betaFeedback

WorksMathematics

A New Perspective on Optimal Power Flow: From Control Theory to Numerical Optimization

Abstract

The optimal power flow (OPF) problem is a classical non-convex optimization problem with nonlinear constraints and multiple potential solutions. Existing methods often fail to find the global optimal solution, motivating the development of a method to identify multiple local solutions. This paper applies the optimal control principle to solve the OPF problem, aiming to systematically explore all feasible solutions within a single run. We first handle the constraints in the OPF problem using the penalty method and the Lagrange multiplier, and then reformulate it as an optimal control problem. The connection between the two problems is established by the maximum principle. Moreover, a numerical algorithm is proposed to solve this optimal control problem. The state trajectory of the system can rapidly and stably converge to the solution of the OPF problem. By designing the initial control sequence, the proposed method can successfully find multiple solutions and escape local solutions. Finally, simulations were conducted on several representative test systems to validate the potential of the proposed method in identifying multiple solutions for non-convex problems.

Introduction

The optimal power flow (OPF) problem aims to find a steady state operating point of a power system that minimizes the cost of electric power generation while satisfying operational constraints [1, 2]. Due to the nonlinear constraint of the power flow equations, the OPF problem is a non-convex optimization problem and may contain many solutions [3, 4, 5, 6, 7].

The solution of the OPF problem began with the gradient method proposed by Dommel and Tinney in 1968 [2]. To effectively handle the constraints, classical approaches such as the Lagrange multiplier method and the penalty function method have been extensively explored [8, 9, 10]. Building on these foundations, the interior point method (IPM), which introduces slack variables and incorporates a logarithmic barrier function, has become a mainstream approach due to its fast convergence rate [11, 12, 13]. However, when dealing with highly nonlinear objective functions or constraints, the effectiveness of these traditional methods gradually decreases, showing limitations and being prone to getting stuck in local optimal solutions [5], [14, 15, 16].

Heuristic methods have been applied to power systems for their ability to handle complex problems [17]. Among them, genetic algorithms (GAs) have attracted considerable attention [3], [18, 19, 20, 21]. However, these methods often suffer from premature convergence in OPF problems [22], making it difficult to capture all possible solutions.

In our recent work, we have successfully applied the optimal control principle (OCP) to optimization problems and achieved notable results [23, 24, 25]. Motivated by the characteristic of optimal controllers to drive system states rapidly and stably to equilibrium, we incorporate OCP into finding the solution of non-convex problems. This paper explores the application of the OCP in the OPF problem and proposes a novel optimization method. This method demonstrates outstanding performance in solving highly nonlinear and non-convex problems. Specifically, the main contributions of this work are as follows:

  1. The proposed method handles inequality constraints using the penalty method and directly solves for the Lagrange multipliers to ensure satisfaction of the equality constraints. Compared with approaches that rely on power flow calculations to satisfy equality constraints during the optimization of the OPF problem [26, 27, 28], this strategy avoids complex calculations and effectively prevents the objective function from becoming ill-conditioned as in traditional penalty methods, thereby enhancing the numerical stability of the algorithm used to solve this problem.

  1. A solution method for the OPF problem based on the OCP is proposed. The OPF problem is transformed into a discrete-time optimal control problem, where the optimal controller adjusts the first-order difference equations to minimize a cost function composed of control energy and the objective function. It is shown that the steady state of the system in the optimal control problem corresponds to the stationary point of the original problem. Furthermore, an optimization algorithm with superlinear convergence rate is proposed to solve this optimal control problem. Compared with the traditional methods, the proposed method is guaranteed to converge rapidly and stably.

  1. More importantly, it can effectively escape local solutions of the OPF problem and identify multiple OPF solutions within a single execution by appropriately designing the initial control sequence. Numerical experiments further confirm the effectiveness.

This work was supported by the Original Exploratory Program Project of the National Natural Science Foundation of China (62450004), the Joint Funds of the National Natural Science Foundation of China (U23A20325), and the National Natural Science Foundation of China (52307115). (Corresponding author: Huanshui Zhang)

Yeming Xu, Qi Liu, Liping Zhang, and Huanshui Zhang are with the College of Electrical Engineering and Automation, Shandong University of Science and Technology, Qingdao 266590, China (e-mail: [email protected], [email protected], [email protected], [email protected]). Huanshui Zhang is also with the School of Control Science and Engineering, Shandong University, Jinan, Shandong 250061, China.

Minyue Fu is with the School of Automation, Guangdong University of Technology, Guangzhou 510006, China (e-mail: [email protected]).

ACOPF Model and Constraint Handling

This section introduces the AC OPF formulation in power systems and the associated operational constraints.

