Prediction of fuel consumption for operational ships based on an improved CNN–BiLSTM–Attention model

In complex marine environments, ship fuel consumption is jointly affected by hull motion, propulsion load, and wind–wave conditions, while operational data are often intermittent and non-stationary, which makes accurate prediction challenging. Based on full-scale data from a container ship, this study develops a multi-source data-driven framework that fuses AIS navigation information, onboard sensor measurements, and ERA5 metocean reanalysis on a unified spatiotemporal grid. Sailing-state filtering and fixed-length sliding windows are used to retain continuous propulsion segments, where multi-step multi-source features form input sequences and fuel consumption at the window end is taken as the prediction target. An improved CNN–BiLSTM–Attention model is then constructed: convolutional layers extract local “sea state–operating condition” patterns, BiLSTM captures temporal dependencies, and a temporal attention mechanism adaptively weights different time steps. Non-negativity of fuel consumption and power/speed consistency are embedded into the loss function as physical constraints. Ablation studies and wind–wave regime analyses show that the proposed model achieves R 2 = 99.28%, RMSE = 0.0178 t/h, and MAPE = 1.35%, and that removing environmental or engine-room features significantly degrades accuracy, while attention weight visualization confirms the model’s interpretability, thereby confirming both the effectiveness and interpretability of the improved CNN–BiLSTM–Attention model for fuel consumption prediction under complex sea conditions.