This work aims to develop a low-fidelity model for a lattice support structure for offshore wind applications. The proposed low-fidelity model consists of a sequence of regular Timoshenko beams, each of them characterized by homogenized mechanical and mass properties representative of the single bays of the reference space-frame structure. The homogenized elastic coefficients of the sequence of beams are then computed by means of two alternative procedures: case (a), via analytical expressions available in the literature and accounting for a partially isotropic behaviour; case (b) by means of an optimization procedure, with ad hoc calibration factors. The suggested methods to derive the homogenized elastic coefficients are then tested for both straight and tapered lattice structures. The prediction performance is evaluated in terms of estimation of the first five natural frequencies and mode shapes, response to dynamic loads, and ability to predict rotor-structure interaction phenomena. A parametric study is then performed to evaluate the potential and limitations of the proposed models. To bypass the optimization procedure (b), a data-driven approach is also proposed for the case of straight lattice structures.Overall, the developed low-fidelity model leads to a computational speed-up factor of at least 60. The prediction reliability of the low-fidelity model is discussed for a tapered and regular straight lattice structure. However, for the latter one, a more detailed comparative study between the various modelling assumptions is performed and discussed. With reference to the straight lattice tower, whenever an optimization procedure is used (case (b)), and with reference to a typical subset of the investigated geometrical parameter space, the mean prediction error of the first five natural frequencies is lower than 1%. On the other hand, for case (a) and for the same investigated subset, the mean prediction errors for the first two bending modes and the torsional mode are, 5.2%, 13.3% and 18.8%, respectively. These results are improved in case a data-driven regression model is used to predict the calibration factors, leading to mean prediction errors below 5% for the entire investigated parameter space.
Read full abstract