基于长短时记忆递归神经网络的电动车电池状态评估

Electric Vehicle Battery Evaluation Based on Long-Term Memory Recurrent Neural Network

针对现有新能源电池状态评估算法存在的评估精度低等问题,设计了一种基于长短时记忆递归神经网络的评估方法。首先提取动力电池工作中的健康度相关指标和工作电压相关指标,将其作为评估电池状态的原始数据;其次构建递归神经网络框架,利用长短时记忆模块取代神经元,但保留神经元之间的纵向通信能力;根据长短时记忆模块的门控机制,确定模块单元的实时状态并输出最终的评估结果。实验结果显示:提出的动力电池评估算法模型迭代训练能力相较于递归神经网络和前馈神经网络更强,对电池状态评估的准确度接近理论曲线,训练集和测试集的电压峭度值偏差分别为0.002和0.004。

Aiming at the low evaluation accuracy of existing new energy battery state evaluation algorithms,an evaluation method based on long-term memory recurrent neural network was designed.Firstly,the indicators related to the health degree and the working voltage of the power battery were extracted and used as the original data set to evaluate the state of the battery.Secondly,a recurrent neural network framework is constructed to replace neurons with short and long term memory modules,but the longitudinal communication ability between neurons is retained.According to the gating mechanism of the long and short time memory module,the status of the module unit is determined and the final evaluation result is output.The experimental results show that the model iteration training capability of the proposed power battery evaluation algorithm is stronger than that of recurrent neural network and feedforward neural network,and the accuracy of battery state evaluation is close to the theoretical curve,and the voltage kurtosis value deviations of the training set and the test set is 0.002 and 0.004,respectively.