Submitted by Bárbara Conte (conte.babi@gmail.com) on 2013-09-12T19:53:02Z No. of bitstreams: 1 Bárbara.Conte_DissertaçãoMPFE.pdf: 523048 bytes, checksum: 827713ba2324041acb3fa60f11a232f7 (MD5) / Approved for entry into archive by Suzinei Teles Garcia Garcia (suzinei.garcia@fgv.br) on 2013-09-13T12:28:38Z (GMT) No. of bitstreams: 1 Bárbara.Conte_DissertaçãoMPFE.pdf: 523048 bytes, checksum: 827713ba2324041acb3fa60f11a232f7 (MD5) / Made available in DSpace on 2013-09-13T13:09:37Z (GMT). No. of bitstreams: 1 Bárbara.Conte_DissertaçãoMPFE.pdf: 523048 bytes, checksum: 827713ba2324041acb3fa60f11a232f7 (MD5) Previous issue date: 2013-08-19 / This paper aims to apply a nonlinear model to the Brazilian GDP. To achieve this goal we tested the existence of non-linearity of the data generating process with the methodology suggested by Castle and Henry (2010). The test verifies the persistence of the nonlinear regressors in an unconstrained linear model. Next, the series is modeled as an Autoregressive Model Threshold using the general-to-specifs approach to select the model. Autometrics is the automatic selection algorithm used to choose the nonlinear model. The results indicate that the Gross Domestic Product of Brazil is best explained by a non-linear model with three regime changes that occur in the early '90s, which, in fact, was a period quite volatile. Through modeling nonlinear exists the potential for cycle dating, however the results were not sufficient for such analysis. / O trabalho tem como objetivo aplicar uma modelagem não linear ao Produto Interno Bruto brasileiro. Para tanto foi testada a existência de não linearidade do processo gerador dos dados com a metodologia sugerida por Castle e Henry (2010). O teste consiste em verificar a persistência dos regressores não lineares no modelo linear irrestrito. A seguir a série é modelada a partir do modelo autoregressivo com limiar utilizando a abordagem geral para específico na seleção do modelo. O algoritmo Autometrics é utilizado para escolha do modelo não linear. Os resultados encontrados indicam que o Produto Interno Bruto do Brasil é melhor explicado por um modelo não linear com três mudanças de regime, que ocorrem no inicio dos anos 90, que, de fato, foi um período bastante volátil. Através da modelagem não linear existe o potencial para datação de ciclos, no entanto os resultados encontrados não foram suficientes para tal análise.

Dans différents domaines de recherche, la modélisation est devenue un outil efficace pour étudier et prédire l’évolution possible d’un système, en particulier en épidémiologie. En raison de la mondialisation et de la mutation génétique de certaines maladies ou vecteurs de transmission, plusieurs épidémies sont apparues dans des régions non encore concernées ces dernières années. Dans cette thèse, un modèle décrivant la transmission de l’épidémie de chikungunya à la population humaine est étudié. Ce modèle prend en compte la mobilité spatiale des humains, ce qui est nouveau. En effet, c’est un facteur intéressant qui a influencé la réapparition de plusieurs maladies épidémiques. Le déplacement des moustiques est omis puisqu’il est limité à quelques mètres. Le modèle complet (modèle EDOs-EDPs) est alors composé d’un système à réaction-diffusion (prenant la forme d’équations différentielles partielles (EDPs) paraboliques semi-linéaires) couplé à des équations différentielles ordinaires (EDOs). Nous démontrons pour ce modèle, d’abord l’existence et l’unicité de la solution globale, sa positivité et sa bornitude, puis nous donnons quelques simulations numériques. Dans ce modèle, certains paramètres ne sont pas directement accessibles à partir des expériences et doivent être estimés numériquement. Cependant, avant de rechercher leurs valeurs, il est essentiel de vérifier l’identifiabilité des paramètres pour déterminer si l’ensemble des paramètres inconnus peut être déterminé de manière unique à partir des données. Cette étude permettra de s’assurer que les procédures numériques peuvent être couronnées de succès. Si l’identifiabilité n’est pas assurée, certaines données supplémentaires doivent être ajoutées. En fait, une première étude d’identifiabilité a été effectuée pour le modèle EDOs en considérant que le nombre d’œufs peut être facilement compté. Toutefois, après avoir discuté