An RNN-CNN-Based Parallel Hybrid Approach for Battery State of Charge (SoC) Estimation Under Various Temperatures and Discharging Cycle Considering Noisy Conditions

With the increasing use of lithium-ion (Li-ion) batteries in electric vehicles (EVs), accurately measuring the state of charge (SoC) has become crucial for ensuring battery reliability, performance, and safety. In addition, EVs operate in different environmental conditions with different driving styles, which also cause inaccurate SoC estimation resulting in reduced reliability and performance of battery management systems (BMSs). To address this issue, this work proposes a new hybrid method that integrates a gated recurrent unit (GRU), temporal convolution network (TCN), and attention mechanism. The TCN and GRU capture both long-term and short-term dependencies and the attention mechanism focuses on important features within input sequences, improving model efficiency. With inputs of voltage, current, and temperature, along with their moving average, the hybrid GRU-TCN-Attention (GTA) model is trained and tested in a range of operating cycles and temperatures. Performance metrics, including average RMSE (root mean squared error), MAE (mean absolute error), MaxE (maximum error), and R2 score indicates the model is performing well, with average values of 0.512%, 0.354%, 1.98%, and 99.94%, respectively. The proposed model performs well under both high and low noise conditions, with an RMSE of less than 2.18%. The proposed hybrid approach is consistently found to be superior when compared against traditional baseline models. This work offers a potential method for accurate SoC estimation in Li-ion batteries, which has an important impact on clean energy integration and battery management systems in EVs.