Vol 77 No 1 77101

Abstract

Marine diesel engines play a critical role in ensuring vessel safety and stable operation. Among their operating parameters, cylinder exhaust gas temperature is a key indicator of an engine’s thermal performance and overall health. Accurate forecasting of this parameter is therefore essential for real-time monitoring and condition-based maintenance in intelligent ship propulsion systems. This study proposes a hybrid CNN-LSTM-Attention model for long-range prediction of marine engine cylinder exhaust temperature. The model utilizes real-world sensor data acquired from shipboard monitoring systems, rather than synthetic or simulated datasets. Experimental results demonstrate the model’s superior performance, achieving a Mean Absolute Error (MAE) of 3.9944, a Mean Absolute Percentage Error (MAPE) of 1.3481 %, and a coefficient of determination (R²) of 0.9805. These metrics indicate high prediction accuracy and minimal deviation from actual values compared to conventional algorithms. To further enhance the model’s predictive capability and generalization performance, Particle Swarm Optimization (PSO) was applied. This automatically fine-tune key hyperparameters, including the optimal parameter settings for learning rate and the number of LSTM units. Results confirm that PSO optimization improves the CNN-LSTM-Attention model’s predictive performance. The predicted temperature curves align better with actual measurements. Specifically, the MAE was reduced by 1.5431, the MSE decreased by 14.3743, and the R² value moved closer to 1. Leveraging PSO’s self-adaptive search capability avoided the need for complex manual hyperparameter tuning. Overall, the findings confirm that the optimized model delivers highly accurate predictions of cylinder exhaust temperature in intelligent marine engines.