Accueil du site > Séminaires > Probabilités Statistiques et réseaux de neurones > Artificial Neural Networks for Energy Management System. Applicability and limitations of the main paradigms
Vendredi 29 mars 2002 à 9h30
Artificial Neural Networks for Energy Management System. Applicability and limitations of the main paradigms
Gonzalo Joya (Université de Malaga)
Résumé : Electrical energy has obviously become an essential element for the operation and development of current society. Consequently, the improvement of the set of tasks implicated in its management - what we call Energy Management System (EMS)- constitutes a high-priority research field from the social, economical and human points of view. These tasks, which may be grouped as forecasting, state estimation and security related tasks, present all or most of the following characteristics : 1) their solution involves a high number of noisy and/or incomplete data. 2) Complex relationships exist among the variables implicated in each problem. 3) They are difficult to handle by an operator. 4) It is difficult to find a numerical or algorithmical solution to the problem, and if this solution is found, it presents a high computational cost. 5) They cannot be described by means of a simple set of rules based on the expert’s knowledge. 6) Real time operation is frequently required. These features discourage the application of classical numerical methods, whereas Artificial Neural Networks (ANN) based techniques turn out to be especially well suited for them. Besides, many of these problems may be approached as either a classification or a function approximation problem, and both approaches fit into the different paradigms that ANN techniques comprise. Thus, on one hand, feed-forward supervised neural networks may be used to obtain a particular numerical function. On the other hand, unsupervised neural networks take advantage of their ability to extract unknown criteria from a pattern set to achieve a visual classification of the patterns. Yet ANNs are often improperly used and they are required to solve problems that they are not prepared for. This spurious usage is partly due to the complex internal representation of the network parameters, but these parameters are easily obtained by means of well-established training algorithms. Thus, we are tempted to use ANNs not only as "black boxes" but as some kind of "magic boxes". This risk justifies a deep study of the internal behavior of ANNs.
In this course we review the application of ANNs for EMS from a double perspective. On one hand, we will study the most significant operations on an EMS. From their features and the limitations of the classical solutions, we will justify a neural solution and the choice of the most appropriate neural paradigm. On the other hand, we will use the EMS environment as a "benchmark" to highlight the main features, limitations and usage recommendations of the mostly applied neural paradigms.
Dans la même rubrique :
- Jan 15 2010, 11h00 : Validation de processus ponctuels marqués de Gibbs à travers l’analyse des résidus. , Jean-François Coeurjolly (Université P. Mendes-France, Grenoble)
- Jan 08 2010, 11h00 : Classification de variables qualitatives autour de variables latentes. , Vanessa Kuentz (Universités Bordeaux 1 et 2)
- Dec 18 2009, 11h00 : Formule de représentation pour les EDSR dirigées par une martingale continue et application en Finance. , Anthony Reveillac (Université Humboldt, Allemagne)
- Dec 04 2009, 11h00 : Estimation et sélection en classification semi-supervisée , Vincent Vandewalle (Université Lille 1-Lille 2)
- Nov 27 2009, 11h00 : Equation de la chaleur stochastique avec un bruit fractionnaire de dimension infinie , Raluca Balan (Université d’Ottawa)
- Nov 20 2009, 11h00 : Apprentissage supervisé pour le diagnostic du paludisme à haut-débit : réconcilier des experts en conflit. , Anne-Claire Haury (Mines Paristech/INSERM/Institut Curie)
- Oct 16 2009, 11h00 : Log-periodogram regression on non-Fourier frequencies sets. , Mohamed Boutahar (GREQAM, Université de Marseille-Luminy).
- Oct 09 2009, 11h00 : Sélection de modèles pour la classification non supervisée. , Jean-Patrick Baudry (Université de Paris-Sud)
- Jun 12 2009, 11h00 : Variations and Hurst index estimation for a Rosenblatt process using longer filters. , Frederi Viens (Purdue University, USA)
- Jun 05 2009, 11h00 : Estimation de densité avec des tailles de fenêtres locales : quelques résultats théoriques et des applications possibles. , Catherine Aaron (Université de Clermont-Ferrand)
- Apr 10 2009, 11h00 : Limit theorems for multiple sums of random variables. , Oleg Klesov (University Universität Paderborn, Allemagne and National Technical University of Ukraine)
- Apr 03 2009, 11h00 : Ratio of Generalized Hill’s estimator and its asymptotic normality theory. , Aliou Diop (Université de Saint-Louis, Sénégal)
- Mar 27 2009, 11h00 : Random attractors for stochastic Navier-Stokes equations in some unbounded domains. , Zdzislaw Brzezniak (University of York, UK)
- Mar 20 2009, 11h00 : On the stochastic Landau-Lifshitz’ Equation , Zdzislaw Brzezniak (University of York, UK)
- Feb 27 2009, 11h00 : Un point de vue statistique pour la régularisation de problèmes inverses mal posées et sa connexion avec les méthodes à noyaux. , Anna Karina Firmin (Université Paris X)
- Feb 06 2009, 11h00 : Une extension de l’ACP : les modèles auto-associatifs , Serge Iovleff (Université Lille I)
- Jan 16 2009, 11h00 : Modèles de Markov cachés en météorologie , Pierre Ailliot (Université de Brest)
- Jan 09 2009, 11h00 : Sur l’estimation fonctionnelle par le temps d’occupation , Boris Labrador (L.S.T.A., Université P. et M. Curie)
- Dec 19 2008, 11h00 : Biomarker discovery in MALDI-TOF serum protein profiles using discrete wavelet transformation , Theodore Alexandrov (Université de Breme, Allemagne)
- Oct 31 2008, 11h00 : Joint distribution of the sum and maximum of iid exponential random variables , Anna Panorska (University of Nevada, Reno, USA)
