Attention Mechanism with Adaptive Normalization and Mixture of Experts for Estimating Battery Capacity Degradation

Predicting the remaining useful life (RUL) of lithium-ion batteries (LIBs) with precision is essential for ensuring the reliable performance and timely maintenance of battery systems. However, the capacity of batteries continuously degrades with increased charge and discharge cycles. Statistical properties such as mean and variance often change over time, leading to distribution drift. Furthermore, the complex patterns in the data make it challenging for many existing methods to capture information effectively due to insufficient model capacity. To tackle this issue, we propose an attention mechanism that integrates a symmetric adaptive normalization layer and a Mixture of Experts model. The normalization layer removes the non-stationary information before inputting it into the model and restores the original distribution characteristics of the data through denormalization after model output. The mixture of experts (MoE) increases the model’s capacity by utilizing multiple expert experiences. The experimental results validate that the proposed model can effectively predict the RUL of LIBs and achieves up to a 20% improvement in some evaluation metrics.