Source-linked AI summary
Data-driven Optimal Power Flow: A Physics-Informed Machine Learning Approach
Xingyu Lei, Zhifang Yang, Juan Yu, Junbo Zhao, Qian Gao, Hongxin Yu
TL;DR
OPF’s nonlinear, nonconvex formulation makes computation burdensome, and SELM’s limited learning ability requires adjustment to learn OPF features accurately. The paper proposes a physics-informed, three-stage SELM framework with active-constraint pre-classification; benchmark results report best accuracy among the compared cases and easy extension with few parameter adjustments.
Problem
OPF’s nonlinear, nonconvex formulation creates computational burden, while SELM’s limited learning ability makes accurate learning of OPF features challenging.
Source: §I · p. 1Method
A physics-informed, three-stage SELM framework decomposes OPF features and uses active-constraint-based sample pre-classification to reduce learning complexity.
Source: §I · p. 2Results
The proposed approach achieves the best accuracy among the reported OPF cases on IEEE and Polish benchmark systems.
Source: C. Comparison Results with Existing… · p. 8, §V · p. 8Takeaways & Limitations
With little hyperparameter adjustment, the approach can extend to different systems while retaining similar accuracy.
Source: C. Comparison Results with Existing… · p. 8
Abstract
from arXiv · showhide
This paper proposes a data-driven approach for optimal power flow (OPF) based on the stacked extreme learning machine (SELM) framework. SELM has a fast training speed and does not require the time-consuming parameter tuning process compared with the deep learning algorithms. However, the direct application of SELM for OPF is not tractable due to the complicated relationship between the system operating status and the OPF solutions. To this end, a data-driven OPF regression framework is developed that decomposes the OPF model features into three stages. This not only reduces the learning complexity but also helps correct the learning bias. A sample pre-classification strategy based on active constraint identification is also developed to achieve enhanced feature attractions. Numerical results carried out on IEEE and Polish benchmark systems demonstrate that the proposed method outperforms other alternatives. It is also shown that the proposed method can be easily extended to address different test systems by adjusting only a few hyperparameters.
I. INTRODUCTION
OPF’s nonlinearity and nonconvexity make computation burdensome, while direct SELM learning is challenged by the complex mapping from system status to OPF solutions. The paper proposes a physics-informed, three-stage SELM framework and active-constraint pre-classification to address that complexity.
I. INTRODUCTION
- OPF’s nonlinear, nonconvex model creates a computational burden, especially for real-time operation and analyses requiring repeated solutions.Model-based approaches may require numerous iterations, while simplifications can sacrifice accuracy.Source: §I · p. 1
- SELM randomly generates hidden-layer input weights and analytically determines output weights, improving training speed and reducing hyperparameter tuning relative to traditional neural networks.Its limited learning ability still requires further adjustment to learn OPF features accurately.Source: §I · p. 1
- The approach replaces time-consuming OPF iterations with direct SELM mapping and does not require system topologies or parameters.The resulting solution may provide guidance to accelerate model-based OPF computing.Source: §I · p. 2
- The proposed physics-informed SELM framework decomposes OPF features into three stages to reduce learning complexity and correct learning bias.A reinforcement mode in the hidden layer is used to improve each stage’s learning ability.Source: §I · p. 2
- Pre-classification based on active-constraint identification separates samples with different OPF features, aiming to extract features more effectively and reduce learning complexity.The rationale is that different active-constraint combinations yield different features under multi-parametric programming theory.Source: §I · p. 2
II. A FRAMEWORK OF DATA-DRIVEN OPF BASED ON SELM
The framework uses SELM to learn OPF model features, drawing on stacked ELMs whose layer outputs are reduced and passed forward. ELM analytically computes output weights from hidden-layer outputs and training targets, with regularization balancing training error and weight size.
II. A FRAMEWORK OF DATA-DRIVEN OPF BASED ON SELM
- The section introduces a SELM-based data-driven OPF regression framework and a network designed to learn OPF model features.SELM is built by stacking smaller ELM networks and passing input information from layer to layer.Source: §II · p. 2, A. A Brief Introduction of SELM · p. 2
A. A Brief Introduction of SELM
- ELM forms hidden-layer outputs using randomly generated input weights and biases, then calculates output weights analytically from the hidden outputs and target matrix.The target matrix gathers the target vectors for all samples.Source: A. A Brief Introduction of SELM · p. 2
- The output-weight optimization uses L2 regularization, with penalty factor C trading off training error against the norm of the output weights.The output weights are obtained by solving the optimization’s first-order condition.Source: A. A Brief Introduction of SELM · p. 2
- SELM uses PCA to reduce hidden-layer output dimensions, retaining the top l eigenvectors so later iterations need to generate only L − l new hidden neurons.The retained neurons are used with newly generated neurons to form the next layer’s hidden output matrix.Source: A. A Brief Introduction of SELM · p. 2
B. Proposed Learning Framework of Data-Driven OPF
The framework splits data-driven OPF learning into three stages to reduce learning difficulty and correct errors. Each stage targets distinct OPF features, with later stages refining earlier outputs.
