Fuel-Saving Route Planning with Data-Driven and Learning-Based Approaches - A Systematic Solution for Harbor Tugs

Shengming Wang, Xiaocai Zhang, Jing Li, Xiaoyang Wei, Hoong Chuin Lau, Bing Tian Dai, Binbin Huang, Zhe Xiao, Xiuju Fu, Zheng QinIJCAI 2024

Introduction

Maritime transport is facing increasing pressure from the International Maritime Organization (IMO) to decarbonize, with ambitious targets for reducing carbon intensity by 2030 and 2050. While much of the industry's focus remains on large ocean-going vessels, harbor tugs—the essential "workhorses" of port operations—represent a significant opportunity for emission reduction. Operating in crowded and complex environments like the Port of Singapore, these vessels often prioritize operational speed over fuel efficiency, leading to suboptimal consumption patterns.

The primary challenge in optimizing tugboat operations is the highly non-linear nature of fuel consumption. It is influenced by a complex interplay of navigational states (speed, course) and environmental factors (tides, wind, water depth). Traditional routing methods often rely on human intuition or simple geometric logic, which fails to account for the hidden fuel costs of fighting currents or maintaining inefficient speed profiles. This paper presents a systematic, AI-driven solution to minimize fuel consumption and carbon emissions for harbor tugs.

Methodology: A Two-Tiered AI Architecture

To address the complexities of maritime navigation, the researchers proposed a two-tiered architecture that combines high-fidelity fuel prediction with intelligent route planning.

1. Fuel Consumption Estimation (The Ensemble Model)

Standard machine learning models often struggle with the inherent data imbalance in tugboat operations; vessels spend the majority of their time at low speeds, making high-fuel consumption scenarios "rare events" in training datasets. To overcome this, a hybrid ensemble model was developed:

  • Long Short-Term Memory (LSTM): Used for low-consumption states where time-dependent historical features, such as acceleration trends, are most predictive.
  • Multilayer Perceptron (MLP): Utilized for high-consumption states to capture the immediate, "instantaneous" relationship between high speed and fuel flow.

2. Route Planning via Reinforcement Learning

For path optimization, the system utilizes the Deep Deterministic Policy Gradient (DDPG) algorithm. DDPG is particularly well-suited for this application as it handles continuous action spaces, allowing the agent to select precise headings and speeds rather than discrete steps.

System architecture combining the fuel consumption prediction ensemble (LSTM + MLP) with the DDPG-based reinforcement learning framework.

The framework incorporates a comprehensive state space—including position, Course Over Ground (COG), wind, current, and water depth—and a specialized reward function. To ensure navigational safety, the reward function integrates a waterway distance API based on the Theta* algorithm, preventing the agent from proposing routes through land or piers.

Key Insights: The "8-Knot" Pivot

Data analysis of tug operations in the Port of Singapore revealed a critical threshold in fuel efficiency. As shown in the figure below, fuel consumption remains relatively stable at lower speeds but experiences a significant, non-linear spike once the Speed Over Ground (SOG) exceeds approximately 8 knots.

Relationship between Speed Over Ground (SOG) and Fuel Oil (FO) rate, highlighting the non-linear spike after approximately 8 knots.

This "8-knot pivot" served as a foundational insight for designing the reinforcement learning reward structures, encouraging the agent to find the optimal balance between transit time and fuel economy. Furthermore, the study found that traditional oversampling techniques were insufficient for LSTMs in this context, as they amplified noise in time-series sequences, justifying the need for the hybrid LSTM-MLP ensemble approach.

Experimental Results

The proposed system was rigorously evaluated against historical human-operated transits within the Singapore port environment. The ensemble model achieved a Mean Squared Error (MSE) of 2.50, significantly outperforming standalone LSTM (3.76) and MLP (5.77) models.

In real-world case studies, the AI-optimized routes demonstrated dramatic improvements over human-guided paths:

  • Route Irregularity: In scenarios with inefficient historical paths, the AI reduced fuel consumption by ~74% (from 306.8L to 79.7L) by suggesting more direct, terminal-adjacent paths.
  • Path Streamlining: By reducing unnecessary turns and optimizing the curvature of the route, fuel savings of ~15% were achieved.
  • Speed Optimization: By leveraging favorable environmental conditions and adhering to efficient speed profiles, the system reduced fuel use by ~35%.

Conclusion

This research demonstrates that AI-driven route planning can offer a systematic solution for decarbonizing harbor operations. While the current results are based on retrospective simulations, they provide a robust foundation for real-time bridge-support systems. Future work will focus on integrating dynamic obstacle avoidance for moving traffic and adapting the model for vessels utilizing alternative "green" fuels like methanol or ammonia. While currently tuned for the Port of Singapore, the methodology is designed to be generalizable to other global maritime hubs through retraining on local environmental and traffic data.

Summary updated

BibTeX
@inproceedings{wang2024fuel,
  author = {Shengming Wang and Xiaocai Zhang and Jing Li and Xiaoyang Wei and Hoong Chuin Lau and Bing Tian Dai and Binbin Huang and Zhe Xiao and Xiuju Fu and Zheng Qin},
  title = {{Fuel-Saving Route Planning with Data-Driven and Learning-Based Approaches - A Systematic Solution for Harbor Tugs}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2024},
}
Figure