Performance Comparison of Machine Learning Methods with Distinct Features to Estimate Battery SOC

As the world is moving towards using renewable energy sources for a greener environment, battery energy storage systems (BESS) has become an important issue because of the intermittent nature of renewable sources. The quantity associated with the amount of energy stored in battery at a certain time is called the battery’s State of Charge or SOC. Typically it is estimated with the assumption that it changes linearly with the quantities such as voltage, current etc. which can be measured effortlessly. But actually no such linear relationship exists. Machine learning techniques have proven to be useful for measuring SOC. Though the features that are conventionally used often suffer from a phenomenon called multicollinearity. Ridge and lasso regression methods have shown to counter the effects of multicollinearity. In this paper, these methods have been used along with linear regression, support vector machine (SVM) and artificial neural network (ANN). Besides readily available conventional features some nonconventional features are explored too that can be derived from conventional features. Results exhibit that nonconventional features boost the performance of all the methods except for SVM. Ridge regression method with nonconventional features has shown to generate the best outcome among all the scenarios taking both accuracy and speed of computation into account.