An Adaptive Kalman-Guided Soft Sensor Using Feedforward Neural Networks for SOC Estimation in Lithium-Ion Batteries

Accurate State-of-Charge (SOC) estimation is crucial for ensuring the battery's safe operation and prolonged lifespan, but it remains a challenge due to sensor noise and system nonlinearity. This paper proposes a novel covariance-adaptive hybrid approach that interprets neural network outputs as virtual measurements and dynamically integrates them with the a priori state predictions from a Discrete Kalman filter to yield refined a posteriori estimates. The introduced adaptive scheme uses exogenous inputs derived from spectral analysis to provide past information to the Feedforward Neural Network (FNN) to perform soft sensing and generate network-aided measurements. The novel processing occurs after this step by adaptive integration with Kalman filter results. The developed hybrid architecture is tested on real experimental battery datasets for LG 18650 HG2, and the results prove its superior performance against existing state-of-the-art algorithms. Experimental results averaged over twenty Monte Carlo trials demonstrated nearly 28% and 39% improvement in the mean absolute error and root mean square error, respectively, compared to the baseline results.