Harmonic Distortion (THD) State Classifier

Predicts whether Total Harmonic Distortion (THD) on a power system is in a HIGH or LOW state, based on simultaneous real power, voltage, and current readings. Built on real power-quality measurement data (harmonics 1-49, 3 phases).

Live Demo

[Try the live app here]](https://harmonic-distortion-e5bzwwfkatyc5knbaw9jrm.streamlit.app/)

Project Story

This project started as an attempt to forecast tomorrow's THD from the last few days of readings a natural first approach for time-series data.

That approach didn't work, and the notebook documents why:

• Lag features (1-3 days), rolling averages, rolling volatility, and day-of-week were all tested as predictors
• Every feature showed near-zero correlation with next-day THD (all |r| < 0.11)
• The tuned SVR model converged to its most conservative setting (C=0.1) and scored a negative R² - meaning it performed no better than simply guessing the average value every time

Rather than force a result that wasn't there, the project was reframed: instead of predicting the future, classify the present given a moment's power, voltage, and current readings, is the system in a HIGH or LOW THD state right now?

This approach worked:

• Real power showed a real difference between HIGH and LOW THD states (~28% lower during high-distortion periods)
• An SVM classifier (RBF kernel) reached 61% accuracy on held-out test data - a real, if modest, signal
• A Random Forest comparison scored slightly lower (58%), so the SVM was kept as the final model

What's in the Notebook

Section	Content
1. Setup	Imports
2. Load & Prepare Data	Read raw Excel export, remove duplicate columns
3. Feature Engineering	Real power (3 phases), THD (3 phases), RMS voltage/current, power factor
4. Modeling	THD-state classification: SVM (final model) vs. Random Forest, saved with joblib
Appendix	The abandoned day-ahead forecasting attempt, kept for transparency

🛠️ Tech Stack

• Data & modeling: pandas, numpy, scikit-learn
• Demo app: streamlit
• Model persistence: joblib

Run It Locally

git clone https://github.com/AlmaFatima/Harmonic-Distortion.git
cd harmonic-distortion
python -m venv venv
source venv/bin/activate   # Windows: venv\Scripts\activate
pip install -r requirements.txt
streamlit run App.py

📁 Files

• harmonic-distortion-final.ipynb - full analysis and modeling notebook
• App.py - Streamlit demo app
• thd_state_classifier.pkl / thd_state_scaler.pkl - the trained model and its feature scaler
• historical_data.csv - exported data used for the app's context chart

Results Summary

Approach	Result
Day-ahead THD forecast (SVR, 7 engineered features)	No signal - negative R²
Same-moment THD-state classification (SVM)	61% accuracy - final model
Same-moment THD-state classification (Random Forest)	58% accuracy

💡 What I Learned

• How to calculate real power and THD from raw harmonic voltage/current data
• That a "negative result" (no forecastable signal) is still a valid, useful finding - and how to recognize one rather than force a model to fit
• How to reframe a problem (forecasting → classification) when the original framing doesn't hold up
• Feature scaling matters for SVM - an unscaled RBF kernel initially collapsed to predicting a single constant value
• Building an end-to-end pipeline from raw data through to a deployable, interactive demo