Introduction According to [1], with approximately 1300 TW of installed capacity and about 4200 TWh of annual production, hydropower is currently the most important renewable energy source in the world. Salient-pole generators are key components in hydropower conversion. Although very efficient machines (for large units the efficiency exceeds 98%) the inevitable losses that occur in the machine (mostly electromagnetic in nature) need to be carried away from machine to keep the temperature of the active parts within the allowable limits. This is usually achieved by means of the circulating air that collects excessive heat and transfers it to the cooling water in air-water heat exchangers. Air circulation is achieved by mounting an axial or a centrifugal fan on the rotor and/or by using existing rotor structural elements that act as primitive fans, i.e., rotor spider, interpolar space and radial ducts in the rotor rim. The inevitable consequence of air circulation through the generator are aerodynamic losses, which are also called windage losses. According to [2], windage losses are defined as: "the power taken from the generator shaft (without friction in the bearings) when the rotor rotates at a synchronous speed, at a certain flow of coolant (air)". Windage losses participate in the total losses of salient poles generators with approximately 10% to 25%. For old, high-speed and/or aerodynamically non-optimized generators, this value can reach up to 30%. Windage losses importance is illustrated by the fact that, due to the synchronous speed of rotation, their amount is constant regardless of the load at which the machine operates. The motivation for this research is unreliable models used to calculate windage losses. Prerequisite for windage loss calculations is solving the airflow networks (also called lumped parameter models), that are used to calculate the distribution of the air flow through the generator. The calculation of windage losses is based on the assumption that the total windage losses consist of the sum of friction losses of the rotor side faces and rotor end faces and the power taken by the rotor pressure elements from the turbine shaft to achieve the required cooling air flow through the generator. The aim of this research was to investigate aerodynamic phenomena within salient poles generators, to distinguish the influence of layout and geometric characteristics of the most important structural elements on generating windage losses and to determine new correlations for calculating resistance coefficients of most influential structural elements depending on basic geometric features. Furthermore, the research was aimed at defining guidelines for the aerodynamic design of those parts of the generator that are unavoidable in construction and actively participate in the generation of windage losses. Materials and methods This research is based on a statistical analysis of the basic geometric and aerodynamic parameters of the 27 manufactured and installed generators and on a subsequent detailed aerodynamic analysis of the cooling air flow through a CFD models of subset of six selected generators. Analysis of air flow through the generator was performed using the commercial CFD software ANSYS Fluent®. Detailed CFD models have been made for each selected generator. Fully predictive approach was used to model the flow within the generator meaning no inlet or outlet boundary conditions were needed. The basic idea behind this principle is to incorporate a closed air circle in the model while the airflow was driven only by the rotor rotation. Thus, the airflow created through the generator is a result of the solution and not of the boundary conditions. The RANS k-ω SST turbulence model was used in all simulations. Its implementation in Fluent® utilized the y+ insensitive wall treatment and as long as enough prism layers covered the boundary layer high quality numerical results could be achieved [3]. Since weak stator-rotor interaction was expected, rotation of the rotor domain was modelled using multiple reference frame method. The rotational periodic boundary condition on the side faces of the domain and “no slip” boundary condition on all walls were applied. The non-conformal interface approach [3] was used to connect the rotor domain with the stator. Special attention was paid to the quality of the mesh and the resulting minimal orthogonal quality of all meshes was higher than 0.14. To evaluate and quantify the numerical uncertainty arising from the discretization of the domain, the grid convergence index (GCI) method [4] was performed for validation case 3 on two generator examples with three different density meshes. Three validation cases were used to validate the selected simulation approach. First two cases were simple examples from the literature: smooth rotating disc [5] and rotating disc with protrusion [6]. Validation showed good agreement with published data (within ±3,5% for the first case, and ±5,0% for the validation second case). The third validation case was a subset of the 6 generators tested on-site. A measurements of the windage losses were performed during the comissioning of the generators according