Abstract
This study presents a hybrid deep learning architecture designed to predict the operational state of a passenger ferry, integrating an Informer (Transformer) with a Gated Recurrent Unit (GRU). By modeling fuel consumption and navigation auto-regressively, the proposed approach effectively addresses the challenge of cumulative error drift in long-sequence forecasting. Experimental results using two years of real-world maritime data demonstrate significant improvements in predictive accuracy, with the AR+GRU model achieving an $R^2$ of 0.701 for fuel consumption and 0.899 for longitude. Furthermore, the authors introduce an open-source reinforcement learning environment to facilitate the development of AI-driven navigational optimization.
Introduction
The maritime industry is under significant pressure to reduce its carbon footprint, with fuel consumption identified as the primary target for optimization. However, modeling vessel fuel consumption is inherently difficult because existing approaches often ignore the "human in the loop"—real-time operator inputs—or fail to account for the complex temporal dependencies of a ship at sea.
Static models frequently struggle with the "drift" of real-world navigation, where small errors in prediction accumulate over time, leading to significant deviations. Furthermore, there has been a lack of standardized, reality-based simulation environments where reinforcement learning (RL) agents can be trained to improve upon human decision-making in a maritime context. This research addresses these gaps by proposing a robust sequential model and an accompanying open-source simulation framework.
Methodology
The authors propose a sequential model architecture designed to handle the noisy and time-sensitive nature of marine data. The architecture consists of three primary components:
- The Informer (Transformer): The core of the model utilizes an "Informer" architecture, a type of Transformer optimized for long-sequence time-series forecasting. It performs feature fusion, integrating dynamic states (e.g., current speed), static features (e.g., docking area), and external disturbances such as wind and weather.
- GRU Refinement: To solve the problem of cumulative error in auto-regressive models, the authors added a Gated Recurrent Unit (GRU). The GRU is tasked with predicting the residual (the error) of the Transformer's output, effectively "cleaning up" the prediction at each step.
- Auto-Regressive (AR) Modeling: Unlike non-auto-regressive models that predict a whole window at once, this model uses its own previous outputs to inform the next step, mimicking how a ship actually moves through time.
Datasets and Preprocessing
The study utilizes a robust, real-world dataset sourced from a passenger ferry operating on the west coast of Canada (BC Ferries). The dataset covers over 3,000 transits over two years, with sensor data recorded at one-minute intervals.
Beyond raw data usage, the authors applied domain-driven preprocessing. They categorized operating modes (e.g., autopilot vs. docking) and applied power transformations to skewed data distributions to improve neural network performance. To encourage further research, the authors have open-sourced a reinforcement learning-compatible version of this dataset (formatted for the D4RL framework) and a corresponding Gym environment.
Key Results
The AR + GRU (Auto-Regressive with GRU refinement) approach significantly outperformed standard models across all key metrics, including Fuel Consumption (FC), Speed Over Ground (SOG), Latitude (LAT), and Longitude (LON).
As shown in the results table, the addition of the GRU refinement improved the $R^2$ and lowered the Root Mean Square Error (RMSE) compared to the standalone Transformer model in every category. Notably, for Fuel Consumption, the model achieved an $R^2$ of 0.701, compared to a dismal -0.023 for the non-auto-regressive baseline. For Longitude, the $R^2$ reached 0.899, indicating a very high degree of precision in tracking the vessel's path.
Discussion and Future Work
The most significant architectural insight is the use of the GRU to mitigate cumulative errors. In time-series forecasting, small inaccuracies in predicting latitude or longitude can snowball; the GRU acts as a continuous error-correction mechanism. By framing the problem as an RL environment, the authors move beyond just predicting what will happen to creating a sandbox where AI can eventually decide the best way to navigate.
While the current work focuses on the predictive architecture, future research will leverage the newly created RL environment to discover navigational best practices. The ultimate goal is to deploy these models as decision-support tools that provide real-time feedback to crew members, minimizing environmental impact without sacrificing schedule reliability.
Summary updated