Explicit Ensemble Learning Surrogate for Joint Chance-Constrained Optimal Power Flow
Amir Bahador Javadi, Amin Kargarian · 2025
Read Explicit Ensemble Learning Surrogate for Joint Chance-Constrained Optimal Power Flow by Amir Bahador Javadi, Amin Kargarian (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2511-03515.txt.
Introduction
Amir Bahador Javadi and Amin Kargarian replace a costly risk rule in power grid planning with a fast learned stand-in. Wind and solar uncertainty makes fixed planning unsafe, while joint chance rules that cap total risk are slow to solve. Their ensemble of support vector machines learns from simulated best plans.
Trained classifiers become simple flat constraints inside the planning problem through a standard big-number reformulation. The stand-in traces a faceted outline of the true risk boundary, staying readable and scalable. On a large standard test grid, costs stay near optimal while risk breaches average out negligible. Grid operators gain a way to honor joint risk caps without grinding through the full probabilistic problem.
Catalog record
| CREATOR | Amir Bahador Javadi, Amin Kargarian |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 36 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2511-03515
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Explicit Ensemble Learning Surrogate for Joint Chance-Constrained Optimal Power Flow Amir Bahador Javadi, Amin Kargarian Abstract The increasing penetration of renewable generation introduces uncertainty into power systems, challenging traditional deterministic optimization methods. Chance-constrained optimization offers an approach to balancing cost and risk; however, incorporating joint chance constraints introduces computational challenges. This paper presents an ensemble support vector machine surrogate for joint chance constraint optimal power flow, where multiple linear classifiers are trained on simulated optimal power flow data and embedded as tractable hyperplane constraints via Big–M reformulations. The surrogate yields a polyhedral approximation of probabilistic line flow limits that preserves interpretability and scalability. Numerical experiments on the IEEE 118-bus system show that the proposed method achieves near-optimal costs with a negligible average error of 0.03 % 0.03\% . These results demonstrate the promise of ensemble surrogates as efficient and transparent tools for risk-aware optimization of power systems. \thanksto This work is supported by the National Science Foundation under Grant Number 1944752. I Introduction The growing integration of renewable energy sources (RES), particularly wind and solar generation, has profoundly reshaped modern power system operations. Renewable generation offers substantial environmental and economic benefits. However, its inherent variability and uncertainty introduce significant challenges to system reliability, economic efficiency, and secure operation [ 12 ] . Traditional deterministic optimization methods are often inadequate in this context, as they fail to capture the stochastic nature of RES and demand fluctuations. Chance-constrained programming [ 3 ] has emerged as an approach for explicitly incorporating uncertainty into optimization, allowing for controlled levels of risk. In the context of power systems, chance-constrained optimal power flow (CC-OPF) ensures that operational constraints, such as generator and line flow limits, are satisfied with a prescribed probability [ 1 ] . This enables operators to balance risk and cost rather than relying on worst-case assumptions that may lead to overly conservative and uneconomic dispatch solutions. However, the practical application of CC-OPF remains challenging. Exact reformulations are typically tractable only under assumptions, such as Gaussian uncertainty and linearized system models [ 19 , 6 ] . Joint chance constraints (JCCs), which enforce simultaneous satisfaction of multiple limits, are inherently non-linear and often NP-hard [ 14 ] . Approximate techniques, such as Boole’s inequality or scenario approximation, mitigate complexity but often introduce excessive conservatism or computational burdens [ 8 ] . To overcome these challenges, recent research has focused on surrogate models and machine learning (ML) methods that approximate the feasible region induced by network constraints under uncertainty. By learning from data, surrogates can replace intractable chance constraints with tractable approximations, reducing computational complexity while maintaining acceptable accuracy. Examples include deep quantile regression for estimating line flow distributions and constructing probabilistic constraints [ 4 ] , convex relaxations that provide tractable reformulations of probabilistic constraints [ 19 ] , and classification-based surrogates such as support vector machines (SVMs) [ 18 ] , which separate feasible from infeasible operating points with scalable performance. Embedding SVM-derived constraints into the optimal power flow (OPF) formulation allows operators to enforce network security constraints with high fidelity while retaining computational efficiency, and ensemble strategies further reduce the risk of misclassification [ 9 ] . Although these black-box ML approaches are powerful, they often lack interpretability and fail to reveal explicit structural relationships among system variables. Symbolic regression (SR) has recently gained attention as a promising middle ground between applied mathematics and implicit ML [ 10 ] . Unlike purely black-box models, SR searches a space of mathematical expressions to discover