FINITE ELEMENTS

published by WSEAS Press

54

Domain decomposition for coupled fields in electrical engineering

ION CRSTEA Department of Computer Engineering and Communication University of Craiova Str. Doljului nr. 14, bl. C8c, sc.1, apt.7, Craiova ROMANIA incrst@yahoo.com http://ioncarstea.webdrisign.net
Abstract: This chapter presents some computational aspects for coupled magnetic and thermal fields in the context of the finite element method, with emphasis on the reduction of the computing resources. We present a coupled model for the two fields: magnetic field and thermal field. The mathematical model for magnetic field is based on time-harmonic Maxwell equations in vector magnetic potential formulation for axisymmetric fields. The model for the heat transfer is the heat conduction equation. A numerical model based on the finite element method is developed. We propose simplified models for coupled fields in electromagnetic devices with emphasis on the induction heating devices. The dynamic interfaces of the subdomains in the induction heating simulation can be exploited in the sense of reduction of the computational effort. The analysis domain is divided into two overlapping subdomains for the two coupled-fields considering physical significance of the pseudo-boundary of the two subdomains. Key-Words: - Domain decomposition; Coupled fields; Finite element method.

1 Introduction
In many engineering applications the general configuration of the field domain can be viewed as a union of a finite number of regions with distinct physical properties. For example, in 3D eddy-current analysis we can identify the following sub-domains: A region represents the free space with supply currents in it A region consists of non-conducting magnetic material whose permeability may have linear or non-linear characteristics A region with eddy-currents, where electric conductivity is finite and the field problem is linear Recently, there were developed valorous papers in the area of decomposition techniques for Maxwell equations, especially for time-harmonic case [7]. The motivations of this interest are based on the inherent advantage of the decomposition techniques for the parallel computing. We try to present the basis of this technique with emphasis on coupled problems in 2D applications. To analyse a large complex system it is necessary for economic reasons in terms of computer resources, to divide the system into a number of small components or subsystems. Each component is considered as though it was separate problem and it is solved independently. The subsystems are finally

coupled together in such way that interface conditions are satisfied at common boundaries. When dividing a major system into n subsystems, for each individual subsystem any constraints are imposed. In a finite element program, the nodes which are on common boundaries of the subsystems are called as interface nodes while all others are referred to as internal or non-interface nodes. Having developed all the necessary conditions for the individual subsystem, the mathematical model is solved by a numerical method. The physical systems have the mathematical models a set of partial-derivatives equations (PDEs) and only in a first approximation we use lumpedparameter models. Although a tremendous variety of numerical methods have been proposed for simulation of these systems, the most recently invented parallel computational strategies are largely based on the finite element (FE) and multigrid methods. The programs for the simulation of the distributed-parameter systems have an inherent parallelism when FE method is used. The fact that the finite element method is central to many modern engineering simulations and the existence of a large amount of software products exists in this area are motivated by some important reasons: algorithmic issues raised by the FE method in comparison with other methods like the difference finite methods

FINITE ELEMENTS

published by WSEAS Press

55

handling discrete meshes with an irregular or complicated distribution of points. The matrix of linear equation coefficients has not a regular, predictable structure characteristic of the finite difference method the facility with which the FE method can be used to handle many physical problems described by partial differential derivatives equations The technique of dividing a large physical system into a system of components is very old and is still used extensively. In this way different components are designed in parallel by different groups of researchers or companies. It is obviously that this traditional approach can be used with parallel computers if the complete finite element system is subdivided so that each group of elements within a small domain is assigned to one processor. We shall limit our discussion to the solution of partial differential equations (PDE) in 2D space using a distributed unstructured mesh of triangles with linear approximations within an element (piecewise linear elements), although the ideas presented here do extend to higher order elements. Although we present some target examples, the issues of the parallel implementation of FE programs are not strictly dependent upon these particular cases.

What kinds of boundary conditions are set up on the pseudo-boundaries of the sub-domains What kind of domain decomposition is useful for a particular problem: static or dynamic decomposition?

