Skip to content

Abstract

This tutorial provides a comprehensive guide to building a robust, interpretable model for forecasting daily electricity consumption in France for the next day (D+1).

Moving beyond simple black-box approaches, we prioritize methodological rigor and physical intuition. We demonstrate how to combine classic time-series analysis with weather-driven corrections to achieve performance competitive with official transmission system operator (TSO) proxies.

The Learning Path

The tutorial follows an incremental logic to isolate the impact of each modeling component:

  1. The Baseline: Establishing a starting point with a persistence model (MAE ≈ 56,400 MWh).
  2. Calendar & Autoregression: Integrating one-hot encoded temporal variables and historical consumption patterns (MAE ≈ 26,300 MWh).
  3. The Hybrid Architecture: Using a two-stage approach where a base model is supplemented by a Residual Corrector driven by ERA5 thermal aggregates (HDD/CDD, 14-day lag memory).
  4. Final Evaluation: Achieving a final MAE of approximately 24,300 MWh—a 7.6% improvement over weather-free models.

Key Methodological Pillars

The proposed methodology combines three complementary ideas. A strict time-series validation strategy trains the model on 2012–2021 data and evaluates it on 2022–2024, reproducing realistic forecasting conditions and testing its robustness to recent energy market dynamics.

At the same time, physically motivated feature engineering converts raw meteorological variables into spatially aggregated thermal indicators that better reflect the thermosensitivity of the French grid.

Finally, the model is deliberately designed in a modular way, separating baseline consumption from weather-driven corrections to preserve interpretability and simplify future maintenance.

This project is tailored for students, data scientists, and energy professionals looking to master a structured, real-world approach to time-series forecasting.