As the wind energy sector rapidly expands both in popularity and affordability the paper presents an attempt to improve the design of a drag-based Savonius like vertical axis wind turbine (VAWT) design. The utilised optimisation methodology involved high accuracy computational fluid dynamics model of an operating wind turbine combined with differential evolution optimisation algorithm (DE). Evaluations of the wind turbine performance has been carried out using unsteady two-dimensional numerical model. In order to reduce computational time needed to evaluate turbine performance and make the turbine shape optimisation possible, reduced order method (ROM) in a form of repeatedly trained neural network (NN) was developed. The design of the wind turbine blades has been parametrised with low order Bezier curves to allow unobstructed exploration of the design space and ensure the smoothness of the blades shape. Contributing to the aimed holistic approach to newly designed VAWT the optimisation runs have been performed for various operating regimes of the wind turbine, known commonly to be suboptimal for the Savonius type wind turbines. The performance of the optimised wind turbine designs have been verified against previously known optimised Savonius like wind turbine designs. As a result of the performed optimisation, significant increase in the value of energy conversion efficiency has been predicted, reaching as high as 35%.