ML & Data Science

Time-series forecasting

Household Power Forecasting

Forecasting daily household power consumption, and finding which family of models actually wins.

Role
Data Scientist
Year
2025
Focus
Time-series forecasting
Stack
5 technologies
  • Prophet
  • SARIMAX
  • LSTM
  • XGBoost
  • LightGBM
Code

  • 0.0125

    Test RMSE (XGBoost)

  • 0.2614

    Best classical RMSE (LSTM)

  • 7

    Models compared

Overview

Classical statistical models are compared against deep learning and gradient boosting on the same engineered daily series. Every model's artifacts and metrics are saved with timestamps, and the best model is recorded for deployment.

XGBoost won on the test set. Prophet is kept for its uncertainty intervals, and ensembling with Random Forest is the next step.

Architecture

Workflow

  1. 01

    Prepare

    Clean and resample to daily consumption.

  2. 02

    Engineer

    Calendar, lag, and rolling features.

  3. 03

    Classical models

    SARIMA, Prophet, and LSTM.

  4. 04

    Advanced models

    XGBoost, LightGBM, SARIMAX, and Random Forest.

  5. 05

    Evaluate

    Robust RMSE and MAPE, then ranked comparison tables.

  6. 06

    Select

    Best model saved with metadata for deployment.