基于改进门控循环单元神经网络的锂电池组荷电状态预测

准确预测锂电池组的荷电状态(state of charge, SOC)能够有效防止电池过度充电或者放电，是储能设备安全运行的重要保障。为了解决SOC无法通过测量直接获得的问题，提出了一种基于猎人猎物优化算法(hunter prey optimization, HPO)优化门控循环单元(gated recurrent unit, GRU)神经网络的预测模型。在GRU的基础上添加Dropout机制，来增强模型的泛化能力，并通过HPO算法优化GRU的超参数，使锂电池的数据特征与网络拓扑相匹配。为了验证HPO-GRU模型的有效性，以某储能公司现场采集的储能锂电池组历史数据进行仿真实验，并与反向传播神经网络(back propagation, BP)、长短期记忆网络(long short term memory, LSTM)和GRU 3种预测模型的预测结果进行对比分析。可得HPO-GRU模型预测值与真实值的误差最小，在5%以内。可见HPO-GRU模型的预测精度最高，具有良好的鲁棒性以及较强的泛化能力。

State of charge prediction of the lithium-ion battery pack based on improved gated recurrent unit

Accurate prediction of the state of charge (SOC) of the lithium-ion battery pack can effectively prevent overcharging or discharging of the battery, which is an essential guarantee for the safe operation of energy storage devices. In order to solve the problem that SOC cannot be obtained directly by measurement, a prediction model is proposed to optimize the gated recurrent unit (GRU) neural network based on the hunter-prey optimization (HPO) algorithm. The Dropout mechanism is added to GRU to enhance the model''s generalization ability, and the HPO algorithm optimizes the hyperparameters of GRU to match the data features of lithium batteries with the network topology. In order to attest to the effectiveness of the HPO-GRU model, simulation experiments were conducted on historical data of energy storage lithium battery pack collected on-site by an energy storage company, and corresponding results were compared with those of three prediction models, back propagation (BP), long short-term memory (LSTM), and GRU. T The error between the predicted value of the HPO-GRU model and the real value is the smallest, within 5%. It can be seen that the HPO-GRU model has the highest prediction accuracy, excellent robustness, and strong generalization capability.