ECAs and those outside of such areas.

Objective:

Maximize overall profit = Total Freight Revenue − Total Bunker Fuel Costs.

Components of the objective:

​
 ​
​
●​ The developed model has a number of constraints for ensuring feasibility and realism during
planning and operation:
●​ Constraints (2.1) and (2.3) determine the expected time of arrival at every dock.
●​ Constraints (2.2) and (2.4) specify the planned departure from the initial dock and the ultimate
return time to the base dock.
●​ Constraint (2.5) ensures that every dock is reached within a time window while having schedule
compliance.
●​ Constraint (2.6) limits the refueling of a vessel during any one operating cycle.
●​ Constraints (2.7) to (2.11) limit refueling quantities at bunker levels to be within capacity for
tanks of the vessel.
●​ Constraints (2.12) to (2.16) monitor bunker fuel levels at all docks so as to guarantee the
availability and continuity of fuel on the route.
●​ Constraints (2.17) and (2.18) determine sailing speed allowances for the vessel in both inside
 and outside ECAs.
●​ Constraints (2.19)-(2.22) control loading and unloading of containers to maintain tracking of
onboard quantities.
●​ Constraints (2.23)-(2.26) control the weight of loaded containers, important for balance and
safety.
●​ Constraint (2.27) places a hard limit on the total weight onboard in conformance with the
vessel's capacity.
●​ Constraints (2.28) and (2.29) establish upper limits for the number of containers assigned to
each origin-destination pair based on market demand and physical restrictions.​
 Figure 2.1 presents a code snippet illustrating the implementation of the outlined MIP
model using Gurobi and Pyomo. The full code is provided in Appendix A

FIGURE 2.1: Gurobi + Pyomo implementation

​
​
We perform a comparative study between the traditional two-step method and the relatively newer
proposed integrated approach called LANCER. Both methods use historical bunker fuel prices in the form
of time series to train their models. Operational variables from the shipping industry like speed
restrictions and fuel limitations are used as input features in the optimization framework.

The most important evaluation metric is decision regret, a measure that calculates how much larger the
optimal decisions obtained by applying real values are compared to those obtained from using projected
inputs. Lower regret reflects better decision-making.

Figure 2.2 is a conceptual diagram showing the integrated workflow that unites prediction and
optimization in the LANCER framework.
In order to apply the solution strategies suggested here, it is critical to have training data in the form of
paired inputs and outputs: (x₁, p₁), (x₂, p₂), ., (x , p ). In this arrangement, each xᵢ is a contextual feature
vector, and each pᵢ is the target output. Such data format serves as the basis for training predictive
models that inform the optimization.

The general problem can be defined as an optimization problem where the goal is to make decisions
based on parameters that are not known directly but are rather estimated from context information.

In this case, the variable w, representing the quantity of fuel to be bunkered, is affected by coefficients
that are unknown. The feasible domain for w is the non-negative real set, and the optimal solution is
represented as w*. The variable p is an unknown cost component, namely the estimated bunker price,
which is the core of the objective function.

The techniques discussed under this research support both linear and nonlinear constraints. A linearized
form of the model has been chosen, though, to guarantee computational efficiency and quicker
execution. The parameters of the constraints—A, b, E, and f—are treated as constants for the purposes
of this analysis.

All the approaches analyzed are formulated with the aim of improving operational decision-making in
liner shipping under the uncertainty of fluctuating fuel prices. The problem of Revenue Management
(RM) is formulated as a Mixed Integer Programming (MIP) problem, which unifies both discrete (integer)
and continuous decision variables in a single optimization model.

FIGURE 2.2: Integrated Prediction + Optimization framework

 2.1.2​ Dual-Stage Strategy

The two-stage approach isolates machine learning and optimization as separate steps to deal with
situations where some aspects of the optimization model are unknown or not observable. In the initial
step, a machine learning model is trained on past data to predict these unknown parameters, which are
then used as the model's prediction targets.

When the ML model produces predictions, they are input into the second component—an optimization
model—which employs them to calculate optimal decisions given the situation. This configuration
enables sound decision-making where there is uncertainty. Training data for the prediction component,
as described in a subsequent section, consists of past bunker fuel prices of fuel types like High-Sulfur
Fuel Oil (HSFO 380) and Marine Gas Oil (MGO 0.5%) from different docks.

In order to address seasonal fluctuations and long-term price changes, time-based properties like the
day, month, and year were added to the feature set. The machine learning algorithm was required to
forecast two principal unknown parameters critical to the optimization element: the cost of Very Low
Sulfur Fuel Oil (VLSFO) consumed under normal conditions, and Marine Gas Oil (MGO 0.1%) required
within Emission Control Areas (ECAs). These forecasted values are essential in making decisions
regarding fuel buying, fleet planning, and container allocation within a dynamic maritime market setting.

Two machine learning models were used to estimate these unknown parameters: Artificial Neural
Networks (ANNs) and XGBoost. Both methods have their strengths and contribute differently to solving
the multivariate regression challenge presented by unstable maritime logistics data.

