基于门控神经网络的储能电站荷电状态估计研究

Storage battery SOC estimation based on gated recurrent unit neural network and dropout

电站荷电状态(SOC)估算是电池管理系统中重要的功能之一,同时SOC也是保障储能电站正常运行和安全的重要参数。对于储能电站,SOC的准确估计能防止电池的过充或过放,延长电池使用寿命,准确估算SOC也是当前的难点。本文提出使用门控神经网络(GRU)结合Dropout正则化算法估算SOC,实现电池SOC的准确估计与良好泛化能力。GRU网络用来学习电压、电流测量值与SOC的非线性关系。训练时,Dropout将隐藏层中的所有神经元以一定的概率随机置为0,避免对隐藏层中任一节点产生依赖,防止过拟合,从而增加模型的泛化能力。实验结果显示,本文提出的方法可实现SOC准确估计,最优结果为平均绝对误差1.38%,均方根误差1.66%。

SOC estimation is one of the most important functions in battery management system and it is also an important parameter to ensure the normal operation and safety of energy storage power station.For energy storage stations,accurate estimation of SOC can prevent battery overcharge or overdischarge,thus prolonging battery life,and it is also the difficult point to be solved.In this paper,the GRU neural network combined with Dropout regularization algorithm is proposed to estimate the SOC,so as to realize the accurate estimation and great generalization ability.The GRU network is used to learn the nonlinear relationship between the voltage,current and SOC.During training,dropout randomly sets all neurons in the hidden layer to 0 with a certain probability to avoid dependence on any unit in the hidden layer and prevent overfitting,thus increasing the generalization ability of the model.The results show that the method proposed in this paper can achieve accurate SOC estimation with the optimal result of MAE of 1.38%and RMSE of 1.66%.