avec les chercheurs épidémiologistes, il apparaît que c’est le nombre de larves qui peut être estimé semaines par semaines. Ainsi, nous ferons une étude d’identifiabilité pour le nouveau modèle EDOs-EDPs avec cette hypothèse. Grâce à l’intégration de l’une des équations du modèle, on obtient des équations plus faciles reliant les entrées, les sorties et les paramètres, ce qui simplifie vraiment l’étude d’identifiabilité. A partir de l’étude d’identifiabilité, une méthode et une procédure numérique sont proposés pour estimer les paramètres sans en avoir connaissance. / In different fields of research, modeling has become an effective tool for studying and predicting the possible evolution of a system, particularly in epidemiology. Due to the globalization and the genetic mutation of certain diseases or transmission vectors, several epidemics have appeared in regions not yet concerned in the last years. In this thesis, a model describing the transmission of the chikungunya epidemic to the human population is studied. As a novelty, this model incorporates the spatial mobility of humans. Indeed, it is an interesting factor that has influenced the re-emergence of several epidemic diseases. The displacement of mosquitoes is omitted since it is limited to a few meters. The complete model (ODEs-PDEs model) is then composed of a reaction-diffusion system (taken the form of semi-linear parabolic partial differential equations (PDEs)) coupled with ordinary differential equations (ODEs). We prove the existence, uniqueness, positivity and boundedness of a global solution of this model at first and then give some numerical simulations. In such a model, some parameters are not directly accessible from experiments and have to be estimated numerically. However, before searching for their values, it is essential to verify the identifiability of parameters in order to assess whether the set of unknown parameters can be uniquely determined from the data. This study will insure that numerical procedures can be successful. If the identifiability is not ensured, some supplementary data have to be added. In fact, a first identifiability study had been done for the ODEs model by considering that the number of eggs can be easily counted. However, after discussing with epidemiologist searchers, it appears that it is the number of larvae which can be estimated weeks by weeks. Thus, we will do an identifiability study for the novel ODEs-PDEs model with this assumption. Thanks to an integration of one of the model equations, some easier equations linking the inputs, outputs and parameters are obtained which really simplify the study of identifiability. From the identifiability study, a method and numerical procedure are proposed for estimating the parameters without any knowledge of them.

In this thesis I propose methods and strategies for the design of advanced model predictive control designs. The contributions are in the areas of data-driven model based MPC, model monitoring and explicit incorporation of closed-loop response considerations in the MPC, while handling issues such as plant-model mismatch, constraints and uncertainty. In the initial phase of this research, I address the problem of handling plant-model mismatch by designing a subspace identification based MPC framework that includes model monitoring and closed-loop identification components. In contrast to performance monitoring based approaches, the validity of the underlying model is monitored by proposing two indexes that compare model predictions with measured past output. In the event that the model monitoring threshold is breached, a new model is identified using an adapted closed-loop subspace identification method. To retain the knowledge of the nominal system dynamics, the proposed approach uses the past training data and current input, output and set-point as the training data for re-identification. A model validity mechanism then checks if the new model predictions are better than the existing model, and if they are, then the new model is utilized within the MPC. Next, the proposed MPC with re-identification method is extended to batch processes. To this end, I first utilize a subspace-based model identification approach for batch processes to be used in model predictive control. A model performance index is