PaperlayerView exact revision on arXiv

Source-linked AI summary

DeepOPF: A Feasibility-Optimized Deep Neural Network Approach for AC Optimal Power Flow Problems

Xiang Pan, Minghua Chen, Tianyu Zhao, Steven H. Low

arXiv:2007.01002v6eess.SYcs.LG

TL;DR

Ensuring DNN-generated AC-OPF solutions satisfy nonconvex power-flow and operational constraints is non-trivial. DeepOPF reconstructs variables through power flow, penalizes inequality violations, and uses zero-order gradient estimation; simulations report two-orders-of-magnitude speedups with minor cost differences.

  • Problem

    Ensuring generated solutions satisfy nonconvex power-flow balance equations and operation limits simultaneously is non-trivial.

  • Method

    DeepOPF predicts independent variables, reconstructs remaining variables by solving AC power flow equations, and uses penalty training with zero-order optimization for inequality constraints.

  • Results

    DeepOPF speeds up computing time by two-orders of magnitude compared with conventional optimization-based solvers, with minor cost difference.

  • Takeaways & Limitations

    The penalty approach improves DNN solution feasibility, while zero-order gradient training achieves similar or better feasibility rates than implicit-function-theorem training.

  • Takeaways & Limitations

    The formulation sets P-Q bus generator output and operating bounds to zero and uses bus 0 as the slack bus.

Abstract

from arXiv · show

High percentage penetrations of renewable energy generations introduce significant uncertainty into power systems. It requires grid operators to solve alternative current optimal power flow (AC-OPF) problems more frequently for economical and reliable operation in both transmission and distribution grids. In this paper, we develop a Deep Neural Network (DNN) approach, called DeepOPF, for solving AC-OPF problems in a fraction of the time used by conventional solvers. A key difficulty for applying machine learning techniques for solving AC-OPF problems lies in ensuring that the obtained solutions respect the equality and inequality physical and operational constraints. Generalized the 2-stage procedure in [1], [2], DeepOPF first trains a DNN model to predict a set of independent operating variables and then directly compute the remaining dependable ones by solving power flow equations. Such an approach not only preserves the power-flow balance equality constraints but also reduces the number of variables to predict by the DNN, cutting down the number of neurons and training data needed. DeepOPF then employs a penalty approach with a zero-order gradient estimation technique in the training process to preserve the remaining inequality constraints. As another contribution, we drive a condition for tuning the size of the DNN according to the desired approximation accuracy, which measures the DNN generalization capability. It provides theoretical justification for using DNN to solve the AC-OPF problem. Simulation results of IEEE 30/118/300-bus and a synthetic 2000-bus test cases show that DeepOPF speeds up the computing time by up to two orders of magnitude as compared to a state-of-the-art solver, at the expense of $<$0.1% cost difference.

NOMENCLATURE

AC-OPF optimizes system objectives subject to power-flow and operational constraints, while renewable-driven variability increases the need for faster repeated solutions. DeepOPF addresses AC-OPF feasibility and speed using predict-and-reconstruct, penalty training, and zero-order optimization.

  • AC-OPF optimizes objectives such as generation cost subject to power-flow balance and generation, voltage, and branch-flow constraints.
  • Renewable and distributed energy resources introduce intermittent fluctuations, increasing the need for efficient OPF solutions for reliable and economical operation.
  • DeepOPF predicts independent operating variables and reconstructs the remaining variables with AC power-flow equations, guaranteeing power-flow balance equality constraints.
  • A penalty approach targets violations of generation, voltage, and branch-flow limits, while zero-order optimization computes penalty gradients when nonlinear reconstruction prevents explicit gradients.
  • The paper characterizes the load-to-solution mapping and derives a DNN-size condition tied to approximation accuracy, providing theoretical justification for DNN-based AC-OPF.
  • DeepOPF simulations on IEEE 30/118/300-bus and synthetic 2000-bus cases achieved up to two orders of magnitude speedup with less than 0.2% cost difference versus Pypower.

III. AC-OPF PROBLEM

The standard AC-OPF problem minimizes generation cost under nonlinear power-flow and operating constraints. DeepOPF learns a load-to-solution mapping while reconstructing dependent variables through AC power flow to preserve equality feasibility and reduce prediction complexity.

