Characterizing the convective heat transfer on stator ventilation ducts for large hydro generators with a neural network

• M. Schrittwieser ^[1] ; O. Bíró ^[1] ; E. Farnleitner ^[2] ; G. Kastner ^[2]
  1. [1] Graz University of Technology

     Graz University of Technology

     Graz, Austria

     
  2. [2] Andritz Hydro GmbH
• Localización: Compel: International journal for computation and mathematics in electrical and electronic engineering, ISSN 0332-1649, Vol. 34, Nº 5 (Special Issue: IGTE 2014), 2015, págs. 1522-1536
• Idioma: inglés
• Enlaces
  • Texto completo
• Resumen

  • Purpose – The purpose of this paper is to approximate the convective heat transfer using a few non-dimensional parameters in the design process of large synchronous machines. The computed convective wall heat transfer coefficient can be used in circuit models or can be defined in numerical heat conduction (HC) models to compute the thermal field in the solid domains without the time consuming computation of the fluid domain.

    Design/methodology/approach – Computational fluid dynamics (CFD) has been used to include a wide range of different designs, operating conditions and cooling schemes to ensure accurate results for a wide range of possible machines. Neural networks are used to correlate the computed heat transfer coefficients to various design parameters. The data set needed to define the weights and bias layers in the network has been obtained by several CFD simulations. A comparison of the evaluated solid temperatures with those obtained using the conjugate heat transfer (CHT) method has been carried out. The results have also been validated with calorimetric measurements.

    Findings – The validation of the HC model has shown that this model is capable of yielding accurate results in a few minutes, in contrast to the several hours needed by the CHT solution. The workflow to determine the convective heat transfer in a specific part of an electrical machine has been also been established. The approximation of the convective wall heat transfer coefficient is shown to be obtainable in sufficient detail by using a neural network.

    Research limitations/implications – The paper describes a method to approximate the convective heat transfer accurately in a few seconds, which is very useful in the design process. The heat convection can then be characterized in a HC model including the solid domains only. The losses can be defined as sources in the solid domains, e.g. copper and iron, obtained by electromagnetic calculations and the thermal field can hence be easily computed in these parts. This HC model has the main advantage that the time consuming computation of the fluid domain is avoided.

    Originality/value – The novelty in this work is the approximation of the convective heat transfer by using a neural network with an accuracy of less than 5 percent as well as the development of a HC model to compute the temperature in the solid domains. The investigations presented pinpoint relevant issues influencing the thermal behavior of electrical machines.