Multi-stage deep probabilistic prediction for travel demand

摘要

Accurate demand prediction is an essential component of any decision support system for smart vehicle dispatching. However, predicting real time demand at the micro-geographical level is a challenging task due to its underlying complexity, and prior work has focused largely on predicting one-step-ahead demand at a macro-geographical level, often using classical times series models that do not consider exogenous factors or quantify the prediction uncertainty in a probabilistic setting. In this paper, we propose an end-to-end deep learning-based framework with a novel architecture to predict multi-step-ahead real time travel demand, along with uncertainty estimation. The model is multi-stage and captures both spatial and temporal aspects. We employ the encoder-decoder framework with variations of convolution and LSTM units, and incorporate an attention mechanism into the model to quantify the interdependence between the input and output elements. We illustrate our model using two real data sets: a taxi and a ride-sharing service (Uber) for the city of New York.