Comparative study of global pooling for pre-departure ship fuel consumption prediction: Advantages over per-ship and adaptation-based approaches

Heungjun Im; Bongmin Cha; Agung Nugraha; Donghyun Kim; Seongphil Kang; Minwook Choi; Jeongmin Park; Jihwan Lee

cs.LGNov 2, 2026

Comparative study of global pooling for pre-departure ship fuel consumption prediction: Advantages over per-ship and adaptation-based approaches

Authors: Heungjun Im, Bongmin Cha, Agung Nugraha, Donghyun Kim, Seongphil Kang, Minwook Choi, Jeongmin Park, Jihwan Lee

Organizations: Department of Industrial and Data Engineering, Major in Industrial Data Science and Engineering, Pukyong National University, Busan, 48513, Republic of Korea · DX-Solutions Co., Ltd., Busan, Republic of Korea · Division of Mechanical, Shipbuilding & Air Conditioning Engineering, Pukyong National University, Busan, 48513, Republic of Korea · LAB021 Co., Ltd., Busan, Republic of Korea

Abstract

Most fuel oil consumption (FOC) models for voyage planning are per-ship models and thus generalize poorly beyond each vessel's records. Transfer and federated learning have been reported to perform better, but under different protocols that often use inputs unavailable before departure or random data splits. It therefore remains unclear which approach performs best. In this study, we compare 11 per-ship, adaptation-based, and global methods on 21 vessels. To prevent leakage, all methods receive identical inputs known or forecast before departure, excluding engine RPM and shaft power, and are tested on intact voyage legs after all training records of each vessel. Global pooling, in which one model describes each vessel by its ship identity and seven ship specifications, performs best overall in every setting examined. For in-fleet vessels, Global LightGBM reaches an overall MAPE of 8.92%, against 11.14% for Per-Ship LightGBM, and outperforms every per-ship and adaptation-based model. Global models of five architectures, including a Transformer, all outperform the per-ship and federated models. Fine-tuning a global model for each vessel, even by low-rank adaptation of the Transformer, adds no accuracy. Without any records of a vessel, the global model reaches a median MAPE of 11.20% and outperforms a per-ship model trained on that vessel's records for 14 of 21 vessels. Adding its first records to the pooled data is as accurate as vessel-specific adaptation. Ship identity and specifications are largely interchangeable for in-fleet vessels, but only the specifications support zero-shot prediction. Its error grows as a vessel's specifications differ from the fleet's. Per-ship models and vessel-specific adaptation are therefore unnecessary under fair comparison. One global model can serve a whole fleet, including newly delivered vessels with similar specifications.

Figures & tables

Study	Approach	$N_{\text{v}}$	Period	Interval (R/P)	$N_{\text{raw}}$	$N_{\text{obs}}$	Plan. $\dagger$	Temp. $\ddagger$	Evaluation
Luo et al. (2024)	Per-ship	1	$\sim$ 18 mo.	daily / daily	335	168	$\checkmark$	–	8:2 hold-out
Zhang et al. (2024)	Per-ship	1	29 mo.	60 s / –	1.0M	–	$\times$	$\checkmark$	5-fold CV + temporal holdout
Wang et al. (2023b)	Per-ship	1	$\sim$ 7.5 mo.	10 min / 1 h	–	3,361	$\checkmark$	$\times$	interleaved split
Chen et al. (2024)	Per-ship	1	14–17 mo.	30 s / 1 min	1,237,562	65,560	$\times$	$\checkmark$	70/30 voyage hold-out
Son and Kim (2026)	Per-ship	1	18 mo.	10 s / 10 s	–	2.4M	$\checkmark$	$\times$	5-fold voyage CV
Mohamed et al. (2026)	Per-ship	1	12 mo.	daily / daily	296	266	$\times$	$\checkmark$	80:20 random split + chronological split

Table 1: Methodological comparison of FOC prediction studies.

\checkmark

\times

: criterion met/not met; –: not reported or not applicable.

