Embedded Dense Neural Networks for Battery Cyclability Prediction on Automotive Microcontroller Devices

Abstract

Recent increase in energy demand has resulted in an upsurge of interest in developing efficient battery storage systems and designing scalable machine learning models that predict the performance of these batteries. This paper explores the development of a testing methodology in the context of accurately predicting future battery performance using machine learning. In particular, I propose a Dense Neural Network Model that can predict battery life using data from the first 50 cycles in the battery life. The model has an accuracy of 91.43% and is 1 − 2 orders of magnitude smaller in size than baseline models proposed in the literature, which allows it to be deployed on a SPC584B-DIS board that runs on a 32-bit e200z4 power architecture and 2MB flash memory. This paper also describes and implements extensive feature extraction based on Principal Component Analysis, using knowledge of the chemical and physical processes that occur during battery degradation. I suggest a methodology for reducing the model size using quantization to fit the memory constraints of the board, and for deploying and running the model on-device for energy- and memory-efficient inferences.