Distributed modeling of smart parking system using LSTM with stochastic periodic predictions

Abstract

Parking in contemporary cities is a time- and fuel-consuming process. It affects daily stress levels of drivers and citizens. To design the future cities, parking process should be handled efficiently to improve drivers’ time comfort and fuel economy toward a green smart city (SC) ecosystem. In this paper, we propose to model smart parking (SP) with multiagent system (MAS) using long short-term memory (LSTM) neural network. Our model outperforms similar approaches as evidenced from the presented results using an online dataset from the SC of Aarhus, Denmark. We use LSTM for stochastic prediction based on periodic data provided by parking sensors. A SP provides such data on daily basis over a short period of time in the SC. We evaluate the proposed MAS with the prediction accuracy metric and compare it with other approaches in the literature. The proposed system achieves higher prediction accuracy per daily basis than the compared approaches due to our stochastic periodic prediction design and input to the proposed MAS and LSTM model. In addition, LSTM is used more efficiently under the proposed architecture of MAS, which enables online scaling thanks to dynamic and distributed nature of MAS.