Rationalizing machine learning models for real-time health prognosis of lithium-ion batteries

Accurate, personalized, and real-time health forecasts for lithium-ion batteries are essential to ensure the safety and reliability of energy storage systems. Two critical aspects that must be jointly addressed are the State of Health (SoH) and the Remaining Useful Life (RUL). Current data-driven approaches often rely on deep neural network models, which are computationally demanding and depend on the interplay between SoH tracking and RUL prognosis.To improve both efficiency and performance, this study introduces a conceptual framework grounded in four key concepts: adaptability, complexity, memory, and synergy. We analyze and refine a state-of-the-art deep transfer learning model, identifying which components are most significant concerning these concepts. Building on this, we propose several deep neural network variants and lightweight Echo State Networks with a novel reservoir structure.Extensive experiments on three diverse battery datasets demonstrate that our models consistently outperform deep transfer learning approaches for both RUL and SoH. Notably, efficiency and performance gains are achieved through strategies that reinterpret the conventional synergy between degradation dynamics leveraged in deep transfer learning. On the largest dataset, our best models achieve reductions in root mean square error of 40% for RUL and 91% for SoH.