Enhancing Fleet Management with GPS Data in Port Terminals: A Deep Learning Architecture for Truck Position Prediction

Abstract

Effective truck fleet management in port terminals is crucial for improving operational efficiency and reducing delays. However, GPS-based positioning systems often face limitations in port environments due to signal obstructions and network issues. This study explores the utilization of advanced methods for deep learning, particularly Recurrent Neural Networks (RNNs) for predicting truck positions when GPS signals are unavailable or unreliable. The study utilizes GPS data from an operation terminal, transformed into local coordinate systems, to train and evaluate the models. We propose a novel approach that integrates convolutional layers (CNN) to extract temporal features such as speed and acceleration, followed by RNN-based models to capture sequential dependencies, and finally, a Multi-layer Perceptron (MLP) layers to refine the position estimates. The model that is founded on Gated Recurrent Unit (GRU) performed the best with a minimum Mean Position Error (MPE) of 1.56 meters, outperforming LSTM (Long Short-Term Memory) and Simple RNN models in terms of both reliability and computational efficiency. The results highlight the model's capability to predict future truck positions, offering significant potential to improve operational efficiency in port terminals.