This paper is a further study of two papers [1] and [2], which were related to Ill-Conditioned Load Flow Problems and were published by IEEE Trans. PAS. The authors of this paper have some different opinions, for example, the 11-bus system is not an ill-conditioned system. In addition, a new approach to solve Load Flow Problems, E- ψtc, is introduced. It is an explicit method; solving linear equations is not needed. It can handle very tough and very large systems. The advantage of this method has been fully proved by two examples. The authors give this new method a detailed description of how to use it to solve Load Flow Problems and successfully apply it to the 43-bus and the 11-bus systems. The authors also propose a strategy to test the reliability, and by solving gradient equations, this new method can answer if the solution exists or not.
In the early 1980’s, there were two papers which studied ill-conditioned Load Flow Problems ( [
Since 1960’s, people have taken the direct method as a major tool to solve load flow equations; among them, Newton-Gaussian elimination (LU) is a typical representation [
Recently people have turned their attention to iterative method, e.g., Inexact-Newton-GMRES method, ( [
The extent of ill-conditioned problem is determined by the conditioned number of the system. From [
For the conclusion (3) of [
Brown method was used by [
Considering that the two papers were published in 1980’s, we should not make excessive demands for authors of [
Before the description of the new method, let us first review the derivation process of the two classical methods:
1) Newton method:
From
We have
Or, for
Take
That is to say, we can use both Static State and Dynamic State approaches to derive Newton method. Despite the Newton method can be recognized as an explicit ODE Euler method, the right side function includes an inverse matrix of
If the Newton method fails to converge, two approaches can be taken to improve the convergence:
a) Using damp factor: Replacing
b) Altering initial values: If the initial value
2) About the fixed iteration
For
will have the same solution of
method. In fact, for differential equation
Take h = 1, we can get classical fixed iteration:
It is well known that the convergence condition of fixed iteration is:
If h = 1, the convergence condition of classical Fixed Point Iteration Method does not satisfy. We can reduce h, if only all of the eigenvalues of the Jacobian matrix of
enough, the inequality
We know that explicit ODE method only has very small stability region. If the Jacobian matrix of
3) Does the explicit ODE method with the large stability region exist? The answer is positive. In [
Our purpose is to reach the steady point of the dynamic system
our method is within the attracting domain of the steady point. So our method can take the very large step size, and goes to steady point quickly. The traditional explicit ODE methods cannot take large step size. Usually, their trajectory is very close to the solution of the system. In a sense the trajectory approximately reflects the real transient process of the system, however the trajectory produced by our method does not. So, our method can be recognized as “Pseudo-Transient Continuation.”
The existing “Pseudo-Transient Continuation” method can be written as follows:
Various algorithms adopt different strategies to control h. They are known as the “LINEARLY IMPLICIT EULER DISCRETIZATION”, [
Because each iteration step needs to solve a system of linear equations, so all of them are implicit methods.
Our method is the unique Explicit “Pseudo-Transient Continuation” method. In order to express this difference, our method is named by [
From [
For
1)
2)
3)
Here
For
If the steady point
1) Our method is an explicit stiff ODE method, so it has a very strong capability to deal with tough problems. Brown’s almost linear equation is as follows: [
Obviously,
For n = 10, 30, Newton method diverges drastically, but Brown method converges well.
If n > 35, according to [
What happens to our method? For n = 40, 100, 200, 400, we can get the solution easily. In this paper, we just show the result of n = 400:
Virtually, the differential equation to deal with is:
Here the diagonal matrix
As we did in [
Tol3 = 1D−10; h1 = 0.015, h2 = 0.15, h3 = 0.8. After 826 iterations, we got the solution:
Incidentally, the efficiency of Brown method is strongly related to the order of the equations. It is more efficient if the equations are ordered by increasing the degree of nonlinearity. In our method, each component can be calculated independently, so there is no ordering problem.
