Predict-then-Optimize for Seaport Power-Logistics Scheduling: Generalization across Varying Tasks Stream

TL;DR

DFCL tackles the problem of generalizing decision-focused forecasting under evolving seaport task streams by introducing a Fisher information-based regularization to preserve past decision mappings while learning new tasks, and a differentiable surrogate to enable end-to-end gradient-based training through the non-convex scheduling layer. The framework combines a Transformer-based forecasting module, a memory-guided surrogate for convexification, and KKT-based implicit differentiation to propagate regret signals through day-ahead and real-time optimizers. Empirical results on Jurong Port show that DFCL outperforming standard DFL and offline baselines in both decision quality and computational efficiency, with EWC-based variants offering the strongest cross-task generalization and memory protection. The work demonstrates a practical, scalable approach to continual learning in predict-then-optimize settings, with potential applicability to broader planning problems facing evolving task distributions in logistics and energy systems.

Abstract

Power-logistics scheduling in modern seaports typically follow a predict-then-optimize pipeline. To enhance the decision quality of forecasts, decision-focused learning has been proposed, which aligns the training of forecasting models with downstream decision outcomes. However, this end-to-end design inherently restricts the value of forecasting models to only a specific task structure, and thus generalize poorly to evolving tasks induced by varying seaport vessel arrivals. We address this gap with a decision-focused continual learning framework that adapts online to a stream of scheduling tasks. Specifically, we introduce Fisher information based regularization to enhance cross-task generalization by preserving parameters critical to prior tasks. A differentiable convex surrogate is also developed to stabilize gradient backpropagation. The proposed approach enables learning a decision-aligned forecasting model across a varying tasks stream with a sustainable long-term computational burden. Experiments calibrated to the Jurong Port demonstrate superior decision performance and generalization over existing methods with reduced computational cost.

Figures (18)

• Figure 1: Illustration of different learning paradigms for predict-then-optimize pipelines.
• Figure 2: Illustration of the loss landscapes for SBL vs. DFL (left) and DFL vs. DFCL (right). The x-axes and y-axes represent two dimensions of the parameter space, and the differently colored regions denote low-loss areas of distinct loss landscapes. The arrows indicate different learning trajectories.
• Figure 3: End-to-end training pipeline for DFCL in PLS. Each module represents a differentiable computational unit.
• Figure 4: Loss heat maps of a DFL model trained on Task 1 and evaluated on Tasks 1 to 6. The x-axis represents the load forecast error (actual minus predicted), while the y-axis represents the price forecast error. The color intensity indicates the regret value, with colder colors representing higher regret.
• Figure 5: Regret trends on Tasks 1 to 5 and average regret across all tasks after learning each task.