A novel U-type Darrieus Wind Turbine (UDWT) is proposed by this study. During the design and optimization of UDWT, Machine Learning (ML) method based on BPNN and three optimization algorithms, GA, PSO and SA, are adopted. Besides, the quadratic Response Surface (RS) method combined with NSGA - II serves as the comparison of ML method. Meanwhile, Fluid - Structure Interaction (FSI) model, simultaneously considering the aerodynamic and structural loads, is utilized to generate the training data and validate the optimization results of ML and RS method. Moreover, the rotor performance, aerodynamic forces, pressure differential on center of blade span, velocity deficit of near-field fluid as well as the deformation, equivalent stress of UDWT are analyzed and revealed extensively. Results show that optimal UDWT can operate effectively at broader range of tip speed ratio (TSR) and the power coefficient is improved by14.74% and 53.98% at TSR of 4.0 and 4.5, respectively. Meanwhile, the novel strut type can keep the deformation and equivalent stress of rotor within the allowable limits of material when it rotates under high angular velocity. UDWT is proven to improve aerodynamic and structural performances, compared with conventional two- and four-blade wind turbine.