基于深度学习和贝叶斯优化的短程车速预测研究

Short-time Vehicle Velocity Prediction Based on Deep Learning and Bayesian Optimization

针对模型预测控制能量管理策略中自车短程车速预测精度不足和模型优化耗时长,结果不佳的问题,文章建立了带注意力机制的编码解码器进行短程车速的预测,并利用贝叶斯优化来确定模型的超参数。首先,介绍了模型的整体结构和计算单元的选择,然后利用实车工况和UDDS、HWFET标准循环工况建立工况库,并对本文的模型与RBF神经网络和LSTM神经网络进行了精度对比;其次,建立贝叶斯优化的框架,对模型的超参数进行优化,并与网格搜寻和随机搜寻的优化结果进行对比。研究结果表明:相比于RBF神经网络和LSTM神经网络,文章提出的车速预测模型在UDDS工况上精度分别提高了26.5%和12.7%,在HWFET工况上精度分别提高了39.2%和18.8%,同时模型所占空间大小和预测用时都满足车辆实时使用的需求;在使用贝叶斯进行超参数优化后,相较于网格搜寻和随机搜寻精度分别提高了30.3%和27.2%,同时优化用时均减少了67.2%。

To address the problems of insufficient accuracy of the short-time speed prediction of the self-vehicle and the time-consuming and poor results of the model optimization in the model predictive control energy management strategy,this paper uses encoder-decoder model with attention mechanism for short-time speed prediction and Bayesian optimization to determine the hyper parameters of the model.Firstly,the overall structure of the model and the selection of computational units are introduced;then a library of conditions is established using real vehicle conditions and standard cyclic conditions of UDDS and HWFET.Proposed model is compared with RBF neural network and LSTM neural network afterwards.Secondly,the framework of Bayesian optimization is established to optimize the hyper parameters of the model,and the optimization results are compared with those of grid search and random search.The results show that compared with the RBF and LSTM neural networks,the proposed speed prediction model improves the accuracy by 26.5% and 12.7% for UDDS conditions and 39.2% and 18.8% for HWFET conditions respectively,and the space and prediction time occupied by the model meet the requirements of real-time use of real vehicles.After using Bayesian optimization,the accuracy is improved by 30.3% and 27.2% compared with grid search and random search,while the optimization time is reduced by 67.2%.