An adaptive prediction method based on data stream mining for future driving cycle of vehicle

Abstract

Due to complex and changeable driving cycles in urban roads, it is a challenging task for most of the current control strategies utilized in vehicles to adapt to the driving environment. At the same time, hardware requirements for storing and processing a massive amount of streaming data are increasing, which lead to excessive accumulated errors and high computational cost. To deal with this problem, an innovative prediction method, which is based on Markov chain and data stream mining, is proposed to predict the future driving cycle of vehicles. State transition probability matrix is updated in real time with data stream mining technology, and every time a new record arrives, the expired record is replaced by the new arrived one in the memory, and both state division and the sizes of the sliding window can be adjusted adaptively based on prediction accuracy for the changing driving cycles. The results show that the proposed method is more suitable for predicting changing driving cycles, which is able to maintain better prediction accuracy than the traditional method. In addition, based on the proposed method, the memory space utilized for storing temporary records were saved largely, and the calculation resource required was reduce.