Static Stochastic Model-Based Prediction of City Bus Velocity

This paper proposes a static, stochastic, deep feed-forward neural network-based model for prediction of city bus velocity along a regular route. The emphasis is on a proper formation of model outputs to consistently learn the conditional probability distribution of vehicle velocity based on the vehicle position as only input feature. First, a rich set of recorded driving cycles of a representative fleet of ten city buses is statistically analyzed. Next, the recorded dataset is properly downsampled and used for computationally-efficient training and validation of the neural network. Finally, the prediction accuracy is demonstrated on a test dataset by considering different prediction quality indices.