A Comparative Data-Driven Study for State-of-Charge Estimation of Lithium-Ion Batteries Under Variable Operating Conditions

Abstract

The widespread use of lithium-ion batteries in the electric vehicle sector has accelerated the development of data-driven state-of-charge estimation algorithms. This study investigates the state of charge estimation performance of unidirectional deep-learning architectures, namely Long Short-Term Memory and Gated Recurrent Unit, and their bidirectional counterparts, Bidirectional Long Short-Term Memory and Bidirectional Gated Recurrent Unit, on an original dataset obtained from eight different battery cells under various temperature (-10 °C, 10 °C, 20 °C and 40 °C) and variable charge/discharge current (0.2C-5C) conditions. One of the key outcomes of this study is the demonstration that increasing model complexity does not directly improve performance in the SOC estimation problem. The GRU model achieved a root-mean-square error of 1.37% with only 90,000 parameters, whereas the bidirectional architectures, with 230,000–304,000 parameters, provided no additional performance gain despite their higher computational cost. This result reveals that low-cost yet highly accurate models are more suitable for real-time battery management systems.