  • The standard bus injection AC-OPF minimizes total active-generation cost subject to power-flow balance, generation, voltage-magnitude, and branch-flow limits.
  • The bus injection and branch flow models are equivalent, but this paper focuses on the bus injection model.
  • AC-OPF is nonconvex and NP-hard, motivating methods that reduce solution time for repeated stochastic operating scenarios.
  • DeepOPF learns the load-to-solution mapping by predicting selected independent variables and reconstructing dependent variables through AC power-flow equations.
  • The predict-and-reconstruct design reduces the mapping dimension and consequently the DNN size and training-data requirements.
  • The method assumes a unique solver-generated load-to-solution mapping, although nonconvex AC-OPF may admit multiple optimal solutions.
  • DNN approximation error can still violate generation, voltage, and branch-flow limits even though power-flow equalities are preserved.

C. Prediction and Reconstruction

DeepOPF predicts bounded independent variables and reconstructs the remaining AC-OPF variables through nonlinear power flow. Its training loss combines prediction error with penalties for reconstructed inequality-constraint violations.

  • For P-V buses and the slack bus, the DNN predicts selected voltage and generation variables, while other variables are reconstructed.
  • The framework replaces the DC power-flow equations of earlier PR2 methods with nonlinear AC power-flow equations.
  • Each predicted variable is mapped from a scaling factor constrained to (0, 1) through its operating bounds, using a Sigmoid output activation.
  • Newton’s method reconstructs dependent variables from predicted independent variables by solving the nonlinear AC power-flow equations.
  • Penalty-based training: The prediction-error loss uses the squared ℓ2 norm between the predicted scaling-factor vector spred and reference vector sref.
  • Penalty-based training: The penalty term captures violations of generation limits, voltage-magnitude constraints, and branch-flow limits for reconstructed variables.
  • Penalty-based training: The overall training objective averages prediction error and reconstructed-variable penalties using positive weighting factors.

F. Zero-order Optimization Technique for Penalty Approach

DeepOPF addresses the difficulty of computing penalty gradients by estimating them with a two-point zero-order method, then uses back-propagation to train the DNN. The estimator requires only two AC power-flow solves and is empirically observed to converge in this setting.

  • The penalty gradient is difficult to compute because reconstructed variables lack an explicit expression in terms of the DNN predictions.Implicit-function differentiation is possible but requires non-trivial Jacobian and singular-value calculations.
  • DeepOPF estimates the penalty gradient with a two-point zero-order optimization technique using random unit-ball directions.The smoothing parameter δ is positive, and the estimator is unbiased when the penalty function is differentiable.
  • Two-point estimation requires solving the AC power-flow equations twice, regardless of the input-space dimension.Uniform-sampling-based numerical estimation would require many power-flow solves and is therefore more computationally expensive.
  • The estimated penalty gradient is combined with the prediction-loss gradient before back-propagation updates the DNN parameters.Algorithm 1 summarizes the gradient computation and parameter-update process.
  • The proposed training method is essentially a stochastic-gradient-descent variant, but convergence conditions involving estimation variance and gradient smoothness remain difficult to verify.The unbiased-gradient condition holds by Proposition 1; the other two conditions are left as future work.

G. Feasibility Test and Recovery

DeepOPF tests inequality feasibility after reconstructing an AC-OPF solution and invokes a conventional solver as a warm start when violations occur. The paper also establishes approximation-theoretic support for learning the load-to-solution mapping when the optimum is unique.

  • G. Feasibility Test and Recovery: DNN approximation error can leave generation, voltage, and branch-flow inequality constraints violated even when power-flow equalities are satisfied.The Predict-and-Reconstruct framework guarantees the equality constraints, but not the operational limits.
  • G. Feasibility Test and Recovery: DeepOPF outputs a solution only after checking inequality feasibility; violations trigger conventional-solver recovery using the DNN solution as a warm start.If recovery fails, the load input is classified as not supportable.
  • G. Feasibility Test and Recovery: More than 99% of DeepOPF solutions are feasible and avoid feasibility recovery in the reported simulations.The recovery step is usually faster than solving AC-OPF directly when the warm-start point is close to optimal.
  • A. Understanding the AC-OPF Load-Solution Mapping: When the optimal solution is unique over a compact load domain, the AC-OPF load-to-solution mapping is continuous and has a Hessian almost everywhere.The exceptional set has Lebesgue measure zero.
  • A. Understanding the AC-OPF Load-Solution Mapping: Under the uniqueness condition, DNNs can approximate the continuous load-to-solution mapping arbitrarily closely as the number of neurons increases.The paper uses this result to justify learning the mapping with a DNN and to guide network-size selection by approximation accuracy.

