⚡ electrical engineering

Decision-Focused Continual Learning for Seaport Power-Logistics Scheduling: Generalization across Varying Tasks

This paper proposes a decision-focused continual learning framework that utilizes Fisher-information-based regularization and a differentiable convex surrogate to enable seaport power-logistics scheduling models to adapt online to varying vessel arrival tasks, thereby improving cross-task generalization and decision quality while maintaining sustainable computational and memory requirements.

Original authors: Chuanqing Pu, Feilong Fan, Nengling Tai, Yan Xu, Wentao Huang, Honglin Wen

Published 2026-06-09

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Chuanqing Pu, Feilong Fan, Nengling Tai, Yan Xu, Wentao Huang, Honglin Wen

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: The Seaport as a Busy Kitchen

Imagine a modern seaport is like a giant, high-tech kitchen.

The Ingredients: Electricity (power) and ships (logistics).
The Goal: The kitchen manager needs to decide exactly how much electricity to buy and how to use it to power the cranes and charge the ships, all while trying to save money.
The Problem: The manager has to make these decisions before they know exactly how many ships will arrive, how much cargo they have, or what the electricity price will be tomorrow. They have to guess (forecast) first, then plan (optimize).

The Old Way: The "Perfect Score" Chef

Traditionally, the kitchen manager used a computer to guess the future.

The Strategy: The computer was trained to make the "statistically perfect" guess. If the real electricity price was \ $100 and the computer guessed \$ 102, that was a "bad" guess. If it guessed $98, that was also a "bad" guess. The computer tried to be right on average, minimizing the distance between the guess and reality.
The Flaw: In a seaport, being "statistically right" doesn't always mean being "financially smart."
- Analogy: Imagine you are buying groceries. If you guess you need 5 apples but actually need 4, you waste money. If you guess 4 but need 5, you might have to buy the last one at a huge emergency markup. The cost of being "too low" is different than the cost of being "too high."
- The old computer didn't care about this difference; it just wanted to be close to the number. This led to expensive mistakes in the final schedule.

The First Upgrade: "Decision-Focused" Learning

The researchers introduced a smarter approach called Decision-Focused Learning (DFL).

The Strategy: Instead of training the computer to just be "close" to the real number, they trained it to make the best possible decision based on that number.
The Analogy: Instead of asking the computer, "How close is your guess to the truth?" they asked, "If you make this guess, how much money will we save or lose?" The computer learned to make guesses that might be slightly "wrong" statistically but would result in the cheapest electricity bill.
The New Problem: This worked great for one specific day with a specific set of ships. But seaports are chaotic. Tomorrow, a different number of ships might arrive, or they might carry different cargo. This changes the "rules of the game" (the optimization problem).
The Failure: When the researchers tried to use the "Decision-Focused" model on a new day with different ships, it forgot everything it learned about the previous days. It was like a chef who memorized the recipe for Tuesday's menu but couldn't cook Wednesday's menu because the ingredients changed slightly.

The Solution: "Continual Learning" with a Safety Net

The paper proposes a new system called Decision-Focused Continual Learning (DFCL). Think of this as a chef who learns every day but never forgets the basics.

1. The "Fisher" Safety Net (Elastic Weight Consolidation)

The Concept: As the chef learns a new menu (a new task), they might accidentally unlearn how to cook the old menu. To stop this, the system uses a "Fisher-information-based regularization."
The Analogy: Imagine the chef's brain is a sponge. When they learn a new recipe, they usually soak up new water and squeeze out the old. This new method puts elastic bands around the parts of the sponge that hold the most important "old knowledge."
How it works: The system identifies which parts of the computer's "brain" (parameters) are critical for making good decisions on past days. When learning a new day, it puts "elastic bands" on those parts so they don't stretch too far. This allows the model to adapt to the new ships without forgetting how to handle the old ones.

2. The "Smooth Surrogate" (Differentiable Convex Surrogate)

The Concept: The math behind scheduling ships and cranes is incredibly complex and "bumpy" (non-convex), making it hard for the computer to learn from its mistakes.
The Analogy: Imagine trying to roll a ball down a mountain with jagged rocks and cliffs. The ball gets stuck, and you can't tell which way to push it to get to the bottom.
The Fix: The researchers created a "smoothed-out" version of the mountain (a differentiable surrogate). It's like putting a smooth sheet of plastic over the jagged rocks. The computer learns on this smooth surface, figuring out the best direction to push. Once it learns the direction, it applies that logic to the real, rocky mountain. This makes the learning process stable and fast.

The Results: Why It Matters

The researchers tested this on data from Jurong Port in Singapore.

The Test: They simulated a stream of 6 different days, each with different ship arrival patterns.
The Winner: The new DFCL system was the champion.
- It saved more money than the old "statistical" methods.
- It saved more money than the "Decision-Focused" method that didn't use the "elastic bands" (it forgot too much).
- Efficiency: It didn't need to store massive amounts of old data or retrain from scratch every day. It kept a small, manageable memory footprint, making it practical for real-world use.

Summary

This paper solves a problem where smart computers get "amnesiac" when the rules of a seaport change. By adding elastic bands to protect old knowledge and using a smoothed-out map to guide learning, the new system allows a seaport to continuously learn and adapt to changing ship schedules without forgetting how to save money. It turns a "one-trick pony" into a "seasoned veteran" that gets smarter every day.