developed for batch process, then in the case of poor prediction, re-identification is triggered to identify a new model. In order to emphasize on the recent batch data, the identification is developed in order to increase the contribution of the current data. In another direction, the stability of data driven predictive control is addressed. To this end, first, a data-driven Lyapunov-based MPC is designed, and shown to be capable of stabilizing a system at an unstable equilibrium point. The data driven Lyapunov-based MPC utilizes a linear time invariant (LTI) model cognizant of the fact that the training data, owing to the unstable nature of the equilibrium point, has to be obtained from closed-loop operation or experiments. Simulation results are first presented demonstrating closed-loop stability under the proposed data-driven Lyapunov-based MPC. The underlying data-driven model is then utilized as the basis to design an economic MPC. Finally, I address the problem of control of nonlinear systems to deliver a prescribed closed-loop behavior. In particular, the framework allows for the practitioner to first specify the nature and specifics of the desired closed-loop behavior (e.g., first order with smallest time constant, second order with no more than a certain percentage overshoot, etc.). An optimization based formulation then computes the control action to deliver the best attainable closed loop behavior. To decouple the problems of determining the best attainable behavior and tracking it as closely as possible, the optimization problem is posed and solved in two tiers. In the first tier, the focus is on determining the best closed-loop behavior attainable, subject to stability and tracking constraints. In the second tier, the inputs are tweaked to possibly improve the tracking of the optimal output trajectories given by the first tier. The effectiveness of all of the proposed methods are illustrated through simulations on nonlinear systems. / Dissertation / Doctor of Philosophy (PhD)

La sédimentation des suspensions aqueuses de particules minérales microniques, polydisperses et concentrées a été analysée à l’aide du Turbiscan MA 2000 fondé sur la diffusion multiple de la lumière, en vue d’établir la procédure qui permet de déceler la présence d’une morphologie fractale, puis de déduire les règles de comportements des suspensions fractales par la modélisation avec les réseaux de neurones. Le domaine des interactions interparticulaires physicochimiques (0 à 10% volumique en solide) a été privilégié.La méthodologie de détermination de la structure multifractale des agglomérats et de la suspension a été proposée. La modification structurale des agglomérats qui est à l’origine de comportements non linéaires des suspensions et qui dépend des propriétés cohésives des particules primaires, est interprétée par la variation de la mobilité électrophorétique des particules en suspension. Une approche d’estimation de ces modifications structurales par les réseaux de neurones, à travers la dimension fractale, a été présentée. Les limites du modèle à assimiler ces comportements particuliers ont été expliquées comme résultant du faible nombre d’exemples et de la grande variabilité des mesures aux faibles fractions volumiques en solide. / The sedimentation of aqueous suspensions of micron-sized mineral particles, polydisperses and concentrated, was analyzed using the Turbiscan MA 2000 based on the multiple light scattering in order to establish the procedure to detect the presence of a fractal morphology, and then to deduce the set of laws of fractal behavior of suspensions by modeling with neural networks. The methodology for determining the multifractal structure of agglomerates and the suspension was proposed. The structural modifications of the agglomerates at the origin of the nonlinear behavior of suspensions and which depends on cohesive properties of primary particles, is interpreted by the change of the electrophoretic mobility of suspended particles. The estimation by neural networks of these structural changes, through the fractal dimension has been presented. The limits of the model to learn these specific behaviors have been explained as resulting from the low number of examples and the great variability in the measurements at low volume fractions of solid.

