State-of-charge estimation of lithium batteries using compact RBF networks and AUKF

Abstract

A novel framework for the state-of-charge (SOC) estimation of lithium batteries is proposed in this paper based on an adaptive unscented Kalman filters (AUKF) and radial basis function (RBF) neural networks. Firstly, a compact off-line RBF network model is built using a two-stage input selection strategy and the differential evolution optimization (TSS_DE_RBF) to represent the dynamic characteristics of batteries. Here, in the modeling process, the redundant hidden neurons are removed using a fast two-stage selection algorithm to further reduce the model complexity, leading a more compact model in line with the principle of parsimony. Meanwhile, the nonlinear parameters in the radial basis function are optimized through the differential evolution (DE) method simultaneously. The method is implemented on a lithium battery to capture the nonlinear behaviours through the readily measurable input signals. Furthermore, the SOC is estimated online using the AUKF along with an adaptable process noise covariance matrix based the developed RBF neural model. Experimental results manifest the accurate estimation abilities and confirm the effectiveness of the proposed approach.