How Akvelon developed an ML-based discharge forecast model using real terminal data.
Project Overview
A large logistics hub operator sought to make its resource scheduling more predictable and efficient.
With complex workflows involving vessel discharge and container movements, even small timing delays led to measurable cost and throughput impact.
The Challenge
Inconsistent vessel discharge times created uncertainty in berth allocation and resource planning.
This unpredictability triggered a chain reaction:
- Berths sometimes sat idle, wasting valuable capacity.
- Staffing levels didn’t always match workload, leaving gangs overworked or underutilized.
- Yard congestion slowed down container movements.
- Missed SLAs created a real risk of financial penalties.
Manual estimates were inconsistent and rarely accounted for key operational factors such as crane count, gang assignments, and yard congestion.
The operations team needed a reliable, data-driven method to accurately predict discharge duration, enabling more practical and efficient resource and scheduling planning.
The Solution
Akvelon’s data science team collaborated with port operations experts to design a machine learning prediction pipeline ready for integration into existing planning tools.
Service: Predictive Modeling, Feature Engineering, ML Model Development
Tech Stack: Oracle SQL, Python, Scikit-Learn, Docker
Before diving into details, here’s an overview of the end-to-end workflow — from raw data extraction to production-ready integration.
Detailed Breakdown of Implementation Steps
The following section describes how each stage was implemented in this project.
- Data Foundation
Extracted raw operational event data from terminal systems and defined vessel visits using advanced SQL sessionization logic. - Feature Engineering
Aggregated factors such as container counts, crane and gang assignments, shift patterns, and holidays into a reproducible Python pipeline. - Model Development
- Tested multiple regression models to predict discharge duration.
- Linear Regression achieved a mean absolute error (MAE) of 5.14 hours and an R² of 0.65, outperforming more complex models while remaining interpretable.
- The model was trained and validated on over 2,200 vessel visits spanning 18 months.
- Deployment-Ready Design
Created a modular architecture capable of incorporating real-time data (e.g., weather, GPS, tides) and integrating with terminal planning dashboards, laying the groundwork for future production deployment.
Results
Even at prototype stage, the model provided actionable insights for planners:
- Key drivers identified: Discharged container counts and gang assignments had the greatest impact on discharge time.
- Data quality improvements: The modeling process exposed event-logging inconsistencies, leading to better data hygiene.
- Operational potential: Early testing indicated strong readiness for real-time updates and simulation use.
- Planning value: The model helped forecast high-duration vessels before arrival, enabling smarter staffing and yard allocation.
What’s Next
The next iteration aims to boost accuracy by integrating:
- Real-time weather and wind data
- Vessel GPS feeds for delay tracking
- Tide and marine condition indicators
We've also prepared automated daily feature updates to include in terminal dashboards.
Why It Matters for Business
Accurate discharge forecasts help marine logistics operators reduce idle time, optimize resource use, and maintain service reliability.
Akvelon combines AI and ML expertise with domain experience in complex operations to deliver solutions that turn raw data into measurable efficiency gains.