Final Results and Interpretation

At the end of this journey, it is time to draw a clear assessment of the performances achieved. Over the 2022–2024 test period, we obtain the following results:

• Persistence Model (Naive Baseline): \(\approx\) 56,400 MWh MAE
• Autoregressive Model + Calendar Variables: \(\approx\) 26,300 MWh MAE
• Final Model (Linear Base + Weather-Based Residual Correction): \(\approx\) 24,300 MWh MAE
• RTE Day-Ahead Proxy Forecast: \(\approx\) 25,900 MWh MAE

The cumulative gain compared to the initial baseline is substantial: the average error is cut by more than half relative to the persistence model.

While the comparison with RTE’s proxy forecast must be interpreted with caution—mainly due to the use of ERA5 reanalysis data in our experimental pipeline—achieving a performance level within the same order of magnitude using a relatively simple model built solely from public data is a highly encouraging result.

What stands out as particularly interesting in this study is not just the final score, but the mechanism through which the improvement was achieved.

The 14-day autoregressive history delivers by far the largest gain: it captures the natural inertia of electricity consumption and drastically reduces the error of the naive baseline. The calendar variables (weekdays, weekends, public holidays) then follow up to better capture the system's behavioral regularities.

The contribution of weather data is more subtle. When injected directly into a global model, the thermal variables provided relatively little extra information. Their contribution becomes truly effective only when the problem is reframed as a residual correction. This approach allows the weather model to focus specifically on the discrepancies that the AR + calendar structure fails to explain properly.

This two-tier decomposition—a simple, interpretable base model for the general dynamics, paired with a dedicated corrector for meteorological variations—proved particularly efficient. It illustrates a core principle in modeling: a well-structured problem can sometimes yield greater rewards than a brute-force increase in algorithmic complexity.

However, several major limitations should be kept in mind

First, our model remains an "offline" system. The ERA5 data used here consists of reanalyses produced ex-post by assimilating real-world weather observations. In a true day-ahead operational forecasting system, these data would not be available in this form. They would have to be replaced by operational weather forecasts, which are inherently more uncertain and noisier.

Second, certain long-term structural shifts are not yet explicitly modeled. These include the impact of the COVID-19 pandemic, the evolution of the French energy mix, the rapid rise of solar photovoltaics, the gradual electrification of energy uses, and shifts in consumer behavior. In a more advanced framework, these phenomena might require adding new explanatory variables or frequently updating the model.

Despite these limitations, this tutorial demonstrates that it is entirely possible to build a robust, interpretable, and competitive energy forecasting system using relatively simple tools—provided that the structure of the problem is carefully thought out.