An experiment has been done: We place the nonlinear equation on the top of the
2) Our method is as simple as fixed iteration, so it can handle very large systems. Here we give an ill-condi- tioned, non-symmetric linear equations with dimension n = 1,000,000:
Here
Tolerance | Step size | NFE | 2-Norm of |
---|---|---|---|
1D−0 | 0.015 | 9 | 0.5258D−00 |
1D−5 | 0.15 | 55 | 0.7321D−05 |
1D−10 | 0.8 | 826 | 0.8798D−10 |
We ask the solution is
If E-Ψtc method is used, taking ε = 0.5. For any step size the numerical integration is stable.
The only possible comparing method is fixed iteration (Euler) method. The maximum step size is h = 0.8. In order to get a good stability, we take h = 0.75. The initial Euclid norm is 0.7217D+03 After 150,000 fixed iterations, we have the following results:
We want to make the Euclid norm less than 1D−08, it seems impossible.
How about the Explicit Pseudo-Transient Continue” method. (E-Ψtc)?
The following data can give the answer: (Variable step size)
3) Another outstanding feature is that our method is very suitable for parallel calculation.
Now we turn our focus to the main topic of this paper―Load Flow Calculation.
1) The Construction of the Differential Equations.
The load flow equations can be formulated by a system of nonlinear.
Equations:
Here
In rectangular coordinate system, for PQ buses, the power equations are:
For PV buses, the voltage magnitude values Vis are specified and have the relation:
(3.1) and (3.3) constitute the power equations for PV buses.
For elimination method, the optimal ordering of the buses is the first task we need to solve. In our method, this problem does not need to be considered, because the memory of space only depends on the number of buses and not related to the ordering. For the calculating process, as we have pointed out in Section 3, in our method, each component can be calculated independently.
But for every bus, which corresponds to two real number equations. If the nonlinear equations
The corresponding differential equation should be arranged as follows:
Because in the ith bus, the active power Pi is mainly concerned with fi and reactive power Qi is mainly concerned with ei ( [
In fact, for elimination method, every bus corresponds to a 2-dimension sub-block and which is also arranged just the same.
In order to let the readers understand the discussion above more clearly, let us give a further statement by a simple example:
Consider a 2-dimensional linear equations:
The matrix-vector form is:
Here
If we use a fixed point iteration method to solve this linear equations
This is because the two eigenvalues of the matrix A
If we alter the sequence of x and y, consider equations
Here
The two eigenvalues of the matrix B,
From the differential equations’ point of view:
For differential equations:
Choosing a proper step size h, Euler method will make
However, for the differential equations:
No matter what h is,
The matrix B is diagonal dominant, but matrix A is not. This is why for every bus, we arrange the sequence according to
Here we would like to emphasize this matter again. The algebraic equations and differential equations were all the same, just the unknown variables have different order―one is
Now, let us see what is in the reference [
[
(Ex.)
(Ex.)
This expression is not correct. It will make the Jacobian matrix loose diagonal dominant. That is to say, for each row of the Jacobian matrix, the element with the largest absolute value will not be diagonal element. Obviously, for PV-bus, the one of diagonal elements will be “−2f”. If flat start is used, then for the first step, the diagonal element will be zero. It will be the worst thing for any algorithm to solve the linear algebraic equations. For elimination method, almost all code, column pivot technique was taken. It automatically adjusts the order of equations so people usually ignore the “order problem”. But for any iterative method, the “order problem” is crucial.
The correct expression should be:
(Ex.)
(Ex.)
Or the matrix form:
For the details, please see [
We can also change the order of equations;
(Ex.)
(Ex.)
2) Preconditioning
We will give two preconditioning strategies:
a) point diagonal preconditioning: The preconditioner is a diagonal matrix of the Jacobian. If you do not want to calculate the diagonal elements, according to the observation of the calculated data, we can further simplify this strategy.
