1.1. Motivation
In power system analysis, PF is probably the most important computational tool, playing a vital role in a wide variety of applications such as power system planning and operation, economic dispatch or security analysis, among others. From a mathematical point of view, PF consists on solving the set of equations that model the nonlinear relationships between nodal voltages and power injections. Since the PF equations are nonlinear, this problem cannot be directly solved. In this sense, iterative solvers have been customary used for solving the PF equations. From a PF solver one expects two main characteristics:
Robustness: for effectively solving ill-conditioned systems, which are mainly encountered in presence of cascading failures, heavy loading conditions or badly initialization of the iterative procedure.
Efficiency: this aspect is crucial for online applications, since it determines the ability of a PF solver for quickly finding the solution of the PF equations.
Many efforts have been devoted during decades on developing different PF solvers (see Literature Review). Developed approaches have been mainly focused on separately enhancing one of the characteristics enumerated above. However, in a modern power system paradigm, it is necessary to use tools that are simultaneously efficient and robust enough to, for example, properly handle large-scale ill-conditioned cases. In this context, it is very difficult to find a PF solver which gathers both robust and efficient features. For instance, the conventional Newton-Raphson (NR) can be considered an efficient technique, however, it fails in ill-conditioned systems being so its universal applicability questionable. Recently, some studies have been conducted on studying the applicability of various high-order Newton-like (HONL) methods for PF analysis [
1,
2,
3,
4], manifesting good performance in well-conditioned systems. However, these works lack of a formal analysis of the mathematical properties of such solvers, more specifically, their stability and robustness properties were not properly studied. This work aims at being a first step on filling this gap.
1.2. Literature Review
During decades, PF problem has been extensively investigated by the engineering community. In this context, the use of NR for PF solution at early 60’s [
5] can be considered as one of the main milestones. By far, NR is nowadays the most widely employed solver in PF analysis, however, inversion or factorization of the Jacobian matrix supposed an important barrier for former computers. Thereby, many efforts were conducted on reducing this computational burden either by using decoupled techniques [
6,
7] or sparsity routines [
8]. Decoupled techniques gained popularity during 80’s, nonetheless, the emergence of very efficient computational machines and the lack of effectiveness of such techniques in ill-conditioned systems limit their usage nowadays.
On the other hand, ill-conditioned systems began to be studied at 80’s due to the difficulty of the existing techniques to successfully solve such cases [
9]. Firstly, ill-conditioning in power systems was most related with electric characteristics of the networks (i.e., high R/X ratio or heavy loading conditions). This definition has evolved towards being more related with the starting guess of the iterative procedure, according to the definition given in [
10]. However, ill-conditioning in power systems is currently related with mathematical issues. During decades, solution of ill-conditioned PF problems has supposed as a main concern for engineering researchers. The very first steps in this direction [
9] were devoted on developing robust techniques which resulted very inefficient in most cases (e.g., see results in [
10]).
With the advent of 90’s, PF was revisited in order to address current issues. In this sense, PF analysis in distribution networks (radial or meshed) attracted huge attention. In this kind of networks, conventional (especially decoupled) solvers offered badly performance due to high R/X ratios or natural bad condition of such systems. To address such issues, huge efforts were conducted on developing specific solvers [
11,
12] and formulations [
13,
14,
15]. However, this kind of techniques are very far to be considered universal methods as their main area of application lie in distribution systems. Actually, some of them cannot be directly applied in meshed networks [
11]. In addition, some specific PF solvers such as the well-known backward-forward algorithm are totally inefficient in comparison with Newton-like methods (see results for radial networks in [
3]).
