Step-adaptive residual attention network for multi-step battery temperature prediction

To address the limitations of existing lithium-ion battery temperature prediction models, including insufficient utilization of input features, suboptimal allocation of gradient resources across prediction steps, and the limited representational capacity of convolutional modules, this paper proposes a novel multi-step temperature prediction model named Step-Adaptive Residual Attention Network (SARA-Net). First, a Broad Learning System (BLS) is introduced to efficiently expand the input feature space through mapped nodes and enhancement nodes, enhancing the extraction of deep information from limited monitoring variables. Second, a Gated Sample Convolution and Interaction Network (GSCINet) is employed to extract multi-scale local temporal features via odd-even subsequence splitting, independent processing with dual convolution kernels, and gated interaction learning. Third, a Bidirectional Long Short-Term Memory network (BiLSTM) is used to model global temporal dependencies, capturing long-term dependencies and lagged responses in battery temperature evolution. Finally, a Step-Adaptive Residual Attention Loss function (SARA Loss) is designed to dynamically adjust the loss weights of different prediction steps, effectively achieving an optimal allocation of limited gradient resources. Experiments are conducted on a dataset generated by Fluent finite element simulation under 12 charging/discharging conditions. The results show that under seen conditions, the proposed model reduces Mean Squared Error (MSE) by more than 71.2% compared to traditional models such as Temporal Convolutional Network (TCN) and Long Short-Term Memory network (LSTM), and by more than 63.2% compared to advanced models such as DeepTimeNet. Under unseen conditions, the MSE is reduced by more than 48.2% compared to BIATCN-Informer and other advanced models. Under noisy conditions, the proposed model still outperforms all the other compared models. Ablation studies further verify the effectiveness and necessity of each module. The proposed model provides reliable technical support for battery thermal runaway warning and health status assessment.