âš¡ BijliCast

Overview

BijliCast is a weather-aware household electricity consumption forecasting system developed for Pakistan.

It uses historical household electricity consumption, weather conditions, calendar information and household characteristics to predict next-hour electricity usage.

The prediction target is next_hour_energy_kwh.

Model Information

Property Value
Model HistGradientBoostingRegressor
Framework Scikit-learn
Task Time-series regression
Artifact deployment_model.joblib
Model size Approximately 1.4 MB
Input features 42 raw features

The saved artifact includes the fitted preprocessing pipeline and the trained estimator.

Dataset

BijliCast combines two independent public data sources:

  1. LUMS REWD-P household electricity dataset.
  2. Open-Meteo ERA5 historical weather data.

Dataset characteristics:

  • 59 households
  • Six Pakistani cities
  • 515,786 hourly electricity records

The weather information is joined with household electricity measurements using temporal and geographic alignment.

Methodology

The ML lifecycle includes:

  • Exploratory Data Analysis
  • Missing-value handling
  • Feature engineering
  • Lag and rolling statistics
  • Chronological train/validation/test splitting
  • Scikit-learn preprocessing Pipeline
  • Baseline comparisons
  • Multiple regression algorithms
  • Five-fold TimeSeriesSplit cross-validation
  • Randomized hyperparameter optimization

The final deployment pipeline was fitted using training and validation data. The held-out test set was excluded from model fitting and selection.

Model Selection

Random Forest achieved a validation MAE of 0.3216 kWh, while HistGradientBoosting achieved 0.3273 kWh.

Their cross-validation variation indicated practically comparable performance.

HistGradientBoosting was selected for deployment because it offered comparable predictive quality with a substantially smaller serialized artifact.

Random Forest: approximately 270 MB HistGradientBoosting: approximately 1.4 MB

This made HGB considerably more practical for GitHub-based and lightweight deployment.

Evaluation Results

Held-out test results:

Metric Result
MAE 0.2499 kWh
RMSE 0.4273 kWh
R² 0.7291

These are final evaluation results, not the criteria used to select the model.

Prediction Interval

The application provides a 90% prediction interval using residual-based offsets estimated from out-of-sample validation predictions.

Intended Use

  • Academic research
  • Household electricity consumption analysis
  • Short-term residential load forecasting
  • Educational machine learning demonstrations

Limitations

  • The dataset covers 59 households.
  • Coverage is limited to six Pakistani cities.
  • The demo uses historical replay data rather than live smart-meter readings.
  • The model is not intended for electricity billing or grid operational decisions.

Repository

Full source code, notebooks, reports and deployment:

https://github.com/muhammad-umar-218980/BijliCast

Data Sources

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support