A dynamic ensemble for estimating state-of-charge of interchangeable robot batteries

This paper presents a unique machine learning model that estimates battery state-of-charge (SOC) for robotic applications. Unlike earlier approaches, this study investigates the problem of estimating SOC for several interchangeable batteries that can be used to power a robot. Robots commonly have a reserve pool of batteries available to be swapped for the purpose of extending operational time, but swapping batteries complicates the SOC estimation problem due to parameter variation. The proposed state-based ensemble is novel in that it exceeds the accuracy of traditional ensemble methods by dynamically changing estimation algorithms and predictors based on a preliminary (i.e., rough) state estimate of the battery. Experimental results show statistically significant improvement, on average, of 4 percent for our proposed state-based ensemble.