Computing Framework for SoC Estimation using On-device Learning

Battery Energy Storage Systems (BESSs) are the most widespread technology used in smart grids to promote grid reliability and store excess energy. The BESSs rely on a single, non-measurable parameter called State-of-Charge (SoC) for decision-making and is challenging to obtain. The existing SoC estimation methods have drawbacks, including a lack of real-time correction features and generalizability across different systems. Working to mitigate the aforementioned challenges, this paper presents a complete SoC estimation study with scalable hardware design, machine learning-based SoC estimator, and application of different computing paradigms, i.e., node, edge, and cloud devices. Firstly, a two-cell lithium-ion battery kit is designed for charging-discharging, facilitating data recording, and utilizing a Pi Pico device to send data over WiFi. An LSTM model is deployed in Raspberry Pi 4, which saves upcoming data from Pico, runs inference, and sends the inference to the local build cloud server at the monitoring station over WiFi. Lastly, the On-device learning feature has also been incorporated to tackle inaccuracy, promoting the robustness, adaptability, and reliability of the inference system. The results demonstrated high accuracy, achieving a R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> score of 0.9969 on the testing dataset, and also highlighted the importance of the adaptability feature incorporated in the framework.