Abstract :
[en] Energy demand forecasting at large scale is a critical challenge for modern power grid operators, especially as the transition to all-electric systems, the integration of renewables, and evolving consumption behaviors introduce greater variability and complexity. In an industrial partnership with Creos Luxembourg S.A., this PhD thesis addresses the needs of grid management through a comprehensive exploration of scalable, accurate, and practical solutions to electricity consumption forecasting at the household level.
The first axis of this thesis targets scalability in forecasting solutions. Existing state-of-the-art machine learning models often suffer from significant computational costs, rendering them impractical for large-scale and frequent deployment. This manuscript introduces Transplit, a novel transformer-based forecasting architecture that exploits the inherent seasonality of consumption data by encoding entire seasonal cycles (e.g., days) as compact vectors. This approach dramatically reduces the sequence length processed by the model, resulting in a lightweight method capable of training on thousands of households with minimal hardware (including CPUs), while achieving competitive accuracy against deep learning baselines, paving the way for fine-grained, frequent grid state estimation at population scale.
The second research direction focuses on peak demand forecasting, essential for operational safety, economic dispatching, and load balancing, yet under-addressed by classical error metrics (MSE, MAE) that flatten out local maxima. To overcome this limitation, we propose DeDiPeak, a framework comprising new _Peak Prediction Performance_ (P3) metrics -- specifically designed to assess forecasted peaks on both timing and amplitude -- with a deblurring diffusion model. This diffusion post-processor can be plugged into any forecasting model to selectively enhance peaks, leading to significant improvements in peak-specific accuracy without degrading the overall error. Empirical results show up to a 36% gain in peak forecasting quality over benchmarks, not only on electricity consumption, but also on a broad spectrum of real-world datasets.
The third axis investigates the leverage of external factors, such as weather, holidays, and special events, in global forecasting models. While traditional models often degrade when incorporating such exogenous variables, we show that a hypernetwork-based architecture can bridge this gap by dynamically generating consumer-specific model weights conditioned on compact embeddings that represent both household identity and contextual factors. Evaluated on multi-year, multi-household datasets containing both consumption records and aligned external signals, our hypernetwork approach consistently outperforms classical, global, and expert-mixture models, achieving individual-level accuracy without a costly per-consumer model maintainance, and enabling fast adaptation to changing consumer populations.
Finally, to a lesser extent, this thesis explores the emergent use of Large Language Models (LLMs) in power system optimization, and especially their ability to solve the Optimal Power Flow (OPF) problem by querying LLMs with graph- and table-based representations of realistic power grids. We show that LLMs can be adapted to reliably solve OPF instances, with their performance improving as model size and tailored training increase, thereby opening new perspectives for cross-domain modeling in the energy sector.
Together, these contributions advance the state of the art in scalable grid analytics, robust peak event prediction, multi-factor consumption modeling, and cross-disciplinary energy system optimization. The resulting frameworks are directly applicable within the industrial context of Creos Luxembourg S.A., encouraging the deployment of next-generation forecasting tools, which are critical for tomorrow's smart grids.