Problem Formulation

The AC OPF problem can be formulated as

min⁡Fobj=∑g∈NGfg(PGg),s.t.PGg,min⁡≤PGg≤PGg,max⁡,QGg,min⁡≤QGg≤QGg,max⁡,g∈NG,Vb,min⁡≤Vb≤Vb,max⁡,Pb=PGb−PDb,Qb=QGb−QDb,b∈NB.(1)\begin{aligned} \min\quad&F_{\mathrm{obj}}=\sum_{g\in\mathcal{N}_{G}} f_g(P_{G_g}),\\ \mathrm{s.t.}\quad&P_{G_g,\min}\leq P_{G_g}\leq P_{G_g,\max},\\ &Q_{G_g,\min}\leq Q_{G_g}\leq Q_{G_g,\max},\quad g\in\mathcal{N}_{G},\\ &V_{b,\min}\leq V_b\leq V_{b,\max},\\ &P_b=P_{G_b}-P_{D_b},\\ &Q_b=Q_{G_b}-Q_{D_b},\quad b\in\mathcal{N}_{B}. \tag*{(1)} \end{aligned}

where FobjF_{\mathrm{obj}} is the objective function; NB\mathcal{N}_{B} denotes the set of all buses, and NG\mathcal{N}_{G} is the set of buses where generators are located. PGbP_{G_b} and QGbQ_{G_b} represent the active and reactive power outputs of the generator at bus bb, while PDbP_{D_b} and QDbQ_{D_b} denote the corresponding active and reactive power demands. For buses without generators, i.e., b∈NB∖NGb\in\mathcal{N}_{B}\setminus\mathcal{N}_{G}, we set PGb=QGb=0P_{G_b}=Q_{G_b}=0. The voltage magnitude, net active and reactive power injections at bus bb are denoted by VbV_b, PbP_b, and QbQ_b, respectively. The power flow equations in (1e) and (1f) can be expressed as

Pb=Vb∑j∈NBVj(Gbjcos⁡θbj+Bbjsin⁡θbj),P_b=V_b\sum_{j\in\mathcal{N}_{B}}V_j\left(G_{bj}\cos\theta_{bj}+B_{bj}\sin\theta_{bj}\right),
Qb=Vb∑j∈NBVj(Gbjsin⁡θbj−Bbjcos⁡θbj),Q_b=V_b\sum_{j\in\mathcal{N}_{B}}V_j\left(G_{bj}\sin\theta_{bj}-B_{bj}\cos\theta_{bj}\right),

where θbj=θb−θj\theta_{bj}=\theta_b-\theta_j, and θb\theta_b denote the voltage angle at bus bb; GbjG_{bj} and BbjB_{bj} denote the series conductance and susceptance between buses bb and jj, respectively. In this paper, the generation cost function is selected as the optimization objective.

Handling Constraint

For the inequality constraints in problem (1), we employ the penalty method by directly incorporating penalty terms corresponding to (1b) - (1d) into the objective function FobjF_{\mathrm{obj}}. The resulting equality-constrained optimization problem is then expressed as

min⁡Fpen=Fobj+Ppen,s.t.(1e) and (1f),(2)\begin{aligned} \min\quad&F_{\mathrm{pen}}=F_{\mathrm{obj}}+P_{\mathrm{pen}},\\ \mathrm{s.t.}\quad&(1e)\text{ and }(1f), \tag*{(2)} \end{aligned}

where FpenF_{\mathrm{pen}} is the penalized objective function, and the penalty term is formulated as

Ppen=λ1∑b∈NBϕ(Vb,Vb,min⁡,Vb,max⁡)+λ2∑g∈NGϕ(PGg,PGg,min⁡,PGg,max⁡)+λ3∑g∈NGϕ(QGg,QGg,min⁡,QGg,max⁡),\begin{aligned} P_{\mathrm{pen}}={}&\lambda_1\sum_{b\in\mathcal{N}_{B}}\phi(V_b,V_{b,\min},V_{b,\max})\\ &+\lambda_2\sum_{g\in\mathcal{N}_{G}}\phi(P_{G_g},P_{G_g,\min},P_{G_g,\max})\\ &+\lambda_3\sum_{g\in\mathcal{N}_{G}}\phi(Q_{G_g},Q_{G_g,\min},Q_{G_g,\max}), \end{aligned}

where λ1\lambda_1, λ2\lambda_2, and λ3\lambda_3 are the penalty coefficients.

We define the Lagrangian of problem (2) as

L(y,μ,ν)=Fpen+∑b∈NBμb(PGb−PDb−Pb)+∑b∈NBνb(QGb−QDb−Qb).(3)L(y,\boldsymbol{\mu},\boldsymbol{\nu})=F_{\mathrm{pen}}+\sum_{b\in\mathcal{N}_{B}}\mu_b(P_{G_b}-P_{D_b}-P_b)+\sum_{b\in\mathcal{N}_{B}}\nu_b(Q_{G_b}-Q_{D_b}-Q_b). \tag*{(3)}

