Skip to Main content Skip to Navigation

Apprentissage multitâche pour la prévision de la consommation électrique

Abstract : We study in this manuscript the day-ahead electricity load forecasting problem, at the level of the substations, based on exogenous calendar information, weather forecasts and recent endogenous values of the electricity demand. This work is part of a broader research field participating in the modernization of the French energy system. The emergence of new production means and the evolution of electricity uses have indeed strengthened the need to anticipate the variations of the electricity demand. The Transmission System Operator (TSO), as a central actor of the electricity sector in charge of the supply-demand equilibrium and the management of the resulting energy flows, is particularly affected by these evolutions. Its decision-process relies on the ability to forecast accurately the spatial distribution of both the production and the demand. The advent of modern Machine Learning forecasting tools, in association with the improvement of computing capabilities and the gathering of rich weather and electricity datasets give rise to new opportunities.Data exploration and the dynamic literature about electricity load forecasting serve as a basis for the extension to local forecasts of the more classical models designed for the aggregated loads. We describe a generic bivariate linear model and compare its behavior at the national and the local levels. This allows us to identify both the similarities and the heterogeneous aspects of the substations. At the local level, the data exploration and the experiments are organized around a dichotomy between models learned independently for the different substations and a coupled modeling of the loads. In particular, we motivate a multi-task approach to load forecasting with a characterization of a common structure encountered in the local models, that we intend to leverage for the benefit of the latter, in terms of computational speed and generalization performance.We address several questions related to the multi-task approach. Namely, what to expect from a coupling of the local models ? Which parts of the model should be coupled and how ? How to assess the evolution and the relevance of the multi-task framework ?We study three coupling assumptions, based either on a clustering of the model parameters, an optimization problem with a low-rank constraint that we analyze in details, or on the consistency between the forecasts at different aggregation levels. Thereby, we prove empirically that the number of parameters of the independent local models is unnecessarily large and we confirm the interest of sharing the parameters and the losses during the learning process.
Document type :
Complete list of metadatas

Cited literature [230 references]  Display  Hide  Download
Contributor : Abes Star :  Contact
Submitted on : Friday, July 24, 2020 - 5:13:12 PM
Last modification on : Friday, January 15, 2021 - 5:24:27 PM
Long-term archiving on: : Tuesday, December 1, 2020 - 6:53:46 AM


Version validated by the jury (STAR)


  • HAL Id : tel-02906441, version 1



Benjamin Dubois. Apprentissage multitâche pour la prévision de la consommation électrique. Traitement du signal et de l'image [eess.SP]. Université Paris-Est, 2019. Français. ⟨NNT : 2019PESC1031⟩. ⟨tel-02906441⟩



Record views


Files downloads