混合动力电动汽车在线学习控制策略研究

项目名称： 混合动力电动汽车在线学习控制策略研究

项目编号： No.61273139

项目类型： 面上项目

立项/批准年度： 2013

项目学科： 自动化技术、计算机技术

项目作者： 李卫民

作者单位： 中国科学院深圳先进技术研究院

项目金额： 80万元

中文摘要： 与传统汽车相比，混合动力汽车(HEV)在节能减排方面具有很大优势。而能量管理策略(EMS)直接影响 HEV的性能。由于车辆未来行驶工况的不确定性，现有的EMS通常不能充分发挥HEV的优势。为此，本项目提出三种智能EMS：基于正交小波基神经元动态规划的EMS、基于Q学习的在线自调整模糊EMS以及基于随机模型预测控制的EMS。我们将HEV能量管理建模为马尔科夫决策过程，利用神经网络与强化学习算法在线估计状态的值函数，进而求出最优控制律。所提三种算法具有如下特点：（1）优化效果好；（2）具备在线学习功能，使得EMS的设计不依赖于未来行驶工况的信息，且能适应外部环境的缓慢变化；（3）实时性好，不需要将连续系统离散化，避免了"维数灾难"问题。所提算法的有效性将通过台架试验和整车道路测试进行验证。本项目的研究有助于进一步挖掘HEV节能减排的潜力，提高性能，推动其产业化进程。

中文关键词： 混合动力电动汽车；能量管理策略；神经元动态规划；模糊Q学习；随机模型预测控制

英文摘要： Hybrid electric vehicles(HEVs) have great advantages in improving fuel economy and reducing emissions compared with conventional vehicles. The performance of HEV heavily depends on an efficient energy management strategy (EMS). However, due to the complexity of their powertrain structure and the uncertainty of future driving conditions, many existing EMSs act upon fixed parameters usually fail to fully explore the potential of these advanced vehicles. Three intelligent EMSs are proposed in this project: EMS based on orthogonal wavelets basis functions neuro-dynamic programming(NDP); Self-tuning fuzzy EMS based on Q-learning algorithm; EMS based on stochastic model predictive control(SMPC). The HEV energy management problem is modeled as a Markov decision process. To solve it, we apply neural network and reinforcement learning algorithms to approxiate the system states' value functions online, and accordingly the optimal actions. The proposed EMSs possess such characteristics as: (1)Good optimization effect; (2)These EMSs do not rely on prior information related to future driving conditions, and can self-tune with a wide variance in operating conditions;(3)Good real time property. It needn't discrete the continous-time system, which avoids the problem of "curse of dimensionality";The proposed EMSs will be impleme

英文关键词： Hybrid electric vehicle；energy management strategy；neuro-dynamic programming；fuzzy Q-learning；stochastic model predictive control