where μ=[μb]b∈NB\boldsymbol{\mu}=[\mu_b]_{b\in\mathcal{N}_{B}} and ν=[νb]b∈NB\boldsymbol{\nu}=[\nu_b]_{b\in\mathcal{N}_{B}} are the Lagrange multipliers associated with power flow equality constraints. The optimization variable y=[PG,QG,V,θ]∈Rmy=[P_G,Q_G,V,\theta]\in\mathbb{R}^{m} includes the generator active power outputs PG=[PGg]g∈NGP_G=[P_{G_g}]_{g\in\mathcal{N}_{G}}, generator reactive power outputs QG=[QGg]g∈NGQ_G=[Q_{G_g}]_{g\in\mathcal{N}_{G}}, bus voltage magnitudes V=[Vb]b∈NBV=[V_b]_{b\in\mathcal{N}_{B}}, and bus voltage angles θ=[θb]b∈NB∖r\theta=[\theta_b]_{b\in\mathcal{N}_{B}\setminus r}, where rr is the index of the reference bus and m=2∣NG∣+2∣NB∣−1m=2|\mathcal{N}_{G}|+2|\mathcal{N}_{B}|-1.

For clarity in the subsequent discussion, we report the following well-known lemma. A comprehensive overview can be found in Section 11.3 of [29].

Lemma 1 (First-order necessary conditions): Let y∗=[PG∗,QG∗,V∗,θ∗]y^*=[P_G^*,Q_G^*,V^*,\theta^*] be a local minimum of FpenF_{\mathrm{pen}} subject to the constraints (1e) and (1f). Then, there exist unique Lagrange multipliers μ∗=[μb∗]b∈NB\boldsymbol{\mu}^*=[\mu_b^*]_{b\in\mathcal{N}_{B}} and ν∗=[νb∗]b∈NB\boldsymbol{\nu}^*=[\nu_b^*]_{b\in\mathcal{N}_{B}} such that

∇yL(y∗,μ∗,ν∗)=0.(4)\nabla_yL(y^*,\boldsymbol{\mu}^*,\boldsymbol{\nu}^*)=0. \tag*{(4)}

In the rest of the paper, we refer to any triple (y∗,μ∗,ν∗)(y^*,\boldsymbol{\mu}^*,\boldsymbol{\nu}^*) that satisfies (4) and the constraints (1e) and (1f) as a stationary point of problem (2).

Solution Method for ACOPF via Optimal Control

Design of the Optimal Control Problem

As stated in Lemma 1, the stationary points of problem (2) satisfy

∇xL(x)=0,(5)\nabla_xL(x)=0, \tag*{(5)}

where x=[y,μ,ν]T∈Rnx=[y,\boldsymbol{\mu},\boldsymbol{\nu}]^{\mathsf{T}}\in\mathbb{R}^{n} and n=m+2∣NB∣n=m+2|\mathcal{N}_{B}|. Nonlinear equations (5) are typically addressed via Newton’s method

xk+1=xk+αkdk,x_{k+1}=x_k+\alpha_kd_k,

where xk=[yk,μk,νk]Tx_k=[y_k,\boldsymbol{\mu}_k,\boldsymbol{\nu}_k]^{\mathsf{T}} represents the vector of the variable xx at the kk-th iteration, αk>0\alpha_k>0 is step size and dkd_k is the update direction obtained by solving

H(xk)dk=−∇xL(xk),(6)H(x_k)d_k=-\nabla_xL(x_k), \tag*{(6)}

where H(x)H(x) is the Jacobian matrix of ∇xL(x)\nabla_xL(x). We designed the following optimal control problem

min⁡uy,uzmax⁡uxJ=∑k=1N(L(xk)+12(uky)TRyuky−12(ukz)TRzukz)+L(xN+1),s.t.xk+1=xk+[ukyukz].(7)\begin{aligned} \min_{u^y,u^z}\max_{u^x}\quad J={}&\sum_{k=1}^{N}\left(L(x_k)+\frac{1}{2}(u_k^y)^{\mathsf{T}}R_yu_k^y-\frac{1}{2}(u_k^z)^{\mathsf{T}}R_zu_k^z\right)\\ &+L(x_{N+1}),\\ \mathrm{s.t.}\quad&x_{k+1}=x_k+\begin{bmatrix}u_k^y\\u_k^z\end{bmatrix}. \tag*{(7)} \end{aligned}

where (7b) is the discrete-time system, xk=[yk,zk]Tx_k=[y_k,z_k]^{\mathsf{T}}, zk=[μk,νk]Tz_k=[\boldsymbol{\mu}_k,\boldsymbol{\nu}_k]^{\mathsf{T}} is the system state, JJ is the cost function, and uk=[(uky)T,(ukz)T]Tu_k=[(u_k^y)^{\mathsf{T}},(u_k^z)^{\mathsf{T}}]^{\mathsf{T}} is the control input. The parameter NN represents the time horizon, while Ry∈S++mR_y\in\mathbb{S}_{++}^{m} and Rz∈S++n−mR_z\in\mathbb{S}_{++}^{n-m} are control weighting matrices. The control sequence uku_k in problem (7) is designed to drive the state trajectory {xk}\{x_k\} toward the stationary point of L(x)L(x), as described below.