For PQ bus:
For PV bus:
The precondition matrix can be simplified as a constant diagonal matrix. The elements are just Bii, and −2.0. It was named as D1. The differential equation that we deal with is:
b) 2-dimension block diagonal preconditioning: We know every bus corresponds to a 2-dimension diagonal sub-block of the Jacobian matrix. These sub-blocks constitute a block diagonal matrix D2:
The differential equation that we deal with is:
The preconditioning not only raises efficiency, but also helps us to determine the parameter ε = 0.5. Because the Jacobian matrix diagonal elements of
trix is in the diagonal place of each row and the sparse property of the matrix) and
spectral radius. (See the reference [
For ordinary load flow problem, if point diagonal preconditioning is used, h is in the range of 5 to 7 is suggested. If an iteration cannot converge, we suggest to reduce h. If h is small enough, say h < 1, the iteration still cannot converge, the iteration should be stopped, then check the whole process to see what mistake was taken.
For this system, 8 calculating results are given, we take the Largest Absolute Values of the Mismatch (LAVM), less than 1D−05 as the stopping condition (double precision):
1) Normal Newton method: 7 Newton iteration, The LAVM = 0.2123D−08.
2) Optimal Multiplier Newton method: 6 Newton iteration. The LAVM = 0.4134D−08. (The values of optimal multiplier µ are taken from [
3) Euler method: step size h = 0.00065 (maximum step-length to meet the stability requirements), 157006 function evaluations. The LAVM = 1D−05.
4) No preconditioning E-Ψct: ε = 0.0005, h = 0.00085, 120076 function evaluations. The LAVM = 1D−05.
5) Point diagonal preconditioning E-Ψct: ε = 0.5, h = 0.35, 42505 function evaluations. The LAVM = 0.9998D−05.
6) 2-dimension block diagonal preconditioning E-Ψct: 2008 function evaluations. The LAVM = 0.9995D−05.
7) Variable step size 2-dimension block diagonal preconditioning E-Ψct: ε = 0.5, h1 = 6.5, h2 = 7.0, h3 = 7.5, h4 = 8.0, h5 = 9.0, h6 = 10.0.
1335 function evaluations. The LAVM = 0.9968D−05.
What conclusion that we can get based on the 1)-7) results?
From 3) we can see that this system is surely an ill-conditioned system, but from 1), the ill-conditioned extent is not severe, using double precision, only 7 Newton iterations are needed to make the LAVM = 0.2123D−08. Compare 1) and 2), Optimal Multiplier Newton method saved one iteration, if the stopping condition is 1D−08. This is the contribution of [
Compare 3) and 7): the proportion of numbers of function evaluation is 157006/1335 ≈ 118. For 7), besides function evaluations, forty-two 2-dimension sub-matrix need to be inverted in each iteration. The workload of 2-dimension matrix inverse is relatively very small, so we can say that for ill-conditioned Load Flow Problems, using 7) is a good choice, especially when the system is large.
For 7), if someone considers that the workload of 1335 function evaluations is too big to be accepted, here we give another choice. This choice is based on the following observation: For Newton method, the quadratic convergence superiority can be utilized only when the norm of
8) At the beginning, we use 6) (ε = 0.5, h = 6.0), after 132 function evaluations, the LAVM = 0.9821D−01, then we switch to normal Newton iteration, after 4 iterations, the LAVM = 0.3366D−12. By directly using the normal Newton method, if LAVM < 1D−12 is required, 8 iterations are needed. That is to say, at the cost of 132 function evaluations, we can save 4 normal Newton iterations.
Here, what we want to say is as following:
For any existing Newton program, you can use our method as a tool to choose a good starting point. The textbook tells us: “The HOMOTOPY method can be used to generate a good starting value”. So our method can be considered as a homotopy method. It can be used independently and can also be embedded into the Newton program as a tool to choose starting value.