Problem Definition

Modern seaports operate as Power–Logistics Systems (PLS) where energy procurement and logistics operations (e.g., vessel berthing, quay-crane scheduling) are co-optimized. The prevailing industrial practice follows a predict-then-optimize pipeline: forecasting models predict net load and electricity prices, which are then fed into an optimization scheduler.

The paper identifies a critical limitation in existing Decision-Focused Learning (DFL) approaches. While DFL aligns forecasting training with downstream decision outcomes (minimizing operational regret rather than statistical error like MSE), standard DFL is inherently task-specific. In seaport operations, the optimization task structure evolves dynamically with vessel arrivals. Each new arrival configuration introduces distinct constraints, decision variable dimensions, and feasible region geometries. Consequently, a DFL model trained on a specific task configuration fails to generalize to new, unseen task structures, leading to performance degradation when applied sequentially to evolving operational scenarios.

Methodology

The authors propose a Decision-Focused Continual Learning (DFCL) framework designed to adapt online to a stream of varying scheduling tasks without catastrophic forgetting or the need for data replay.

1. Problem Formulation:
The framework treats the scheduling problem as a stream of tasks $\{\tau_1, \dots, \tau_K\}$ . The objective is to learn forecasting parameters $\Theta$ that minimize the cumulative regret across all encountered tasks. The regret is defined as the difference between the actual cost incurred using predicted inputs and the cost achievable with perfect foresight.

2. Fisher-Information-Based Regularization:
To prevent forgetting knowledge from previous tasks while adapting to new ones, the authors derive a regularization method inspired by Elastic Weight Consolidation (EWC).

Pseudo-Likelihood Assumption: They assume the regret loss induces a Gibbs-type pseudo-likelihood.
Laplace Approximation: The posterior distribution of parameters from previous tasks is approximated as a Gaussian distribution.
Fisher-Like Curvature: The precision matrix of this Gaussian is approximated using the empirical Fisher information matrix, calculated as the outer product of regret gradients.
Objective Function: The learning objective for a new task $k$ minimizes the current task's regret plus a quadratic penalty term: $\lambda (\Theta - \Theta^*_{k-1})^\top \Lambda_{k-1} (\Theta - \Theta^*_{k-1})$ . This constrains updates to parameters critical for prior tasks.

3. Differentiable End-to-End Pipeline:
A major challenge is that the PLS scheduling problem involves mixed-integer constraints (discrete decisions for vessel assignment and crane scheduling), making it non-differentiable.

Surrogate Model: The authors introduce a differentiable convex surrogate based on a K-Nearest Neighbors (KNN) regression. This model predicts discrete logistics variables (quay-crane power and vessel assignment) based on upstream forecasts.
Convexification: By fixing these predicted discrete variables, the remaining scheduling problem becomes a convex Linear Programming (LP) problem.
Implicit Differentiation: The gradients of the regret loss with respect to the forecasting parameters are computed by differentiating through the convexified optimization layer using implicit differentiation (via KKT conditions).

Key Contributions

Novel Framework: The paper proposes the first task-adaptive predict-then-optimize framework for PLS scheduling that enables continual learning of decision-aligned forecasting models over an unbounded stream of tasks.
Regularization Mechanism: It introduces a Fisher-information-based regularization method that enhances cross-task generalization without the prohibitive computational overhead of replaying historical data or retraining on past tasks.
Differentiable Pipeline: It develops an end-to-end learning pipeline that utilizes a differentiable surrogate to convexify the embedded mixed-integer scheduling problem, stabilizing gradient backpropagation via implicit differentiation.

Experimental Results

The method was evaluated using a benchmark calibrated to Jurong Port, Singapore, combining real vessel arrival data with synthetic load and price data. Six distinct task scenarios were created to simulate evolving vessel arrival patterns.

Performance vs. Baselines: The proposed DFCL (EWC) consistently outperformed:
- Statistical Baseline Learning (SBL): Which minimizes MSE but ignores decision value.
- Task-Specific DFL: Which fails to generalize to new task structures.
- Naïve Continual DFL (Adam): Which suffers from catastrophic forgetting.
- Freezing-based Continual DFL (Adam-F): Which offers limited adaptability.
Metrics: DFCL achieved the lowest cumulative regret and regret ratio (RR) across the task stream. It reduced cumulative regret by up to 2.5% and the forgetting measure (FM) by up to 27.7% compared to naïve continual learning.
Generalization: The model maintained low regret on earlier tasks even after learning subsequent, structurally different tasks, demonstrating robust cross-task generalization.
Efficiency: DFCL demonstrated bounded computational and memory costs. Unlike multi-task learning or task-wise retraining, DFCL's training time and memory footprint remained nearly constant as the number of tasks increased, avoiding the need to store historical data or multiple models.
Sensitivity: The method remained robust to imperfect real-time updates and showed stable performance across different task ordering sequences.

Significance and Claims

The paper claims that the proposed DFCL framework addresses a specific gap in energy systems literature: the inability of current decision-focused methods to handle evolving task structures common in real-world logistics.

Practicality: The approach allows seaport operators to sustain high-quality scheduling decisions under changing operational conditions (e.g., varying fleet sizes and cargo profiles) without the computational burden of retraining bespoke models for every new scenario.
Relative Advantage: The authors modestly state that the results demonstrate relative gains under controlled task-structure shifts using a mixed-source benchmark. They do not claim the assembled dataset perfectly reproduces all specific Singapore market conditions but argue that the algorithmic advantage in handling structural task variations is the primary contribution.
Future Work: The paper notes that future validation should involve geographically co-located port datasets and extensions to rolling-horizon or multi-stage corrective scheduling with partial recourse.