Stochastic Modeling of an Inhomogeneous Magnetic Reluctivity
Transcription
Stochastic Modeling of an Inhomogeneous Magnetic Reluctivity
XXIV Symposium Electromagnetic Phenomena in Nonlinear Circuits June 28 - July 1, 2016 Helsinki, FINLAND ______________________________________________________________________________________________________ Stochastic Modeling of an Inhomogeneous Magnetic Reluctivity Using the Karhunen-Loève Expansion and the Hierarchical Matrix Technique Radoslav Jankoski1,2 , Ulrich Römer1,2 and Sebastian Schöps1,2 1 2 Technische Universität Darmstadt, Institut für Theorie Elektromagnetischer Felder, Darmstadt D-64289, Germany Technische Universität Darmstadt, Graduate School of Computational Engineering, Darmstadt D-64293, Germany {jankoski, roemer, schoeps}@gsc.tu-darmstadt.de Abstract—In this paper the magnetic reluctivity of the core of a single phase transformer has been modeled as a random field by using the Karhunen-Loève expansion (KLE). The KLE enables representing random fields with a minimal number of random variables but on the other hand it requires computing eigenvalues and eigenvectors of a generalized eigenvalue problem with dense matrices. To overcome this difficulty the Lanczos algorithm has been combined with the hierarchical matrix technique in order to reduce the complexity of performing arithmetical operation with dense matrices. and it can be used for uncertainty quantification (UQ). One can study the impact of the uncertainties introduced in the magnetic reluctivity on the output parameters of the electrical machines. One example, which is presented in this paper, is the uncertainty of the spatial distribution of the magnetic reluctivity in a single phase transformer where, for instance, its inductance can be an output parameter of interest. I. INTRODUCTION For optimization of electrical machines that contain magnetic materials, as a part of their construction, having an accurate knowledge of the magnetic behaviour law expressed via the magnetic reluctivity plays an important role. However, in practice there is a lack of knowledge of the magnetic reluctivity because the industrial process of production introduces uncertainties in this material property. To obtain reliable results from numerical simulations it is necessary to study the impact of those uncertainties. The uncertainties in the material properties are typically modeled as discrete random variables as it is done in [3]. In this paper we propose to use the Karhunen-Loève expansion (KLE), to discretize the random field, due to its error minimizing property, i.e., the truncation error is minimal when compared to any other mterm approximation in the mean square sense [4]. The KLE of a random field requires the solution of a computationally expensive generalized eigenvalue problem with a fully populated (dense) matrix on the left hand side of the matrix equation. In order to reduce the computational costs for solving the generalized eigenvalue problem, authors in [5] have combined the iterative Krylov subspace (Lanczos) method with the hierarchical matrix technique. The Lanczos eigenvalue solver requires only matrix-vector multiplications with a dense matrix as an arithmetical operation. By using the hierarchical format of matrices the complexity of the matrixvector multiplications is reduced. Once the KLE is computed the magnetic reluctivity can be represented via minimal number of a random variables The system of interest is shown in Fig. 1. The problem has been simplified by analyzing the system in a 2D representation in a domain Ω with boundary ∂Ω. II. P ROBLEM D ESCRIPTION Fig. 1: Single phase transformer The behaviour of the system is described with Ampere’s law in the 2D magnetostatic formulation, −∇ · (ν(·, ω)∇Az (ω)) = Jz Az (ω) = 0 in Ω, on ∂Ω, (1) where Az is the z component of the magnetic vector potential and ν is the magnetic reluctivity of the core that depends on the outcome of a random event ω. The current sources are given by the current density Jz . To obtain the KLE for the magnetic reluctivity the eigenvalues and the eigenfunctions need to be computed efficiently in the domain where the core of the single phase transformer is