Adaptive Region Distillation: A lightweight knowledge distillation method for ship fuel consumption prediction

Accurate ship fuel consumption prediction is essential for energy efficiency optimization, route planning, and emission reduction. However, interactions among operational and environmental factors produce local fluctuations and heteroscedasticity across operating conditions, making it difficult for lightweight onboard models to preserve predictive structures learned by high-capacity models. Existing regression knowledge distillation methods mainly align point predictions or global predictive distributions, providing limited representation of local and relative structure in continuous output spaces. To address this limitation, this paper proposes Adaptive Region Distillation (ARD), a knowledge distillation framework with region decoupling for ship fuel consumption regression. ARD constructs adaptive proxy distributions around observed targets and decouples knowledge transfer into target-neighborhood and complementary-neighborhood objectives. ARD was evaluated on operational records from two container ships and one cargo ship using MLP and LSTM backbones. Relative to the standalone Student, ARD increased mean R 2 by 0.0308–0.0870 for MLP and 0.0070–0.0954 for LSTM. Relative MAPE reductions were 16.2%–33.5% for MLP, while absolute reductions for LSTM were 0.7256–2.1721 percentage points. Ablation and sensitivity analyses characterized component contributions and hyperparameter responses. ARD introduces no additional parameters or branches at inference time, retaining parameter compression ratios of 167.67 × and 252.17 × for the MLP and LSTM students, respectively.