There are a lot of references related to “HOMOTOPY”, we merely list three of them [
For ordinary load flow problem, e.g. IEEE-39 IEEE-118, our method is also suitable. You can use the method by two ways: One is directly solve nonlinear equations, like what we did in this paper. [
For nonlinear equations, the solution is not unique, as a research paper, we should give the final numerical results or some description for large systems, but neither did [
At the beginning, the initial LAVM = 0.1232E+01, we give the following results:
1) As [
0.3046E+00, 0.1330E+00, 0.8979E−01, 0.1883E−01, 0.2195E−02, 0.7789E−03, 0.2557E−03, 0.8268E−01, 0.2074E−01, 0.5251E−02, 0.1383E−02, 0.4276E−03, 0.2818E−03, 0.1217E−02, 0.3891E−03, 0.3055E−03, 0.5773E−03, 0.2670E−03.
2) Optimal multiplier method (single precision), 18 iterations. Results are:
0.2858E+00, 0.1326E+00, 0.3429E−01, 0.1691E−01, 0.5123E−02, 0.4473E−03, 0.2464E−03, 0.2164E−03, 0.2155E−03, 0.2155E−03, 0.2155E−03, 0.2154E−03, 0.2155E−03, 0.2155E−03, 0.2154E−03, 0.2155E−03, 0.2155E−03, 0.2154E−03.
If we use double precision, we have almost the same results as above 1) and 2). Based on this observation, we can see:
a) Oscillation occurred after 7 normal Newton method iterations, but for the optimal multiplier method, oscillation did not appear.
b) The LAVM cannot go down to zero and this is not caused by the defect of numerical precision of the computer and/or programming code.
c) After 7 iterations, the LAVM of the optimal multiplier method is 0.2464E−03. After 7 iterations the LAVM of the normal Newton method is 0.2557E−03.
3) How about the E-Ψtc method? We use point diagonal preconditioning strategy. Taking ε = 0.5, h = 1.5, after 178 function evaluations, the LAVM is less than 10−4 (0.995493 × 10−4).
Unless it has a very small oscillation, the LAVM basically keeps monotone descending. (Both single and double precision have the same results). The LAVM is about 1/2 of the optimal multiplier method.
4) When using Inexact Newton-GMRES (15), we use single and double precision to calculate, the results are almost the same. Here we give the double precision result: (For every GMRES loop, we calculate the initial residue R0 and the final residue R1 of the linear equations, and take Rl/R0 < 0.1, or Maximum Number of Restart = 100 as the stop condition for each Newton iteration.)
That is to say, after 7 Newton iterations, there are totally (1 + 3 + 4 + 100 + 72 + 100 + 100 = 380) 380 × 15 = 5700 GMRES iterations, finally the LAVM reaches 0.2796D−02, then it no longer descends.
To summarize the results above:
1) Using the normal Newton method, the oscillation occurs;
2) After 9 optimal multiplier iterations, the LAVM keeps 0.2155 × 10−3;
3) With E-Ψtc method, after 178 function evaluations, the LAVM reaches 0.995493 × 10−4;
4) By using Inexact-Newton-GMRES (15), after 5700 GMRES iterations, LAVM = 0.2796 × 10−2.
All results we got show that for the 11-bus system, the LAVM cannot go down to zero, in another word, for this system, the solution of
The first choice should be the Gauss-Newton method:
if the Jacobi is a
We can also solve the gradient equations by using Newton method. In order to compare the two results, re-solve it again by LINPACK SUBROUTINE DGECO. By this way, we can get the condition number of the matrix sequence
Comparing the two results, we can see that before oscillation occurs, except 6th-iteration has tiny difference, both results are exactly same. It shows, even if the condition number of the
What about the Point Diagonal Preconditioning E-Ψct for the gradient equation?
