EnergyWise
Intelligent Energy Consumption Forecasting, Anomaly Detection & Predictive Analytics Platform
Production-grade electricity telemetry platform featuring causal multi-horizon demand forecasting (1h, 6h, 24h), Conformalized Quantile Regression (90% CQR), a 5-detector unsupervised anomaly consensus ensemble, and Time-of-Use tariff optimization.
Why it exists
Monthly power bills hide baseload creep, appliance inefficiencies, and post-outage transients. I wanted to turn millions of high-frequency smart-meter readings into verifiable forecasts and actionable electrical anomaly signatures.
How it works
- 01Ingestion and physical consistency validation of 2.07M 1-minute smart-meter records from the UCI household power dataset (S = V * I / 1000 vs sqrt(P^2 + Q^2), r = 0.99964)
- 02Strictly causal feature engineering across 87 temporal, lag, rolling-window, EWMA, Fourier harmonic, and circuit-interaction features with automated unit tests asserting bitwise historical invariance
- 03Multi-horizon benchmarks across statistical baselines, Ridge, Random Forest, XGBoost, and LightGBM across 1h, 6h, and 24h horizons with 3-fold expanding-window walk-forward validation
- 04Conformalized Quantile Regression (CQR) delivering distribution-free 90% prediction intervals calibrated on validation data
- 055-detector unsupervised anomaly consensus ensemble (Rolling Z-Score, Robust IQR, Isolation Forest, LOF, Forecast Residual) categorizing 7 physical signatures with synthetic test injection verification (100% precision, 0% FPR)
- 06Interactive 6-tab Streamlit analytics platform with permutation feature importance, drift tracking (PSI / KS), and a Time-of-Use (ToU) tariff load-shifting simulator
What was hard
Enforcing strict temporal non-leakage across rolling window computations while ensuring multi-detector anomaly consensus reliably differentiates real appliance transients from true electrical faults.