2.1 Decomposition techniques

The desire of the scientific community for faster processing on lager amounts of data has driven the computing field to a number of new approaches in this area. The main trend in the last decades has been toward advanced computers that can execute operations simultaneously, called parallel computers. For these new architectures, new algorithms must be developed and the domain decomposition techniques are powerful iterative methods that are promising for parallel computation. Ideal numerical models are those that can be divided into independent tasks, each of which can be executed independently on a processor. Obviously, it is impossible to define totally independent tasks because the tasks are so intercoupled that it is not known how to break them apart. However, algorithmic skeletons were developed in this direction that enables the problem to be decomposed among different processors. The mathematical relationship between the computed sub-domain solutions and the global solution is difficult to be defined in a general approach. In the area of the coupled fields we define two levels of decomposition, that is we define a hierarchy of the decompositions: One at the level of the problem The other at the level of the field In other words, we decompose the coupled problem in two sub-problems: an electromagnetic problem and a thermal problem, each of them with disjoint or overlapping spatial domains. This is the first level of decomposition. At the next level, we decompose each field domain in two or more subdomains. The decomposition is guided both by the different physical properties of the materials, and the difference of the mathematical models. At this level of decomposition the Steklov-Poincar operator can be associated with field problem [7]. This operator reduces the solution of the coupled subdomains to the solution of an equation involving only the interface values. One efficient and practical solution of elliptical partial differential equations is the dual Schur complement method [7]. For simplicity we limit our discussion at the electromagnetic problem. Domain decomposition for the time-harmonic Maxwell equations in the case of low frequency is presented in many works [3]. This

2 Problem formulation
The domain decomposition method is the best among three possible decomposition strategies for the parallel solution of PDEs, namely, operator decomposition, function-space decomposition and domain decomposition. This is one of the motivations to present the principles of the domain decomposition methods in this section. The domain decomposition could be determined from mathematical properties of the problem (real boundaries or interfaces between subdomains), or from the geometry of the problem (pseudoboundaries). For elliptic partial differential equations, there exists a mathematical approach based on the ideas given earlier in 1890 by Schwarz [1]. In Schwarz procedure there is an inherent parallelism with a data communication time for the passage of pseudo-boundary data between processors. There is no general rule for the domain or/and operator decomposition. It is defined in a somewhat random fashion. The problems and questions that appear in the decomposition technique are: do domain decomposition or the operator decomposition Which approach is the best: disjoint or overlapping sub-domains?

FINITE ELEMENTS

published by WSEAS Press

56

approach is applied for some practical cases with direct application in the high voltage equipment where the conductor and insulation form the parts of the electromagnetic device. For example, in a high voltage (HV) electric cable the medium that defines the device consists of a conductor and an insulator. Finite element domain decomposition can be used and an iterationby-subdomain procedure is applied. A complete physical description of electromagnetic field is given by Maxwells equations in terms of five field vectors: the magnetic field H, the magnetic flux density B, the electric field E, the electric field density D, and the current density J. In low-frequency formulations, the quantities satisfy Maxwells equations [2]: H = J (1) B E = (2) t div B = 0 (3)
(4) div D = c with c the charge density, the electric conductivity, and the magnetic permeability. For simplicity we give up to the bold notations for vectors. The second set of relationships, called the constitutive relations, is for linear materials: B = H ; D = E ; J = E The B-H relationship is often required to represent nonlinear materials. The current density J in Eqn. (1) must represent both currents impressed from external sources and the internally-generated eddy currents. The formulation with vector and scalar potentials has the mathematical advantage that boundary conditions are more often easily formed in potentials than in the fields themselves. The magnetic vector potential is a vector A such that the flux density B is derivable from it by the curl ( ) operation: B = A We present in brief the domain decomposition for the heterogeneous medium composed by two parts: one is a homogeneous isotropic conductor, and the other one is a perfect insulator [3]. This case corresponds to the HV cables used in the electric transport networks. In other words the conductivity is a strictly positive constant in the conductor, say c, of , while is equal to zero in the complementary part i=\c. The mathematical model with E and H the electric and magnetic field is defined by Eqn. (1) and Eqn. (2). The Dirichlet boundary condition for the electric field is: (n E ) =

depending on the space variable x and the field quantities, and are assumed to be uniformly positive in a conductor. For a perfect insulator is zero. In coupled problems they depend on temperature. Assuming the E and H are time-harmonic with angular frequency , the model for linear problems takes the well-known form in complex quantities [1]: H E = 0 (5) E + jH = 0 (6) where H and E represent the phasors of the field quantities. For constructing an approximation procedure, an appropriate equivalent two-domain formulation is used. In each subdomain one has to solve an equation of simple type and the coupling of the two solutions is done by means of the interface conditions. The algorithm based on the domain decomposition has the following form: 1. Assign interface conditions to subdomains 2. Solve independently the problems in each subdomain 3. Update the interface conditions in each subdomain in accordance with the new solution values 4. Repeat the steps 2-3 until a convergence criterion is fulfilled In each subdomain the finite element approximation can be used. We followed the approaches presented in reference [7]. The advantage of this approach is obviously: in each subdomain one has to solve an independent field problem, that is the initial complex problem is divided in more simple problems that are solved by different solvers. A measure of the simplicity can be defined in terms of the mathematical operations and the computer resources [1].