Ill-conditioned systems were barely studied at the beginning of the 21st century, however, this issue have re-emerged in recent dates due to this kind of systems are becoming more frequent [
16]. In addition, in contrast to the traditional ill-conditioned systems studied at 80’s which comprised very few buses [
8], modern robust solvers require to be efficient and competitive with conventional techniques (e.g., NR) due to most of ill-conditioned cases are, in fact, very large (>1000 buses). In this sense, traditional robust approaches such like [
9] can be considered totally out of date [
10]. This way, multiple works have been recently conducted on developing efficient and robust techniques that could satisfactory addressed both well and ill-conditioned cases. The application of the Continuous Newton’s method for PF analysis [
10,
17,
18,
19] suppose a clear example of this trend. The Continuous Newton’s method establishes a formal analogy between NR and Euler’s techniques by which any other numerical integration method (e.g., the Runge-Kutta formulas [
18]) could be applied for developing robust and efficient PF solvers. However, as reported in [
18], those solvers based on the Continuous Newton’s method are still not competitive with NR in the sense that they are very inefficient. As a sake of example, the solver introduced in [
10] which is based on the 4th order Runge-Kutta method requires four matrix factorizations. Since the factorization of the Jacobian matrix could be considered the heaviest computational part of a PF solver [
10], one could deduce that the technique developed in [
10] is four-times less efficient than NR. To address such issue, the authors developed a combined Runge-Kutta-Broyden paradigm in [
20], which avoids the factorization of the Jacobian matrix. However, this calculation is replaced by the inversion of the Jacobian, which is unaffordable in large-scale systems. Similarly, the dynamic computing paradigm developed in [
16] and its posterior refinements [
21,
22] are not suitable for large-scale cases, as the authors pointed out in [
16].
Regularization-based methods have been recently applied for solving PF equations, especially in ill-conditioned cases [
23,
24]. This approach basically consists of avoiding the singularity of the Jacobian matrix when it is ill-conditioned, by adding the so-called regularizing operator. Due to the Jacobian matrix is modified, the solution of the iterative process may not be the actual solution of the PF problem, as commented in [
24] and reported in [
25]. This issue has limited the applicability of this kind of techniques. Homotopic and holomorphic principles have been also applied for solving PF problem [
26,
27]. This kind of methodologies solve the PF equations by posing a continuous version of them by which their solution is reaching by progressively increasing a well-known homotopic parameter. This kind of approaches are quite robust and have found certain interest on ill-conditioned cases. However, as in the case of regularization-based methods, they may reach a non-physical solution which lacks of interest [
28]. Recently, the authors have developed various self-developed robust solvers [
28,
29], which arise from combination of different approaches. Although this kind of techniques present good computational performance, they are not competitive with other conventional techniques in well-conditioned cases yet.
Recently, HONL methods hav gained popularity for solving PF equations. Basically, a HONL is a nonlinear solved based on NR with a convergence order higher than two. Thus, a HONL should converge employing less iterations than NR. If besides this feature the solver is computationally efficient, the obtained PF solution approach may be competitive with NR. In this context, Pourbagher and Derakhshandeh firstly compared various HONL in [
1], showed very good results. The conclusions in this reference motivated the authors to further study this kind of methods in [
2,
3,
4], confirming that HONL methods may suppose an attractive alternative to NR.
1.3. Contributions and Paper Organization
Motivated by the good results showed by HONL techniques in recent studies. This work aims at further analyzing this kind of techniques for PF analysis. So far, HONL methods have been only employed for solving well-conditioned systems, therefore, their robust properties have not been analyzed from a theoretical and empirical point of view. This work aims at supposing a first step to filling this gap. In this sense, we focus on two well-known family of cubic techniques. In this context, we comprehensively analyze the cubic methods proposed by Weerakoon [
30] (3OW) and Darvishi [
31] (3OD). The stability, efficient and robustness properties of these solvers are firstly theoretically deduced and analyzed and, posteriorly, these theoretical foundations are empirically confirmed by numerical experiments. The two cubic methods are compared with the most standard PF solver (i.e., NR) and the 4th order Runge-Kutta method (RK4) developed in [
10] with the aim of providing a benchmark study for the applicability of these families in both well and ill-conditioned systems. It is worth mentioning that these two cubic methodologies suppose the background for the development of many HONL techniques (e.g., see [
32]). Thereby, this work aims at supposing a valuable guidance for applying other HONL techniques that are in fact modifications of the studied cubic techniques.
Remainder of this paper is organized as follows.
Section 2 outlines the necessary background. Stability of two families of cubic techniques is studied in
Section 3.
Section 4 is devoted on comparing two families of cubic methods with the Newton-Raphson approach using a well-known efficiency index.
Section 5 presents several numerical experiments with results. The main conclusions of this work are duly drawn in
Section 6.