Resumen
This study presents a working concept of a model architecture allowing to leverage the state of an entire transport network to make estimated arrival time (ETA) and next-step location predictions. To this end, a combination of an attention mechanism with a dynamically changing recurrent neural network (RNN)-based encoder library is used. To achieve this, an attention mechanism was employed that incorporates the states of other vehicles in the network by encoding their positions using gated recurrent units (GRUs) of the individual bus line to encode their current state. By muting specific parts of the imputed information, their impact on prediction accuracy can be estimated on a subset of the available data. The results of the experimental investigation show that the full model with access to all the network data performed better in some scenarios. However, a model limited to vehicles of the same line ahead of the target was the best performing model, suggesting that the incorporation of additional data can have a negative impact on the prediction accuracy if they do not add any useful information. This could be caused by poor data quality but also by a lack of interaction between the included lines and the target line. The technical aspects of this study are challenging and resulted in a very inefficient training procedure. We highlight several areas where improvements to our presented method are required to make it a viable alternative to current methods. The findings in this study should be considered as a possible and promising avenue for further research into this novel architecture. As such, it is a stepping stone for future research to improve public transport predictions if network operators provide high-quality datasets.