The global shift towards sustainable energy sources has placed lithium-ion (Li-Ion) batteries at the forefront of technological innovation. These batteries serve as pivotal components in various applications, including renewable energy storage and electric vehicles. Accurate modeling of the electric behavior of Li-Ion cells remains a complex challenge vital for optimizing their performance and durability. The high complexity of the behavior of these types of cells, especially under extreme conditions, results in conventional Equivalent Electrical Circuit Models (EECM) exhibiting limited accuracy. Here, Machine Learning (ML) methods offer a potential solution due to their ability to learn and model nonlinear, dynamic, and highly complex relationships directly from large datasets.

Within this work, a comprehensive investigation is encompassed, involving the identification of promising ML approaches from existing literature and the subsequent development of these models. The primary objective is to enhance the conventional EECM process of Li-ion batteries by leveraging automation and addressing inherent issues, potentially substituting or complementing the existing EECM with ML-based features. By comparing the innovative ML-supported models with conventional EEC models, this research seeks to highlight the potential advantages and disadvantages of the discussed methods.

The scope of the work includes testing data from four different cylindrical Li-Ion cells with different chemistries, namely NCA, NMC and LFP, focusing on different data preparation and training techniques for highly dynamic time series data. Standalone ML algorithms, ML-supported EECMs and conventional EECMs are validated with application profiles at different temperature levels. Using open-circuit-voltage and pulse testing profiles as training datasets and by incorporating end-to-end concatenation and time series cross-validation as ML training techniques, an accuracy improvement of up to 6% is achievable, particularly with ML-supported EECMs.