Unlike conventional horizontal-axis turbines, Darrieus Vertical Axis Wind Turbines (VAWTs) perform effectively in environments where wind direction is highly unpredictable. Thanks to their compact design and ability to capture wind from any direction, they are commonly used in industrial parks, telecommunications stations, offshore platforms, and urban smart grids. However, scaling VAWTs to larger sizes presents challenges due to manufacturing and transportation constraints, limiting their use mostly to small-scale distributed power generation systems. As a result, optimizing turbine design is crucial for enhancing energy output and efficiency. In this work, we compare different machine learning (ML) models trained to predict the aerodynamic forces on a three-bladed VAWT. In particular, we aim to analyze the effects on the power output by perturbing the initial airfoil geometry (NACA0015) in terms of thickness and curvature, with different Reduced Order Models (ROM) approaches. Such data driven models are trained on transient 2D computational fluid-dynamics (CFD) VAWT simulation results. We compare deep-learning (DL) geometric encoding ROMs with non-intrusive multi-fidelity Proper Orthogonal Decomposition (POD) -based ROMs. The DL approach, based on DeepSDF [1] architecture, exploits an implicit representation of the airfoil in terms of signed distance function (SDF) to learn a latent encoding of the airfoil shape, without the need of providing an explicit parametrization. In this work, we leverage this parametrization-agnostic technique to approximate forces acting on blades of arbitrary shape by training a second ML model that maps the learned latent space to the target outputs. To this end, we introduce an improved regularization of the latent geometric space which enforces the coherence between distances in the latent space and those in the SDF space. This industrially relevant test case demonstrates the potential of both the DL-based and POD-based models to approximate vectorized outputs on varying geometries, with the DL approach overcoming the limit of relying on an a priori parametrization. The proposed approaches find valuable application in time-consuming tasks like design and optimization which typically require multiple expensive CFD solutions.