Research of Ternary Lithium/Iron Phosphate Lithium Battery SOC Estimation Based on Data-Driven Model Integrating Self-Attention Mechanism

Abstract To enhance the accuracy of lithium-ion battery state-of-charge (SOC) prediction, this study develops an improved deep learning model optimized by the novel improved dung beetle optimizer (NIDBO). The NIDBO algorithm is derived from traditional dung beetle optimizer by introducing an optimal value guidance strategy and a reverse learning strategy. The deep learning model integrates convolutional neural networks (CNN), bidirectional gated recurrent units (BIGRU), and a self-attention mechanism to form the CNN-BIGRU-SA model. Subsequently, the NIDBO algorithm is employed to optimize the hyperparameters of the model, aiming to improve prediction performance. Discharge data from ternary lithium batteries and lithium iron phosphate batteries were collected. Each type of battery was subjected to 12 operating conditions, totaling 24 sets of battery operating condition data, which were used to test and validate the effectiveness of the model. The results demonstrate that the proposed model exhibits exceptional accuracy in SOC prediction, offering significant advantages over traditional methods and unoptimized models. At the same time, the model was tested under dynamic stress test and federal urban driving schedule conditions. Additionally, the generalization capability of the model is verified by cross-validating the discharge data of the two types of batteries.