Energy Usage Forecasting

Setup

python -m pip install -r requirements.txt

Streamlit dashboard

Run the production-ready dashboard locally:

streamlit run streamlit_app.py

Notes:

• The dashboard loads data/processed/global_active_power_hourly.csv by default. Override with ENERGY_APP_DATA_CSV=/path/to/your.csv.
• For model-based forecasts, run the training scripts at least once to generate outputs/ artifacts (e.g., run_ml_models.py, prophet_forecasting.py, arima_forecasting.py, run_model_comparison.py).
• Plot PNG downloads use kaleido (included in requirements.txt).

Data preparation

Downloads the UCI household power consumption dataset (if missing), builds an hourly target series, engineers time-based features, and saves temporal train/val/test splits under data/processed/.

python prepare_time_series_data.py

Prophet forecasting (train → test)

Runs Prophet with:

• daily + weekly seasonality enabled
• yearly seasonality enabled only if there is at least ~2 years of training data
• calendar regressors derived from existing time features (hour/day/month sin/cos, weekend/business hour)
• optional country holidays (default: France)

python prophet_forecasting.py --output-dir outputs/prophet

Artifacts:

• outputs/prophet/prophet_forecast.csv
• outputs/prophet/prophet_forecast_vs_actual.png
• outputs/prophet/prophet_components.png

ARIMA baseline (SARIMAX, train → test)

Fits a small SARIMAX model on the last N training points (default 1500) and forecasts the test horizon.

python arima_forecasting.py --output-dir outputs/arima --max-train-points 1500

Artifacts:

• outputs/arima/arima_forecast.csv
• outputs/arima/arima_forecast_vs_actual.png

Compare Prophet vs ARIMA

Runs both models, writes a metrics table, and saves an overlay plot.

python compare_prophet_arima.py --output-dir outputs --max-arima-train-points 1500

Artifacts:

• outputs/compare/metrics_comparison.csv
• outputs/compare/forecast_overlay.png

LSTM neural network (24-hour lookback → next hour)

Trains an LSTM on sequences of the previous 24 hours to predict the next hour. It:

• scales inputs/targets with MinMaxScaler (fit on train only)
• uses EarlyStopping + ModelCheckpoint
• saves loss curves and test predictions/metrics

Note: TensorFlow support depends on your Python version. If pip install tensorflow fails on your interpreter, use a Python version supported by TensorFlow (commonly 3.10–3.12) or install a compatible TensorFlow build for your environment.

python run_lstm.py --lookback-hours 24 --epochs 50 --batch-size 128

Optional: include calendar/time features (still fed as sequences):

python run_lstm.py --include-time-features --lookback-hours 24

Artifacts:

• outputs/lstm/best_model.keras
• outputs/lstm/training_history.png
• outputs/lstm/lstm_test_predictions.csv
• outputs/lstm/lstm_pred_vs_actual.png
• outputs/lstm/model_comparison.csv