Cette thèse porte sur l'optimisation et la commande prédictive des centrales de production d'énergie en utilisant des modèles physiques des installations. Les modèles sont réalisés à l'aide du langage Modelica, un langage équationnel adapté à la modélisation de systèmes multi-physiques. La modélisation de systèmes physiques dans ce langage est présentée dans une première partie, ainsi que les traitements symboliques réalisés par les compilateurs Modelica pour mettre les modèles sous une forme adaptée à l'optimisation. On présente dans une seconde partie le développement d'une méthode d'optimisation dynamique hybride pour les centrales de production d'énergie, qui fournit une trajectoire optimisée de l'installation sur un horizon long. Les trajectoires calculées incluent les trajectoires des commandes continues ainsi que les décisions d'engagement des différents équipements. L'algorithme d'optimisation combine la méthode de collocation et une méthode nommée Sum Up Rounding (SUR) pour la prise en compte des décisions d'engagement. Un algorithme de commande prédictive (MPC) est enfin introduit afin de garantir le suivi des trajectoires optimales et de prendre en compte en temps réel la présence de perturbations et les erreurs du modèle d'optimisation. L'algorithme MPC utilise des modèles linéarisés tangents générés automatiquement à partir du modèle non linéaire. / This thesis deals with hybrid optimal control and Model Predictive Control (MPC) of power plants by use of physical models. Models of the facilities are developped with Modelica, an equation based language tailored for modelling multi-physics systems. Modeling of physical systems with Modelica is introduced in a first part, as well as some of the symbolic processing done by Modelica compilers that transform the original model to a form suited for optimization. Then, a method to solve optimal control problems on hybrid systems (such as power plants) is presented. This methods provides an optimal trajectory for the power plant on a long horizon. The optimal trajectory computed by the method includes the trajectories of continuous inputs as well as switching decisions for components in the plant. The optimization algorithm combines the collocation method and a method named Sum Up Rounding (SUR) for dealing with switches. Finally, a Model Predictive Controller is developped in order to follow this optimal trajectory in real time, and to cope with disturbances on the actual system and modelling errors. The proposed MPC uses tangent linear models of the plant that are derived automatically from the nonlinear model.

Os modelos de mistura são muito eficazes para analisar dados compostos por diferentes subpopulações com alocações desconhecidas ou que apresentam assimetria, multimodalidade ou curtose. Esta tese propõe relacionar a distribuição de probabilidade beta e a técnica de ajuste de modelos mistos à metodologia de modelos de mistura para que sejam adequados na análise de dados que assumem valores em um intervalo restrito conhecido e que também são caracterizados por possuírem uma estrutura de agrupamento ou hierárquica. Foram especificados os modelos de mistura beta mistos linear, com dispersão constante e variável, e não linear. Foi considerada uma abordagem bayesiana com uso de métodos de Monte Carlo via Cadeias de Markov (MCMC). Estudos de simulação foram delineados para avaliar os resultados inferenciais destes modelos em relação à acurácia da estimação pontual dos parâmetros, ao desempenho de critérios de informação na seleção do número de elementos da mistura e ao diagnóstico de identificabilidade obtido com o algoritmo data cloning. O desempenho dos modelos foi muito promissor, principalmente pela boa acurácia da estimação pontual dos parâmetros e por não haver evidências de falta de identificabilidade. Três bancos de dados reais das áreas de saúde, marketing e educação foram estudados por meio das técnicas propostas. Tanto nos estudos de simulação quanto na aplicação a dados reais se obtiveram resultados muito satisfatórios que evidenciam tanto a utilidade dos modelos desenvolvidos aos objetivos tratados quanto a potencialidade de aplicação. Ressaltando que a metodologia apresentada também pode ser aplicada e estendida a outros modelos de mistura. / Mixture models are very effective for analyzing data composed of different subpopulations with unknown allocations or with asymmetry, multimodality or kurtosis. This work proposes to link the beta probability distribution and the mixed models to the methodology of mixture models so that they are suitable to analyse data with values in a restricted and known interval and that also are characterized by having a grouping or hierarchical structure. There were specified the linear beta mixture models with random effects, with constant and varying dispersion, and also the nonlinear one with constant dispersion. It was considered a Bayesian approach using Markov Chain Monte Carlo (MCMC) methods. Simulation studies were designed to evaluate the inferential results of these models in relation to the accuracy of the parameter estimation, to the performance of information criteria in the selection of the number of elements of the mixture and to the diagnosis of identifiability obtained with the algorithm data cloning. The performance of the models was very promising, mainly due to the good accuracy of the point estimation of the parameters and because there was no evidence of lack of identifiability of the model. Three real databases of health, marketing and education were studied using the proposed techniques. In both the simulation studies and the application to real data had very satisfactory results that show both the usefulness of the models developed to the treated objectives and the potentiality of application. Note that the presented methodology can also be applied and extended to other mixing models.

