CNN-based prediction of multi-variables for lithium-ion batteries optimized with TinyML and deployed on edge devices

Estimation or prediction of several battery cell attributes can be very useful in developing operation strategies to prevent battery failures or to schedule battery replacements or updates in electronic devices including electrical vehicles (EVs) that use lithium-ion batteries. Examples of such attributes include state of charge (SoC), remaining useful life (RUL), and temperature. While one can develop individual AI/ML models to predict such attributes, the models may require too much memory and computational capacities that may not be available on resource constrained edge devices that use microcontrollers or small processors. That is why, developing AI/ML models that predict multiple such attributes with a single model may offer a better trade-off between required resources and performance when deployed on edge devices. Therefore, in this paper, we investigate a novel AI/ML model to predict both battery RUL and cell temperature. The proposed novel deep learning model combines a temporal convolutional network (TCN) with a multi-head attention layer that can output two or more variables simultaneously. Simulation results demonstrate that prediction of both RUL and cell temperature can be achieved accurately with a single model and that accuracy is on par with models from existing literature that focused on models for individual predictions. Moreover, we optimize the proposed model using tinyML technologies and deploy it for real time inference on a Raspberry Pi 4 edge device. Experiments show that the model performance is still very good even after such optimizations of the model to make it fit on the edge device.