B. DNN Approximation Error of Load-to-Solution Mapping

The section derives an approximation-error lower bound for DNNs learning AC-OPF load-to-solution mappings and evaluates DeepOPF on standard and synthetic networks.

  • Approximation-error analysis: Theorem 4 analyzes ReLU DNN approximation of a unique scalar load-to-solution mapping over a normalized load domain.The input domain is a multidimensional unit hypercube, and the load distribution has a continuous density bounded below by a positive constant.
  • Approximation-error analysis: The theorem assumes nonzero curvature on a connected subset, and AC-OPF mappings usually have nonzero Hessians, making the assumption practically plausible.The bound’s constant depends on the minimum load-density value and the minimum absolute Hessian-eigenvalue magnitude.
  • Approximation-error analysis: The approximation-error lower bound decreases exponentially with hidden-layer depth L but polynomially with the maximum neurons per layer m.The result supports using deeper architectures for approximating nonlinear AC-OPF mappings.
  • Experiment setup: The experiments use IEEE 30-, 118-, and 300-bus cases plus a synthetic 2000-bus mesh network, with independent and correlated load datasets.Independent loads vary uniformly within ±10% of default values, while correlated profiles use an interpolated California demand curve.
  • Limitations: Larger DNNs can overfit and degrade performance, so the experiments select architecture size through educated guesses and iterative tuning.The minimum training-sample requirement for strong performance remains future work.

B. Performance Evaluation under Test Datasets

DeepOPF improves AC-OPF feasibility and achieves substantial speedups with small optimality loss across test datasets. Zero-order gradient training performs similarly to implicit-function-based training, while penalty weighting trades feasibility against optimality.

  • 80% higher feasibility is achieved with the penalty approach compared with DNN-E.
  • >99% feasibility is achieved before feasibility recovery, demonstrating the penalty approach’s effectiveness.
  • ×123 speedup and less than 0.2% average optimality loss are obtained versus conventional AC-OPF solvers on two datasets.
  • Penalty weighting factors: Larger penalty weight w2 can increase optimality difference because training prioritizes reducing constraint penalties.The compared variants use w1 = 1 with w2 equal to 1 or 0.1.
  • Gradient computation methods: DeepOPF-ZO matches DeepOPF-IF overall and achieves a better feasibility rate in simulations.DeepOPF-ZO estimates the penalty gradient with a zero-order technique, whereas DeepOPF-IF uses the implicit function theorem.
  • DeepOPF develops one model for varying loads but fixed network topology and physical or operational parameters.A contingency such as line outage or transmission switching may therefore require a pre-trained model for the changed network.

APPENDIX A PROOF OF THEOREM 3

The proof establishes continuity of the unique AC-OPF load-to-solution mapping by contradiction. It uses feasibility, optimality, sequence convergence, and continuity of the problem functions to rule out discontinuity.

  • The proof relies on the AC-OPF objective and constraint functions being continuous and on bounded optimal solutions.These assumptions support passage to limits in the feasibility and objective arguments.
  • The proof assumes discontinuity at a load point and constructs a convergent sequence of optimal solutions approaching that point.The contradiction argument uses Xk = f*(Dk) with Dk converging to D0.
  • Continuity of the inequality constraints shows the limiting solution remains feasible at the limiting load.The proof considers cases based on whether strictly feasible points exist and whether constraints are active.
  • The proof compares objective values across convergent subsequences to establish that the limiting feasible solution is optimal.It treats sequences whose objective values are above or no greater than the objective at the limiting point.
  • An additional optimal solution at D0 contradicts uniqueness, so the load-to-solution mapping must be continuous.This contradiction is the proof’s decisive step.

APPENDIX B PROOF OF THEOREM 4

The appendix proof extends a one-dimensional approximation-error lower bound to general nonlinear functions and input distributions.

  • The proof begins with piecewise-linear approximation of quadratic functions on [0, 1] under a density bounded below by a positive constant.It then extends the resulting lower bound to general nonlinear functions and distributions.
Loading 2007.01002v6…