SOC estimation for lithium batteries based on the full parallel nonlinear autoregressive neural network with external inputs

Abstract

Lithium-ion batteries are the core element in electric-vehicle systems. Because they provide a power source for the entire electric vehicle, their state of charge (SOC) is crucial to ensure the proper operation of the vehicle. Thus, a more precise estimate of the SOC should significantly improve the safety and utility of these batteries. Toward this end, we propose herein novel nonlinear autoregressive with exogenous input (NARX) and full-parallel (dual-close loop) structure NARX (FPNARX) networks to overcome the disadvantage of the traditional series-parallel (open-closed loop) structure NARX (SPNARX) network for estimating the SOC. To test the performance of the FPNARX network, we use the dynamic stress test and the urban dynamometer driving schedule discharge profiles in battery-discharge experiments. Furthermore, feed-forward neural networks and other time-series neural networks, including the time-delay neural network and SPNARX, are also implemented for the same experiments. The experimental results show that the time-series networks can estimate the SOC more precisely than the feed-forward network under non-constant-current discharge. Moreover, the applicability of the time-series networks depends on the current amplitude and frequency. However, the FPNARX network always gives the best results.