Made available in DSpace on 2015-03-26T13:32:11Z (GMT). No. of bitstreams: 1 texto completo.pdf: 446864 bytes, checksum: 1ba8efc18a08b922adc7e2c5eb3dc55c (MD5) Previous issue date: 2011-02-24 / The aim of this study was to evaluate the efficiency of spatial analysis, which considers spatially dependent errors, for classification of common bean families in relation to traditional analysis in randomized blocks and lattice that assuming independent errors. Were considered different degrees of spatial dependence and experimental precision. Were taken as reference to simulate the results of seven experiments carried out in simple square lattice for genetic evaluation of yield (g/plot) of families and bean cultivars of winter crops and water used in 2007 and 2008. From the results presented in the simulation, it was possible to assess the quality of their experiments based on different analysis (Block, lattice and Spatial) and simulated average of 100 families in different scenarios for Spatial Dependence (DE) and Accuracy Selective (AS). In the process of simulation, the average yield (645 g/plot) and the residual variance (7744.00), was defined based on the analysis results of the tests in blocks of bean breeding program at UFV. To make up the four simulated scenarios were considered magnitude of spatial dependence (null, low, medium and high), corresponding to ranges of 0, 25, 50 and 100% of the maximum distance between plots. Were also simulated three classes of selective accuracy (0.95, 0.80 and 0.60), corresponding to the experimental precision very high, high and average, respectively. The actual classification of families was used to evaluate the efficiency of analysis methods tested by Spearman correlation applied to orders and genotypic classification of Selection Efficiency between classifications based on tested methodologies and the actual classification for the selection of 10, 20 and 30% of the best families. To compare the efficiency of adjustment of the models tested, was used the Akaike information criterion (AIC), based on likelihood. Spatial analysis has provided estimates of residual variance very close to the simulated residual variance and higher selective accuracy estimated in all scenarios, indicating greater experimental accuracy. With the reduction in the accuracy and selective increase in spatial dependence, there was greater influence of analysis on the classification of families, and the spatial analysis showed the best results, providing more efficient selection of bean families than traditional analysis of randomized blocks and lattice, mainly for the selection of fewer families. The results for selective accuracy estimated on the basis of F statistics were very close to those obtained with the Spearman correlation between estimated and simulated averages for families, indicating that the accuracy should be used selectively as a measure of experimental precision tests of genetic evaluation. / O objetivo deste trabalho foi avaliar a eficiência da análise Espacial, que considera erros dependentes espacialmente, para classificação de famílias de feijoeiro em relação às análises tradicionais em blocos casualizados e em látice que assumem erros independentes. Considerou-se diferentes graus de dependência espacial e de precisão experimental. Foram tomados como referência para simulação os resultados de sete ensaios instalados em látice quadrado simples para avaliação genética da produtividade de grãos (g/parcela) de famílias e cultivares de feijoeiro das safras de inverno e das águas de 2007 e 2008. A partir dos resultados apresentados na simulação, foi possível avaliar a qualidade dos respectivos experimentos com base nas diferentes análises (Bloco, Látice e Espacial) e médias simuladas das 100 famílias nos diferentes cenários para Dependência Espacial (DE) e Acurácia Seletiva (AS). No processo de simulação, a média de produção (645 g/parcela), bem como a variância residual (7744,00), foi definida com base nos resultados de análises em blocos de ensaios do programa de melhoramento do feijoeiro da UFV. Para a composição dos cenários simulados foram consideradas quatro magnitudes de dependência espacial (nula, baixa, média e alta), correspondendo aos alcances 0, 25, 50 e 100% da distância máxima entre parcelas. Também foram simuladas três classes de acurácia seletiva (0,95, 0,80 e 0,60), correspondente a precisão experimental muito alta, alta e média, respectivamente. A classificação real das famílias foi utilizada para avaliar a eficiência das metodologias de análise testadas através da correlação de Spearman aplicada às ordens de classificação genotípica e da Eficiência de Seleção entre classificações com base nas metodologias testadas e na classificação real, para a seleção de 10, 20 e 30% das melhores famílias. Para comparar a eficiência de ajuste dos modelos testados, foi utilizado o critério de Informação de Akaike (AIC), baseado em verossimilhança. A análise Espacial apresentou estimativas de variância residual muito próxima da variância residual simulada e maior acurácia seletiva estimada em todos os cenários, indicando maior precisão experimental. Com a redução na acurácia seletiva e aumento na dependência espacial, observou-se maior influência do tipo de análise sobre a classificação das famílias, sendo que a análise espacial apresentou os melhores resultados, proporcionando seleção mais eficiente das famílias do feijoeiro do que as análises tradicionais em Látice e em Blocos casualizados, principalmente, para seleção de menor número de famílias. Os resultados para acurácia seletiva estimada em função da estatística F foram muito próximos aos obtidos para a correlação de Spearman entre médias estimadas e simuladas para as famílias, indicando que a acurácia seletiva deve ser utilizada como medida de precisão experimental nos ensaios de avaliação genética.

