The capacity knee point represents a critical transition in the performance degradation of lithium-ion batteries and serves as a crucial indicator for assessing the safety of electrochemical energy storage systems. To address the issues of inadequate multi-source monitoring data fusion in existing knee point prediction methods

which result in low prediction accuracy

poor generalization capability

and limited application scenarios

this study proposes a lithium-ion battery capacity knee point prediction method based on multi-parameter data fusion decision-making. Firstly

the knee point is identified by integrating the Bacon-Watts method with a sliding window approach

and the correlation between the temporal degradation characteristics of current

voltage

and temperature parameters with the knee point is analyzed. Secondly

a Transformer-AGCN-MoE prediction model is constructed. The Transformer module extracts temporal features from multiple parameters

and an attention-enhanced graph convolutional network (AGCN) mines nonlinear correlations across different parameters and fuses the multi-parameter features. Then

a mixture of experts (MoE) system dynamically optimizes the decision-making for knee point prediction. Finally

the model’s prediction performance is validated using multiple public battery datasets

including those from Massachusetts Institute of Technology(MIT) and Xi’an Jiaotong University. The experimental results demonstrate that

by using only four consecutive cycles of data

the proposed method achieves a coefficient of determination (R²) exceeding 0.98 and a root mean square error (RMSE) of less than 27 cycles for knee-point prediction

indicating high prediction accuracy along with good applicability and interpretability.