Expensive Dynamic Multi-Objective Optimization Problems (EXDMOPs) pose significant challenges due to their dynamic and costly nature, as both the objective and constraint functions evolve over time with limited fitness evaluations. Traditional methods treat EXDMOPs as multiple independent and static expensive multi-objective problems, which often ignore the experiences obtained in solving the EXDMOPs or reuse experiences within a single EXDMOP, resulting in inefficient optimization performance. Taking this cue, in this paper, we introduce a novel approach leveraging cross-problem knowledge to enhance the ability to solve EXDMOPs. Unlike existing knowledge transfer methods within a single dynamic problem, our proposed method leverages archived solutions from well-optimized tasks to construct an effective initial population through cross-problem task selection and knowledge transfer. Specifically, a multivariate Gaussian distribution and the Wasserstein distance are adopted to choose the most appropriate task for knowledge transfer. Cross-problem knowledge transfer devises a Support Vector Machine (SVM) classifier to transfer solutions, enabling the construction of high-quality initial populations when dynamic occurs. To assess the performance of the proposed method, extensive empirical studies were conducted on commonly used EXDMOP benchmarks. The results confirm the method’s ability to enhance both the efficiency and effectiveness in solving EXDMOPs.