The Power Hour

XGBoost hourly energy demand forecaster for PJM East - 145,366 observations across 16 years, with temporal feature engineering, hyperparameter tuning, and SHAP interpretability.

Watts next? This model knows.

---

Results

Metric	Validation	Test
MAPE	0.77%	0.84%
RMSE	346 MW	360 MW
MAE	246 MW	267 MW

Sub-1% MAPE on an out-of-sample test year, with no weather data, no holiday calendars, and no external features. Just timestamps and lagged demand. The tight val/test gap (0.77% vs 0.84%) confirms the model generalizes across time periods without overfitting.

---

What It Does

Takes 16 years of hourly electricity consumption data for the PJM East region, engineers 20+ temporal features (calendar, cyclical encoding, lag values, rolling statistics), and trains a hyperparameter-tuned XGBoost regressor to predict next-hour demand. The model is evaluated on a held-out test year (2017-2018) with detailed error analysis by hour and month, and explained with SHAP feature importance.

Dataset

Hourly Energy Consumption - PJM East Interconnection hourly demand in megawatts (2002-2018). CC0 Public Domain license.

How to Run

1. Clone and install:

git clone https://github.com/lenamonj/the-power-hour.git
cd the-power-hour
pip install -r requirements.txt

2. Download PJME_hourly.csv from Kaggle and place it in data/.

3. Run:

jupyter notebook the_power_hour.ipynb

Engineered Features

Category	Features
Calendar	hour, dayofweek, month, dayofyear, weekofyear, quarter, is_weekend
Cyclical	hour_sin/cos, month_sin/cos, dow_sin/cos
Lag	1h, 24h, 48h, 168h (1 week)
Rolling	24h mean/std, 7-day mean/std
Trend	days_since_start

Design Decisions

• XGBoost over LSTM/Prophet. Gradient-boosted trees with engineered features consistently outperform deep learning and statistical models on tabular time series. Simpler, faster, more interpretable.
• Temporal splits only. Train on 2002-2016, validate on 2016-2017, test on 2017-2018. No random shuffling - that would leak future information into training.
• No external data. Weather, holidays, and economic indicators would improve accuracy but add dependencies. This notebook demonstrates what pure time series feature engineering can achieve.
• Low learning rate + early stopping. Tuned from [0.005, 0.01, 0.02] with up to 5,000 trees. The model finds its own stopping point on validation MAPE.
• SHAP over feature importance. XGBoost's built-in importance is biased toward high-cardinality features. SHAP values are theoretically grounded.

Project Structure

.
├── the_power_hour.ipynb    # Full pipeline - one notebook
├── data/                   # Kaggle data (not tracked in git)
├── requirements.txt        # Python dependencies
├── LICENSE                 # MIT License
└── README.md               # You are here

License

This project is licensed under the MIT License.

---

_{Built with Python, XGBoost, and public data. No proprietary tools required.}