Development of a Machine Learning Technique to Accurately Estimate Battery State of Charge

Depleting fossil fuel reserves, concerns over air pollution and global warming position electric vehicles as compelling alternatives to conventional vehicles with internal combustion engines. A large portion of the commercially available electric vehicles today are of the battery-electric type, which use batteries to store energy. Accurate computation of the battery state of charge is critical to ensure that a reasonable estimate of the vehicle range is available before the batteries need to be recharged. Conventional approaches to estimate battery state of charge rely on the battery equivalent circuit models and state observers. However, the non-linearity of the battery properties combined with their dependence on temperature and driving cycle render these approaches inaccurate, necessitating the development of robust techniques that can overcome these drawbacks and accurately estimate the state of charge. This paper first reviews the conventional techniques to estimate the battery state of charge and identifies their limitations. Next, a machine learning technique is developed to accurately estimate the battery state of charge and it's performance is evaluated over drive cycle data that was experimentally collected from a commercial electric vehicle under different test conditions. The results show the suitability of the developed technique to accurately estimate the battery state of charge with RMS errors less than 3% under most operating conditions and standard driving cycles.