Power System Neural Network
Deep learning model using PyTorch to predict branch overloads in electric power grids, implementing neural networks for critical infrastructure monitoring.
Live in-browser demo
This is the trained PowerSystemNN running entirely in your browser as plain JavaScript matmul — no server, no API call. Adjust the load and generation sliders and watch the overload probability update in real time.
Model architecture
Architecture diagram
Overview
A deep learning project using neural networks to predict branch overloads in electric power systems. The project implements a multi-layer neural network in PyTorch to analyze power flow data and predict potential grid failures, critical for maintaining power grid stability and preventing cascading blackouts.
Technologies
Key Features
- ✓Custom neural network architecture design
- ✓Multi-layer fully connected network with ReLU activation
- ✓Feature scaling and normalization pipeline
- ✓Binary classification for overload prediction
- ✓Model training with validation monitoring
- ✓Performance evaluation (accuracy, precision, recall)
Challenges
Designing network architecture for power system data characteristics, handling class imbalance in overload events, selecting optimal hyperparameters, and ensuring model generalizes across different grid configurations.
Results
Trained PowerSystemNN (Linear→ReLU→Linear→ReLU→Linear→Sigmoid, 37 selected bus/generator features → 46 per-branch overload probabilities) and deployed it as a live in-browser demo on this project page — visitors adjust load and generation sliders and watch the model predict overloads across all 46 grid branches in real time, running entirely client-side with no server.
© 2026 Maxwell Vaglica. All rights reserved.