Battery health prediction under data scarcity: A cross-domain physics-informed 5-shot framework with GRU-Transformer

Battery health prediction faces multi-factor challenges. Significant differences exist in degradation patterns across material systems and usage scenarios. Traditional models relying on homogeneous data struggle with complex operating conditions. This study innovatively proposes a 5-shot meta-learning framework to accurately predict battery health under data scarcity. Experiments employ dual-threshold truncation for multi-source data processing, eliminating sensor noise interference. By analyzing capacity decay dynamics, we design a physics-informed loss function that embeds electrochemical mechanisms into a gated recurrent unit-Transformer hybrid architecture. Combined with cross-domain transfer strategies, the model achieves high-precision, robust predictions through physical law constraints and meta-learning fast adaptation. The constructed multi-physics synergistic model integrates diffusion-reaction-bias dynamics, while dynamic masking enhances small-sample adaptability. 349,443 cross-material query tests demonstrate model MAPE as low as 0.45 % with R2 reaching 0.9634, validating the effectiveness of physical constraints and meta-transfer mechanisms.