<p>Data-driven evolutionary algorithms (DDEA) have proven to be highly effective in solving expensive and computationally intensive problems. However, offline DDEA heavily relies on the availability of historical data. Additionally, existing methods that construct a single surrogate model are well-suited for specific issues but may not generalize well to others. Therefore, this paper proposes a data-driven evolutionary optimization method with clustering-aided ensemble learning and Taylor polynomial-based data generation (CELDDEA-TPDG). In CELDDEA-TPDG, we incorporate a data synthesis mechanism based on the Taylor polynomial. This method is inspired by the difference approach in the finite element method (FEM). By leveraging the first-order difference scheme to assist in derivative estimation and utilizing the principles of the Taylor polynomial, synthetic data are generated within the neighborhood of historical data. Furthermore, a clustering module is incorporated into the ensemble learning process to selectively filter base learners, thereby enhancing the performance of the ensemble learning model. To evaluate the performance of CELDDEA-TPDG, we first constructed multiple variants of the CELDDEA-TPDG method to investigate the effectiveness of its components. Furthermore, we demonstrated the advantages of CELDDEA-TPDG by comparing it with ten other methods (both traditional and State-of-the-art (SOTA)) on benchmark functions. The results indicate that, compared with the six other state-of-the-art algorithms, CELDDEA-TPDG consistently ranks first with higher stability. Finally, in a real-world hydrodynamic performance optimization problem, we compared CELDDEA-TPDG with the classical DDEA-SE method to illustrate its superiority in solving practical engineering problems. The results show that CELDDEA-TPDG can achieve better performance. The data used in this work is available at <a href="https://github.com/Tenhree/Marine-Propeller-Data.git">https://github.com/Tenhree/Marine-Propeller-Data.git</a>.</p>