Fast Power Flow Prediction with Machine Learning & cGAN

Fast Power Flow Prediction Using Machine Learning

M.Sc. Research Overview

Omar Shadafny

M.Sc. Student – Electrical Engineering | Technion

Introduction

Today's energy landscape is evolving rapidly. We focus on fast power-flow prediction in modern grids characterized by:

Modern power grids are changing fundamentally<br><br>Renewable energy integration increases uncertainty<br><br>Fast decision-making is required for stability

Motivation

Traditional power systems were predictable, but modern requirements challenge existing tools.

Renewable sources are intermittent.

Classical solvers are too slow for rapid fluctuations.

Real-time control demands millisecond speed.

Research Problem

The core objective is to accurately predict grid states across the entire network topology.

Predict Active Power (P)

Predict Reactive Power (Q)

For all buses in the grid simultaneously

Limitations of Existing Methods

A trade-off between accuracy and computational speed.

Numerical Solvers (e.g., Newton-Raphson)

Highly accurate but computationally expensive and slow.

Standard Machine Learning

Fast inference but often ignores underlying grid structure and physics.

Proposed Solution

We propose a Conditional Generative Adversarial Network (cGAN) that integrates topology into the learning process.

Conditional GAN model architecture.

Explicitly includes grid topology.

Learns physical relationships between buses.

Model Structure

Input Layers

• Past P and Q values<br>• Bus and line features<br>• Grid Connectivity Matrix

Output Layers

• Future P predictions<br>• Future Q predictions<br>• System State Estimation

Data and Simulation

IEEE Benchmark Grids

Utilized standard test systems widely recognized in the industry.

Power-Flow Simulations

Generated extensive datasets covering multiple grid configurations and load profiles.

Research Results

The topology-aware model demonstrated significant improvements over baselines.

Low Prediction Error

High accuracy achieved for both Active (P) and Reactive (Q) power.

Stable Performance

Consistent results across various load conditions.

Good Generalization

Effectively handles unseen grid configurations.

Practical Applications

<div style='margin-bottom:30px'><strong>Real-time Grid Control</strong><br/>Enables immediate response to renewable fluctuations.</div><div style='margin-bottom:30px'><strong>Faster Optimization</strong><br/>Reduces calculation time for Optimal Power Flow (OPF) tasks.</div><div><strong>Smart Energy Systems</strong><br/>Supports advanced decision-making for smart cities.</div>

Conclusion

Combining machine learning with physical grid information is the key to future grid stability.

Fast power-flow prediction achieved successfully.

Topology-aware learning outperforms standard methods.

Promising potential for real-time industrial deployment.