Evolutionary multitasking (EMT) utilizes positive cross-domain knowledge transfer to achieve better convergence performance compared to traditional evolutionary algorithms. However, identifying beneficial knowledge and designing efficient transfer mechanisms remain key challenges. Therefore, this paper proposes an evolutionary multitasking algorithm that integrates dynamic cross-domain knowledge screening and non-negative subspace alignment (EMT-DNSA). First, a dynamic cross-domain knowledge selection mechanism is designed, where the Mahalanobis distance between the source and target task distributions is used to calculate the threshold for direct migration of beneficial solutions. Second, for untransferred solutions with large cross-domain knowledge differences, a non-negative subspace alignment strategy is proposed, using non-negative matrix factorization to extract low-dimensional representations of high-dimensional features and minimize subspace differences to reduce negative transfer. Third, the complete framework of the proposed multi-task optimization algorithm is presented, based on the complementary mechanisms of knowledge selection and subspace alignment. Finally, ablation, comparison, and verification experiments on a multi-task benchmark suite and a real-world aluminium electrolytic energy consumption problem validate the algorithm’s effectiveness. The results demonstrate significant competitive advantages over five other advanced algorithms.