Class	Type	$N_{\text{v}}$	DWT	MCR	LBP	$B$	Age
			( $\sim$ kDWT)	( $\sim$ kW)	( $\sim$ m)	( $\sim$ m)	(yr)
A	Bulk carrier	8	400	29,000	350	65	9–15
B	Bulk carrier	11	320	21,000	330	62	4–7
C	Bulk carrier	1	180	19,000	280	45	$\sim$ 15
D	Oil tanker	1	110	13,000	240	44	$\sim$ 8
	Total	21					4–15

Table 2: Fleet composition by vessel class. Specifications are approximate values representative of each class. All values are rounded to preserve confidentiality.

Figure 1: Data preprocessing pipeline. The raw 10-minute sensor records (about 4.0 million) are cleaned in three stages, averaged over 3-hour windows within voyage legs, and extended with derived features, resulting in 130,369 observations with 47 features.

Vessel	Class	$n_{\text{train}}$	$n_{\text{test}}$	Vessel	Class	$n_{\text{train}}$	$n_{\text{test}}$
Ship 1	B	4,304	1,304	Ship 12	B	5,715	1,842
Ship 2	B	7,392	1,917	Ship 13	A	600	224
Ship 3	D	5,877	1,530	Ship 14	A	5,983	1,878
Ship 4	B	4,826	1,523	Ship 15	A	1,774	501
Ship 5	B	5,918	1,631	Ship 16	A	2,877	856
Ship 6	B	7,717	2,134	Ship 17	B	5,776	1,713

Table 3: Number of observations per vessel.

n_{\text{train}}

: observations in the training and validation partitions;

n_{\text{test}}

: observations in the temporal test set (Section 4.1 ). Vessels are anonymized as Ship 1–21.

Ship descriptors (8)
Group	$n$	Features	Source before departure / description
Ship identity	1	Ship name	Vessel registry
Ship specifications	7	MCR, DWT, LBP, breadth, propeller diameter, propeller pitch, ship age	Vessel registry and design documents
Operating state & environment (21)
Operational	5	SOG, course (sin/cos), mean draft, cargo quantity	Voyage plan (speed, course); loading plan (draft, cargo)
Environmental	12	Wind speed & dir. (sin/cos), current speed & dir. (sin/cos), wave height, dir. (sin/cos) & period, water depth, SST	Weather and ocean forecasts along the planned route; charts (water depth)

Table 4: The 47 features by group. For raw inputs, the last column gives where the value comes from before departure; derived features are computed from the raw inputs. Main-engine RPM and shaft power are excluded. Ship age is computed at 1 March 2026. The groups are those analyzed in the ablation (Section 5.3 ).

Figure 2: Training procedures of the compared methods: global pooling (a), the adaptation-based approaches of transfer (b) and federated learning (c), and per-ship modeling (d). (a) Global: a single model is trained on pooled fleet data. (b) Transfer: centralized pretraining followed by per-ship fine-tuning. (c) Federated: local training with weight averaging across

R

communication rounds, optionally followed by per-ship fine-tuning. (d) Per-ship: independent models trained on each vessel’s records with no data sharing.

Figure 3: Architecture of the global model. Four data sources feed a feature design stage that produces the 47 features of Table 4 , namely 29 raw inputs (the ship descriptors and the operating and environmental variables) and 18 derived features (the within-observation physics features and the cross-ship ratios). One model, encoding each vessel by its ship identity and ship specifications, is trained on the pooled records of all 21 vessels and predicts FOC without a vessel-specific stage. For an unseen vessel, the identity is unknown and the prediction rests on its ship specifications and operating inputs (zero-shot prediction).

Figure 4: Data partition. (a) The records of each vessel are split by voyage leg. The latest legs of each vessel form the test set, and the earlier legs are assigned at random, by whole leg, to training or validation. (b) Training procedure of the main comparison.

Partition	Voyage legs	Observations	Use
Training	548	82,472	Fitting during model selection
Validation	137	18,753	Hyperparameter and duration selection
Training + validation	685	101,225	Final fit with the selected configuration
Test	176	29,144	Evaluation only, latest legs of each vessel

Table 5: Data partitions. Voyage legs are never split between partitions; the test legs are the latest legs of each vessel.

