Recurrent Neural Networks applied to Forecasting of Speed of Freight Transport in Dense Areas of Santiago, Chile

The speed prediction of freight transport is a growing and important task since in an ideal transport system the drivers could make optimal decisions about the route to follow. A computational method for predicting the the vehicle speeds is using autoregressive or other statistical models. Nonetheless, given the widespread success of modern neural networks, we believe that it could be feasible to have high prediction quality using this type of model. In this work, we propose the use of deep recurrent neural networks considering temporal multiple inputs for the speed prediction of freight vehicles. In particular, we use three models of neural networks for this task in specific dense areas of the city of Santiago of Chile. Standard metrics are reported to measure the models quality. Experiments showed that one of the proposed models using the information from the previous seven days give better results than other two proposed models. From the results, we conclude that these models have a good predictive capacity being able to predict the speed in different points on the Santiago map. As future work, we hope to test our models over a larger area, in addition to the incorporation of new variables that can influence and improve our models such as economic variables or relevant dates with high expected traffic.