The differential equation should be:
Here DD is the diagonal matrix of
One may think 8858 gradient function evaluations are too expensive, but please note the step size is 20.0 - 600.0. If we use fixed iteration (Euler method), the step size is not greater than 0.76. We try this method, the
Iteration | Rcond | LAVM |
---|---|---|
1 | 0.5843D−06 | 0.3046D+00. |
2 | 0.6465D−06 | 0.1330D+00 |
3 | 0.3566D−06 | 0.8979D−01 |
4 | 0.1804D−06 | 0.1883D−01 |
5 | 0.4104D−07 | 0.2195D−02 |
6 | 0.9671D−08 | 0.7329D−03 |
7 | 0.1611D−08 | 0.2557D−03 |
8 | 0.1187D−11 | 0.8330D−01 |
23 | 0.9184D−12 | 0.1070D+00 |
28 | 0.1452D−08 | 0.2654D−03 |
49 | 0.3995D−11 | 0.2469D−01 |
50 | 0.1886D−06 | 0.6238D−02 |
Iteration | Step size | Norm of gradient | Norm of | LAVM |
---|---|---|---|---|
0 | 0.2215D+03 | 0.2621D+01 | 0.1232D+01 | |
930 | 20.0 | 0.9991D−02 | 0.8234D−02 | 0.3678D−02 |
1734 | 50.0 | 0.9565D−04 | 0.2209D−02 | 0.1125D−02 |
3955 | 200.0 | 0.8513D−05 | 0.2647D−03 | 0.1712D−03 |
8858 | 600.0 | 0.9891D−07 | 0.2089D−03 | 0.1027D−03 |
step size is from 0.73 − 0.76, it needs 12,902,751 evaluations. The proportion of the number of the evaluating gradient functions is 12902751/8858 ≈ 1456. In the later period of iteration, these two iterations keep monotone descending. When
Usually we evaluate the accuracy of a solution by the mismatch of the equation, but for ill-conditioned problem, it might be misleading. Here we give an extreme example.
Consider a 2-dimension linear algebraic equations:
If substitute x1 by 0.9911, and x2 by −0.4870, the mismatches will be −10−8 and 10−8. People might think the most accurate solution has been obtained. Unfortunately, s/he is wrong. The exact solution of these equations is 2 and −2. From a mathematical theory, discriminating the accuracy of a linear equation, we need to resort the knowledge of the condition number [
We suggest taking the following easy ways to identify if the system is ill-conditioned or not:
Using both single and double precision to do the calculation, if the results basically keep coincidence, we can consider the system is well-conditioned.
If using a Normal Newton method, for 43-bus system, single precision cannot make the LAVM less than 1E−5, but for double precision, the LAVM can go down less than 1D−12 easily, so 43-bus system is surely an ill-conditioned system.
For 11-bus system, we do the same process above, the results of single and double precision are almost the same, the LAVM cannot go down to 1E−4 and 1D−4. After 7 Newton iterations, the LAVM are 0.255709E−03 and 0.255653D−03, then oscillation occurs for both iterations. So we confirmed that the 11-bus system is not an ill-conditioned system, but the LAVM cannot go down to zero.
In order to comprehend this phenomenon, consider the equation:
The solution does not exist, but at x = 0.0,
If
In engineering design, making
Strictly speaking, for the engineering design problem, the solution of
BUT, finding the minimum of
To summarize our opinion, the 11-bus system is not ill-conditioned, the norm of
Now let us discuss the reliability of the algorithm. In many literature, there has been too much emphasis on measuring the efficiency of the algorithm proposed by the author(s) but not enough on reliability. For the nonlinear equations, the non-uniqueness of the solution makes measuring the reliability being an important subject. Reference [
For the power system, the authors of this paper propose the following strategy to test the reliability:
1) Using single and double precision to solve the equation
2) If single and double precision are used, the iterations do not converge, the doubts about the algorithm and the existence of the solution might be raised. We suggest that people change the algorithm first to see what happens. If the iteration still does not converge, the existence problem should be considered. We observe all the iteration processes, if all of the iterations diverges, we can affirm that the solution does not exist. If the solution around some value oscillates and this value is relatively small, like the 11-bus system, we can recognize the solution exists, in engineering it is called “operational”.