3 A parabolic-elliptic problem
The coupled problem in the induction heating devices is partly parabolic and partly elliptic so that this special case can be called a parabolic-elliptic problem. A discrete formulation can be obtained by the finite element method. The existence and the uniqueness of the approximate solution are proved in references so that we present only the final results with emphasis on the 2D problems for the electromagnetic problem [4]. The motivation for studying this case is provided by one of the engineering approaches to the computation of the eddy-currents where the electromagnetic field in different regions of the field domain has different behaviours described by elliptic and parabolic equations. For example, in the workpiece we have a parabolic equation and in the air or insulation the

where is a given tangential vector field defined on and n is the unit outward normal to . The quantities and are, in general, symmetric matrices,

FINITE ELEMENTS

published by WSEAS Press

57

electromagnetic field is described by an elliptic equation [2]. Let us consider the field domain is divided in two disjoint subdomains P and E where the subdomains correspond to parabolic and elliptic models, respectively. All quantities from the two subdomains will be indexed with the subscript P and E respectively. We consider two disjoint subdomains, that is,

= P E ;

P E = .

We

will

set

= . The problem in discussion can be P E formulated as [4]: Find a function u: x (0,T) R1 such the equations u u 2 P = ( P P ) + f P in P (0, T ) (7) t xi i =1 xi u 2 ( E E ) + f E = 0 in E (0, T ) (8) xi i =1 xi where 0<T< and (x)>0, M=(x, |grad u(x,t)|), fM(x,t) are given functions. In our case has the meaning of the electric conductivity, is the magnetic reluctivity (the inverse of the magnetic permeability) and the right-hand terms f is the one component of the current density. The unknown function u is the component of the magnetic vector potential A, that is u=Az or u=A in planeparallel model and axisymmetric case, respectively. Equation (7) is of parabolic type and Eqn. (8) is of elliptic type. These equations are solved with specified boundary conditions on the boundary and initial condition for the parabolic equation. For example, we can have homogeneous Dirichlet and Neumann conditions on different parts of the boundary by the form: u = 0 on 1 (0, T )
= 0 on 2 (0, T ) n where n is the unit outward normal to =1 U 2. The initial condition is prescribed on P in the form: u P ( x,0) = u P 0 ( x) x P The two subdomains are separated by a common boundary so that on this boundary the function u has to satisfy so called transition condition for any time t. This means that at every x from the common boundary B = PE , the limit value of uP is equal to the limit of uE , and u

on B n n where n denotes the unit normal to the common boundary oriented in a unique way. The two field equations can be written in a compact form: 2 u u ( ) + f in 0 (0, T ) (9) = t i =1 xi xi where the field domain is 0=-B and =0 in E. In engineering applications the electric conductivity is considered piecewise constant. It is a positive function in conductive parts and is equal to zero in non-conductive parts of the electromagnetic device. The reluctivity is dependent on the unknown u in the form: ( x, s ) = M ( u ) x M M = P, E with M(s) a function of class C1([0, )). The above equations take particular form for Maxwell equations (1) and (2). In the A-formulation for 2D problems, u represents one component of A. In 2D problems the vector A= (0,0,A), and the vector J=(0,0,J) so that the unknown function u =A and f=J. Transition conditions take the following particular form: E = B P on H tE = H tP ; Bn n B where subscripts t and n refer to the tangential components of the magnetic field strength, respectively normal components of the magnetic induction B. The superscripts refer to the problem type (E for elliptic and P for parabolic problem). In the two-dimensional eddy-current case, the mathematical model in A-formulation for timeharmonic case is [1]: ( A) jA = J s (10) with Js the source current density. Any continuous problem has a weak form obtained by multiplying the PDE by a function v, integrating over the domain , and performing integration by parts [6]. The integral statement is a kind of average and hence weaker than the pointwise requirement of PDE. A numerical model based on finite element method can be developed starting with the variational formulation of the problem [6]. The variational form associated to Eqn. (10) is given by [1]: F ( A) = [ (A) 2 + jA2 2 J s A]d (11) where represents the whole of the twodimensional problem space. Because the eddy-current density is J=-jA, a boundary condition for A can be viewed as a

u P

=E

u E

FINITE ELEMENTS

published by WSEAS Press

58

boundary condition for the eddy currents. This idea can be exploited in numerical model where the solution is driven by directly specifying the eddy currents on the boundary. The thermal field is described by the heat conduction equation [5]: [(c )(T ) T ] + [ k (T ) T ] = q (12) t where: T(x,t) is the temperature in the spatial point x at the time t; point k is the tensor of thermal conductivity; is mass density; c is the specific heat that depends on T; q is the density of the heat sources that depends on T. In the coupled problems we use the formula [1]: q = (T ) J 2 (13) with the electrical resistivity of the material. Equation (8) is solved with boundary and initial conditions. The boundary conditions can be of different types: Dirichlet condition for a prescribed temperature on the boundary; convection condition; radiation condition, and mixed condition [1].

