State of charge estimation of lithium-ion batteries using a deep learning approach

In this dissertation, a novel operating-state-sensitive modeling approach based on deep learning is proposed for the state of charge (SOC) estimation of lithium-ion batteries. While conventional methods typically process all operational data through a single model, the proposed approach first classifies the instantaneous operating state of the battery (charging, discharging, idle), and then routes the data to a dedicated long short-term memory (LSTM) network trained specifically for that state. Thanks to this parallel architecture, each network is trained solely on context-specific data, which enhances estimation accuracy and makes the learning process more efficient. The effectiveness of the proposed model is evaluated using two different datasets. In the first stage, an open-access dataset provided by the University of Wisconsin, which includes various driving cycles (US06, UDDS, HWFET, LA92, etc.) for the Panasonic NCR18650PF cell, is utilized. The parallel LSTM architecture trained with this dataset outperforms the conventional single-network structure in terms of mean absolute error (MAE) and mean square error (MSE) and also demonstrates suitable inference performance for online deployment in embedded systems. In the second stage, a custom experimental dataset is constructed. For this purpose, various driving cycles are applied to the Sony US18650VTC6 lithium-ion cell, and current-voltage measurements are recorded at a sampling rate of one second using GW Instek programmable power supplies and electronic loads. These experiments are conducted under different ambient temperatures (e.g., 0 °C, 25 °C and 40 °C), allowing the impact of environmental conditions such as temperature on SOC estimation to be analyzed. The results indicate that the proposed structure is robust and adaptable not only to varying load profiles but also to environmental changes. In conclusion, the state-of-charge estimation approach proposed in this thesis, which is sensitive to the battery operating condition, provides advantages such as high accuracy, low computational cost, and environmental flexibility. It has been demonstrated that the state-of-charge estimation of lithium-ion batteries can be successfully performed using a deep learning approach.