TempTPI: Informer-Based trajectory prediction for maritime vessels
Kevin Ferneding, Veronika Lietavcova, Aleksandra M. Blachowiak, Peder Heiselberg
cs.LG
Quick verdict
Informer+temporal encoding for AIS vessel prediction claims 55% lower MSE than TPTrans at 5h; promising, but evaluation is figures-only, single baseline.
Summary
The paper proposes TempTPI, a vessel trajectory prediction framework combining an Informer encoder (ProbSparse self-attention, self-attention distilling) with a multi-channel temporal encoding that expands cyclic features (hour of day, day of week, month of year) into sine/cosine frequency terms. Trained on two weeks (1-14 October 2025) of Danish Maritime Authority AIS data resampled at 6-minute intervals, it is compared against the CNN-Transformer hybrid TPTrans using MSE over 1-5 hour prediction horizons with a 5-hour input window. TempTPI reportedly outperforms TPTrans at all tested horizons, with roughly 55% lower MSE at 5 hours; an ablation over input length (240-480 min), vessel volume (100-1000 MMSIs), and stride shows TempTPI needs more data than TPTrans. Strengths include a detailed preprocessing pipeline and a systematic 21-test ablation design. However, evidence is figures-only with no numeric tables, error bars, seeds, or significance tests; TPInform ablation results are omitted; TPTrans is the only external baseline; hyperparameters (e.g., frequency range K, layer sizes) are unreported and no code is released. The two-week, single-month dataset cannot support the claimed seasonal-cycle modeling, and excluding vessels below 6 km/h removes anchored traffic. A promising direction with weakly supported evaluation.
Review complete
Assessment at a glance
6
risk flags identified
● high 0
● medium 4
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Replacing TPTrans's Transformer encoder with an Informer (ProbSparse) encoder significantly reduces MSE across 1-5 hour prediction horizons.
- Adding the multi-channel temporal encoding on top of the Informer encoder yields a further consistent improvement at every tested horizon.
- At a 5-hour prediction horizon, TempTPI achieves roughly 55% lower MSE than TPTrans on Danish-waters AIS data.
- TempTPI requires a larger vessel sample volume than TPTrans to reach its best performance, and longer input windows reduce performance because usable data shrinks.
Strengths
Well-designed two-factor ablation
Section II-C separates the Informer-encoder contribution (TPInform intermediate model) from the temporal-encoder contribution, with a systematic 21-test grid: 3 input lengths (240/360/480 min) x 5 horizons (60-300 min) plus 6 data-volume/stride tests (k = 100/500/1000 MMSIs, stride 15/30 min).
Explicit, reproducible data curation rules
Section II-A specifies concrete thresholds: 15-min gap segmentation, 6-min resampling, Web Mercator projection, step <=50 km, gap <=4 h, speed <=100 km/h, minimum duration/track-length filters - a level of preprocessing detail many AIS papers omit.
Candid limitations and negative results
Consistency of gains across settings
Public data source
Weaknesses
Single external baseline; key ablation omitted
No statistical reporting whatsoever
Dataset contradicts the temporal-encoder motivation
Underspecified model and evaluation protocol
Efficiency motivation not exercised
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
TempTPI (Sections I-II) is a vessel trajectory prediction framework for AIS data making two changes over TPTrans [10]: (1) an Informer encoder [12] with ProbSparse self-attention and self-attention distilling replaces the standard Transformer encoder; (2) a temporal encoder expands cyclic calendar features - hour of day, day of week, month of year - into sine/cosine terms across a frequency range K, each with a linear projection and bias, fused with the data channel before a positional encoder. Predictions come from separate fully connected heads for latitude and longitude. Central claims: consistent superiority over TPTrans at 1-5 h horizons on Danish AIS data; roughly 55% lower MSE at 5 h; and higher data appetite than TPTrans.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 6 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is TempTPI and how does it predict vessel trajectories?
TempTPI is a deep learning framework that predicts ship positions 1 to 5 hours ahead from AIS (Automatic Identification System) data. It replaces a standard Transformer encoder with an Informer encoder using ProbSparse sparse attention, and adds a temporal encoder that turns hour-of-day, day-of-week, and month-of-year into sine/cosine frequency features. On Danish AIS data the authors report lower error (MSE) than the TPTrans baseline at every tested horizon, about 55% lower at 5 hours.
What data was the model trained on?
Two weeks of AIS messages (1-14 October 2025) from Danish waters, published by the Danish Maritime Authority. The data was filtered to valid MMSI identifiers of commercial vessels, segmented when transmission gaps exceeded 15 minutes, resampled to fixed 6-minute intervals, and cleaned with physical plausibility limits; vessels moving slower than 6 km/h were excluded, which removes anchored and moored traffic.
Can I reproduce the TempTPI results?
Not straightforwardly from the paper alone: results are reported only as figures without numeric tables, architecture hyperparameters (including the frequency-expansion size K) are not listed, and no code repository is provided. The underlying AIS data is publicly available from the Danish Maritime Authority, so an independent reimplementation is possible but requires filling in unreported details.
Why does a 55% MSE improvement at 5 hours matter for maritime safety?
Trajectory forecasts typically degrade quickly at long horizons, and unreliable 5-hour-ahead positions limit vessel coordination and collision-avoidance planning. A substantially lower error at those horizons would be operationally valuable - though in this paper the gain is demonstrated against a single baseline on a two-week, single-region dataset, and some predicted trajectories still cross land boundaries.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- All quantitative results exist only as figures (Figs. 3-7); the extraction contains captions and prose claims (e.g., 'roughly 55%') but no plotted values, so numeric verification of results was impossible from the text.
- The value of K (temporal frequency range) and most architecture hyperparameters are genuinely absent from the entire provided text - a paper-level omission, not truncation.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.