Theorem 1: The state trajectory {xk}\{x_k\} generated by optimal control problem (7) converges to the stationary point of problem (2).

Proof: The maximum principle [30] is applied in problem (7) to derive the following equations:

pk−1=pk+∇xL(xk),(8)p_{k-1}=p_k+\nabla_xL(x_k), \tag*{(8)}
0=Ruk+pk,(9)0=Ru_k+p_k, \tag*{(9)}

where (8) is the costate equation with the terminal condition pN=∇xL(xN+1)p_N=\nabla_xL(x_{N+1}), R=blkdiag⁡(Ry,−Rz)R=\operatorname{blkdiag}(R_y,-R_z), and (9) is the equilibrium condition. It follows from (7b) that as the system approaches steady state, for any given tolerance ϵ>0\epsilon>0, there exists an integer k0>0k_0>0 such that ∥uk∥<ϵ\lVert u_k\rVert<\epsilon for all k>k0k>k_0. Combining with (9), we obtain ∥pk∥<ϵ∥R∥\lVert p_k\rVert<\epsilon\lVert R\rVert for all k>k0k>k_0. It then follows from (8) that

∥∇xL(xk)∥<2ϵ∥R∥,∀k>k0+1,\lVert\nabla_xL(x_k)\rVert<2\epsilon\lVert R\rVert,\qquad\forall k>k_0+1,

which implies that {xk}\{x_k\} converges to the stationary point of problem (2). The proof is now completed.

Design of OCP-based Method

We derive the gradient and Hessian matrix of JJ with respect to {uk}\{u_k\} to clearly elucidate the numerical algorithm. The gradient and Hessian matrix of JJ with respect to {uk}\{u_k\} are given by

∇UJ=(IN+1⊗R)U+L~[∇xL(x1)⋮∇xL(xN+1)],(10)\nabla_UJ=(I_{N+1}\otimes R)U+\widetilde L \begin{bmatrix} \nabla_xL(x_1)\\ \vdots\\ \nabla_xL(x_{N+1}) \end{bmatrix}, \tag*{(10)}
∇U2J=IN+1⊗R+L~DL~T,(11)\nabla_U^2J=I_{N+1}\otimes R+\widetilde L D\widetilde L^{\mathsf T}, \tag*{(11)}

where U=[u0T,…,uNT]TU=[u_0^{\mathsf T},\ldots,u_N^{\mathsf T}]^{\mathsf T}, L~=TN+1⊗In\widetilde L=T_{N+1}\otimes I_n, TN+1∈R(N+1)×(N+1)T_{N+1}\in\mathbb{R}^{(N+1)\times(N+1)} is an upper-triangular matrix with all entries on and above the diagonal equal to 1, and DD is a block-diagonal matrix whose ii-th diagonal block is given by H(xi)H(x_i), i=1,…,N+1i=1,\ldots,N+1.

Denote Ut=[u0T,…,uNT]TU_t=[u_0^{\mathsf T},\ldots,u_N^{\mathsf T}]^{\mathsf T} as the vector of the variable UU at the tt-th iteration. The iterative scheme of the proposed numerical algorithm is given as

Ut+1=Ut−αtgt(Ut),U_{t+1}=U_t-\alpha_tg_t(U_t),
gj(Ut)=βt(∇UJ(Ut)+Mgj−1(Ut)),j=1,…,t,(12)g_j(U_t)=\beta_t\left(\nabla_UJ(U_t)+Mg_{j-1}(U_t)\right),\qquad j=1,\ldots,t, \tag*{(12)}

where βt=(M+∇U2J(Ut))−1\beta_t=(M+\nabla_U^2J(U_t))^{-1}, g0(Ut)=βt∇UJ(Ut)g_0(U_t)=\beta_t\nabla_UJ(U_t), M∈S+n(N+1)M\in\mathbb{S}^{n(N+1)}_+ is a adjustable matrix. The convergence rate analysis of the proposed numerical algorithm is as follows.

Theorem 2: Assume that ∇UJ(U∗)=0\nabla_UJ(U^*)=0 and ∇U2J(U∗)∈S++n(N+1)\nabla_U^2J(U^*)\in\mathbb{S}^{n(N+1)}_{++}. If the sequence {xk}\{x_k\} generated by (12) converges and the step size is chosen as αt=1−δt+1\alpha_t=1-\delta^{t+1} with δ∈[0,1)\delta\in[0,1), then it admits superlinear convergence.

Proof: By direct derivation, we obtain

gt(U∗)=[In(N+1)−((M+∇U2J(U∗))−1M)t+1]∇U2J(U∗)−1∇UJ(U∗)=0n(N+1),g_t(U^*)=\left[I_{n(N+1)}-\left((M+\nabla_U^2J(U^*))^{-1}M\right)^{t+1}\right]\nabla_U^2J(U^*)^{-1}\nabla_UJ(U^*)=0_{n(N+1)},
∇Ugt(U∗)=In(N+1)−((M+∇U2J(U∗))−1M)t+1.\nabla_Ug_t(U^*)=I_{n(N+1)}-\left((M+\nabla_U^2J(U^*))^{-1}M\right)^{t+1}.