Method	Approach	Base model	Training data	Per-ship stage	$N_{\text{model}}$
Global LGBM (ours)	Global	LightGBM	All fleet (pooled)	None	1
Global CatBoost	Global	CatBoost	All fleet (pooled)	None	1
TabM-FiLM	Global	TabM	All fleet (pooled)	None	1
Global ANN	Global	ANN $[256,128,64]$	All fleet (pooled)	None	1
F2V-Former	Global	Transformer (F2V)	All fleet (pooled)	None	1
ANN Transfer	Transfer	ANN $[256,128,64]$	All fleet $\to$ per-ship FT	Fine-tuning	21

Table 6: The 11 evaluated methods. “Training data” is the data seen in the main training phase; “Per-ship stage” indicates whether a separate per-ship fine-tuning or training step follows.

N_{\text{model}}

: number of models used at prediction time.

Method	Approach	MAPE ${}_{\text{overall}}$	MAPE ${}_{\text{mean}}$	$R^{2}$	RMSE	MAE
Global LGBM (ours)	Global	8.92%	7.97%	0.774	0.228	0.168
F2V-Former	Global	9.09%	8.01%	0.790	0.219	0.162
TabM-FiLM	Global	9.11%	7.94%	0.794	0.218	0.163
Global CatBoost	Global	9.12%	8.24%	0.790	0.220	0.165
Global ANN	Global	9.69%	8.71%	0.757	0.236	0.175
F2V-Former + LoRA	Transfer	9.42%	8.24%	0.760	0.235	0.169

Table 7: Method comparison on the temporal test set (29,144 observations across 21 vessels). Methods are grouped by approach: global models (no per-ship stage), the adaptation-based transfer and federated models (pretraining

\pm

per-ship fine-tuning), and per-ship models. Best in bold.

Figure 5: Overall MAPE comparison across all 11 methods, sorted in ascending order. Global LightGBM outperforms all other methods, and every global model outperforms the per-ship and federated models.

Figure 6: Per-ship MAPE heatmap for four key methods across 21 vessels. Global LightGBM achieves lower MAPE than Per-Ship LightGBM on 17 of 21 vessels, with the largest gains among data-scarce vessels.

	Record MAPE			Leg-total error
Perturbation	Global LGBM	Per-ship LGBM	Gap (pp)	Global LGBM	Per-ship LGBM
None (recorded inputs)	8.92	11.14	2.22	9.32	10.97
Environment, $s=10\%$ , $d=10^{\circ}$	9.01	11.18	2.17	9.33	10.98
Environment, $s=20\%$ , $d=20^{\circ}$	9.28	11.35	2.07	9.42	11.05
Environment, $s=30\%$ , $d=30^{\circ}$	9.64	11.58	1.94	9.55	11.17
Speed, $\sigma=0.25$ kn	9.10	11.25	2.15	9.17	10.98

Table 8: Robustness of the global advantage to input perturbations (MAPE, %, mean of five draws; standard deviation across draws

\leq 0.4

pp). Environment: wind speed, wave height, wave period, and current speed scaled by

1+\varepsilon

\varepsilon\sim\mathcal{N}(0,s^{2})

, with

\mathcal{N}(0,d^{2})

added to their directions. Speed:

\mathcal{N}(0,\sigma^{2})

added to speed over ground. Draws are independent per record unless marked per leg. Leg total: mean absolute percentage error of each test leg’s total FOC (176 legs).