B. Proposed Learning Framework of Data-Driven OPF
- Stage 1 maps active and reactive power inputs to active and reactive branch flows, using historical power-flow data to decompose branch flow from the complicated OPF model for SELM learning.Source: B. Proposed Learning Framework of… · p. 3
- Stage 2 maps branch flows to voltage magnitudes and angles, while Stage 3 uses the demands, branch flows, and voltages to learn generator controls and objective value.The passage describes Stage 2 as covering physical power-flow information, including line parameters and physical laws; Stage 3 serves as error correction.Source: B. Proposed Learning Framework of… · p. 4
- The framework decomposes OPF learning into three stages, using later stages as error-correction steps to reduce learning difficulty.Although an individual stage may lack sufficient accuracy, the passage says learning error gradually reduces across the stages until accuracy requirements are met.Source: B. Proposed Learning Framework of… · pp. 3–4
C. Architecture Design of SELM Network
The three-stage SELM architecture separates learning targets to correct errors and adds a reinforcement mode in the hidden layer to improve SELM's learning ability.
C. Architecture Design of SELM Network
- The architecture separates the three stages' learning targets to correct errors in learning the OPF model.Its hidden layer uses a reinforcement mode intended to improve SELM's learning ability.Source: C. Architecture Design of SELM Network · p. 4
- The reinforcement process adds multiple supervised layers to the SELM regression network.Stage 1 is used as the example; a single SELM output may deviate from the real branch-flow value, and reinforcement is used to decrease that error.Source: C. Architecture Design of SELM Network · p. 4
- The reinforcement layer count is a design trade-off between learning performance and network complexity.The proposed architecture is described as reducing learning complexity while strengthening SELM through reinforcement.Source: C. Architecture Design of SELM Network · p. 4
III. SAMPLE PRE-CLASSIFICATION STRATEGY
The section introduces OPF-model-based sample pre-classification as a strategy to improve learning performance.
III. SAMPLE PRE-CLASSIFICATION STRATEGY
- The section proposes a sample pre-classification method based on the OPF model.The method is intended to achieve better learning performance.Source: §III · p. 4
- Sample pre-classification is the method introduced in this section to improve learning performance.Source: §III · p. 4
- The proposed pre-classification strategy is grounded in the OPF model.Source: §III · p. 4
A. Basic Idea of the Sample Pre-Classification
OPF input-output mappings are divided into critical regions determined by active constraints. Within a region, samples share those constraints, simplifying the nonlinear relationship and its KKT representation.
A. Basic Idea of the Sample Pre-Classification
- A critical region groups inputs whose corresponding OPF outputs share active constraints, yielding a fixed nonlinear input-output relationship within that region.Grouping samples by active constraints is intended to mitigate the complex mapping from random power demand to optimal solutions.Source: A. Basic Idea of the Sample… · p. 4
- The OPF model is expressed as a nonlinear program with an objective, power-flow equalities, and inequality constraints.Its KKT conditions include stationarity, multiplier sign, complementary slackness, and primal feasibility conditions.Source: A. Basic Idea of the Sample… · pp. 4–5
- For samples sharing active constraints, the KKT condition is modified, reducing the problem in a specific region to nonlinear equations whose sensitivity can be computed with a Jacobian.The passage states this formulation relieves SELM learning difficulty.Source: A. Basic Idea of the Sample… · p. 5
B. Proposed Framework
The pre-classification method addresses the difficulty of identifying test-sample active constraints amid many possible combinations. It clusters training samples by similar active constraints, then uses a SELM classifier to assign test samples to classes.
