Conference Paper

A Novel Hybrid System for Dynamic Control.

DOI: 10.1007/11903697_77 Conference: Simulated Evolution and Learning, 6th International Conference, SEAL 2006, Hefei, China, October 15-18, 2006, Proceedings
Source: DBLP


In this paper we propose a hybrid model which includes both first principles differential equations and a least squares support
vector machine (LS-SVM). It is used to forecast and control an environmental process. This inclusion of the first principles
knowledge in this hybrid model is shown to improve substantially the stability of the model predictions in spite of the unmeasurability
of some of the key parameters. Proposed hybrid model is compared with both a hybrid neural network(HNN) as well as hybrid
neural network with extended kalman filter(HNN-EKF). From experimental results, proposed hybrid model shown to be far superior
when used for extrapolation compared to HNN and HNN-EKF.

5 Reads
  • [Show abstract] [Hide abstract]
    ABSTRACT: Methods are presented for the design of a static estimator which infers unmeasurable product qualities from secondary measurements. The secondary measurements are selected so as to minimize the number of such measurements required to obtain an accurate estimate which is insensitive to modeling errors. Unlike previous work, the number of secondary measurements can be fewer than the number of unmeasured disturbances. If the statistics of the disturbances and/or measurement noise are available, this information can be incorporated into the design procedure to obtain an optimal static estimator.The design procedure is illustrated by application to a simulated industrial debutanizer. Data for the simulation were supplied by the Marathon Oil Company. Deviations in bottoms product quality are compared for the current control policy (maintenance of a stage temperature at its set point) and the inferential control system with the column subjected to representative feed composition disturbances. Results show that inferential control based on four, five, or six tray temperature measurements improves the steady state control performance by as much as 400%.
    AIChE Journal 05/1978; 24(3):485 - 492. DOI:10.1002/aic.690240313 · 2.75 Impact Factor
  • [Show abstract] [Hide abstract]
    ABSTRACT: The last few years have seen the proliferation of inferential estimation methodologies. These are techniques which are capable of providing estimates of `difficult-to-measure' fermentation process variables from other easily available process measurements. In the literature, two avenues tend to predominate: on being based upon the use of physicochemical models, while the other makes use of an input-output or `black-box' model structure. Clearly, those based upon an accurate physical description will offer the best performance. In reality, however, the best possible situation will be the availability of a `reasonable' process model. Alternatively, at the other end of the spectrum of process knowledge, is the use of `adaptive' generically structured observers. Firstly, a generally structured linear estimator will be introduced. Successful applications of this estimator to an idustrial mycelial fermentation will be presented. Such success is due to the continuous nature of the process, which undergoes smaller changes in dynamics than, say, a batch fermentation. For such cases it is felt that more representative nonliner observers are required. A technique now receiving recognition is that of artificial neural networks. The neural network paradigm is discussed and a comparison with the performance of the adaptive linear estimator is presented.
    American Control Conference, 1990; 06/1990
  • [Show abstract] [Hide abstract]
    ABSTRACT: A neural-network based model of the kinetics in a fermentation process is presented. In the development and application of the model, the state of the process was estimated by an independent model based on elemental balances and an electroneutrality condition. The aim of the work was a model that describes the overall kinetic relations well enough to be used for on-line simulation and prediction. Both feed-forward and partially recurrent networks were investigated. The networks were trained off-line, but in the on-line application they are adapted to the process conditions by adjusting a set of correction factors. The approach is illustrated on a number of batch runs of Saccharomyces cerevisiae, where the networks are demonstrated to be able to accurately describe the biphasic growth.On présente un modèle de cinétique basé sur un réseau neuronal pour un procédé de fermentation. Dans la mise au point et l'application du modèle, l'état du procédé a été estimé par un modèle indépendant basé sur les équilibres élémentaires et une condition d'électroneutralité. Le but de travail était de trouver un modèle qui décrive suffisamment bien les relations de cinétique globales pour leur utilisation dans la simulation et la prédiction en ligne. Les réseaux à boucle de régulation aval et les réseaux partiellement récurrents ont été étudiés. Les réseaux ont été entrainées hors ligne, mais pour l'application en ligne, on peut les adapter aux conditions du procédé en ajustant un jeu de facteurs de correction. La méthode est illustrée pour plusieurs essais discontinus de Saccharomyces cerevisiae, où on démontre que les réseaux peuvent décrire avec précision la croissance diphasique.
    The Canadian Journal of Chemical Engineering 02/1996; 74(1):124 - 131. DOI:10.1002/cjce.5450740116 · 1.23 Impact Factor