Made available in DSpace on 2014-12-17T14:56:17Z (GMT). No. of bitstreams: 1 GasparFDJ_DISSERT.pdf: 3388304 bytes, checksum: 086a8f61099f69978a8b9f477f351d24 (MD5) Previous issue date: 2014-01-23 / Petr?leo Brasileiro SA - PETROBRAS / Slugging is a well-known slugging phenomenon in multiphase flow, which may cause problems such as vibration in pipeline and high liquid level in the separator. It can be classified according to the place of its occurrence. The most severe, known as slugging in the riser, occurs in the vertical pipe which feeds the platform. Also known as severe slugging, it is capable of causing severe pressure fluctuations in the flow of the process, excessive vibration, flooding in separator tanks, limited production, nonscheduled stop of production, among other negative aspects that motivated the production of this work . A feasible solution to deal with this problem would be to design an effective method for the removal or reduction of the system, a controller. According to the literature, a conventional PID controller did not produce good results due to the high degree of nonlinearity of the process, fueling the development of advanced control techniques. Among these, the model predictive controller (MPC), where the control action results from the solution of an optimization problem, it is robust, can incorporate physical and /or security constraints. The objective of this work is to apply a non-conventional non-linear model predictive control technique to severe slugging, where the amount of liquid mass in the riser is controlled by the production valve and, indirectly, the oscillation of flow and pressure is suppressed, while looking for environmental and economic benefits. The proposed strategy is based on the use of the model linear approximations and repeatedly solving of a quadratic optimization problem, providing solutions that improve at each iteration. In the event where the convergence of this algorithm is satisfied, the predicted values of the process variables are the same as to those obtained by the original nonlinear model, ensuring that the constraints are satisfied for them along the prediction horizon. A mathematical model recently published in the literature, capable of representing characteristics of severe slugging in a real oil well, is used both for simulation and for the project of the proposed controller, whose performance is compared to a linear MPC / A golfada ? um regime inst?vel do fluxo multif?sico, com oscila??es de press?o e vaz?o abruptas no processo de produ??o de petr?leo, podendo ocasionar problemas tais como vibra??o na tubula??o e alto n?vel de l?quido nos separadores. Pode ser classificada de acordo com seu local de ocorr?ncia. A mais severa destas, conhecida como golfada no riser, ocorre na tubula??o vertical que alimenta a plataforma. Conhecida tamb?m como golfada severa, ela ? capaz de causar bruscas oscila??es na press?o, nas vaz?es do processo, vibra??o excessiva, inunda??o dos tanques separadores, produ??o limitada, parada n?o programada da plataforma, entre outros aspectos negativos que motivaram a produ??o deste trabalho. Uma solu??o vi?vel para lidar com tal problema seria projetar um m?todo efetivo para a remo??o ou diminui??o deste regime, como um controlador. De acordo com a literatura, o controlador convencional PID n?o apresenta bons resultados devido ao alto grau de n?o linearidade do processo, o que impulsionou o desenvolvimento de t?cnicas avan?adas de controle. Dentre estas, o controlador preditivo, cuja a??o de controle resulta da solu??o de um problema de otimiza??o, al?m de ser uma t?cnica que apresenta robustez e pode incorporar restri??es f?sicas e/ou de seguran?a. O objetivo deste trabalho ? estudar a aplica??o de uma t?cnica de controle preditivo n?o linear ao controle de golfada severa, visando controlar a quantidade de massa l?quida no riser atuando na v?lvula de produ??o e, indiretamente, suprimir as oscila??es de