A structural pruning method for lithium-ion batteries remaining useful life prediction model with multi-head attention mechanism

Abstract

The utilization of data-driven models in predicting the remaining useful life (RUL) of lithium-ion batteries has gained traction. However, deploying these models on edge devices presents challenges attributed to their limited generalization capacity and substantial computational demands amidst varying operational conditions. In this study, we formulate a RUL prediction model for lithium-ion batteries by leveraging a multi-layer convolutional network alongside a multi-head attention mechanism. Through the incorporation of a domain adapter and a residual structure, we bolster the model's generalization and anti-overfitting capabilities. Additionally, we introduce an innovative global pruning strategy that hinges on scaling coefficients and adaptive sparse pruning. This strategy outperforms existing pruning methods in terms of model parameter reduction, prediction accuracy, and operational efficiency. The amalgamation of our proposed model and pruning strategy proffers a viable avenue for implementing data-driven models within embedded battery management systems.