Time series machine learning augmented with social network events to improve national electricity consumption profile estimation
Résumé
Assessing the electricity demand is essential for the grid stability, as energy production must always meet energy demand. It is also important for the electricity market settlement. Depending on that, energy load investigation is fundamental in enhancing the reliability, efficiency and fairness of the electrical grid. In general, two main factors impact the energy load. The first one is the meteorological conditions. The second one is human activities. Nevertheless, human activities are hard to anticipate and evaluate. Yet this work focuses on identifying and quantifying important human events (regarding their impact on the energy consumption) from social media. In order to achieve that, this work proposes a complete method to detect when an event might have occurred. It develops a mechanism to collect data from social media, more particularly from Twitter. Then it presents how different events can be picked and labeled. It shows how those events can be quantified and integrated into different models to enhance the load profile estimation. Eventually, it validates the whole process with real data from the French electrical grid. It also demonstrates how the proposed approach can minimize the error of the profile load estimation up to 60% the root means square error. In addition, it does illustrate how artificial intelligence techniques like transfer learning can be helpful in extracting knowledge from a standard energy profile like a residential profile and project it on another one like an industrial profile, especially when there is a lack of historical data.
Domaines
Energie électriqueOrigine | Fichiers produits par l'(les) auteur(s) |
---|