Advanced health state intelligent diagnosis of lithium-ion batteries based on CNN-WNN-WBiLSTM model with attention mechanism

Abstract

Efforts to enhance Battery Management Systems (BMS) for Li-ion batteries in electric vehicles are crucial for ensuring safety and performance. Our study introduces CNN-WNN-WBiLSTM-AM, a novel time series model designed specifically for forecasting Li-Ion Battery State of Health (SOH). Integrating CNN, WNN, WBiLSTM, and AM methodologies, our model effectively captures battery aging dynamics and extracts relevant time series features. We employ the RMSprop optimizer for performance optimization and global convergence during training. Experimental validation using multi-battery datasets from NASA and the CALCE dataset demonstrates the accuracy and robustness of our model. Initially, training and testing involve batteries B0005, B0006, and B0007 at 24°C, with B0018 for testing. To enhance model robustness, four Li-ion batteries operating at 43°C (B0029, B0031, and B0032 for training, B0030 for testing) are included. Additionally, batteries B0053, B0054, B0055, and B0056 operating at 4°C are used, along with the CALCE dataset (CS2-35, CS2-36, CS2-37, and CS2-38). Our results underscore the efficacy of our algorithm in Li-ion battery health management, offering higher prediction accuracy and precise multi-step SOH prediction compared to other neural network methods.