Multistage prediction approach of EVs charging performance in smart transportation systems by deep learning technique

Abstract

Electric Vehicles (EVs) are increasingly recognized as a fundamental component of intelligent transportation systems within smart city frameworks. Therefore, several studies in recent decades have been trying to improve the performance of EVs to maximize the benefits from their connection to the network. Machine Learning (ML) and data-driven methods are used for analyzing EV charging behavior to maintain significant improvements in the prediction and scheduling fields. Although many of these studies have relied on historical charging data to predict the EVs' State of Charge (SoC) and Charging Available Time (CAT), influential features have often been overlooked. These features are represented in real-time distance, road characteristics (road type, traffic pattern, and events data), and weather data. This study proposes a novel multistage approach, based on a Feedforward Deep Neural Network (FDNN) that combines historical charging data with these influential features to predict both SoC and CAT. The proposed approach outperforms existing literature with SMAPE scores of 0.00044, 0.00018 and 0.00014, 0.00012 for initial, required SoC and CAT predictions, respectively. Through comparative analyses with prior studies on the same dataset, this research highlights substantial improvements in predictive accuracy. It underscores the significance of integrating influential features for the precise prediction of EV charging behaviors within smart transportation systems.