State of Charge Estimation for Li-Ion Batteries: An Edge-Based Data-Driven Approach

The traditional machine learning approach requires substantial computational resources which are scarce in the embedded devices. Recently, the confluence of Edge computing with IoT has enabled resource constrained embedded devices to implement machine learning algorithms. TinyML, with its emphasis on integrating machine learning into embedded systems, seeks to move the end users away from high-performance machines, towards devices with limited resources and power. The present work investigates the estimation of state of charge (SoC) of Lithium Ion (Li-ion) batteries focussing on data-driven methodologies developed during the last five years. This paper mainly focusses on the relationship between dataset characteristics and data stationarity, exploring battery behaviour prediction and related dataset comprehension techniques. A Long Short-Term Memory (LSTM) network, a variant of Recursive Neural Networks (RNN), is utilised for SoC estimation. A 1C rating standard is implemented to comprehend the charge and discharge properties of a Li-ion battery. This study includes hardware evaluation as well as Monte Carlo simulation analysis in circuit component design. The experimental results provide a Mean Absolute Error (MAE) of 0.0630, a Mean Squared Error (MSE) of 0.0107, and a Root Mean Squared Error (RMSE) of 0.1033, confirming the efficacy of the proposed methodology. These results illustrate the accuracy and reliability of SoC estimation model. In conclusion, the proposed data-driven technique can lead to improved data security, reduced latency and cost in the estimation of SoC for Li-ion batteries.