A Multi-Scale Traffic-State Gated Attention Network for Bus Arrival Time Prediction

Abstract

Accurate bus arrival time prediction supports passenger information services and transit dispatching, but irregular sampling and time-varying traffic conditions make the task strongly nonlinear. This paper proposes a Spatial Attention Multi-Scale Traffic-State Gated Long Short-Term Memory (SA-MS-TG-LSTM) network. Seven trajectory features are constructed from Global Positioning System (GPS) records, including acceleration, congestion index, and speed volatility. Scale-preserving Spatial Feature Attention (SFA) first reweights the feature dimensions, and three Traffic-State Gated Long Short-Term Memory (TG-LSTM) branches subsequently encode the latest 5, 10, and 15 time steps. Their outputs are concatenated and mapped to the remaining travel time. Experiments are conducted on 8,834 GPS records from 293 trips on a single bus route (Route 86). Under a controlled setting in which all recurrent models use a hidden size of 16, one recurrent layer, and at most 50 training epochs, SA-MS-TG-LSTM achieves the best aggregate performance, with a root mean square error (RMSE) of 1.056 min, mean absolute error (MAE) of 0.762 min, mean absolute percentage error (MAPE) of 12.06%, and coefficient of determination ($R^2$) of 0.9631. Compared with the Convolutional Neural Network--Gated Recurrent Unit (CNN-GRU), it reduces RMSE, MAE, and MAPE by 6.30%, 5.57%, and 11.26%, respectively. Trip-level analysis further confirms statistically significant MAE improvements over a standard Long Short-Term Memory (LSTM) model and TG-LSTM. These results demonstrate the effectiveness of the proposed method on the evaluated route, while broader multi-route validation remains an important direction for future work.