We do not suggest readers to change the initial values to improve the convergence, because for the power system, flat start is good enough, these initial values are very close to the solution.
Even though oscillation can also be regarded as divergence, in this paper, we discriminate those two different phenomena. Here, so-called divergence means drastic oscillation happens in the iteration process.
3) As we have mentioned before, the solution of a nonlinear equation is not unique. The small LAVM does not guarantee the solution we searched is the solution what we want to get. The reliability asks us to list the results produced by different methods and to compare them, but none of [
The value of the angle of the bus-2 is probably wrong.
From the three results above in
In this paper, we give a review about references [
We find out that the benefits of the optimal multiplier method can only occur in the initial period/stage of Newton iteration. For the 43-bus system, if the stopping condition is LAVM < 10−8, it can save one Newton iteration. For the 11-bus system, it can eliminate the oscillation which occurs in the normal Newton method. During the later period, we should take care of the loss of quadratic convergence. The applicability of the optimal multiplier method is only for the Newton iterative method. Other methods, such as steepest descent method, are not available.
43-bus system is an ill-conditioned, using double precision can overcome “convergence trouble”.
11-bus system is not an ill-conditioned. It does not have convergence trouble. The problem is that the LAVM of
page 3657 | Optimal multiplier | E-Ψtc | ||||||
---|---|---|---|---|---|---|---|---|
Bus-1 | 1.024 | 0.000 | 1 | 1.024 | 0.000 | 1 | 1.024 | 0.000 |
2 | 1.0552 | 1.215 | 2 | 1.0546 | −2.469 | 2 | 1.0568 | −2.413 |
3 | 1.0444 | 4.608 | 3 | 1.0436 | −4.170 | 3 | 1.0461 | −4.106 |
4 | 1.0306 | 2.830 | 4 | 1.0275 | −2.877 | 4 | 1.0312 | −2.841 |
5 | 1.0329 | 4.830 | 5 | 1.0320 | −4.906 | 5 | 1.0349 | −4.855 |
6 | 1.0501 | 2.916 | 6 | 1.0473 | −2.961 | 6 | 1.0511 | −2.925 |
7 | 0.8310 | 12.066 | 7 | 0.7743 | −12.569 | 7 | 0.8082 | −12.566 |
8 | 0.9131 | 14.857 | 8 | 0.8588 | −15.626 | 8 | 0.9072 | −15.475 |
9 | 1.1264 | 15.911 | 9 | 1.1266 | −16.675 | 9 | 1.1912 | −16.419 |
Bus-10 | 0.8266 | 21.033 | 10 | 0.7487 | −22.753 | 10 | 0.8129 | −21.998 |
Bus-11 | 1.0186 | 23.996 | 11 | 0.9778 | −25.992 | 11 | 1.0666 | −24.750 |
Note: For the data on page 3657, in the front of angle should have a negative sign.
strictly mathematical theory, we should solve the gradient equation. We surely did it. The result tells us, for the 11-bus system, LAVM = 0.0001027, the minimum of
[
For the 11-bus system, applying the Newton method cannot make LAVM down to zero. This is determined by the parameters of the system, and is not caused by the defect of Newton method. “Divergent” and “cannot make LAVM down to zero” are two different concepts. [
In this paper, we introduce a new method, E-Ψtc. By two examples, one of them is a tough nonlinear equation; another one is super lager scale ill-conditioned linear equations. These results fully demonstrate the capability of the new method. For Load Flow Problems, the preconditioning technique is needed. If the system is ill-condi- tioned, using 2-dimension block diagonal preconditioning. For well-conditioned problems, use point diagonal preconditioning. As for the reliability, we compare our results with the results of existing methods and find that they are basically coincident (we did not list the results of the 43-bus).
