Research map / Time series / Causal and generative time series
Sequential Modeling of Complex Marine Navigation: Case Study on a Passenger Vessel (Student Abstract)
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Builds an auto-regressive Informer-plus-GRU time series model that forecasts ferry states (LAT, LON, SOG, FC) from actions and disturbances, and releases an offline RL dataset with a Gym environment for a passenger ferry.
Problem
Reducing vessel fuel consumption by modeling and forecasting ferry states from sensor logs, which prior models handle poorly due to lack of operator input, environmental factors, and temporal structure.
Approach
Combines a pretrained Informer transformer for auto-regressive multi-step forecasting with a GRU module that predicts residuals to curb cumulative auto-regressive error, and packages the two years of ferry data as an offline RL dataset with a three-stage reward curriculum Gym environment.
Key results
- AR + GRU achieves the best FC RMSE of 0.0608 and R^2 of 0.701, versus 0.0777 RMSE and 0.298 R^2 for AR alone.
- AR + GRU attains the best LON RMSE of 0.0604 and LAT R^2 of 0.874, compared with 0.0960 RMSE for AR on LON.
- Auto-regressive modeling generally outperforms non-auto-regressive prediction across the four ferry states, per the reported table.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add GRU refinement to autoregressive forecasting (ablation) | Informer forecasting model · Ferry FC | RMSE: 0.0777 -> 0.0608 | Helped |
| Add GRU refinement to autoregressive forecasting (ablation) | Informer forecasting model · Ferry SOG | RMSE: 0.0435 -> 0.0304 | Helped |
| Add GRU refinement to autoregressive forecasting (ablation) | Informer forecasting model · Ferry LAT | RMSE: 0.0613 -> 0.0385 | Helped |
| Add GRU refinement to autoregressive forecasting (ablation) | Informer forecasting model · Ferry LON | RMSE: 0.0960 -> 0.0604 | Helped |
| Use autoregressive forecasting with GRU rather than non-auto-regressive forecasting with GRU (ablation) | Informer forecasting model · Ferry FC | RMSE: 0.0769 -> 0.0608 | Helped |
| Use autoregressive forecasting with GRU rather than non-auto-regressive forecasting with GRU (ablation) | Informer forecasting model · Ferry SOG | RMSE: 0.0322 -> 0.0304 | Helped |
| Use autoregressive forecasting with GRU rather than non-auto-regressive forecasting with GRU (ablation) | Informer forecasting model · Ferry LAT | RMSE: 0.0684 -> 0.0385 | Helped |
| Use autoregressive forecasting with GRU rather than non-auto-regressive forecasting with GRU (ablation) | Informer forecasting model · Ferry LON | RMSE: 0.1097 -> 0.0604 | Helped |
Surprising findings
- The plain AR approach has a lower LAT standard deviation (0.0133) than AR + GRU (0.0137), and NAR + GRU has the lowest SOG standard deviation, so the full model is not best on every metric.
Limitations
- Evaluation is limited to a single ferry on a single constant route.
- The RL environment and reward stages are described but no RL agent results are reported.
- Std. and R^2 results are mixed, with the plain AR model winning on some metrics.
Experimental setup
Datasets: two-year ferry operational dataset (minute-level sensor records, west coast Canada)
Benchmarks: ferry operational dataset held-out prediction
Metrics: RMSE standard deviation R^2
Baselines: non-auto-regressive Informer non-auto-regressive Informer + GRU auto-regressive Informer
Builds on: Informer D4RL GRU
Closest papers by meaning
- Common Task Framework For a Critical Evaluation of Scientific Machine Learning Algorithms (NeurIPS 2025 · similarity 0.80)
- FuXi-Ocean: A Global Ocean Forecasting System with Sub-Daily Resolution (NeurIPS 2025 · similarity 0.80)
- From Generalization Analysis to Optimization Designs for State Space Models (ICML 2024 · similarity 0.80)
- Amortized Equation Discovery in Hybrid Dynamical Systems (ICML 2024 · similarity 0.79)
- Structured Matrix Basis for Multivariate Time Series Forecasting with Interpretable Dynamics (NeurIPS 2024 · similarity 0.79)
- Neptune-X: Active X-to-Maritime Generation for Universal Maritime Object Detection (NeurIPS 2025 · similarity 0.79)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.