In the usual procedure we express the temperature field and its gradients within each element as
T e = N iTi = [N]{T }
i =1 M

{T} = [B]{T} where [N] is the row-vector of the field interpolation, and [B] is the field-gradient interpolation matrix: [B] = {N} Ti is the value of T at node i and {T} is the column vector of element nodal values. In practice the shape interpolation functions Ni are polynomials with the degree 1 or 2. We use the symbol [ ] for a rowvector, and { } for a column vector. The Galerkins method is used to derive the element equation. Thus, after some manipulation, the resulting element equations are [1]:
([K C ] + [K h ]){T} = {R f } + {Rq } + {Rh }

where [ K C ] = [ B ]T [k ][ B ]d
e

3.1. Numerical models based on the finite element method

In a numerical solution of the mathematical models for coupled problems, we determine an approximate solution for the unknown function at a finite number of discrete points in the domain. The finite element method (FEM) is presented in a large professional literature so that we limit the discussion to a brief description. FEM involves an important concept, the element, a domain area delimited by line segments or curves between grid points. Every grid point is a part of least one element. As example let us discuss the steady-state case of the thermal field [1]. We consider Dirichlet, Neumann and convection boundary conditions on the boundary C=C1 U C2 U C3 by the form: T ( x , y ) C = T d ( x, y )
1

[ K h ] = h{N }[ N ]ds
C3

{R f } =

q{N}d ;
e

{Rq } = q n {N }ds
C2

{Rh } = T {N }ds
C3

[k [k

T +q ] =0 n n C2 T + h(T - T )] = 0 n C3

where Td is a prescribed temperature (Dirichlet condition), q n is the value of the heat flux on the boundary (Neumann condition), h is the convection coefficient, and T is the environment temperature. In the general case of FE method, the solution domain is divided into M elements of r nodes each.

The superscript T denotes the transposed matrix. The coefficient matrices [Kc] and [Kh] are element conductance matrices and relate to conduction and convection, respectively. The convection matrix is computed only for elements with surface convection. The vectors {Rf}, {Rq} and {Rv} are heat load vectors arising from specified internal heat generation, specified surface heating and surface convection, respectively. The convection heat load vector is computed only for elements with surface convection. In the FE formulation there exist nonzero matrix elements connecting any two unknowns Ti and Tj whose points share a common element. It is obviously that the task of computing matrix entries may be broken up into computation of element matrix contributions that may be sequentially assembled into the global matrix system. It is the proper computation of the element matrices that is one of the principal topics of the finite element treaties and determines in part the accuracy of the solution [1]. To obtain the global system equations, the assembly of the element equations follows the standard procedure of the finite element method. Note that since the nodal unknowns Ti are scalars, no

FINITE ELEMENTS

published by WSEAS Press

59

transformations of matrices computed in local coordinates are necessary before assembly of the global matrices. The transient thermal fields involve the computation of a matrix C called capacitance matrix. The elemental contribution of this matrix is defined by: [C ]e = (c )[ N ]T [ N ]d
e

For the time-harmonic field the stationary functional defined by the relationship (11) leads to a matrix system. Decomposing the field domain into finite elements e, the function A is sought in terms of a polynomial approximation: n A = N j ( x , y ) A j (t ) j =1 with Nj interpolation functions, n - the number of nodes of the element, and Aj are unknown coefficients. In the case of the time-harmonic field in 2-D space, a matrix equation is obtained [1]: [S]{A} + j[T]{A= {J} where the elements of the matrices [S], [T] and the column vector {J} are given by: N N j N i N j + ) d S ij = ( i x x y y e

In general case, the heat transfer is described by the following equation with the boundary conditions incorporated [2]: d [C] {T} + [K]{T} + {R} = 0 dt The numerical solution can be obtained by an iterative procedure using the well known the -rule in different versions [1]. In general, the time dependent problems after a spatial discretization can lead to a lumped-parameter model. For example, Maxwells equations in differential form for low frequency in 2-D case, after spatial discretization, lead to a system of ordinary differential equations by the form [2]:

Tij = N i N j d e J i = J s Ni d e

A [ S ] + [ R]{A} + {b} = 0 t
where [R] and [S] are matrices and the column vector {b} is the vector of the free terms. We can use the decomposition at the level of the field. To simplify the computation, one approach is to separate the spatial domain of the magnetic problem in conducting and non-conducting parts, such that A1 is the solution vector in conducting regions, and A2 is the solution vector in the non-conducting regions. By reordering the matrices, the system of equations is divided in two systems [1]: A [R11 ]{A1 } + [R12 ]{A2 } + [S] 1 + {b1 } = 0 (14) t [R21 ]{A1 } + [R22 ]{A2 } + {b2 } = 0 (15)
The second system (15) is formed of algebraic equations; the first system (14) is formed of differential equations. These systems are solved by an iterative procedure in time [1]: 1. Choose the starting value for {A1} at t=0; 2. Compute {A2} from the system (15) 3. Compute {A1/t} from the system (14) 4. Compute the values of {A1} at t=t+t; 5. Repeat the steps 2-4 for each time step until the final time is reached. This approach is expensive in terms of computation time. An efficient approach is based on Galerkins procedure for the time domain [2].