to the IEC 60034-2-2:2012 standard. Agreement of ±6,5% between calculated and measured windage losses were achieved. Results and discussion Basic statistical analysis of measured windage losses on a set of 27 built and tested generators showed that the average relative windage losses (reduced to the rated power of the generator) is 0.293% with a standard deviation of 0.1157%. Slightly less than 80% of the considered generators have relative windage losses in the range between 0,15% and 0,40%. No visible trend was observed when windage losses were correlated to the rated power of the generator. On the other hand, a linear trend is visible between measured windage losses and expressions ω3·D5 and ω3·D4·L which denote the measure of rotor end and side faces losses. Achieved coefficients of determination were R2=0,903 and R2=0,908, respectively. In order to improve the fit of the data, least squared fit to expression a·ωb·Dc·Ld was proposed, where a, b, c and d are fitting coefficients. Achieved coefficient of determination was R2=0,974, which represents an acceptable empirical model for first estimation of the windage losses in the early stage of the design process. The overall analysis of the calculated windage losses for selected set of six generators showed that an average of 80% of the losses can be attributed to the pressure rise elements used to establish airflow through the generator. Of these, more than 50% belongs to interpolar space, which makes them main single source of windage losses in salient poles generators. It is worth noting that the highest share of the poles in generation of windage losses is observed in ventilation system with axial fans (80%), and the lowest in radial ventilation without fans (30%). On average, 20% of the total windage losses is associated with end faces of the generator. These losses are called parasitic losses and are associated with the roughness of the end faces. For rough rotors, with large protrusions, they can reach more than 30%. A detailed analysis of the air flow through the generator models indicated significant interactions that occur between the different structural elements installed on the rotor. Two main issues detected are existance of different air recirculations that are main source of unreliability of existing calculation models and distribution of tangential velocity profile within the generator. Some airway details have been clarified leading to proposed changes to existing ventilation networks which are expected to increase the reliability of the ventilation calculation model. Based on the calculated aerodynamic forces on stiffeners and nuts mounted on end faces of the rotor, empirical model for calculation of the drag coefficient is proposed. In order to improve existing models for calculation of characteristics of the rotor pressure rise elements (rotor spider, radial channels in rotor rim and interpolar space) it has been shown that simplified CFD models of these rotor structures can lead to more reliable characteristics. Due to the simple geometry of these rotor elements, the construction of such models can be automated and integrated into new, improved models for calculating ventilation. Finally, changes in the existing models for the calculation of ventilation and windage losses are proposed, which are expected to lead to more reliable models for their calculation. Scientific contribution 1. An empirical model for the preliminary calculation of windage losses in the early design phase of a salient poles generators is proposed. 2. By comparing the results of measurements and numerical simulations on six generator models, it was determined that using a detailed 3d CFD generator model it is possible to calculate windage losses and airflow within engineering acceptable limits of ± 6.5%. 3. Design elements of construction that significantly affect the creation of windage losses in salient poles generators have been determined, and the mechanisms of their action in the air throughflow the generator have been clarified. 4. An empirical expressions for the calculation of the drag coefficients for stiffeners and nuts often mounted on the rotor end faces are proposed. 5. Guidelines to redefine existing airflow networks in order to reduce deviations of the calculation of ventilation and windage losses from those obtained by existing models are proposed. Conclusion The conducted research confirmed the initial hypotheses and achieved the research goals. Analysis of airflow through different models of the generators revealed the roles of main rotor components in generation of windage losses. A new correlation for estimation of the windage losses in the early stage of the design process is proposed. Interactions between different rotor components and their influence on the formation of vortices that affect the energy conversion pressure elements used to establish cooling airflow through the generator are explained. The air paths through the generator are explained and based on that, changes in the existing models are proposed in order to obtain more reliable models for calculating the air distribution through the generator and thus a more reliable calculation of windage losses.