Latent Surgical Interventions in Residual Neural Networks

Benjamin Donnot 1 Isabelle Guyon 2, 1 Zhengying Liu 1 Marc Schoenauer 1 Antoine Marot 3 Patrick Panciatici 3
1 TAU - TAckling the Underspecified
LRI - Laboratoire de Recherche en Informatique, UP11 - Université Paris-Sud - Paris 11, Inria Saclay - Ile de France, CNRS - Centre National de la Recherche Scientifique : UMR8623
Abstract : We propose and study a novel artificial neural network framework, which allows us to model surgical interventions on a physical system. Our approach was developed to predict power flows in power transmission grids, in which high voltage lines are disconnected and reconnected with one-another from time to time, either accidentally or willfully. However, we anticipate a broader applicability. For several exemplary cases, we illustrate by simulation that our methodology permits learning from empirical data to predict the effect of a subset of interventions (ele-mentary interventions) and then generalize to combinations of interventions never seen during training. We verify this property mathematically in the additive perturbation case. In terms of transfer learning, this is equivalent to training on data from a few source domains then, with a zero-shot learning, generalizing to new target domains (super-generalization). Our architecture bears resemblance with the successful ResNets, with the simple modification that interventions are encoded as an addition of units in the neural network. For applications to real historical data, from the French high voltage power transmission company RTE, we evaluate the viability of this technique to rapidly assess curative actions that human operators take in emergency situations. Integrated in an overall planning and control system, methods deriving from our approach could allow Transmission System Operators (TSO) to assess in real time many more alternative actions, reaching a better exploration-exploitation tradeoff, compared to presently deployed physical system simulator. 1 Background and motivations In this paper, we are interested in speeding up the computation of power flows in power transmission grids using artificial neural networks, to emulate slower physical simulators. Key to our approach is the possibility of simulating the effect of actions on the grid topology. Such neural networks may then be used as part of an overall computer-assisted decision process in which human operators (dispatchers) ensure that the power grid is operated in security at all times, namely that the currents flowing in all lines are below certain thresholds (line thermal limits). We describe our application setting for concreteness, but anticipate a broader applicability of the techniques developed in this paper in various domains of physics, chemistry, manufacturing, biomedicine and others, in which some actions can be combined with each other, but running extensive simulations for each possible combination of such actions is computationally untractable. Electric power generated in production nodes (such as power plants) is transmitted towards consumption nodes in a power grid. The power lines enable this transmission through substations interconnecting them. Each pattern of connections is referred to as a grid topology. This topology is * Benjamin Donnot corresponding authors: benjamin.donnot@inria.com 32nd Conference on Neural Information Processing Systems (NIPS 2018), Montréal, Canada.
Type de document :
Pré-publication, Document de travail
2018
Liste complète des métadonnées

Littérature citée [10 références]  Voir  Masquer  Télécharger

https://hal.archives-ouvertes.fr/hal-01906170
Contributeur : Zhengying Liu <>
Soumis le : vendredi 26 octobre 2018 - 14:52:51
Dernière modification le : jeudi 7 février 2019 - 10:26:02
Document(s) archivé(s) le : dimanche 27 janvier 2019 - 14:08:22

Fichier

latent-surgical-interventions-...
Fichiers produits par l'(les) auteur(s)

Identifiants

  • HAL Id : hal-01906170, version 1

Citation

Benjamin Donnot, Isabelle Guyon, Zhengying Liu, Marc Schoenauer, Antoine Marot, et al.. Latent Surgical Interventions in Residual Neural Networks. 2018. 〈hal-01906170〉

Partager

Métriques

Consultations de la notice

234

Téléchargements de fichiers

107