It follows from (12) that

∥Ut+1−U∗∥=∥Ut−U∗−αtgt(Ut)∥=∥Ut−U∗−αt[gt(U∗)+∇gt(U∗)(Ut−U∗)+rt]∥≤∥Ut−U∗−αt∇gt(U∗)(Ut−U∗)∥+∥rt∥=∥[δt+1In(N+1)+αt((M+∇U2J(U∗))−1M)t+1](Ut−U∗)∥+∥rt∥≤∥δt+1In(N+1)+αt((M+∇U2J(U∗))−1M)t+1∥⋅∥Ut−U∗∥+∥rt∥,\begin{aligned} \lVert U_{t+1}-U^*\rVert &=\lVert U_t-U^*-\alpha_tg_t(U_t)\rVert\\ &=\lVert U_t-U^*-\alpha_t[g_t(U^*)+\nabla g_t(U^*)(U_t-U^*)+r_t]\rVert\\ &\leq\lVert U_t-U^*-\alpha_t\nabla g_t(U^*)(U_t-U^*)\rVert+\lVert r_t\rVert\\ &=\left\lVert\left[\delta^{t+1}I_{n(N+1)}+\alpha_t\left((M+\nabla_U^2J(U^*))^{-1}M\right)^{t+1}\right](U_t-U^*)\right\rVert+\lVert r_t\rVert\\ &\leq\left\lVert\delta^{t+1}I_{n(N+1)}+\alpha_t\left((M+\nabla_U^2J(U^*))^{-1}M\right)^{t+1}\right\rVert\cdot\lVert U_t-U^*\rVert+\lVert r_t\rVert, \end{aligned}

where ∥rt∥=o(∥Ut−U∗∥)\lVert r_t\rVert=o(\lVert U_t-U^*\rVert) denotes the remainder term. Noted that ρ((M+∇U2J(U∗))−1M)<1\rho((M+\nabla_U^2J(U^*))^{-1}M)<1, we obtain

lim⁡t→∞∥Ut+1−U∗∥∥Ut−U∗∥=0,\lim_{t\to\infty}\frac{\lVert U_{t+1}-U^*\rVert}{\lVert U_t-U^*\rVert}=0,

i.e., (12) is superlinearly convergent.

The numerical algorithm is summarized as Algorithm 1.

Algorithm 1 OCP-based method.
1: Initialization: U0U_0, x0x_0, M⪰0M\succeq0, t←0t\leftarrow0, δ\delta and ϵ\epsilon
2: repeat
3: Forward update {xkt}\{x_k^t\};
4: Calculate ∇UJ(Ut)\nabla_UJ(U_t) and ∇U2J(Ut)\nabla_U^2J(U_t) via (10) and (11);
5: Update Ut+1U_{t+1} using (12);
6: t←t+1t\leftarrow t+1
7: until ∥∇UJ(Ut)∥<ϵ\lVert\nabla_UJ(U_t)\rVert<\epsilon
Output: {xkt}\{x_k^t\}

Table 1.

Algorithm 1 is initialized with an initial control sequence U0U_0. Consider the case where the initial state is initialized as x0=x∗x_0=x^*, with x∗x^* denoting a stationary point of problem (2), and the initial control sequence U0U_0 is chosen as the zero sequence. In this case, the state trajectory generated by U0U_0 satisfies all the conditions in (7b), (8), and (9), causing the algorithm to terminate immediately and output xk0=x∗x_k^0=x^* for k=1,…,N+1k=1,\ldots,N+1. This means that all states remain fixed at x∗x^*. Such a zero-sequence initialization leads to an initial trajectory difficult to explore other potential solutions. Therefore, to facilitate a more extensive exploration of the feasible domain, the strategic design of U0U_0 is crucial. In the following section, unless otherwise specified, U0U_0 is generated as follows: First, the feasible domain of the variable yy is partitioned into several subregions, and the central points of these subregions are selected to form a set S\mathcal{S}, ensuring that S\mathcal{S} contains at least N+1N+1 elements. Then, for each time step k=1,…,N+1k=1,\ldots,N+1, y0ky_0^k is assigned by sequentially selecting a stream from S\mathcal{S}, while y00y_0^0 is fixed as the given initial value y0y_0. The Lagrange multipliers μ0k\mu_0^k and ν0k\nu_0^k are initialized as zero vectors for all index kk. Finally, U0U_0 can be computed from uk0=xk+10−xk0u_k^0=x_{k+1}^0-x_k^0.

Numerical Examples

The effectiveness of the proposed method is evaluated through numerical simulations on several benchmark multi-modal ACOPF systems from [5], including WB2, WB3, and WB5.

