Abstract— This paper presents a neural network-based state of charge (SOC) observer for Li-ion batteries. Using the single particle model (SPM) as a plant model, voltage and SOC data were collected for varying current profiles. Multiple neural network observers were compared, varying factors such as number of inputs, number of hidden layers, and experimentation of simple moving average filters and exponential moving average filters. These observers were compared to as an SPM based extended Kalman filter (EKF), as a control. The best performing model was a four-input, 3 layer (5 neurons each) neural network with an exponential moving average filter. After fine-tuning, a model-best root mean square error of 0.5803% was achieved.
Keywords— Neural Networks, SOC Estimation, Observer Design, Electric Vehicles, Li-Ion Battery
Full paper here.
