Microcontroller architectures for battery state of charge prediction with tiny neural networks

Technology development for batteries has been one of the major challenges to tackle in the progress of electric cars. In particular for the battery management system state-of-charge estimation is of primary importance. In fact, battery features have complex time-varying and non-linear properties which are difficult to modelize analytically. Therefore, accurate estimation of the state of charge is a challenging task especially if the resulting approach shall be implemented on a resource constrained, low power micro controller. In this paper, a comparison between different artificial neural networks (ANNs) topologies for predicting the state of charge of the battery will be presented. The precision, complexity and robustness of the proposed architectures were comprehensively validated through comparative analysis of many sets of hyperparameters and input measurements. The experiments resulted in precise estimations at different aging states and time-varying temperature conditions with a range of loss mean squared error around 3e-04 (±5.8e-04). Finally the computation complexity has been profiled on automotive microcontrollers by using the toolsets SPC5-Studio.AI, which automatically converts pre-trained ANNs and generates optimized and validated ANSI C code for SPC58 processors. Those profiles allow to estimate complexity needed when deploying the ANNs for 2 battery pack case studies; Formula Student FSAE and BMWi3.