Investigation of different LSTM-based encoder-decoder neural networks for vehicle speed prediction

In this work, different Long Short-Term Memory (LSTM) encoder-decoder artificial neural networks are investigated. These networks differ in their complexity. The aim of this work is to evaluate whether complex networks are necessary for vehicle speed predictions or whether simple and less computing power intensive networks can handle this task also with sufficient accuracy. For this task, simulatively generated data is used, which is created with a Simulation of Urban Mobility (SUMO) traffic simulation from an urban traffic scenario from the city center of Darmstadt, Germany. The data generating process is described as well as the data handling and processing. For the investigated network architectures, grid searches are executed to investigate the sensitivity to four main hyperparameters, the mini batch size, the learning rate, the weight decay and the number of LSTM cells within each layer. The results are then evaluated based on their accuracy regarding a test data set and based on the computing power required for training. The results presented in this work indicate that also less complex models can handle the task of speed predictions and, at least for these applications, simple models should be considered in order to save computing power and, as a consequence, also energy.