Publicação
Controlo Difuso Predictivo para Aplicação em Processos Industriais
| Resumo: | This dissertation studies the application of predictive control, and models based in recurrent fuzzy neural networks, to the control of non-linear processes. The studied control system is based in predictive control where it is used the GPC (Generalized Predictive Control) algorithm, already known in the industry. This algorithm is part of the model based controllers family, as it needs to have a model of the process under control. The predictive part is due to it's use of the model to estimate the future behaviour of the process, and then determine the best control effort based on a cost function that penalizes the future error of the process output, and also penalizes the global control effort.A discrete-time model updated in real time, is fed to the controller algorithm. This model is implemented by a Recurrent Fuzzy Neural Network (RFNN). This network is built with a set of fuzzy rules where the consequent of each rule is a linear model. These linear models are activated by the fuzzy rules and are aggregated in order to produce a resultant non-linear model of the process to be used by the GPC. The antecedent parts of the rules in the recurrent fuzzy neural network are updated with the back-propagation algorithm, and the consequent parts are updated with the recursive least squares algorithm.In this dissertation two controllers are tested based on this structure. The first one implements simplifications on the control law in order to eliminate the need to determine the inverse of a matrix in real-time.The second controller, is a new controller proposed in this dissertation. This controller replaces the recursive least squares algorithm consequent adaptation law, with a new one that is designed to adjust the model parameters, while limiting that the system output following error is bounded. The proposed controller was published in a conference paper.The developed controllers are tested in simulation environment, and also in a laboratory environment. In the simulation environment tests, where used three models. One represents a water tank, where the system input is the DC motor input voltage, and the system output is the water level inside the water tank. The other two models used represent abstract systems.The real world process consists of a electrical motor coupled to a generator. The system input is the motor supply voltage, while the system output is the angular velocity of the motor shaft. The Generator serves the purpose of being a perturbation to the system. |
|---|---|
| Autores principais: | Sousa, Nuno Miguel Santos de |
| Assunto: | Sistemas Difusos Sistemas neuro-difusas Controlo Preditivo Sistemas Não Lineares Identificação Fuzzy Systems Neuro-Fuzzy Systems Predictive Control Nonlinear Systems Identification |
| Ano: | 2016 |
| País: | Portugal |
| Tipo de documento: | dissertação de mestrado |
| Tipo de acesso: | acesso aberto |
| Instituição associada: | Universidade de Coimbra |
| Idioma: | português |
| Origem: | Estudo Geral - Universidade de Coimbra |
| Resumo: | This dissertation studies the application of predictive control, and models based in recurrent fuzzy neural networks, to the control of non-linear processes. The studied control system is based in predictive control where it is used the GPC (Generalized Predictive Control) algorithm, already known in the industry. This algorithm is part of the model based controllers family, as it needs to have a model of the process under control. The predictive part is due to it's use of the model to estimate the future behaviour of the process, and then determine the best control effort based on a cost function that penalizes the future error of the process output, and also penalizes the global control effort.A discrete-time model updated in real time, is fed to the controller algorithm. This model is implemented by a Recurrent Fuzzy Neural Network (RFNN). This network is built with a set of fuzzy rules where the consequent of each rule is a linear model. These linear models are activated by the fuzzy rules and are aggregated in order to produce a resultant non-linear model of the process to be used by the GPC. The antecedent parts of the rules in the recurrent fuzzy neural network are updated with the back-propagation algorithm, and the consequent parts are updated with the recursive least squares algorithm.In this dissertation two controllers are tested based on this structure. The first one implements simplifications on the control law in order to eliminate the need to determine the inverse of a matrix in real-time.The second controller, is a new controller proposed in this dissertation. This controller replaces the recursive least squares algorithm consequent adaptation law, with a new one that is designed to adjust the model parameters, while limiting that the system output following error is bounded. The proposed controller was published in a conference paper.The developed controllers are tested in simulation environment, and also in a laboratory environment. In the simulation environment tests, where used three models. One represents a water tank, where the system input is the DC motor input voltage, and the system output is the water level inside the water tank. The other two models used represent abstract systems.The real world process consists of a electrical motor coupled to a generator. The system input is the motor supply voltage, while the system output is the angular velocity of the motor shaft. The Generator serves the purpose of being a perturbation to the system. |
|---|
Atividades financiadas
Não foram encontrados projetos financiados.