Figure 7: Pooling gain (per-ship MAPE

-

global MAPE) of each vessel against its training set size, colored by vessel class. Positive values favor the global model. The three smallest training sets receive the three largest gains, and the four labeled vessels below zero are served better by their per-ship models. The fleet-wide rank correlation is weak (Spearman

\rho=-0.39

p=0.08

Ship	Class	$d_{\text{NN}}$	$n_{\text{test}}$	LOSO MAPE	Per-Ship LGBM	Global LGBM
Ship 3	D	2.856	1,530	38.37%	9.39%	9.26%
Ship 15	A	0.767	501	30.74%	50.75%	4.88%
Ship 7	C	2.856	1,970	21.36%	9.40%	11.72%
Ship 1	B	0.074	1,304	15.48%	20.67%	16.06%
Ship 5	B	0.074	1,631	14.36%	15.04%	14.52%
Ship 14	A	0.189	1,878	14.30%	3.90%	4.81%

Table 9: Zero-shot prediction in leave-one-ship-out (LOSO) tests for all 21 vessels.

d_{\text{NN}}

: specification distance, that is, the Euclidean distance to the nearest other vessel in standardized specification space (Section 4.5 ). LOSO MAPE: error of Global LightGBM fitted without any records of the vessel. Per-Ship LGBM and Global LGBM: in-fleet errors of Per-Ship LightGBM and Global LightGBM on the same test legs (Table 7 ). Vessels are sorted by LOSO MAPE. Summary rows give the median and mean per-ship MAPE and the overall MAPE.

Figure 8: Zero-shot (leave-one-ship-out) error and specification distance. (a) Zero-shot MAPE of each vessel against its specification distance

d_{\text{NN}}

to the nearest other vessel (log scale); the four vessels with

d_{\text{NN}}\geq 0.5

are labeled. (b) Zero-shot MAPE of the 17 vessels with

d_{\text{NN}}<0.5

and of the four vessels with

d_{\text{NN}}\geq 0.5

; bars mark group medians. The 0.5 boundary is descriptive, not an applicability threshold.

	LightGBM			F2V-Former
$n$ (records)	Pooled refit	Adaptation	Per-ship	Pooled refit	LoRA
0 (zero-shot)	11.20 (12.48)			12.19 (13.01)
100	12.88 (12.67)	11.48 (12.37)	17.46 (21.99)	10.65 (10.31)	12.45 (13.10)
200	11.59 (12.78)	11.14 (12.78)	15.47 (20.20)	10.33 (9.96)	10.64 (11.19)
400 $\dagger$	10.42 (11.80)	11.31 (11.85)	16.70 (18.68)	10.92 (10.31)	10.28 (10.99)
800 $\dagger$	11.02 (10.95)	11.22 (11.29)	17.05 (18.90)	9.59 (9.51)	10.67 (10.99)

Table 10: Few-shot prediction for unseen vessels with their first

n

records: median (mean) per-ship MAPE (%) over the 21 vessels, evaluated on the original test legs of each vessel. Pooled refit: the

n

records are added to the pooled data and the global model is refitted with the fixed configuration of the LOSO tests. Adaptation: the zero-shot model is boosted further on the

n

records. Per-ship: a model fitted on the

n

records alone. F2V-Former (seed 42): the

n

records are added to the pooled data before refitting, or used for low-rank adaptation (LoRA) of its zero-shot model. Single seed.

\dagger

Ship 8 (247 records) and, for

n=800

, Ship 13 (600 records) use all their records.

Feature ladder (objective fixed: direct MAPE)
Configuration	$n_{\text{feat}}$	In-fleet MAPE (%)	$\Delta$ (pp)
Base features	34	12.00 $\pm$ 0.12	—
+ ship identity (categorical)	35	9.75 $\pm$ 0.10	$-$ 2.25 $\pm$ 0.10
+ ship specifications	42	9.16 $\pm$ 0.04	$-$ 0.60 $\pm$ 0.13
+ cross-ship ratios	47	9.06 $\pm$ 0.04	$-$ 0.10 $\pm$ 0.06
Objective ladder (features fixed: full 47)

Table 11: In-fleet ablation: overall MAPE, mean

\pm

standard deviation over five seeds, fixed hyperparameters.

\Delta

: seed-paired change from the preceding row. The 35-, 42-, and 47-feature rows equal F4, F2, and F1 of Table 12 .

\dagger

MAE with sample weights

1/y_{i}

Figure 9: Ablation results. (a) In-fleet feature ladder: overall MAPE as each feature group is added, training objective fixed at direct MAPE; numbers inside bars give the step change. (b) Feature-group contribution to accuracy (

\Delta

MAPE, positive = the group helps), in-fleet versus zero-shot. Both identity bars compare the base features with and without identity; with specifications present (F2 versus F2 - ), identity contributes

+0.19

pp in-fleet and

-0.06

pp under LOSO. Five seeds; error bars

\pm 1

standard deviation.

