Study Of Energy Management Strategy Of Hybrid Electric Vehicle Based On Deep Learning

Hybrid Electric Vehicle(HEV)is a new type of vehicle driven by multiple energy sources.The energy management strategy adopted greatly affects the performance of HEV.The energy management strategy distributes torque reasonably among different energy sources according to dynamical system characteristics and driving cycle,thereby gain the optimal fuel economy performance,the lowest pollution emission and the comfortable driving performance.Aiming at the existing problems of energy management strategy of hybrid electric vehicle,combining fuzzy control,deep neural network,multi-objective optimization,vehicle communication and other related technologies,this dissertation conducts indepth research on energy management strategy and proposes optimization and improvement.The main research work is as follows:Firstly,this dissertation regards Wuhan's road traffic as the research object,build a driving cycle acquisition system with on-board real-time monitoring technology.Then build a representative driving cycle in Wuhan.by an auto encoder.This method uses the characteristics of feature extraction and data compression of encoder and the function of reconstructed data of decoder,which to make reconstructed data is as close as possible to the actual speed and variation rule of the vehicle.According to the representative driving conditions,a double trapezoid model is proposed to modalize the reconstructed data output of decoder.Then the validity of the proposed model is verified through evaluation indexes and feature analysis.Secondly,a fuzzy energy management strategy based on improved pigeoninspired optimization is proposed.On this basis,a driving cycle recognizer based on deep belief network is constructed.The fusion of the two makes the fuzzy torque distribution controller to determine its corresponding membership function and control rules according to the type of driving cycle.The simulation results show that the integration of driving cycle identification technology into energy management strategy effectively solves the problem that the selection and determination of membership function and fuzzy rules are too dependent on subjective engineering experience.At the same time,it reduces the disturbance effect of driving cycle diversity on the system and more effectively improves the performance of energy management strategy.Thirdly,a short-term driving cycle prediction method under vehicle communication is proposed and applied to energy management strategy based on model predictive control.The method collects state information broadcast from surrounding vehicles and/or roadside units and processes the information to set parameter values of the hybrid deep learning model.The hybrid model outputs short-term driving cycle for a given prediction horizon,and the model predictive control is used to incorporate the disturbance of the control time domain into rolling optimization at each sampling instant so as to improve the performance of the global solution.In addition,the influence of vehicle networking communication on prediction accuracy is analyzed.Finally,the simulation experiments verify that the optimization and improvement made in this study have important practical significance for improving the energy utilization of hybrid electric vehicles.Fourthly,a energy management strategy based on NAF deep reinforcement learning is proposed.This method introduces normalized advantage functions on the basis of deep Q network to ensure that the corresponding Q value is the maximum value while outputting actions,thus realizing that the input of state space and the output of action space are both high dimensional and continuous.In order to adapt to different driving conditions online,an online reinforcement learning framework is further proposed based on V2 X communication.The simulation results show that the energy management strategy based on deep reinforcement learning has achieved good oilsaving effect in both offline and online application scenarios.