located. ______________________________________________________________________________________________________ 143 III. K ARHUNEN -L O ÈVE E XPANSION AND H IERARCHICAL M ATRICES A. Karhunen-Loève Expansion The correlation length of the random field is denoted as d and d = 1 for this computations. The first eigenfunction is shown in Fig. 2 for illustration purposes. The KLE is given for an inhomogeneous magnetic reluctivity as follows: ~ ν(~x, ω) ≈ ν(~x, ξ(ω)) = ν(~x) + m p X λi fi (~x)ξi (ω), (2) i=1 where ν is the mean value of the random field, ~x is a vector of spatial coordinates and ξ~ is a vector of mutually uncorrelated random variables. The random variable ξi is evaluated as follows: Z 1 (3) ξi (ω) = √ (ν(~x, ω) − ν(~x))fi (~x)d~x. λi Ωc The eigenfunctions fi and the eigenvalues λi that appear in the KLE are obtained by solving the Fredholm integral equation: Z Cov(~x, ~y )fi (~y )d~y = λi fi (~x), Fig. 2: First eigenfunction Fig. 3 depicts the magnitude of the largest 30 eigenvalues. (4) Ωc where Cov is the covariance function, which is usually deduced from measurements, and Ωc is the domain of the transformer core. The domain Ωc is triangulated and the solution for the eigenfunction fi is approximated as piecewise constant within one triangle. By applying the Galerkin method the equation (4) results in a generalized eigenvalue problem, Af = λBf , (5) with matrices A, B ∈ Rn×n and eigenvector f ∈ Rn . The number of triangles is denoted as n. Matrix A is a dense and symmetric and matrix B is diagonal. B. Hierarchical Matrices The idea of the hierarchical matrix technique is to find subblocks of the matrix and perform a low-rank approximation that enables matrix vector multiplication with almost linear complexity O(n log n). A certain subblock à in the matrix A is low rank approximated if it satisfies the so called admissibility condition which is related to the geometry of the system. If two blocks of the domain of interest are well separated by a distance their mutual interaction is negligible and thus the low-rank approximation is allowed. Fundamental theory of hierarchical matrices can be found in [2]. IV. R ESULTS For verification of the numerical algorithm the generalized eigenvalue problem given with the equation (5) has been solved by using the open source H2Lib library [1] and our own implemented Lanczos eigenvalue solver. The covariance function is assumed to be given in advance instead of deducing it from measurements, ! ||~x − ~y ||l1 Cov(~x, ~y ) = exp − . (6) d Fig. 3: Eigenvalues A rather slow decay of the eigenvalues is observed, i.e., only one order of magnitude. V. C ONCLUSION AND O UTLOOK An efficient approach for computing the KLE based on the Lanczos algorithm and the hierarchical matrix technique has been presented in this paper. The KLE has been computed in a domain where the core of a single phase transformer is located. The computational benefits will be discussed in the full paper where the presented approach is compared with the standard eigs MATLAB function (also based on Lanczos algorithm). In future work the KLE will be applied for stochastic modeling of magnetic materials with a nonlinear, inhomogeneous and anisotropic magnetic reluctivity. VI. ACKNOWLEDGMENT The work in this paper has been supported by the Graduate School of Computational Engineering (GSCE) at the Technische Universität (TU) Darmstadt. R EFERENCES [1] H2lib public repository. https://github.com/H2Lib. [2] M. Bebendorf. Hierarchical Matrices: A Means to Efficiently Solve Elliptic Boundary Value Problems. Springer, 2008. [3] R. Ramarotafika, A. Benabou and S. Clénet. Stochastic modeling of soft magnetic properties od electrical steels: Application to stator of electrical machines. IEEE Transaction on Magnetics, 48(10):2573–2584, 2012. [4] R. G. Ghanem and P. D. Spanos. Stochastic Finite Element Method: a spectral approach. Springer-Verlag New York, Inc., 1991. [5] B. N. Khoromskij, A. Litvinenko and H. G. Mathies. Application of hierarchical matrices for computing the Karhunen-Loève expansion. Computing, 84:49–67, 2009. ______________________________________________________________________________________________________ 144 Proceedings of EPNC 2016, June 28 - July 1, 2016 Helsinki, FINLAND