A multi-variable multi-step Seq2seq networks for the state of charge estimation of lithium-ion battery

Abstract

Due to the complexity and changeable of lithium-ion batteries, we propose a multi-variable and multi-step Temporal neural network to cover this task. Specially, a novel multi-step training strategy is applied to deal with long time sequences, and multi-variables is added to supervise the prediction. In addition, the Seq2seq net with long short term memory (LSTM) modules is employed to improve the prediction accuracy in various lithium-ion battery operating environments of −10 °C, 10 °C, 25 °C and 40 °C. The A123 lithium-ion batteries’ database collected by University of Maryland for the real scenarios has been used in the training process. The ablation studies indicate that the multi-step module is necessary under the root mean square error, mean absolute error and mean absolute percentage error index. The prediction error of comparison experiment under our proposed algorithms is 0.45 %@ root mean square error (RMSE), 0.30 %@ mean absolute error (MAE) and 1.11 %@ mean absolute percentage error (MAPE) at 10 °C. Our study introduces a novel and accurate SOC prediction method, which can provide accurate SOC estimates in a wide range of scenarios, emphasizing its novelty and relevance in addressing real-world challenges.