Research on an Improved Transformer-Based Model for Inland Vessel Fuel Consumption Prediction

To tackle the challenges of low accuracy and inefficiency in predicting inland vessel fuel consumption, an improved Transformer-based model is proposed. The model integrates Locally Linear Embedding (LLE), a patch embedding strategy, and a Two-Stage Attention (TSA) mechanism to enhance both prediction accuracy and computational efficiency. LLE reduces data dimensionality while preserving critical features, easing the training process. The patch embedding strategy streamlines time-series data processing, minimizing memory usage and accelerating training. The TSA mechanism, combining Content-Based Attention (CBA) and Self-Attention (SA), improves the model’s ability to capture intricate patterns. Experimental results show that the proposed model significantly outperforms traditional BP neural networks and LSTM models in fuel consumption prediction, achieving superior MSE and MAE metrics with shorter training times. This enhanced model offers a precise and efficient tool for fuel consumption prediction, supporting inland shipping companies in achieving sustainable development.