Simulation Configuration

We selected two benchmark solvers, i.e., the MIPS solver based on the IPM and the fmincon solver based on SQP, to compare with our proposed method. The initial values for both solvers are sequentially set as yk0y_k^0 for k=0,…,N+1k = 0,\ldots,N+1. Ry=10−10ImR_y = 10^{-10}I_m and Rz=10−10In−mR_z = 10^{-10}I_{n-m} are is used in all test systems. To obtain stationary points, the matrix MM is constructed as a block-diagonal matrix with N+1N+1 identical blocks, each being an n×nn \times n diagonal matrix. In each block, the first mm diagonal entries are set to cc, while the remaining diagonal entries are zero. We set ε=10−8\varepsilon= 10^{-8} as the tolerance for all algorithms.

2-, 3-, and 5-bus Systems

The test systems WB2, WB3, and WB5 are employed to validate the effectiveness of the proposed method in finding multiple solutions in this subsection. For the two-bus system WB2 and the three-bus system WB3, the local and global optimal solutions are very close, with only minor differences in their objective function values. In contrast, the five-bus system WB5 includes two generators with distinct costs, leading to a larger disparity in objective function values. All three systems have two feasible solutions, denoted as S1S_1 and S2S_2. The solutions for each test system are summarized in Table I (PDF p. 4).

SystemSolutionFobjF_{\mathrm{obj}}BusVbV_b (p.u.)θb\theta_b (deg)PGbP G_b (MW)QGbQ G_b (Mvar)
WB2S1S_1877.7810.9520.00438.8994.44
WB2S1S_121.050-57.14--
WB2S2S_2905.7310.9500.00452.86164.32
WB2S2S_220.976-64.94--
WB3S1S_1417.2510.9510.00208.628.93
WB3S1S_120.950-27.25--
WB3S1S_130.981-30.36--
WB3S2S_2418.1410.9500.00209.07-20.79
WB3S2S_221.011-26.36--
WB3S2S_231.050-29.23--
WB5S1S_1946.5811.0500.00181.43124.09
WB5S1S_120.960-3.46--
WB5S1S_130.950-3.21--
WB5S1S_140.98037.65--
WB5S1S_151.05045.48220.87-30.00
WB5S2S_21082.3311.0100.00246.0529.27
WB5S2S_220.950-6.52--
WB5S2S_230.950-6.57--
WB5S2S_241.010358.66--
WB5S2S_251.050363.1198.15-30.00

Table I (PDF p. 4). ACOPF SOLUTIONS FOR WB2, WB3, AND WB5 TEST SYSTEMS

For the case of WB2, we set N=4N = 4, initial value y0=[0.9,−0.4,1,1,2.6]y_0 = [0.9,-0.4,1,1,2.6], c=0.001c = 0.001, δ=0.9\delta= 0.9 and λi=103\lambda_i = 10^3, i=1,2,3i = 1,2,3 for our method. Figure 1 shows the results of our proposed method. From Figure 1a, it can be seen that {xk}\{x_k\} successfully converges, with the corresponding gradient norm of L(x)L(x) below 10−810^{-8}. Figure 1b illustrates the values of FobjF_{\mathrm{obj}} along the state trajectory {xk}\{x_k\}. It can be observed that x2x_2, x3x_3, x4x_4, and x5x_5 stay at the global optimal solution S1S_1, while x1x_1

Convergence behavior of the proposed method on WB2

Figure 1. Convergence behavior of the proposed method on WB2. (a) ∥∇xL(x)∥\|\nabla_xL(x)\| along the state trajectory {xk}\{x_k\}. (b) Values of FobjF_{\mathrm{obj}} along the state trajectory {xk}\{x_k\}.

stays at the local solution S2S_2. This indicates that all solutions are successfully identified in WB2. It should be noted that although x0x_0 corresponds to a smaller FobjF_{\mathrm{obj}} for the original problem, it does not satisfy the equality constraints (1e) and (1f).

Figure 2 presents the results for the WB3 with N=3N = 3, y0=[7.5,−3,1,1,0,0]y_0 = [7.5,-3,1,1,0,0], c=0.0001c = 0.0001, δ=0.99\delta= 0.99, and λi=104\lambda_i = 10^4 for i=1,2,3i = 1,2,3. As shown in Figure 2a, the state trajectory converges to the stationary points. Figure 2b demonstrates that all solutions are successfully identified, i.e., x2x_2 corresponds to S2S_2, whereas x1x_1, x3x_3, and x4x_4 correspond to S1S_1.

Convergence behavior of the proposed method on WB3

Figure 2. Convergence behavior of the proposed method on WB3. (a) ∥∇xL(x)∥\|\nabla_xL(x)\| along the state trajectory {xk}\{x_k\}. (b) Values of FobjF_{\mathrm{obj}} along the state trajectory {xk}\{x_k\}.

