RF-LSTM Two-Stage Residual Learning Framework for SOC Estimation in Lithium-Ion EV Batteries: Comparative Benchmarking with Deep Learning Models

With the rapid development of electric vehicles, the accurate estimation of the state of charge (SOC) of lithium-ion batteries has become a critical concern. Most existing data-driven methods use a single-stage approach to predict the SOC values, which limits their performance in many cases. This article proposes a novel two-stage residual learning approach based on the random forest (RF) and long short-term memory (LSTM). In the first stage, RF predicts the SOC using the sliding voltage, current, and temperature values as inputs. Then, the residual from the RF estimation is used to train the LSTM, and the final SOC is obtained by adding the output of the LSTM to the predicted SOC in the first stage. The proposed approach was validated using Panasonic 18650PF cells over four drive cycles and five temperatures, comprising 20 operating scenarios in total. The RF–LSTM method achieved the lowest mean absolute error (MAE) in all cases, with the best value being as low as 0.586%. On average, the MAE was 1.241%, and the highest R² reached 0.9991. Finally, the method was verified using the CALCE INR 18650-20R (NMC) battery data, showing its adaptability to a different cell chemistry.