3.2 Iterative algorithms for coupled problem

A complete mathematical model for coupled fields involves Maxwells equations and the heat conduction equation. Combining these equations yields a coupled system of non-linear equations. In a discrete form the unknowns are the nodal values of the temperature T and the magnetic vector potential A. For electromagnetic field we consider the Aformulation, that is we define the magnetic vector potential A by B = curl A. More, the domain is the same for temperature and the electromagnetic field although in practice the interest is for different field domains. The non-linear equations for T and A are straightforwardly obtained by a Galerkin finite element method. For the 2D steady-state problems we do the approximations at the element level [1]: r T ( x, y ) = N j ( x, y )T j j =1 r A( x, y ) = N j ( x, y ) A j j =1 where the interpolation functions Nj are basis functions in the mesh over , and r is the number of nodes of an element. The usual procedure for the FEM applications leads to a system of 2p equations where p is the total number of the unknowns in each field problem. These nonlinear equations can be solved by two different basic strategies: Solving the equations for Ti and Ai simultaneously

FINITE ELEMENTS

published by WSEAS Press

60

Solving the equations for the two fields in sequence with an outer iteration, technique known as operator-splitting technique (for example Newton-Raphson procedure) In the area of the first strategy, Gauss-Seidel and Jacobi methods are well-known. We present these methods in brief [2]. For this, let us define the two discrete equations derived from the electromagnetic field model and the thermal field model in the form:

f A ( A1 ,..., Ap , T1 ,..., Tp ) = 0 fT ( A1 ,..., Ap , T1 ,..., Tp ) = 0

where the subscript denotes the original problem (A for the magnetic field in the magnetic vector potential formulation; T for the thermal field). The Gauss-Seidel algorithm for coupled fields has the following pseudo-code: For m:=1 , 2, until convergence DO Solve ( m) (m) (m 1) (m 1) f A ( A1 ,..., A p ; T1 ,..., T p )=0 (m) (m) with respect to A1 , Ap Solve ( m) ( m) ( m) ( m) fT ( A1 ,..., A p ; T1 ,..., T p ) = 0 with respect to T1(m) , Tp(m) In other words, the system is solved firstly with respect to A, using the values of T from the previous iteration. Afterwards, the equation derived from the thermal field model is solved using the computed values of A from the current iteration. The equations fA=0 or/and fT=0 are non-linear and must be solved by an iterative procedure (for example Newton-Raphson method). The algorithm Jacobi-type is similar to GaussSeidel method, except that at the iteration m when we must solve the model for T, the values for A are from the previous iteration, that is A(m-1). The algorithm has the following pseudo-code: For m:=1 , 2, until convergence DO Solve (m1) ( m) (m) (m1) f A ( A1 ,..., A p ; T1 ,..., T p )=0 (m) (m) with respect to A1 , Ap Solve ( m1) (m 1) ( m) ( m) fT ( A1 ,..., A p ; T1 ,..., T p ) = 0 with respect to T1(m) , Tp(m)

researchers. It is known software of company Tera Analysis developed for modelling a variety of field problems using the finite element method. It is beyond the goal of this work to present in detail this software so that we limit our discussion to a brief description. More details about the features of the product can be found in user manual [8] and on Internet at the site http://tera-analysis.com/. The last release Quick 5.4 brings new options and has a good compatibility with CAD systems and other professional programs (as Mathlab, Mathcad etc). The most advantage of the product is the minimum effort required for initial training. It uses an important tool named LabelMover that is based on ActiveField Technology. It generates series of QuickField problems with specified changes of the model geometry and the physical properties, solves these problems and presents results as a table. All types of analysis provided by QuickField are supported. Some other useful features are presented in the manual and consist in the following: LabelMover supports the QuickField coupled problems Results of the numerical simulation can be viewed as an X-Y plot, coloured map, vectors or local values All the information about the LabelMover session can be saved in a file to use it later. One of the exciting features of the LabelMover is its ability to perform the tolerance analysis. After the base QuickField problem is formulated and solved, it is possible to analyse the influence of random changes to geometrical sizes and physical properties of some of its parts. LabelMover automatically generates series of QuickField calculations, and performs statistical analysis of the results. This tool may also be applied for estimation of the maximal possible deviation of input parameters providing the needed level of the system performance.