	Configuration	$n_{\text{feat}}$	In-fleet MAPE (%)	LOSO median (%)	LOSO mean (%)
F5 $\dagger$	Raw	22	10.48	15.13	18.19
F3	Raw + specs	29	9.97 $\pm$ 0.13	11.48 $\pm$ 0.56	13.15 $\pm$ 0.20
F4	Raw + within	35	9.75 $\pm$ 0.10	16.00 $\pm$ 0.91	17.45 $\pm$ 0.16
F4 -	Raw + within, no identity	34	12.00 $\pm$ 0.12	14.12 $\pm$ 0.14	15.39 $\pm$ 0.11
F2	Raw + within + specs	42	9.16 $\pm$ 0.04	10.78 $\pm$ 0.25	12.77 $\pm$ 0.42
F2 -	Raw + within + specs, no identity	41	9.34 $\pm$ 0.10	10.72 $\pm$ 0.23	12.44 $\pm$ 0.16

Table 12: Zero-shot ablation: feature-group configurations evaluated in-fleet and leave-one-ship-out (LOSO median and mean over the 21 vessels, each predicted as an unseen vessel). Mean

\pm

standard deviation over five seeds, fixed hyperparameters; F5

\dagger

: single seed. F1–F5 include ship identity; F2 - and F4 - omit it. raw = operational + environmental + maintenance; within = within-observation physics (13); specs = ship specifications (7); ratios = cross-ship ratios (5).

Figure 10: Top 15 features by mean

|\text{SHAP}|

value. Colors indicate feature groups: blue = cross-ship ratios, green = ship specifications, orange = ship identity, gray = base features.

Figure 11: Partial dependence plots: (a) speed

\rightarrow

FOC (increasing), (b) draft

\rightarrow

FOC (monotonically increasing), (c) wave height

\rightarrow

FOC (weakly increasing). The directions are physically plausible.

Quarter	Static	Cumulative	Sliding 4Q	Sliding 2Q
2023Q3	14.87	14.87	13.41	14.35
2023Q4	15.01	11.83	8.93	8.26
2024Q1	15.13	9.72	9.42	10.41
2024Q2	13.86	7.52	8.49	10.30
2024Q3	12.38	7.36	8.50	9.41
2024Q4	14.01	7.22	6.76	7.39

Table 13: Quarterly retraining results (overall MAPE, %) of Global LightGBM with the hyperparameters of the main experiment. Static: trained once on the data up to 2023 Q2. Cumulative: retrained every quarter on all past data. Sliding 4Q/2Q: retrained every quarter on the latest four or two quarters. 2025Q3 contains July 2025 only, the last month of the dataset. Best per quarter in bold.

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Method	Architecture	Optimizer	LR	Loss	BS	Max iter.	ES
Global LGBM	GBDT	—	HPO	Custom MAPE	—	3,000	50
CatBoost	Ordered GBDT	—	HPO	Native MAPE	—	3,000	50
TabM-FiLM	TabM + FiLM	AdamW	HPO	L1	HPO	200 ep.	25
Global ANN	$[256,128,64]$ + emb(16)	Adam	$10^{-3}$	L1	512	200 ep.	25
ANN Transfer	Global ANN $\to$ FT	Adam	$5\!\times\!10^{-5}$	L1	256	30 ep.	10
F2V-Former	Transformer (4 blocks)	Adam	$3\!\times\!10^{-4}$	L1	128	200 ep.	25

Appendix

Table 14: Training configuration for all 11 methods. LR: learning rate; BS: batch size; ES: early stopping patience.