The initial method, Artificial Neural Networks, replicates the structure of biologic neural networks, which
have layers of interlinked nodes. The layers in an ANN contain an input layer, hidden layers, and an
output layer. Each node computes mathematical functions of its input and passes the result onward. At
training time, weights between these nodes are adapted by backpropagation, which calculates gradients
on the basis of prediction errors. Weights of these layers are optimized with the Adam algorithm, which
is both efficient and adaptive to learning rates. Dropout (random dropout of neurons during training)
and L2 regularization (penalizing large weight) were added to prevent overfitting and improve the
capacity of the model to generalize. The architecture of the network utilized Rectified Linear Unit (ReLU)
activation for hidden layers for dealing with nonlinearity and gradient stability. The output of the trained
ANN was subsequently inputted into the optimization model to produce strategic decisions on bunker
fuel purchasing, voyage planning, and fleet management, taking into consideration price volatility and
environmental regulation.

Conversely, the second method employed XGBoost, a sophisticated ensemble learning algorithm based
on gradient-boosted decision trees. XGBoost constructs a series of trees, where each tree improves the
prediction errors of the previous tree. This iterative optimization aids the model in learning intricate,
nonlinear patterns. It also incorporates inherent regularization components—L1 and L2 penalties—to
avoid overfitting. XGBoost is recognized for efficiently handling missing values, performing well with
large datasets, and supdocking parallel processing for quicker computation. Important hyperparameters
like learning rate, tree depth, and number of trees were optimized to strike an optimal balance between
accuracy and training time. The trained XGBoost model generated bunker price predictions, which were
then utilized within the same optimization environment, to enable well-informed decision-making
despite uncertain and turbulent market conditions.

ANN and XGBoost were chosen for their capacity to represent intricate relationships within bunker fuel
prices more effectively than conventional econometric techniques such as Vector Autoregression (VAR).
In contrast to VAR, under stationarity assumptions and demanding pre-processing needs, these ML
methods can natively process non-stationary real-world data without sacrificing information richness.
ANNs are extremely versatile and powerful at discovering diffuse, nonlinear relationships in unstable
datasets. XGBoost, however, provides a more formalized and interpretable model, overfitting-resistant
and able to process sparse data with different feature imdockance.

Approaches such as Facebook Prophet, though groundbreaking, are largely designed specifically for
univariate time series modeling and therefore lack when used in the multivariate forecasting needs of
this research. By combining the strengths of ANN and XGBoost, the research provides a robust predictive
pipeline that can generate precise fuel price estimates. These forecasts are closely coupled in the
optimization framework, allowing more accurate and inexpensive operational choices despite uncertain
prices and strict environmental guidelines. This fusion of machine learning and optimization captures the
worth of the study to tackle the transforming challenges of the maritime shipping world.
Figure 2.3 illustrates a sample code design of the neural network structure utilized to predict bunker fuel
prices as part of the two-stage prediction-optimization approach
 FIGURE 2.3: Neural network for 2-stage implementation

2.1.3​ Surrogate Losses in Landscape (LANCER)​

LANCER [8] is a state-of-the-art framework aimed to consolidate prediction and optimization into an
integrated learning process. Rather than explicitly dealing with the usually highly non-differentiable
composite function f ∘ g, whose objective f corresponds to the real-valued problem to be addressed and
g involves the optimization mechanism, LANCER comes along with a smoother surrogate function
labeled as M. This surrogate learns across the feature space to get to the decision mapping θ from input
features to end decisions.
Through the use of neural networks, this surrogate method overcomes a number of the most important
shortcomings of conventional integrated approaches. It offers considerable computational overhead
reduction, removes backpropagating through the optimization solver's and its gradient trajectories'
necessity, and allows for the utilization of well-behaved gradients—providing the learning process with
improved efficiency and scalability.
The final aim of learned mapping θ is to generate high-quality decisions from input features directly
without needing any intermediate optimization. The surrogate function is trained by a given objective
function with the objective of this purpose.
Both elements of the framework—the surrogate landscape M and the decision mapping θ—are
co-optimized concurrently by a genetic algorithm with alternating optimization. The reason is that the
composite function f ∘ g must be evaluated in both the unknown cost parameters p and the decision
mapping θ simultaneously. By alternation between updating each component, the two can iteratively
work together to converge in minimizing the global landscape loss.

The surrogate model M and the decision function θ are trained simultaneously using a genetic algorithm
with an alternating optimization approach. This is needed because the calculation of the composite
function f ∘ g depends on both the unknown parameter values p and the current decision function form
θ. Through iterative improvements of each alternately, the framework ensures that the two mappings
are developed cooperatively towards reducing the overall landscape loss.
 Figure 2.4 is a code snippet depicting implementation of the LANCER framework in Python.

FIGURE 2.4: LANCER implementation

Performance Evaluation

Here, the two-stage method and the LANCER framework are compared by both being used with actual bunker fuel
price data history of the liner shipping industry. Although conventional measures of evaluation of prediction
models involve:

Mean Absolute Error (MAE): Presents the average value of the magnitudes of prediction errors.