The five-bus system WB5 has two solutions, with the objective value of the local solution being 14% higher than that of the global optimal solution. We set N=3N = 3, y0=[7.5,7.5,8,8,1,1,1,1,1,0,0,0,0]\boldsymbol{y}_0 = [7.5, 7.5, 8, 8, 1, 1, 1, 1, 1, 0, 0, 0, 0], c=0.01c = 0.01, δ=0.99\delta= 0.99, and λi=103\lambda_i = 10^3 for i=1,2,3i = 1, 2, 3. As we can see from Fig. 3a and Fig. 3b, the state trajectory {xk}\{x_k\} generated by the optimal control problem (7) successfully identifies the two solutions of WB5, i.e., x3x_3 and x4x_4 stay at the global optimal solution S1S_1, while x1x_1 and x2x_2 stay at the local solution S2S_2.

Tolerance of the Lagrangian at $x_k$

Figure 3a. ∥∇xL(x)∥\lVert\nabla_x L(x)\rVert along the state trajectory {xk}\{x_k\}.

Objective value at $x_k$

Figure 3b. Values of FobjF_{\mathrm{obj}} along the state trajectory {xk}\{x_k\}.

As shown in Table II, MIPS finds only one solution in WB5 by running with different initial values, and it fails to converge in WB2 and WB3. Meanwhile, fmincon fails to converge for some initial values z0z_0 in WB2, but still identifies two solutions overall. In contrast, it finds only one solution in WB3 and WB5. Our method not only successfully converges stably in all three test systems but also identifies all solutions in a single run. This highlights the effectiveness of the proposed method.

WB2WB3WB5
MIPSfminconMIPSfminconMIPSfmincon
y00y_0^0N/AS2S_2N/AS1S_1S2S_2S2S_2
y01y_0^1N/AS2S_2N/AS1S_1S2S_2S2S_2
y02y_0^2N/AS1S_1N/AS1S_1S2S_2S2S_2
y03y_0^3N/AN/AN/AS1S_1N/AS2S_2
y04y_0^4N/AN/AN/AS1S_1S2S_2S2S_2
y05y_0^5N/AS1S_1----

Table 2. Solutions Obtained by Different Solvers in WB2, WB3, and WB5 Under Different Initial Values. Note: ‘N/A’ indicates the solver failed to converge.

Conclusion

This paper has proposed an optimization method for finding multiple solutions in the OPF problem. Specifically, the original constrained OPF problem has been transformed into an optimal control problem by applying the penalty method, introducing Lagrange multipliers, and leveraging the OCP. The resulting problem has then been solved efficiently using a numerical algorithm with superlinear convergence rate. To validate the effectiveness of the proposed method, simulations were conducted on several representative test systems. The results have shown that the method successfully finds all solutions and escapes local solutions, demonstrating the potential of the OCP for handling non-convex problems.