Parameter	Global LGBM (30 trials)	CatBoost (30 trials)	Per-Ship LGBM (10 trials)
num_leaves / depth	$[31,255]$	depth $[4,10]$	$[15,127]$
max_depth	$[4,12]$	—	$[3,10]$
Learning rate	$[5\!\times\!10^{-3},0.2]$	$[5\!\times\!10^{-3},0.2]$	$[5\!\times\!10^{-3},0.2]$
Subsample	$[0.6,1.0]$	$[0.6,1.0]$	$[0.5,1.0]$
Colsample	$[0.6,1.0]$	$[0.6,1.0]$	$[0.5,1.0]$
min_child_samples	$[10,100]$	$[10,100]$	adaptive $\dagger$

Appendix

Table 15: HPO search space for tree-based models. Learning-rate and regularization ranges are sampled on a log scale; the remaining parameters use their stated integer or fractional bounds. Global and Per-Ship LightGBM share the custom MAPE objective.

Parameter	Range
Ensemble size $k$	$\{32,64\}$
Block dimension	$\{192,256,384\}$
Number of blocks	$[2,4]$
Learning rate	$[5\!\times\!10^{-5},\,2\!\times\!10^{-3}]$ (log)
Dropout	$[0.05,0.45]$
Weight decay	$[10^{-5},10^{-2}]$ (log)

Appendix

Table 16: HPO search space for TabM-FiLM (30 Optuna trials).

#	Layer	Input dim	Output dim	Notes
0	Embedding(21, 16)	ship index	16	Learned ship embedding
1	Concatenate	$n_{c}+16$	62	Continuous features + embedding
2	FC block	62	256	Linear + BN + ReLU + Drop(0.2)
3	FC block	256	128	Linear + BN + ReLU + Drop(0.2)
4	FC block	128	64	Linear + BN + ReLU + Drop(0.2)
5	Linear	64	1	Output (FOC, t/h)

Appendix

Table 17: Global ANN architecture (also used as pretrained base for ANN Transfer).

n_{c}=46

continuous features.

#	Layer	Input dim	Output dim	Notes
1	FC block	$n_{c}=46$	256	Linear + BN + ReLU + Drop(0.2)
2	FC block	256	128	Linear + BN + ReLU + Drop(0.2)
3	FC block	128	64	Linear + BN + ReLU + Drop(0.2)
4	Linear	64	1	Output (FOC, t/h)
Total parameters				54,145

Appendix

Table 18: FedANN architecture (used by FedAvg and FedAvg+FT). No ship embedding is included in this implementation.

n_{c}=46

continuous features.

#	Layer	Input dim	Output dim	Notes
1	FC block	$n_{c}=46$	128	Linear + BN + ReLU + Drop(0.2)
2	FC block	128	64	Linear + BN + ReLU + Drop(0.2)
3	FC block	64	32	Linear + BN + ReLU + Drop(0.2)
4	Linear	32	1	Output (FOC, t/h)
Total parameters				16,833

Appendix

Table 19: Per-Ship ANN architecture. One independent model per vessel; no ship embedding needed.

n_{c}=46

continuous features (no ship identity).

Maritime Transportation

See all 20 papers

Aug 17, 2026

Time-Aware Validation of Machine Learning Fuel Consumption Models: Evidence from 1,Hz Operational Data, CCGS Sir Wilfrid Laurier

Ship fuel consumption (SFC) prediction supports vessel operation optimisation, emissions estimation, and decision support systems (DSS) for sustainable maritime transportation. Numerous data-driven fuel models have been developed over the past two decades, but a critical and often overlooked limitation lies in their validation practices: most studies evaluate performance using random train--test splits, which, applied to high-frequency records, admit temporal leakage and yield optimistic results that do not reflect deployment conditions. This paper examines that gap using time-aware evaluation, specifically Time Series Cross-Validation (TSCV) and Blocked TSCV (BTSCV). Using the Canadian Coast Guard Ship (CCGS) Sir Wilfrid Laurier as a case study, six regression models and a physics baseline are tuned under three time-aware schemes and three feature configurations, then evaluated on a common chronological hold-out set drawn from approximately 3.88 million steady-state 1,Hz records.