3.3 Induction heating

As first target example we consider a long cylindrical workpiece excited by a close-coupled axial coil [5]. The problem is an axisymmetric heating device (see the Fig. 1) with: 1- the workpiece, 2 the air, and 3 the coil. The coil is assimilated with a massive conductor. In this case we can not ignore the eddy currents in the coil. We consider a low-frequency current in the coil so that the penetration depth is large. In this case we can decompose the whole domain of the field problem into overlapped subdomains for the two coupledfields.

3.2 The software product Quickfield

In our numerical simulation we used the software product Quickfield [8]. This software is a powerful instrument for professional engineers, designers and

FINITE ELEMENTS

published by WSEAS Press

61

Fig. 1 Axial section of the coil

The radiation plays an important role in induction heating at high temperature. Convection losses are small in through-heating, as the workpiece is contained in a shell that does not permit air movement. In the case the workpiece is in the open air, the convection losses are very important. The finite element method was used for the numerical results. In the Fig. 3 the temperature versus time in a point on the internal surface of the coil is plotted using the program Quickfield [8]. The initial temperature was 20 0C (293.15 K). The workpiece is a steel cylinder and the coil material is copper. The current intensity is 60000 [A] and the time duration is 600 [s]. We considered both forced convection and radiation conditions. The convection coefficient is 50 and radiation coefficient is 0.8.

3.4 Dynamic interfaces at high frequency

For many eddy-current problems the magnetic flux penetration into a conductor without internal sources of the magnetic field is confined mainly to surface layer. This is the skin effect. The skin depth depends on the material properties , and so that for the small depths all of the effects of the magnetic field is confined to a surface layer. The skin effect can be exploited in two directions: To reduce the space domain in analysis with a fine mesh close to conductor surfaces To reduce the material volume since a significant proportion of the conductor is virtually unused The penetration depth is given by the formula:

Fig. 2 The analysis domain

Fig. 3 Temperature vs. time in coil The domain for the magnetic field is shown in the Fig. 2, that is a quarter of the device bounded by a boundary at a finite distance from the device. For the thermal field we consider the workpiece as the analysis domain. The penetration depth of the magnetic field in the workpiece imposes the overlapping domains for the two fields. The numerical model is considered in a cylindrical co-ordinates with the vertical axis Or and the horizontal axis Oz.

The value of changes during heating and becomes small at high frequency. In the last case the spatial domain for the magnetic field can be reduced. More, a part of the domain boundary varies in time so that at each time step a new mesh must be generated. A solution to avoid the regeneration of the mesh is to estimate the maximum value of from the bounds of the working frequency and to use a static boundary. In any case a substantial reduction of the spatial domain of the magnetic problem is reduced in comparison with the low working frequencies. The thermal field can be limited to the workpiece where the distributed heat sources are in a small layer.

4 Reactors
The second example is from the electrical networks. The reactors or reactance coils are large coils used in electric distribution networks. These components have

FINITE ELEMENTS

published by WSEAS Press

62

only inductive impedance in a quasi steady-state electromagnetic field. Such reactors consist in general of concentric coils with different number of turns and rectangular or cylindrical cross sections. In Fig. 4 an axial section is shown. At high values of the current in the reactor coil, the temperature in different parts can reach dangerous values so that an accurate computation of the temperature is necessary. The reactors are built in different variants. For our discussion we consider fixed reactors where there is no possibility to modify the turns number and the core length. The conductor of the coils is copper or aluminium. For domestic frequency the coil conductors are massive conductors. For high frequencies the coil conductors are stranded conductors with a circular cross section and a uniform distribution of the current density. In the case of the reactors without core, these are total or partial shielded. The shields can be: electromagnetic shields (with electrical conductivity equal to infinity in ideal case) magnetic shields (with magnetic permeability equal to infinity) In all shielded reactors, the magnetic field generated by the winding current penetrates the shield in different zones. The eddy voltage generated in accordance with the law of induction causes eddy currents in the electrically conducting shield which produce corresponding heat losses. These losses have to be kept at safe low values, both in normal operation and temporarily during overload and over currents caused by electrical system short circuits. Eddy currents in electromagnetic shield have a negligible effect on the field in the winding. The main magnetic field source is the prescribed alternating electric current in the reactor coils.

