Distributed Points Data Driven LSTM Neural Networks for Daily Traffic Forecasting *

Daily traffic forecasting is crucial for optimizing traffic operations, providing traveler information, assisting traffic control on events, supporting incident management, etc. However, we are still confronted with many challenges, mainly due to the medium- and long-term forecasting horizon, which requires details and trends, and the difficulty of obtaining complete traffic data for large networks. In this paper, we propose a distributed points data driven model and apply a hybrid graph attention based long-short term memory (GAT -LSTM) architecture to carry out a 24-hour traffic forecasting task. External factors of holidays, extreme weather changes, and vacation habits of local people and other special periods of time are also encoded to improve forecasting accuracy. A case study with real-world traffic datasets from Madrid city is introduced to demonstrate the validity and usability of the model. Experimental results are also used to reveal the characteristics of the medium- and long-term forecasts from the perspective of the spatial and temporal granularities of traffic data.