Ship fuel consumption prediction is critical for green shipping, energy-efficiency management, and intelligent operational decision-making. However, real ship operational data exhibit multiple operating regimes, transitional states, and strong variable coupling, making it difficult for a single global model to achieve both high accuracy and interpretability. To address this challenge, this study proposes a physics-guided operating-regime-aware mixture-of-experts framework for ship fuel consumption prediction. An empirical propulsion-power proxy derived from the speed–power relationship is first introduced into a Gaussian mixture model through a latent physical power representation, consistency loss, and monotonicity loss, enabling physically meaningful operating-regime identification. Based on posterior responsibilities, confidence thresholds, and physical-load ranking, ship states are further categorized into stable, transitional, and complex regimes with clear engineering semantics. The identified regime information and statistical features of input sequences are then incorporated into a gating network to adaptively combine two BiLSTM-Attention experts. Experiments on real LPG-VLGC operational data collected from 2017 to 2019 demonstrate superior performance (MAE = 0.98, RMSE = 2.17, R² = 0.97) compared with XGBoost, GRU, CNN, single-expert models, and a standard GMM-MoE baseline. Results show that the proposed framework improves both prediction accuracy and engineering interpretability.