B. Proposed Framework
- Because active-constraint combinations can be numerous, the method forms regions from only part of the active constraints and clusters samples with similar constraints.The approach is intended to avoid classification precision problems and regions unsuitable for SELM regression.Source: B. Proposed Framework · p. 5
- Training samples are grouped into m classes, a SELM classifier is trained on power-demand inputs and class labels, and that classifier assigns test samples to classes.The training groups are identified using the OPF model; the resulting clusters retain the same constraints and variables within each cluster.Source: B. Proposed Framework · p. 5
- Clustering samples with similar active constraints further improves learning performance by reducing the complexity of the OPF regression mapping.The constraints and variables remain the same for samples within a specific cluster.Source: B. Proposed Framework · p. 5
C. Flowchart of The Proposed Approach
The proposed workflow collects OPF training data, pre-classifies samples, and trains SELM networks by class. New power demands are then passed to the trained network to calculate OPF solutions.
C. Flowchart of The Proposed Approach
- Training samples are collected using Monte Carlo simulation and a traditional interior-point algorithm, with system power demand as input and OPF solutions as output.The listed outputs include generator powers, voltages, angles, and the objective value.Source: C. Flowchart of The Proposed Approach · pp. 5–6
- Training samples are pre-classified, then corresponding SELM networks are trained and tested using samples assigned to each class.The classifier supplies the class labels used in the pre-classification step.Source: C. Flowchart of The Proposed Approach · p. 6
- For data-driven calculation, new power demands are input to the trained SELM network to calculate OPF solutions with high precision and fast computation.Source: C. Flowchart of The Proposed Approach · p. 6
IV. CASE STUDY
The case study evaluates the method on IEEE and Polish systems with renewable generation. It defines an accuracy index based on whether prediction error falls below a variable-specific threshold.
A. Case Setting
- Tests cover IEEE 39-, 57-, and 118-bus systems and the Polish 2383-bus system, with wind and photovoltaic generation connected at different buses.The renewable penetration and load fluctuation rates are listed in Table I.Source: A. Case Setting · p. 6
- Accuracy index p is the probability that learning error is below threshold thr.The predicted and actual values define the error being assessed.Source: A. Case Setting · p. 6
- Thresholds are 0.001 p.u. for voltage magnitude and 0.5° for voltage angle; other thresholds are relative to training-data averages.The thresholds are 1% of the average for PF, QF, PG, and QG, and one-thousandth of the average for objective value F.Source: A. Case Setting · p. 6
B. Evaluation of the Proposed Approach
The three-stage SELM regression framework improves learning across OPF variables; reinforcement and sample pre-classification further enhance learning, while staged corrections target different output errors.
B. Evaluation of the Proposed Approach
- Three-stage regression decomposition relieves the complexity that makes direct original SELM learning intractable and significantly improves learning performance across variables.Table III presents the proposed-method improvement evaluation.Source: B. Evaluation of the Proposed Approach · pp. 6–7, Table III · p. 6
- Reinforcement mode enhances learning performance using two supervised layers to trade off accuracy and network complexity; additional layers offer negligible accuracy gains.Source: B. Evaluation of the Proposed Approach · p. 7
3) Comparing M6 with M5, the learning accuracy of each variable is further improved because the complicated mapping relationship between the input and output is mitigated by the proposed sample pre-classification strategy.
- In the IEEE 39-bus system, voltage magnitudes are compared across Stages 1 and 2 against actual values, and active-power output errors across Stages 2 and 3.Source: 3) Comparing M6 with M5, the learning… · p. 7, Figure 7 · p. 7
- Sample pre-classification mitigates the complicated input-output mapping and further improves learning accuracy across variables.Source: 3) Comparing M6 with M5, the learning… · p. 7
- Stage 2 corrects voltage-magnitude error, while Stage 3 mainly corrects control-variable and objective-function errors.Source: 3) Comparing M6 with M5, the learning… · p. 7
C. Comparison Results with Existing Methods
Across the evaluated OPF cases, the proposed approach achieves the best accuracy, while testing is faster than iterative M5 and the method is easier to extend than deep-learning alternatives. The conclusion identifies realistic-system field-data testing as future work.
C. Comparison Results with Existing Methods
- The proposed approach achieves the best accuracy among the evaluated OPF cases.Source: C. Comparison Results with Existing… · p. 8, §V · p. 8
- Testing for NNs M4, M6, and M7 is acceptable and much faster than the time-consuming iterations in M5.Source: C. Comparison Results with Existing… · p. 8
- Deep-learning methods have higher training costs and require extensive hyperparameter tuning; the proposed approach extends to different systems with little adjustment while retaining similar accuracy.Source: C. Comparison Results with Existing… · p. 8
V. CONCLUSION
- The conclusion reports that direct deep learning for OPF may be intractable and leaves testing on realistic systems with field data for future work.Source: §V · p. 8