State of Health Estimation and Prediction Based on Real-Operating Data of Lithium-Ion Phosphate Power Battery

Abstract

Abstract. The precise assessment and prognostication of lithium-ion battery health status are vital for the holistic lifecycle management and the realization of gradient utilization of batteries. The present prevailing methodologies are predominantly anchored in controlled laboratory settings, which fail to encompass the complexity of the actual operation mode and the inconsistency of the battery cell. This article introduces a novel framework for the state of health (SOH) estimation based on the internal characteristic data of actual operation. Initially, the dataset is preprocessed to determine battery capacity, and the degradation is then qualitatively assessed using generalized additive models. Subsequently, health features are extracted from the operational dataset, with a stringent screening process employing Spearman's rank correlation analysis to identify features with significant correlation. The framework culminates in the construction of a Bayesian-optimized long short-term memory (BO-LSTM) model for the SOH estimation, leveraging Bayesian optimization to fine-tune the LSTM's weights and biases. Experimental results indicate that the proposed BO-LSTM model surpasses the conventional LSTM and gated recurrent unit architectures in battery health prediction tasks, with a peak prediction error maintained below 4%.