Short-term forecasting for ship fuel consumption based on deep learning.

Improving ship energy efficiency and intelligent optimization depend heavily on predictive maintenance of Marine diesel engine performance. For successful Condition-Based Maintenance, a multi-step fuel consumption prediction of ships that is accurate and stable is needed. However, existing methods mainly focus on current time or future single-step forecasts. Therefore, it is essential to investigate the optimum prediction model across various prediction time steps from the perspective of model accuracy and model generalization capability. Based on the 14-month sensor data of bulk carriers, high-quality ship energy consumption data is obtained using the local weighting method to establish a short-term multi-step prediction model of engine fuel consumption based on deep learning. Five real fuel consumption sample sets with different equilibrium levels were determined to evaluate the robustness and generalization of varying prediction models. According to the research, the ensemble empirical mode decomposition-based memory network (EEMD-LSTM) can maintain good stationarity and high accuracy in long-term trend prediction within 30 to 60 steps. In contrast, the bidirectional memory network (BiLSTM) has high accuracy in short-term volatility prediction within 30 steps. An efficient method for ship prediction maintenance and defect diagnosis can be found in a high-precision multi-step forecast method for Marine diesel engine fuel consumption. • The prediction models consider the impact of various dataset equilibrium levels. • Only using historical fuel data for consumption shows predictive solid capability. • We have innovatively identified more appropriate models for different time steps. • Confirmed LSTM and BiLSTM's prediction feasibility through model generalization. • Before step 30, BiLSTM's MAPE is 3.8%; EEMD-LSTM's MAPE is 1.9% from steps 30-60. [ABSTRACT FROM AUTHOR]