parsimonious closed-form relationships that balance interpretability, generalization, and predictive accuracy [ 17 , 5 ] . This makes SR particularly appealing for power system optimization, where tractability and transparency are as critical as predictive performance. By integrating SR into the surrogate modeling for JCC-OPF, one can derive constraints that are not only computationally efficient but also physically meaningful, bridging the gap between theoretical rigor and practical applicability. JCC–OPF \min\;\eqref{eq:obj} s.t. \eqref{eq:power_balance},\;\eqref{eq:gen_limits},\;\eqref{eq:chance_constraint},\;\eqref{eq:agc_weights},\;\eqref{eq:theta_ref} SAA JCC–OPF \min\;\eqref{eq:obj} s.t. \eqref{eq:power_balance},\;\eqref{eq:gen_limits},\;\eqref{eq:active_power_transmission},\;\eqref{eq:big_m_flow},\;\eqref{eq:violation_budget}\;\eqref{eq:binary_var},\;\eqref{eq:agc_weights},\;\eqref{eq:theta_ref} Approx. ( P1c ) with SAA Proposed Method JCC–OPF \min\;\eqref{eq:obj} s.t. \eqref{eq:power_balance},\;\eqref{eq:gen_limits},\;\eqref{eq:svm},\;\eqref{eq:agc_weights},\;\eqref{eq:theta_ref} Approx. ( P1c ) with proposed method Comparison Fig. 1: Conceptual comparison between the conventional SAA and the proposed surrogate-based formulation for solving the JCC–OPF problem under renewable uncertainty. The proposed method replaces the scenario-based approximation with a data-driven linear surrogate to enhance tractability and interpretability. This paper presents a surrogate approach for JCC-OPF under renewable uncertainty by directly integrating ensemble SVM classifier hyperplane expressions into the optimization problem. In this approach, SVMs yield polyhedral approximations of the feasible region among system variables in interpretable algebraic forms. The resulting surrogate constraints are embedded in the OPF model using Big-M formulations. The proposed methodology is tested on the IEEE 118-bus system under wind and load uncertainty scenarios, with comparative analyses against the conventional scenario approximation approach. The results demonstrate that the surrogate framework achieves a favorable balance between cost optimality, interpretability, and reliability. It represents a promising direction for integrating learning techniques with optimization in renewable-dominated power systems. II Mathematical Framework for JCC-OPF This section presents the mathematical formulation of the JCC-OPF problem with affine automatic generation control (AGC) policies. We first introduce the complete optimization model and then provide a detailed exposition of its components, including uncertainty characterization, operational constraints, and the sample average approximation reformulation using a Big-M methodology. Fig. 1 summarizes the reformulation of the JCC–OPF problem under renewable uncertainty. The original JCC–OPF is decomposed into two approaches: the conventional SAA, which replaces probabilistic constraints with scenario-based deterministic ones, and the proposed surrogate-based formulation, which substitutes these constraints with a data-driven linear approximation. This comparison highlights the transition from sampling-based to learning-augmented representations of uncertainty in power system optimization. II-A Problem Formulation Consider a power network represented by a directed graph G(\mathcal{N},\mathcal{L}) , where \mathcal{N} denotes the set of buses and \mathcal{L} denotes the set of transmission lines. Let \mathcal{G}\subseteq\mathcal{N} represent the subset of buses hosting generation units. The JCC-OPF problem is formulated as [ 13 ] : \displaystyle\min_{\{p_{g},\beta_{g},z_{s}\}}\quad \displaystyle\sum_{g\in\mathcal{G}}\left(C_{2,g}p_{g}^{2}+C_{1,g}p_{g}+C_{0,g}+\mathrm{Var}(\Omega)C_{2,g}\beta_{g}^{2}\right) (1) s.t. \displaystyle\sum_{g\in\mathcal{G}}p_{g}+\sum_{n\in\mathcal{N}}\bar{w}_{n}=\sum_{n\in\mathcal{N}}\bar{d}_{n}, (P1a) \displaystyle\underline{p}_{g}\leq p_{g}+\beta_{g}\Omega\leq\overline{p}_{g},\quad\forall g\in\mathcal{G}, (P1b) \displaystyle\mathbb{P}\Big(|f_{l}(\omega)|\leq\overline{f}_{l},\ \forall l\in\mathcal{L}\Big)\geq 1-\alpha, (P1c) \displaystyle\sum_{g\in\mathcal{G}}\beta_{g}=1,\quad 0\leq\beta_{g}\leq 1,\quad\forall g\in\mathcal{G}, (P1d) \displaystyle\theta_{ref}=0, (P1e) where ( 1 ) defines the expected total system cost, where C_{0,g} , C_{1,g} , and C_{2,g} are the quadratic cost coefficients. ( P1a ) enforces power balance. The demand at bus n\in\mathcal{N} is characterized by d_{n}=\bar{d}_{n}+\omega_{n} , where \bar{d}_{n} represents the nominal demand forecast and \omega_{n} denotes a zero-mean stochastic forecast error with variance \sigma_{n}^{2} . The aggregate system-level power imbalance is defined as \Omega=\sum_{n\in\mathcal{N}}\omega_{n} . ( P1b ) ensures that the limits of the generator capacity are met in all executions. To maintain system balance under uncertainty, each generator uses an affine control policy given by \tilde{p}_{g}=p_{g}+\beta_{g}\Omega , where p_{g} denotes the scheduled base-point generation and \beta_{g}\in[0,1] represents the AGC participation factor of generator g . This linear feedback policy ensures real-time compensation of system imbalances. ( P1c ) represents the joint probabilistic transmission constraint, where \alpha\in(0,1) specifies the acceptable risk level and \overline{f}_{l} denotes the thermal capacity of line l . ( P1d ) enforces AGC participation factor requirements, while ( P1e ) fixes the reference angle to zero. II-B Sample Average Approximation (SAA) Reformulation To address the constraint in ( P1c ), the underlying probabilistic