Tianmin Han,Yuhuan Han, (2016) Solving Load Flow Problems of Power System by Explicit Pseudo-Transient Continuation (E-ψtc) Method. Energy and Power Engineering,08,99-115. doi: 10.4236/epe.2016.82009
In conclusion (3) of [
Applying the optimal multiplier, if the values of µ* approaches zero with the iteration count increase the solution does not exist from the initial estimate...”. The authors of this paper can give an example to show this conclusion is incorrect.
For the 43-bus system, if the voltage initial values of all buses are (0.7, j*0). The normal Newton method (double precision) continues to oscillate. The optimal multipliers of the optimal multiplier method are µ = 0.0411, 0.0218, 0.0217, 0.0000, 0.0009, 0.0003, 0.0048, 0.0005, 0.0000, 0.0000, … According to the conclusion (3), the solution does not exist. But in fact, the solution does exist and we can use Damp Newton method to get it. If the damp factor is small enough, i.e., α = 0.007, then after 1957 iterations, the solution can be obtained. The LAVM is less than 10−5 (incidentally α = 0.008, after 2000 iterations, LAVM = 0.1516D+03). That is to say, for initial value (0.7 + j*0), the solution does exist, but the OPTIMAL MULTIPLIER METHOD cannot get it and gives an incorrect information. This is the reason why the authors of this paper think the conclusion (3) is incorrect.
Now let us discuss the statement of reference [
In fact, [
If
What is the benefit of an Optimal Multiplier Method for the 43-bus system? The following figures can answer this question. For flat start, the initial LAVM is 0.5607 × 101, the successive iteration values for the normal Newton method and Optimal Multiplier Method are as follows:
0.3243D+01, 0.6209D+00, 0.1395D+00, 0.2084D−01, 0.1559D−2, 0.1657D−04,
0.2123D−08, 0.1656D−12, 0.4639D−12, 0.3622D−12 (Normal Newton Method)
0.1078D+01, 0.2838D+00, 0.3082D−01, 0.1252D−02, 0.3882D−05,
0.4134D−08, 0.5128D−10, 0.2423D−12, 0.3883D−12, 0.2913D−12
(Optimal Multiplier Method, Optimal Multiplier µ was taken from [
From the figures above, we can see that if the stopping condition is LAVM < 1D−08, the Optimal Multiplier Method needs 6 iterations, it can save one iteration when compare with normal Newton iteration.
But from another side, we see that at Optimal Multiplier Method’s 5th and 6th iteration, the LAVM is 0.3852D−05 and 0.4134D−08 respectively. That is to say, if the values of µ were taken from [
The authors of this paper did the same work of [
Ref. [
0.6522, 1.1043, 1.0740, 1.1333, 1.0006, 1.0011, 0.9876, 1.0011, 1.0001, 0.9876.
Ours:
0.6525, 1.1042, 1.0746, 1.1311, 1.0011, 1.0000, 1.0000, 1.0117, 1.0449, 1.01002
The LAVMs of ours are:
0.1073D+01, 0.2840D+00, 0.3088D−01, 0.1296D−02, 0.4415D−05,
0.4944D−10, 0.1385D−12, 0.1818D−12, 0.2634D−12, 0.3000D−12.
In order to check the values of µ, we use another method to do the calculation.
For 3rd polynomial equation, there must be a real root, using the normal Newton method we can find this root, then to solve a 2nd polynomial equation, three roots can be obtained. Among them, the one which most close to 1, should be the optimal multiplier µ.
The results are as follows:
0.6525, 1.1042, 1.0746, 1.1311, 1.0011, 1.0000, 1.0000, 1.0000, 1.0000
We believe these results are even more reasonable.
The corresponding LAVMs are:
0.1073D+01, 0.2840D+00, 0.3088D−01, 0.1296D−02, 0.4415D−05,
0.4944D−10, 0.2855D−12, 0.2423D−12, 0.2307D−12, 0.2394D−12.
The results above show in our calculation, Optimal Multiplier Method does not lose quadratic convergence nature.