Samarasimha Reddy Chittamuru, Ayhan Akinturk, Allison Kennedy +2

National Research Council Canada, Ocean, Coastal and River Engineering, P.O. Box 12093, St. John’s, A1B 3T5, NL, Canada · Department of Computer Science, Memorial University of Newfoundland, 40 Arctic Avenue, St. John’s, A1B 3X5, NL, Canada

Maritime Transportation Temporal Data Leakage Time Series Forecasting

Ask PDF

Feb 25, 2026

Estimation and Optimization of Ship Fuel Consumption in Maritime: Review, Challenges and Future Directions

To reduce carbon emissions and minimize shipping costs, improving the fuel efficiency of ships is crucial. Various measures are taken to reduce the total fuel consumption of ships, including optimizing vessel parameters and selecting routes with the lowest fuel consumption. Different estimation methods are proposed for predicting fuel consumption, while various optimization methods are proposed to minimize fuel oil consumption. This paper provides a comprehensive review of methods for estimating and optimizing fuel oil consumption in maritime transport. Our novel contributions include categorizing fuel oil consumption & estimation methods into physics-based, machine-learning, and hybrid models, exploring their strengths and limitations. Furthermore, we highlight the importance of data fusion techniques, which combine AIS, onboard sensors, and meteorological data to enhance accuracy. We make the first attempt to discuss the emerging role of Explainable AI in enhancing model transparency for decision-making. Uniquely, key challenges, including data quality, availability, and the need for real-time optimization, are identified, and future research directions are proposed to address these gaps, with a focus on hybrid models, real-time optimization, and the standardization of datasets.

Dusica Marijan, Hamza Haruna Mohammed, Bakht Zaman

Simula Research Laboratory, Kristian Augusts 23, Oslo, 0164, Norway

Maritime Transportation Data Fusion

Ask PDF

May 15, 2026

A Retrieval-Enhanced Transformer for Multi-Step Port-of-Call Sequence Prediction in Global Liner Shipping

Accurate multi-step port-of-call sequence prediction is vital for tactical resource orchestration and logistical efficiency. However, existing methods struggle with unreliable voyage schedules and the inability of AIS data to provide visibility beyond the immediate next port. To address this, this study proposes a Connectivity-Constrained and Retrieval-Enhanced (CCRE) deep learning framework. Inspired by Retrieval-Augmented Generation, CCRE introduces a retrieval-enhanced historical encoder that queries a global maritime database for contextually similar navigational precedents. Transforming these scenarios into candidate-level semantic representations compensates for data sparsity in long-tail routes and resolves routing ambiguities. Integrating this with a Transformer-based trajectory encoder, the architecture executes adaptive "middle fusion" via cross-attention. This dynamically shifts predictive reliance from real-time kinematics for short-term accuracy to historical context for long-term strategic stability. To ensure sequence-level coherence, forecasting is formulated as a joint sequence generation problem using an autoregressive Transformer decoder enriched with Scheduled Sampling and Gumbel-Softmax relaxation. This mitigates error accumulation, while topology masks strictly enforce maritime network reachability to eliminate operationally infeasible routes. Evaluated on a global dataset, CCRE achieves a 72.3% first-destination accuracy and a 61.4% average three-step accuracy, outperforming baselines like CatBoost and LSTM by average margins of 12.6% and 11.3%, respectively. Case studies further corroborate the model's scalability and ability to capture complex routing patterns across diverse international trade lanes.

Yanzhao Su, Fang He, Yineng Wang

Department of Industrial Engineering, Tsinghua University, Beijing 100084, P.R. China · Department of Logistics and Maritime Studies, The Hong Kong Polytechnic University, Hong Kong

Transformer Vessel Trajectory Prediction Maritime Transportation Trajectory Prediction

Ask PDF

Abstract

Maritime Transportation

Explore similar work

Time-Aware Validation of Machine Learning Fuel Consumption Models: Evidence from 1,Hz Operational Data, CCGS Sir Wilfrid Laurier

Estimation and Optimization of Ship Fuel Consumption in Maritime: Review, Challenges and Future Directions

A Retrieval-Enhanced Transformer for Multi-Step Port-of-Call Sequence Prediction in Global Liner Shipping