A comprehensive comparison of Time-Series Models for Battery health diagnostics and prognostics

Abstract

Accurate monitoring of battery health is crucial for ensuring the reliability and longevity of energy storage systems, particularly in applications such as electric vehicles and renewable energy. Traditional methods, like empirical formulas, often struggle to capture the complexities of battery degradation in dynamic systems. This paper investigates the use of advanced time-series forecasting models to predict the State of Health (SOH) and Remaining Useful Life (RUL) of lithium-ion batteries. We examine three sophisticated models capable of capturing both long-term dependencies and short-term fluctuations: (1) Prophet, a statistical model known for its flexibility in handling non-linear trends and complex seasonality and outliers, (2) DeepAR, a machine learning model that excels at modeling high-dimensional dataset and can delivers probabilistic predictions to account for uncertainty, and (3) Temporal Fusion Transformers (TFT), a hybrid model that combines transformer-based architectures with specialized components to enhance forecasting accuracy. These models are applied to a publicly available battery dataset, and their performance is evaluated based on key criteria such as accuracy, robustness, scalability, and interpretability. Additionally, we explore the impact of group training (multi-task learning), supported by both DeepAR and TFT, which allows the models to learn shared patterns across multiple battery systems. By leveraging data from different battery types or systems, group training improves generalization and enhances model accuracy, particularly when individual data is sparse.  This study offers valuable insights into the strengths and weaknesses of each approach, helping identify the most effective model and optimal strategies for applying these models to real-world battery health management.