Driving prediction techniques (DPTs) are used to forecast the distributions of various future driving conditions (FDC), like velocity, acceleration, driver behaviors etc. and the quality of prediction results has great impacts on the performance of corresponding predictive energy management strategies (PEMSs), e.g., fuel economy (FE), lifetime of battery etc. This survey presents a comprehensive study on existing DPTs. Firstly, a review on prediction objectives and major types of prediction algorithms are presented. Then a comparative study on various prediction approaches is carried out and suitable application scenarios for each approach are provided according to their characteristics. Moreover, prediction accuracy-affecting factors are analyzed and corresponding approaches for dealing with mispredictions are discussed in detail. Finally, the bottlenecks of current researches and future developing trends of DPTs are given. In general, this paper not only gives a comprehensive analysis and review of existing DPTs but also indicates suitable application scenarios for each prediction algorithm and summarizes potential approaches for handling the prediction inaccuracies, which will help prospective designers to select proper DPTs according to different applications and contribute to the further performance enhancements of PEMSs for hybrid electric vehicles (HEVs) and plug-in hybrid electric vehicles (PHEVs).

Considering the changeable driving conditions in reality, energy management strategies for fuel cell hybrid electric vehicles should be able to effectively distribute power demands under multiple driving patterns. In this paper, the development of an adaptive energy management strategy is presented, including a driving pattern recognizer and a multi-mode model predictive controller. In the supervisory level, the Markov pattern recognizer can classify the real-time driving segment into one of three predefined patterns. Based on the periodically updated pattern identification results, one set of pre-optimized control parameters is selected to formulate the multi-objective cost function. Afterwards, the desirable control policies can be obtained by solving a constrained optimization problem within each prediction horizon. Validation results demonstrate the effectiveness of the Markov pattern recognizer, where at least 94.94% identification accuracy can be reached. Additionally, compared to a single-mode benchmark strategy, the proposed multi-mode strategy can reduce the average fuel cell power transients by over 87.00% under multi-pattern test cycles with a decrement of (at least) 2.07% hydrogen consumption, indicating the improved fuel cell system durability and the enhanced fuel economy.