4.1 Mathematical modelling

As target example we consider a reactor without core with and without a super-conductor shield on the vertical wall of the tank (parallel to the axis Oz). The axial section of this device without shield is shown in Fig.4. The symmetry in geometrical and physical properties with the standard assumptions considered above lead to a 2D-scalar problem. This example can be treated in a similar manner with the bus bars with the small differences. This is a case of axisymmetric problem so that the differential models for electromagnetic and thermal fields must be expressed in the cylindrical co-ordinates. On the boundary ABCD we can have different boundary conditions in terms of the shield nature: magnetic shield (characterised by )

electromagnetic shield (characterised by ) electromagnetic and magnetic shields At the surface of magnetic shield, the tangential component of the magnetic flux density is zero; at the surface of the electromagnetic shield the normal component of the flux density is zero due to reaction of the eddy currents induced in the shield. If we assume that the steel tank is infinitely permeable we may replace the tank with a Neumann boundary condition in a numerical model. We also assume that the flux is contained the core. In a coupled model we must solve two problems: a magnetic field problem, and a thermal field problem. We consider the geometrical model plotted in the Fig. 4 that illustrates an axisymmetric problem of magneto-thermal coupling. In this application the window of the reactor obtained by an axial section of the device defines the field domain field domain . The analysis domain can be decomposed into two nonoverlapping subdomains with different mathematical models. The coil and the air define one domain; the other is the shield. In this way the original problem is divided into two sub-problems. The first problem involves the computation of the main leakage flux density in the window where the current density of the coil is the single source of the magnetic leakage flux. In the second problem we compute the eddy currents in the shield using the vector electric potential P. A differential model for the eddy-current computation in the shield can be defined using the current vector potential P. The P-vector can be regarded as a generalisation of the H-vector but not necessarily with the same divergence [2]. The relations that introduce it are: P = J e P =0 The eddy current problem is described by the twocoupled equations [2]: A A = J s in 2 D (16) t H ext P 1 ( P) + = in 3D (17) t t ( P ) = 0 (18) with specified boundary conditions for A and P. The equation (17) can be reduced to a 2D problem because of the symmetry of the field problem. More, if the eddy currents in the coil are neglected (an assumption that is valid with the actual technology), the term with A/t disappears and the mathematical model is a classic Poisson equation [2].

FINITE ELEMENTS

published by WSEAS Press

63

In the shield we are interested in the eddy currents computation so that a model for 3D problems must be used. The electromagnetic field problem of such structure is a three-dimensional one but some usual assumptions transform the problem in a 2D one. The standard assumptions are: the coil current density has only an azimuthal component in cylindrical co-ordinates the source current density has a sinusoidal variation in time In a vector potential formulation with A magnetic vector potential, A has only an azimuthal component and the problem is axisymmetric so that the study is made in the plane r-z. In the case of the magnetic shield the boundary conditions are Neumann conditions for the magnetic vector potential A on the lines ABCD (see Fig. 4). In the case of the electromagnetic shields the boundary conditions are Dirichlet conditions. In the case of super-conductor shields we are interested to compute the local heating both in the coil and in the shield. The following assumptions are considered: The primary source of the magnetic field is the prescribed current in the reactor coil and has a sinusoidal time variation The reaction of induced eddy currents is considered both in the coil and the shield In a numerical simulation of the reactor we can use a simplified approach based on the real considerations. The first stage in a numerical solution is to compute the eddy-current losses in the shield as primary heat source in a coupled model. We can divide the electromagnetic field problem into two sub-problems: The computation of the main leakage flux density in the reactor window. The induced eddy currents are not considered; The computation of the shield eddy currents induced by the nominal component of the magnetic flux density.
A. Magnetic field computation For the magnetic field the differential model defined by Eqn. (5) is used. A finite element approach leads to a numerical model. Boundary conditions. In our example we have two types of boundary conditions: at the surface of the magnetic shield (lines AB and CD in the Fig. 4), the tangential component of the magnetic field is zero; at the surface of the electromagnetic shield (the lines BC and AD), the normal component of the magnetic field is zero. Thus, if l is the height of the window of the axial section in reactor, and re is the outer radius of the window, the boundary conditions are:

2 In the A-formulation, the components of the flux density are defined as: A 1 (rA) Br = ; Bz = ; r r z On the basis of the above assumptions the analysis domain can be reduced at the upper-right quarter of the axial section (see Fig.5) defined by the symmetry axis and the centreline. The differential model for the magnetic field is the diffusion equation in cylindrical co-ordinates (r,z) with specified boundary conditions. In this case the following boundary conditions are considered: The Dirichlet boundary condition on the symmetry axis and the outer surface The homogeneous Neumann condition on the centreline The computation of the induced eddy currents can be done in some simplifying assumptions. If the shield has a small thickness, the magnetic field distribution throughout the shield can be considered with a uniform distribution and the normal component of the flux density linearly decreases with the penetration depth. We can use a two-dimensional model defined by the following equations [2]:
J e = Br t

Br = 0

for

r = re

and

z=

Je = 0 where Je is the eddy-current density, Br is the normal component of the magnetic flux density and is the shield conductivity. The finite element method can be used with boundary conditions for the vector potential P. The component Br is defined as: A Br = z The eddy current losses at the level of element are computed by the formula: P = J 2 d e e e B. The computation of the thermal field The losses in a reactor are: The main losses in coil due to Joule Lenzs effect of the electric current The main hysteretic losses in the core Auxiliary losses in the coil due to the radial and axial components of the magnetic field Losses in the shields

FINITE ELEMENTS

published by WSEAS Press

64

The losses in the coil and the core develop local heating of the tank and the coil and represent the heat sources for these components. The mathematical model for the thermal field is the heat conduction equation in cylindrical co-ordinates. The current density J includes both the driving current and skin effect.

