A comparative study of prognostic strategies for battery SoH: Cycle-level estimation vs. sequential forecasting

Lithium-ion batteries are vital for safety-critical applications, particularly in Unmanned Aerial Vehicles (UAVs) where failure can lead to catastrophic asset loss. Accurate prediction of their State of Health (SoH) and Remaining Useful Life (RUL) is essential for robust Battery Management Systems (BMS). However, the field currently lacks a unified benchmark comparing two fundamental prognostic philosophies: (1) feature-based static estimation and (2) raw-data sequential forecasting. This study addresses this gap by comparing seven AI models across these two strategies. Models included classical (Linear Regression, SVM, k-NN), ensemble (Gradient Boosted Trees, Bagging), and sequential (LSTM, and the novel RWKV) architectures. Our results reveal a critical distinction between strategies. We find that classical Linear Regression yields high accuracy ( R 2 = 0.9889) specifically when paired with cycle-aggregated features, highlighting the efficacy of explicit feature engineering for static estimation tasks. Conversely, the novel RWKV architecture – evaluated here for the first time in battery prognostics – achieved second-best overall performance ( R 2 = 0.9166), outperforming LSTM while maintaining linear computational complexity. Crucially, cross-battery validation revealed distinct generalization patterns: ensemble methods demonstrate robust cross-battery performance (mean R 2 = 0.74), while sequential models excel on battery-specific data (mean R 2 = 0.82) but face significant generalization challenges. This work establishes a comparative framework for prognostic strategies, validates the potential of RWKV for on-board BMS, and provides evidence-based model selection guidance for real-world BMS deployment.