TWNN-LSTM-GSTA: A novel hybrid deep learning architecture for enhanced battery state of health prediction

Accurate state-of-health (SOH) estimation is critical to the safe and efficient operation of lithium-ion batteries. However, battery degradation is a nonlinear and non-stationary process with local fluctuations, making it difficult for conventional models to capture both long-term degradation trends and short-term dynamic variations. To address this issue, this study proposes a hybrid SOH prediction framework that integrates a trainable wavelet neural network (TWNN), recurrent neural networks, and a gated spatiotemporal attention (GSTA) mechanism. First, Pearson correlation coefficient and grey relational analysis are jointly used to identify informative health indicators. Based on the screening results, equal-voltage-drop duration, equal-current-drop duration, and maximum discharge temperature are selected as model inputs. TWNN is then employed to extract multiscale degradation features through trainable wavelet filters, while a Frobenius norm constraint is introduced to preserve filter orthogonality. Based on these extracted features, a LSTM architecture is developed to model temporal dependencies in the degradation process. In addition, the GSTA module separately learns temporal attention and feature-wise attention, and adaptively fuses them through a learnable gating mechanism to improve feature representation. Experiments on the NASA battery dataset show that the proposed method achieves accurate SOH prediction and maintains stable performance even when the amount of training data is reduced. Further validation on a laboratory lithium iron phosphate battery dataset with dynamic current profiles suggests that the model also has potential for more complex operating conditions. Nevertheless, the computational complexity of the framework and its generalizability across broader battery datasets still require further investigation.