Enhancing generalization performance of CNN-based state-of-charge estimation for lithium-ion batteries

Lithium-ion batteries are the most important component of electric vehicles. Since it has a chemical structure, the State-of-Charge (SOC) of the batteries cannot be determined precisely, so it is estimated by various methods. However, the generalization capability of the methods to data obtained from experiments at different temperatures and different batteries is still a major challenge. Moreover, the distribution shifting occurring in time series may also reduce the generalization ability. In this paper, the generalization capacity problem has been addressed, and a SOC estimator based on a convolutional neural network framework is proposed. The proposed method reduces the internal covariate shift during training by using batch normalization and improves the generalization performance by normalizing each instance independently by using instance normalization. The results have been compared with state-of-the-art SOC estimation methods and increased accuracy has been observed. Tests were carried out by creating different scenarios. In the experimental results, the benchmark models were outperformed by achieving a 57.1 % increase in MAE accuracy for tests with data obtained at all temperatures (−20 °C, −10 °C, 0 °C, 10 °C, 25 °C), 15.5 % for positive temperatures (0 °C, 10 °C, 25 °C) and 24.9 % for negative temperatures (−20 °C, −10 °C, 0 °C).