Estimating battery state-of-charge with a few target training data by meta-learning

Abstract

This study proposes a meta-learning battery state-of-charge (SOC) estimation approach to reduce the amount of target battery data required for training a Li-ion battery SOC estimation by applying deep-learning. The proposed approach reduces the training data required for the target battery by improving the pre-training performance. Multiple other batteries are used as pre-training data to allow generalization to the target battery. The perfor-mance of meta-learning SOC estimation is compared with that of transfer learning using one battery data for pre-training. The meta-learning performance does not depend on the similarity between the pre-training data and target battery data. The proposed meta-learning SOC estimation accuracy is verified based on battery test data from various driving cycles (US06, urban dynamometer driving schedule (UDDS), and LA92), to reflect actual electric vehicle driving patterns. Using only a small amount of target battery data and nine gradient steps, the proposed meta-learning SOC estimation algorithm achieves a mean squared error (MSE) of 0.0176% and a mean absolute error (MAE) of 1.0075%. These results show that the proposed method can adapt more quickly than transfer learning (with SOC estimation errors of 3.1378% in terms of the MSE and 15.1327% in terms of the MAE under the same conditions).