Abstract: State of Health prediction for lithium-ion batteries is often hindered by low-dimensional monitoring data and the absence of key physical modalities, which restricts the generalization and accuracy of prediction models. To address this, this paper proposes a joint prediction method based on physics-informed heterogeneous transfer learning and an improved WOA-SVR model. First, a heterogeneous physical transfer framework is constructed. By introducing the Oxford battery dataset and UofM expansion dataset as auxiliary source domains, the missing direct current internal resistance and stack pressure features in the target domain are reconstructed through manifold alignment and physical mapping, achieving physical field completion for low-dimensional monitoring data. Second, a Grey Relational Analysis–Mutual Information ?? dual-layer evaluation strategy is proposed to quantify feature importance from both linear trend and nonlinear coupling dimensions, realizing adaptive weighted fusion of multi-source features. Finally, an Improved Whale Optimization Algorithm, incorporating nonlinear dynamic inertia weights and a global memory pool mechanism, is employed to co-optimize feature fusion weights and SVR hyperparameters. Experiments on the NASA battery dataset demonstrate that the proposed method effectively overcomes the feature deficiency problem. The SOH prediction errors across three battery cells are all below 1%, significantly outperforming benchmark models and traditional single-source data-driven methods.