4.2 Numerical modelling

We can use the symmetry of the problem to reduce the analysis domain. The domain plotted in the Fig. 5 is the upper-right quarter. It is defined by the axial axis (Oz), the radial centreline (Or) and the outer surface of the tank. We can enforce the reduction of the analysis domain at the window of the coil including the shield. In this case at the horizontal axis of symmetry we put the boundary condition Ht=0. At the vertical axis of symmetry Bn=0. In accordance with the definition of the magnetic vector potential A, this boundary condition means we have A=constant. Field fades at the infinity so due to continuity we have A=0 also. The numerical model was obtained by the finite element method, both for magnetic field and thermal field. The Quickfield program of the company Tera Analysis was used for the numerical simulation of the numerical models that were solved in cascade. The time constants are different for magnetic and thermal fields. This fact can be exploited in a numerical model in the sense that the electrical properties of the materials are not updated at each time step. We can define some time moments for updating the magnetic field.

Fig. 4 Axial section of the reactor

Fig. 5 The domain for FEM program

4.3 Numerical results

The flux density on the shield surface is plotted in Fig. 6. The values of the flux density on the surface represent the magnetic field source for the shield field. The temperature versus time on the shield surface is plotted in the Fig. 7. The point has co-ordinates r=550 [mm], z=0[mm], and the time interval is equal to 3600 [s]. The heat source is not updated for every time step because the time constant for the thermal field is very large. In the Fig. 8 the final temperature along the centreline is plotted. The large temperature-drop in the conductor modifies the electric conductivity, and consequently the electrical current (the principal heat source for the thermal field). By numerical simulation we can find the optimum value of the current amplitude, or the best

Fig. 6 Variations of B (green), Br (blue) and Bz (red)

FINITE ELEMENTS

published by WSEAS Press

65

material so that an imposed temperature in different parts of the device can be maintained.

In practice the coupled fields are treated independently in some simplified assumptions. The accuracy of the numerical computation is poor. With the new architectures, a multidisciplinary research is possible. Some iterative procedures were presented with emphasis on the coupled magnetic and thermal fields. Domain decomposition offers an efficient approach for large-scale problems or complex geometrical configurations. This method in the context of the finite element programs leads to a substantial reduction of the computing resources as the time of the processor. In coupled problems a hierarchy of decomposition can be defined with a substantial reduction of the computation complexity.

Fig. 7 Shield temperature vs. time at the middle of the height Table 1 Tank 106 100 630 59 7600

Coil [S/m] 6.107 r 1 c [J/Kg.K] 380 k [W/K.m] 389 [Kg/m3] 8900

Shield 5.7 107 1 380 389 8900

Air 0 1 0 0.01 0

Fig. 8 Final temperature on the centerline

5 Conclusions
The problem of coupled fields in electrical engineering is a complex problem in terms of computing resources.

References: [1]. Crstea, D., Crstea, I. CAD of the electromagnetic devices. The finite element method. Editor: Sitech, 2000. Romania. [2]. Crstea, D. CAD tools for magneto-thermal and electric-thermal coupled fields. Research Report in a CNR-NATO Grant. University of Trento. Italy, 2004. [3]. Alonso, A., Valli, A. A domain decomposition approach for heterogeneous time harmonic Maxwell equations. In: Computer methods in applied mechanics and engineering, 143 (1997), pg. 97-112 [4]. Zenisek, A. Nonlinear elliptic and evolution problems and their finite element approximations. Academic Press Limited, 1990. [5]. Marchand, Ch., Foggia, A. 2D finite element program for magnetic induction heating. In: IEEE Transactions on Magnetics, vol. 19, no.6, November 1983. [6]. Rodriguez, A. A., Fernandes, P., Weak and strong formulations for the time-harmonic eddy-current problem in general multiconnected domains. In: Euro. Jnl. of Applied Mathematics, vol. 14. 2003. Cambridge University Press. [7]. Quarteroni, A., Valli, A. Domain Decomposition Methods for Partial Differential Equations. Oxford Science Publication. 1999 [8]. *** QuickField program, version 5.4. Page web: www.tera-analysis.com. Company: Tera analysis.