BatteryMoE: An Improved Sparse Mixture-of-Experts Transformer for Battery SOH Estimation and RUL Prediction

Accurate State of Health (SOH) estimation and Remaining Useful Life (RUL) prediction on uninterruptible power supply (UPS) batteries are critical for ensuring system reliability. However, battery degradation processes exhibit significant non-linearity and non-stationarity, posing challenges for traditional deep learning models in capturing long-term dependencies and handling sensor noise. In this paper, we propose BatteryMoE, a novel normalization-free sparse Mixture-of-Experts (MoE) Transformer designed for robust battery prognostics. Distinct from standard Transformers, BatteryMoE incorporates three key physics-aware architectural innovations: (1) It introduces a dynamic error function mechanism to replace traditional normalization layers, constructing a normalization-free architecture that enhances robustness against high-variance battery data. (2) It adopts an imaginary-enhanced Rotary Positional Embedding that utilizes the full complex-valued attention score to capture both local charge–discharge patterns and global long-term degradation trends, significantly improving RUL extrapolation. (3) A sparse MoE strategy is employed to dynamically activate specialized experts for different aging stages (e.g., linear degradation vs. capacity plunge), balancing model capacity with robust predictive performance. Extensive experiments conducted on five widely used datasets show that BatteryMoE achieved significant results.