Investigation of Data-Driven Models for Robust Lithium-ion Battery Modeling

A battery model plays a crucial role in ensuring the effective use and operation of a battery system. It provides an understanding of the battery's behavior, and by simulating different scenarios, a robust model helps optimize the design and management of battery system. Current modeling methods include model and data-driven based approaches. The former approach can be classified as electrochemical models (EMs) and equivalent circuit models (ECMs), while the latter approach uses machine learning (ML) to develop data-driven models (DDMs) of battery. EMs are highly complex and demand computational resources, making them challenging to implement in real-time applications. ECMs, on the other hand, face challenges in capturing the non-linearities of batteries over time. Although DDMs require large amount of data to train the ML models, they have been shown to address the limitations of EMs and ECMs. However, past research on DDMs often neglects the variations and conditions of batteries, such as the battery types, C-ratings, and ambient temperatures. Therefore, this research aims to investigate various DDMs, focusing particularly on their voltage estimation compared to empirical data under different scenarios, to identify a suitable DDM for robust lithium-ion battery (LIB) modeling. Several DDMs were investigated, and the long short-term memory (LSTM) model outperformed other DDMs, achieving the Mean Absolute Percentage Error (MAPE) of 2.664 % under different ambient temperatures, and 3.475% under various battery types, with C-rating of batteries varying from 0.05C to 5C. Although its MAPE in both cases are not the lowest, its consistency performance in voltage estimations making it the preferable model for LIBs.