References

  1. [1]J. Carpentier, “Contribution à l’étude du dispatching economique,” Bull. Soc. Fr. Elec. Ser., vol. 3, p. 431, 1962.
  2. [2]H. W. Dommel and W. F. Tinney, “Optimal power flow solutions,” IEEE Trans. Power App. Syst., no. 10, pp. 1866–1876, Oct. 1968.DOI
  3. [3]M. S. Kumari and S. Maheswarapu, “Enhanced genetic algorithm based computation technique for multi-objective optimal power flow solution,” Int. J. Electr. Power Energy Syst., vol. 32, no. 6, pp. 736–742, Jul. 2010.DOI
  4. [4]A. G. Bakirtzis, P. N. Biskas, C. E. Zoumas, and V. Petridis, “Optimal power flow by enhanced genetic algorithm,” IEEE Trans. Power Syst., vol. 17, no. 2, pp. 229–236, May 2002.
  5. [5]W. A. Buksh, A. Grothey, K. McKinnon, and P. A. Trodden, “Local solutions of the optimal power flow problem,” IEEE Trans. Power Syst., vol. 28, no. 4, pp. 4780–4788, Nov. 2013.
  6. [6]I. A. Hiskens and R. J. Davy, “Exploring the power flow solution space boundary,” IEEE Trans. Power Syst., vol. 16, no. 3, pp. 389–395, Aug. 2001.
  7. [7]Z.-Y. Wang and H.-D. Chiang, “A necessary and sufficient condition for computed OPF solutions to be locally optimal,” IEEE Trans. Power Syst., vol. 38, no. 6, pp. 5491–5500, Nov. 2023.DOI
  8. [8]A. H. El-Abiad and F. J. Jaimes, “A method for optimum scheduling of power and voltage magnitude,” IEEE Trans. Power App. Syst., no. 4, pp. 413–422, Apr. 1969.
  9. [9]Z. Qiu, G. Deconinck, and R. Belmans, “A literature survey of optimal power flow problems in the electricity market context,” in Proc. IEEE/PES Power Syst. Conf. Expo., 2009, pp. 1–6.DOI
  10. [10]M. Osman, M. A. Abo-Sinna, and A. Mousa, “A solution to the optimal power flow using genetic algorithm,” Appl. Math. Comput., vol. 155, no. 2, pp. 391–405, Aug. 2004.DOI
  11. [11]J. A. Moham, M. El-Hawayary, and R. Adapa, “A review of selected optimal power flow literature to 1993. ii. newton, linear programming and interior point methods,” IEEE Trans. Power Syst., vol. 14, no. 1, pp. 105–111, Feb. 1999.
  12. [12]A. Forsgren, P. E. Gill, and M. H. Wright, “Interior methods for nonlinear optimization,” SIAM Rev., vol. 44, no. 4, pp. 525–597, Jan. 2002.DOI
  13. [13]F. Capitanescu, M. Glavic, D. Ernst, and L. Wehenkel, “Interior-point based algorithms for the solution of optimal power flow problems,” Electr. Power Syst. Res., vol. 77, no. 5-6, pp. 508–517, Apr. 2007.DOI
  14. [14]M. Ebeed, S. Kamel, and F. Jurado, “Optimal power flow using recent optimization techniques,” in Classical and Recent Aspects of Power System Optimization. Amsterdam, The Netherlands: Elsevier, 2018, pp. 157–183.DOI
  15. [15]M. A. Abido, “Optimal power flow using particle swarm optimization,” Int. J. Electr. Power Energy Syst., vol. 24, no. 7, pp. 563–571, Oct. 2002.DOI
  16. [16]S. Sojoudi and J. Lavaei, “Physics of power networks makes hard optimization problems easy to solve,” in Proc. IEEE Power Energy Soc. Gen. Meet., 2012, pp. 1–8.DOI
  17. [17]Z. Michalewicz and M. Schoenauer, “Evolutionary algorithms for constrained parameter optimization problems,” Evol. Comput., vol. 4, no. 1, pp. 1–32, Mar. 1996.DOI
  18. [18]D. Devaraj and B. Yegnanarayana, “Genetic-algorithm-based optimal power flow for security enhancement,” IEEE Proc., Gener. Transm. Distrib., vol. 152, no. 6, pp. 899–905, Nov. 2005.DOI
  19. [19]M. Younes, M. Rahli, and L. Abdelhakem-Koridak, “Optimal power flow based on hybrid genetic algorithm,” J. Inf. Sci. Eng., vol. 23, no. 6, Nov. 2007.DOI
  20. [20]L. L. Lai, J. Ma, R. Yokoyama, and M. Zhao, “Improved genetic algorithms for optimal power flow under both normal and contingent operation states,” Int. J. Electr. Power Energy Syst., vol. 19, no. 5, pp. 287–292, Jun. 1997.DOI
  21. [21]T. Nummonda and U. Annakkage, “Optimal power dispatch in multitude electricity market using genetic algorithm,” Electr. Power Syst. Res., vol. 49, no. 3, pp. 211–220, Apr. 1999.
  22. [22]G. Sen and M. Z. Bilgin, “Optimal power flow using kepler optimization algorithm for active power loss analysis in island mode: A case study,” Heliyon, vol. 11, no. 2, 2025.DOI
  23. [23]H. Zhang, H. Wang, Y. Xu, and Z. Guo, “Optimization methods rooted in optimal control,” Sci. China Inf. Sci., vol. 67, no. 12, p. 222208, 2024.DOI
  24. [24]H. Wang, Y. Xu, Z. Guo, and H. Zhang, “Optimization algorithms with superlinear convergence rate,” IEEE Trans. Autom. Control, vol. 71, no. 1, pp. 684–691, 2026.DOI
  25. [25]C. Lv, X. Yin, H. Li, and H. Zhang, “Optimal control of nonlinear systems using accelerated gradient and hessian computation,” IEEE Trans. Ind. Electron., vol. 73, no. 3, pp. 4645–4653, 2026.
  26. [26]K. Abaci and V. Yamacli, “Differential search algorithm for solving multi-objective optimal power flow problem,” Int. J. Electr. Power Energy Syst., vol. 79, pp. 1–10, Jul. 2016.DOI
  27. [27]B. Mahdad and K. Srairi, “Blackout risk prevention in a smart grid based flexible optimal strategy using grey wolf-pattern search algorithms,” Energy Convers. Manage., vol. 98, pp. 411–429, Jul. 2015.DOI
  28. [28]J.-H. Zhu, J.-S. Wang, X.-Y. Zhang, H.-M. Song, and Z.-H. Zhang, “Mathematical distribution coyote optimization algorithm with crossover operator to solve optimal power flow problem of power system,” Alex. Eng. J., vol. 69, pp. 585–612, Apr. 2023.DOI
  29. [29]D. G. Luenberger, Y. Ye et al., Linear and Nonlinear Programming. New York, NY: Springer, 1984.
  30. [30]D. S. Naidu, Optimal Control Systems. Boca Raton, FL: CRC Press, 2018.

Paper details

vlnc:W_4au6ng32
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 1

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents