PaperlayerView exact revision on arXiv

Source-linked AI summary

Physics-Guided Deep Neural Networks for Power Flow Analysis

Xinyue Hu, Haoji Hu, Saurabh Verma, Zhi-Li Zhang

arXiv:2002.00097v2cs.LGeess.SPstat.ML

TL;DR

Power-flow analysis needs alternatives to numerical solvers when system models are inaccurate or unavailable, while existing data-driven methods have limited performance and generalizability. The paper develops physics-guided neural networks with an auxiliary power-flow-model reconstruction task, and reports higher accuracy, better generalizability, and physically interpretable weight structures. The methods still require adaptation as power grids evolve.

  • Problem

    Dynamic and uncertain power systems can make power-flow models inaccurate or unavailable, while existing data-driven solvers suffer from poor performance and generalizability.

  • Method

    Physics-guided neural networks jointly predict bus voltages and reconstruct power-flow behavior, using physical knowledge to regularize the solver.

  • Results

    4.57 × 10−3 average voltage error, about 4 times smaller than MLPNN, while the proposed solvers also perform better on test samples deviating from training data.

  • Takeaways & Limitations

    The proposed solvers enable power-flow analysis when exact system parameters and control logic are difficult to obtain, while their weight matrices reflect power-system structure.

  • Takeaways & Limitations

    Well-trained neural power-flow solvers need adaptation to future system changes as power grids vary and evolve over time.

Abstract

from arXiv · show

Solving power flow (PF) equations is the basis of power flow analysis, which is important in determining the best operation of existing systems, performing security analysis, etc. However, PF equations can be out-of-date or even unavailable due to system dynamics and uncertainties, making traditional numerical approaches infeasible. To address these concerns, researchers have proposed data-driven approaches to solve the PF problem by learning the mapping rules from historical system operation data. Nevertheless, prior data-driven approaches suffer from poor performance and generalizability, due to overly simplified assumptions of the PF problem or ignorance of physical laws governing power systems. In this paper, we propose a physics-guided neural network to solve the PF problem, with an auxiliary task to rebuild the PF model. By encoding different granularity of Kirchhoff's laws and system topology into the rebuilt PF model, our neural-network based PF solver is regularized by the auxiliary task and constrained by the physical laws. The simulation results show that our physics-guided neural network methods achieve better performance and generalizability compared to existing unconstrained data-driven approaches. Furthermore, we demonstrate that the weight matrices of our physics-guided neural networks embody power system physics by showing their similarities with the bus admittance matrices.

I. INTRODUCTION

Modern power-system dynamics and incomplete models make conventional power-flow analysis unreliable, motivating physics-guided neural solvers that combine voltage prediction with physical reconstruction. On IEEE test systems, the proposed methods improve accuracy, generalizability, and physical interpretability over unconstrained data-driven approaches.

  • Motivation: Dynamic and uncertain power systems can make conventional power-flow analysis ineffective because line parameters and distributed-energy-resource controls may be inaccurate, missing, or difficult to model.These modeling deficiencies create safety threats and complicate distributed-energy-resource adoption.
  • Approach: The proposed architecture jointly minimizes voltage-prediction errors and power-reconstruction errors through an auxiliary task that rebuilds the power-flow model.This multi-task design regularizes the neural solver with physical constraints.
  • Results: On IEEE 57- and 118-bus systems, the proposed solvers achieve an order of magnitude higher accuracy and better performance on test samples deviating from training data.The authors attribute these results to neural networks guided toward physically consistent solutions.
  • Approach: The paper proposes physics-guided neural networks for power-flow analysis when system parameters and power-flow models are inaccurate or unavailable.The approach uses generic physical knowledge rather than impractical assumptions about complete system models.
  • Results: The proposed physics-guided methods consistently achieve better accuracy and generalizability than existing data-driven power-flow solvers.The paper quantitatively evaluates and compares the data-driven alternatives.
  • Interpretation: The regularizer variants produce weight matrices with patterns similar to bus admittance matrices, indicating that the auxiliary task can support physical interpretation and model rebuilding.The paper also discusses corresponding use cases for the encoded physical knowledge.

II. POWER FLOW ANALYSIS PROBLEM FORMULATION

Power-flow analysis can be formulated through nonlinear Kirchhoff-law equations in polar coordinates or polynomial equations in rectangular coordinates. Bus types determine which powers and voltages are specified, while numerical solvers iteratively determine the remaining voltage states.

  • Coordinate formulations: Power-flow equations can be formulated in two coordinate systems: polar coordinates and rectangular coordinates.The two formulations represent the same analysis problem using different voltage variables.
  • Polar formulation: In polar coordinates, active and reactive power injections are expressed through nonlinear trigonometric equations representing Kirchhoff’s laws.The equations use voltage magnitudes, phase-angle differences, and bus-admittance-matrix elements.
  • Polar formulation: In the polar formulation, p_i and q_i denote real and reactive injections, G_ik and B_ik are admittance-matrix components, V_i is voltage magnitude, and N is bus count.θ_ik denotes the phase-angle difference between buses i and k.
  • Bus types: PQ buses specify real and reactive power injections, while voltage magnitudes and angles remain unknown.These buses represent load buses in the basic power-flow formulation.
  • Bus types: PV buses specify real power injections and voltage magnitudes, while reactive power injections and voltage angles remain unknown.These buses represent generation buses.
  • Bus types: The Vθ bus is the reference bus, with specified voltage magnitude and angle but unknown real and reactive power injections.It anchors the voltage reference in the system formulation.
  • Numerical solution: The resulting power-flow equations have equal numbers of equations and unknowns, and conventional numerical methods iteratively solve them to determine all bus voltages.Known injections into PQ and PV buses enter the equation system.
  • Rectangular formulation: In rectangular coordinates, μ_i = V_i cosθ_i and ω_i = V_i sinθ_i represent the real and imaginary voltage-phasor components, making power injections second-order polynomials.The voltage vector is represented by [μ, ω].

B. From Model-based Methods to Data-driven Methods

Data-driven methods address missing or inaccurate power-system models, but parameter rediscovery and numerical solution can accumulate errors and suffer convergence problems. These limitations motivate power-flow solvers that learn directly under more flexible conditions.

  • Model limitations: Numerical power-flow solvers require accurate equations, yet line parameters and distributed-energy-resource models may be missing or unreliable.This is especially challenging in distribution grids and for independently owned distributed-energy resources.
  • Data-driven reconstruction: Data-driven model-rebuilding methods attempt to recover system parameters from historical measurements so conventional numerical solvers can be used again.They calibrate or rebuild power-flow models rather than directly predicting solutions.
  • Error sensitivity: Parameter rediscovery and numerical approximation accumulate errors, while small parameter variations can substantially disturb operating settings and cause errors to blow up.Numerical solutions are sensitive to differences in system parameters.
  • Convergence challenges: Distributed-energy-resource integration can create convergence problems because power-electronic interfaces introduce degrees of freedom into device models.The resulting problem may contain unequal numbers of equations and unknowns, requiring a new power-flow solver.

C. Data-Driven Power Flow Solvers

Data-driven PF solvers learn mappings from specified power-flow inputs to bus voltages, but their generalization is limited when physical models are inaccurate or unavailable. The proposed solver addresses this by jointly predicting voltages and reconstructing power injections, using auxiliary-task regularization and physical structure.

  • Data-driven PF solvers learn mappings from specified load, generator, and reference-bus variables to bus voltages.
  • Avoiding overfitting and improving generalizability are the two main challenges for data-driven PF solvers.
  • Accurate system parameters cannot be assumed in all settings, leaving incorporation of generic physical laws into data-driven methods as an open problem.
  • The proposed architecture combines an encoder for PF solving with a decoder for rebuilding PF models as an auxiliary task.
  • Its loss jointly penalizes voltage-prediction errors and power-injection reconstruction errors, thereby regularizing the under-constrained supervised learning problem.

B. Encode Physical Knowledge in Regularizer

The decoder variants encode progressively more power-system structure while rebuilding mappings from bus voltages to power injections. Bilinear structure reflects Kirchhoff’s laws, and topology pruning imposes the sparsity pattern of the bus admittance matrix.

  • The decoder designs rebuild PF models using different levels of prior physical knowledge.
  • MLPNN Regularizer: MLP decoders separately learn active- and reactive-power mappings from encoder-estimated bus voltages.
  • BNN Regularizer: BNN decoders represent power injections as bilinear functions of rectangular voltage coordinates, following the structure indicated by Kirchhoff’s laws.
  • TPBNN Regularizer: Topology-pruned BNNs use the adjacency matrix as a hard attention mask so nonneighboring buses retain zero corresponding parameters and gradients.

A. Experiment Setup

The experiments evaluate the proposed physics-guided networks on IEEE 57- and 118-bus systems using load measurements simulated from GEFCom 2012 data and MATPOWER-generated voltage responses.

  • Experiments use the standard IEEE 57- and 118-bus systems with real-world load data from GEFCom 2012.
  • MATPOWER generates the associated voltage magnitudes and phase angles for each bus.
  • The default split assigns 60% of samples to training, 10% to validation, and 30% to testing.

B. Power Flow Solution Results

The evaluation compares the proposed PF solvers with existing data-driven methods under matched inputs and outputs, excluding RBFNN because it has similar capability and performance to MLPNN.

  • The proposed solvers are compared with LR, SVR, and MLPNN using the same power-flow inputs and bus-voltage outputs.
  • RBFNN is omitted because it has similar universal-function-approximation capability and performance to MLPNN.

1) Basic Results:

Physics-guided PF solvers achieve lower voltage errors than unconstrained data-driven methods and generalize better to interpolation and extrapolation settings. Their advantage is also reflected in comparisons involving noisy training data and learned physical structure.

  • Basic accuracy: The physics-guided neural networks achieve an average voltage RMSE of 4.57 × 10−3, about four times smaller than MLPNN.The comparison uses LR, SVR, and MLPNN as unconstrained data-driven baselines.
  • Interpolation and extrapolation: Figure 4 reports RMSE distributions for interpolation and extrapolation results, with validation errors shown in shaded areas and testing errors outside them.The distributions are evaluated across the IEEE 57 and 118 cases shown in the associated panels.
  • Interpolation and extrapolation: The average differences between testing and validation errors are 0.0185 for SVR, 0.0145 for MLPNN, and 0.00712 for physics-guided NNs.The lower difference indicates stronger interpolation performance in the reported comparison.
  • Interpolation and extrapolation: Physics-guided PF solvers have interpolation capabilities about 2 times better than SVR and MLPNN methods.The comparison concerns interpolation on inputs divided into distinct value portions.
  • Interpolation and extrapolation: Physics-guided PF solvers have extrapolation capabilities at least 2 times better than other unconstrained PF solvers.The result concerns testing samples that deviate from the training-data range.

3) Robustness to Data Outliers:

The study evaluates robustness to corrupted training data and examines whether decoder parameters reflect power-system structure. Physics-guided networks remain accurate with outliers, while their learned matrices capture features of bus admittance matrices and topology.

  • Robustness to Data Outliers: Physics-guided neural networks continuously perform best and remain robust even when 10% of training data are outliers.Linear regression degrades rapidly as outlier levels increase, whereas SVR and neural-network methods are relatively stable.
  • Physical consistency: The decoder weight matrices of BNN and TPBNN are related to bus admittance matrices G and B.The comparison is made for the IEEE 118-bus system.
  • Physical consistency: Both BNN and TPBNN learn the sparsity of bus admittance matrices, while BNN roughly captures power-system topology.TPBNN represents diagonal nonzero parameters more accurately than BNN.
  • Physical consistency: TPBNN provides a very good approximation of bus admittance matrices when topology prior knowledge is available.The paper links this result to the topology-augmented architecture and the 2nd-polynomial structure.
  • Physical consistency: The decoder RMSEs for LR, SVR, MLPNN, BNN, and TPBNN are 8.84 × 10−2, 5.58 × 10−2, 4.30 × 10−2, 1.09 × 10−2, and 1.95 × 10−3, respectively.The values are average power-injection modeling errors across the two test systems.

V. DISCUSSION

The proposed physics-guided solvers maintain comparable calculation accuracy while targeting different power-system settings based on available physical knowledge. Their applicability is bounded by topology requirements and the need to adapt models as grids evolve.

  • The three physics-guided PF solvers achieve comparable calculation accuracy but suit different situations according to available prior physical knowledge.
  • MLP+MLP is a general method suited to grids with unknown control policies caused by diverse active controllers and increasing DER penetration.
  • MLP+BNN and MLP+TPBNN suit grids whose control policies indicate a quadratic PF model, enabling parameter approximation and PF-equation rebuilding.
  • TPBNN requires accurate topology information, making it a restricted method despite using quadratic structural information.
  • Physics-guided neural-network solvers must be retrained or adapted as power grids vary and evolve over time.
  • The conclusion reports higher accuracy and better physical consistency than existing unconstrained data-driven methods, with BNN and TPBNN promising for rebuilding PF equations.
Loading 2002.00097v2…