vaz?o e press?o. Com a finalidade de obter benef?cios ambientais e econ?micos. A t?cnica de controle preditivo proposta baseia-se no uso de aproxima??es lineares do modelo e na resolu??o repetida de um problema de otimiza??o quadr?tica que proporciona solu??es que melhoram a cada itera??o. No caso em que a converg?ncia desse algoritmo ? satisfeita, os valores preditos das vari?veis do processo s?o iguais ?queles que seriam obtidos pelo modelo n?o linear original, garantindo que as restri??es nessas vari?veis sejam satisfeitas ao longo do horizonte de predi??o. Um modelo matem?tico publicado recentemente na literatura, capaz de representar caracter?sticas da golfada severa em um po?o real, ? utilizado tanto para a simula??o, quanto para projeto do controlador proposto, cujo desempenho ? comparado ao de um controlador preditivo linear

Dans cette thèse, nous étudions le problème de l'identification d'un système à dynamique linéaire. Dans un premier temps, nous répertorions les différentes méthodes qui ont été développées dans la littérature en nous concentrant plus particulièrement sur les méthodes des observateurs adaptatifs. Dans un second temps nous présentons un premier algorithme qui est une approche mixant les méthodes des sous-espaces et celles des observateurs adaptatifs. Ce nouvel algorithme est d'autant plus intéressant qu'il nous permet d'identifier des réalisations de systèmes MIMO dans une base d'état arbitraire. La convergence de cet algorithme est démontrée en utilisant les notions d'excitation persistantes. Dans un troisième chapitre nous introduisons une nouvelle méthode qui s'appuie sur le concept des observateurs de Luenberger non linéaires développés ces dernières années. Ce nouvel algorithme se différencie des algorithmes existants par sa capacité à produire une estimation simultanée des paramètres et de l'état du système. Nous démontrons alors sa robustesse à des perturbations affectant la dynamique interne ou les mesures. La convergence de cet algorithme est obtenue si les entrées du système satisfont une hypothèse d'excitation différentielle. Tous ces algorithmes sont alors évalués et implémentés sur un banc d'expérimentation / In this thesis, we study the problem of identification of a linear dynamical system. First, we survey various methods that have been developed in the literature. We focus more particularly on methods named adaptive observers. Secondly we present an approach which combines subspace identification methods and adaptive observers. This new method is interesting since it allows us to identify MIMO systems in an arbitrary basis. The convergence of this algorithm is demonstrated using the persistent excitation notions. In the third chapter we introduce a new method that is inspired from nonlinear Luenberger observers developed in recent years. This new algorithm is different from the existing algorithms since the parameters and the systemstatus are estimated simultaneously. We demonstrate the robustness of this approach. The convergence of the algorithm is obtained if the system inputs satisfy a differential excitation hypothesis. All these algorithms are evaluated and implemented on an experimental bench

[pt] Esta dissertação de mestrado analisa os efeitos de previsão de séries temporais com redes neurais em conjunto com a técnica de poda, denominada de Regularização Bayesiana. Utilizam-se diversas séries simuladas cujo processo gerador é de fato linear para comparar as previsões feitas por meio de modelos auto-regressivos lineares e redes neurais. Apresenta-se,ao final, uma comparação entre os modelos citados acima, segundo à eficiência preditiva de cada um. / [en] This paper studies the performance of neural networks estimated with Bayesian regularization to model and forecast time series where the data generations process is in fact linear. A simulation experiment is carried out to compare the forecast made by linear autoregressive models and neural networks.