A Data-Driven Lithium-Ion Concentration Estimator for Electrode Solid Phase in Battery Models.

Abstract

This thesis introduces a data-driven approach for lithium-ion concentration estimation in battery cells, leveraging a deep learning model based on Long Short-Term Memory (LSTM) networks. As lithium-ion batteries gain widespread adoption, particularly in electric vehicles and renewable energy storage, accurate and efficient estimation of lithium-ion concentration is pivotal. Existing methods like the polynomial Reduced Order Models (ROMs), although popular, face trade-offs between computational efficiency and accuracy. In this work, we propose an LSTM-based estimator designed to achieve high accuracy without compromising computational speed. The LSTM architecture’s depth allows it to capture intricate relationships within the cell. The estimator’s performance was rigorously evaluated and compared against a polynomial ROM model. Our estimator exhibited re markable accuracy in short term estimation, achieving a Mean Absolute Percentage Error (MAPE) of 1.09% at 1000s and 3.27% at 2000s for estimation of surface concentration for single particle anode. For estimation of average concentration for single particle anode, the MAPE was 1.27% at 1000s and 2.55% at 2000s. The estimator’s swift computation time provides potential for real-time applications. Overall, this thesis provides a comprehensive exploration of LSTM-based lithium-ion concentration estimation, from model design and implementation to performance evaluation. This work underscores the potential of data-driven approaches in battery state estimation, opening up new possibilities for real-time applications and improved battery management systems.