Hybrid Wavelet-PCA and DNN-LSTM-Kalman Framework for Robust State-of-Charge Estimation Using Minimal Sensors

Abstract

An accurate prediction of state of charge (SOC) is essential for the operational performance and durability of lithium (Li)-ion battery systems in electric vehicles. Its accuracy in dynamic operation, however, is often limited by noise sensitivity, drift buildup and rest-conditioning dependence of conventional procedures such as Coulomb counting and open-circuit voltage. This work presents a two-stage hybrid framework for SOC estimation to overcome these limitations. The first stage includes wavelet transform denoising and principal component analysis (PCA) to enhance the quality of signal, eliminate redundant information, and achieve a signal-to-noise ratio (SNR) improvement up to 77.9 dB with 95% variance retention. The second stage combines a deep neural network (DNN)–long short-term memory (LSTM) model with Kalman filtering to capture nonlinear temporal dynamics and generate smoothed SOC predictions. Experimental results show that the root mean squared error (RMSE) of baseline SOC estimation methods was 21.8%, while that of the proposed method was only 0.52%, achieving over forty-fold accuracy improvement. The proposed framework demonstrates strong robustness under noisy and dynamic operating conditions and can be further extended to state of health (SOH) prediction by incorporating battery degradation features, enabling predictive maintenance and enhancing long-term reliability in battery management systems.