Travel time estimation (TTE) is a critical function in intelligent driving systems. Current research and applications related to TTE primarily focus on urban environments. The objective of this study is to develop TTE methods that are applicable to wilderness areas characterized by plateau and mountainous topography. We selected Transformer, which has greater robustness in capturing long-distance dependencies than LSTM, to develop a Transformer-based model. The model simultaneously integrates positional encoding and multi-head self-attention mechanisms with the objective of enhancing the accuracy of travel time predictions based on a substantial number of trajectory points in wilderness settings. A meta-learning strategy was employed to improve the model's generalization ability, thereby ensuring its applicability for accurate travel time estimation across a range of challenging environments. Two datasets were constructed based on measurements from two selected areas in eligible plateau and mountainous regions of western China. For each dataset, two categories of features were defined: terrain-weather features and spatio-temporal features. These categories were established in accordance with the influence of seven specific features on traffic conditions in both urban and wilderness areas. Experiments were conducted on both datasets utilizing terrain-weather features. When evaluated alongside the five models that are most commonly utilized in urban settings, the mean absolute percentage error (MAPE) of our model exhibited a 14.89% improvement in plateau environments and a 12.20% improvement in mountainous environments in comparison with the most effective model, namely MetaTTE-GRU. These findings substantiate the assertion that the proposed model is an effective means of estimating travel times in complex environments, and that it exhibits superior accuracy compared to existing LSTM-based estimation models.
Read full abstract