Cross-domain thermal-health prediction and early warning for lithium-ion batteries via dual-stream network

The joint prediction of thermal states and State of Health (SOH) is vital to ensure the safe operation and reliable management of lithium-ion batteries. However, existing approaches usually model these targets independently. Such separate modeling leads to insufficient coupling awareness, restricts practical warning decision-making, and yields poor cross-domain generalization, which impedes reliable battery lifespan management. To tackle these challenges, this paper proposes a dual-stream multi-task network for joint thermal-health prediction and early warning to improve battery management systems. The architecture contains two parallel branches: a temporal convolutional network captures local thermal dynamics, while a Mamba model characterizes long-range degradation evolution. The outputs of the two branches are fed into adaptive gated fusion modules followed by task-specific prediction heads. An auxiliary temperature difference prediction task is introduced to improve thermal consistency, and a weighted loss function focusing on high-temperature regions is designed to enhance prediction reliability under safety-critical conditions. A domain distribution alignment strategy is further adopted to realize few-shot cross-environment domain adaptation. After alignment-based pre-training, a small number of labeled target-domain samples are utilized for model fine-tuning. Experiments under dynamic conditions on the XJTU dataset demonstrate that the proposed model achieves competitive performance, with a mean absolute error (MAE) of 0.011 for SOH and 0.898 °C for maximum temperature. Cross-environment transfer tests on the NASA dataset obtain an SOH MAE of 0.022. Relying on the prediction results, the proposed warning scheme achieves a strict timely response rate of 92.6%, with an average warning lead time of 48.6 cycles across test cases. These results verify that the proposed framework enables proactive thermal-risk monitoring and holds great application potential for intelligent battery management.