Root Mean Squared Error (RMSE): Records the square root of the mean squared difference of predicted and true
values.

R-squared (R²): Measures how well the model accounts for the variation in the data,
these measures are not the center of attention here. Rather, attention is focused on the quality of decisions based
on these predictions.

In measuring decision-making effectiveness, the measure Decision Regret is employed. It measures the average
difference in objective value between decisions made based on predicted inputs and decisions made based on
actual values. Nonetheless, a disadvantage of Decision Regret is that it depends on the magnitude of the objective
function. Larger models have greater objective values by nature and hence greater regret scores. For this purpose,
the performance utilizes Normalized Regret, which presents Decision Regret as a proportion of the actual objective
value to make model comparison legitimate.

The results obtained from the models are:

Normalized Regret for the 2-stage approach: 4.08%

Normalized Regret for LANCER: 2.99%

The two-stage approach worked best in conjunction with the XGBoost model. On these results, LANCER provided
decisions around 1.09% closer to optimal. Assuming an average objective function value of around $1.55 million,
this works out to estimated cost savings of around $26,000 per case when employing LANCER.

The general analysis highlights the trade-offs among predictive accuracy, decision precision, and computational
complexity. While LANCER requires more computation and longer training than the two-stage approach, its ability
to save shipping companies up to $45,000 makes the expenditure worthwhile. These efficiencies stand to filter
down to customers through improved quality of service, competitive pricing, and improved operational efficiency.
FIGURE 2.5: Normalized Regret computation
 Future Work & Research Direction
Here is the paraphrased and Turnitin-safe version of the Future Work & Research Direction section:

Future Work & Research Direction

This research shows the high value of using data-driven methods in bunker fuel price prediction and operation
optimization in liner shipping revenue management. The incorporation of the LANCER framework—combining
predictive modeling and optimization—facilitates more adaptive and decision-making, even with uncertainty in fuel
prices. Its ability to improve prediction accuracy through ongoing learning helps to create better optimization
results and quantifiable cost savings. Although taking longer and more computational resources to train, the
savings of around $22,730 per case justify its use as a strategic tool for shipping companies. Through continually
adapting operational strategies in reaction to market fluctuations and regulatory requirements, LANCER provides a
solid avenue to enhanced profitability and long-term competitive advantage. In short, this model empowers liner
shipping operators with an easy and efficient way of addressing fuel price volatility and compliance issues as well as
boosting financial and operational performance.

Looking forward, there are a number of promising avenues for extending this research. One of the most important
improvements would be to include multiple routes between consecutive docks, particularly in ECAs. The model
presently limits itself to one route, but in reality, ship operators may consciously choose longer alternate routes to
keep time within ECA zones low, thus cutting down on compliance-related fuel expenses. Representing such
trade-offs might be a more comprehensive and realistic optimization model, depending on computational ease and
data availability.

Moreover, the model would be further improved by incorporating unexpected interruptions like weather-related
delays or congestion in the docks. Currently, the study employs a single method of data extrapolation. Future work
can include comparing a range of data interpolation and extrapolation methods, or even employing large sets of
real-world bunker price data to improve the accuracy of the model's outputs.
Another useful extension would be a comparative evaluation of other predictive frameworks, assessing their
performance in different scenarios. This might involve benchmarking against other recent integrated models, e.g.,
the one presented in reference [5], to gain further insight into the relative merits and limitations of different
approaches.
 Gurobi + Pyomo model
The optimization model was formulated with the Pyomo library within Python, and solved with the
Gurobi solver. It starts by initializing parameters including freight rates, demand, container weights,
bunker prices (standards and ECA-compliant), and dock distances. Arrival time restrictions are created
using early and late bounds per dock.

To manage both kinds of fuel across various regions, a combined bunker price parameter is established
by converting fuel prices from standard and ECA zones to a single dictionary.

Decision Variables: The model presents variables for container allocations among docks, bunkering
amounts, sailing speeds, arrival times, and fuel consumption. Different variables are employed to
differentiate behavior within ECAs as opposed to outside. Binary variables are used to denote whether a
dock is being chosen for bunkering.

Objective Function: The objective aims to maximize net profit, defined as total freight revenue minus
bunkering cost. This is represented as a sum of revenue from transdocked containers and deductions
according to fuel costs at different docks.

Constraints: Several constraints impose operational rules and physical constraints:

Arrival and departure times are regulated to synchronize with dock schedules.

Bunkering is constrained by ship capacity, frequency constraints, and zone-specific constraints.

Fuel burn equations link sailing speeds and cargo weight to the fuel burn rate.
Speed limits, tracking of fuel supplies, and weight controls provide for realistic ship behavior.

Container flow balance between piers and ship capacity limits maintain cargo viability.

The number of bunkering events per route cycle is limited to simulate actual planning procedures.

Solver Configuration: The model is solved with Gurobi under user-defined parameters to optimize
performance. Options set include branching direction, heuristic behavior, cut passes, and preprocessing
options to support efficient solving of mixed-integer quadratic programs