A carregar...
Publicação
Controlo Automático Do Efeito De Magnus Sobre Um Cilindro
| Nome: | Descrição: | Tamanho: | Formato: | |
|---|---|---|---|---|
| 2.04 MB | Adobe PDF |
Autores
Orientador(es)
Resumo(s)
Nos campos da ciência e engenharia, uma parte significativa dos sistemas é governada
por equações diferenciais não lineares, conferindo-lhes uma complexidade substancial. O con-
trolo eficaz desempenha um papel crítico em várias áreas, nomeadamente na Mecânica de
Fluidos devido à sua presença nas tecnologias da produção de energia, dos sistemas de trans-
porte e das aplicações médicas. Contudo, o controlo de escoamentos instáveis representa um
desafio significativo, tornando-se muitas vezes uma tarefa inviável quando se recorre a méto-
dos tradicionais. Neste contexto, os avanços em
Machine Learning (ML) permitem extrair pa-
drões que podem ser explorados sem a necessidade de conhecer todo o estado do sistema.
Nesta pesquisa, o objetivo consiste em definir uma solução eficiente e precisa para mo-
delar, prever e controlar o escoamento sobre um cilindro em rotação para um número de
Reynolds de 100. Em vez de realizar previsões do estado completo do sistema, utilizou-se uma
rede neuronal do tipo
Nonlinear Autoregressive Model with Exogenous Inputs (NARX) capaz
de prever uma série temporal composta por diversos valores de coeficientes de sustentação.
Posteriormente, a abordagem seguida baseou-se em modelos preditivos através da incorpo-
ração da rede numa estrutura de Controlo Preditivo para realizar efetivamente o seu controlo.
Verificou-se que a abordagem centrada na Aprendizagem Profunda teve destaque não só pela
capacidade de previsão através da rede NARX, como também pela integração da estrutura de
Controlo Preditivo num sistema complexo. Assim, foi possível controlar os seguintes conjuntos
de valores de coeficiente de sustentação, 𝐶𝐿= [1, 0, -1]; [0,5; -0,25; -0,5] e [-1]. A procura de
técnicas mais eficientes que substituam simulações computacionalmente dispendiosas é es-
sencial para melhorar o projeto e otimização de sistemas complexos.
In the fields of science and engineering, a significant proportion of systems are governed by non-linear differential equations, giving them substantial complexity. Effective control plays a critical role in several areas, notably fluid mechanics due to its presence in energy production technologies, transportation systems and medical applications. However, the control of unsta- ble flows represents a significant challenge, often becoming an unfeasible task when traditional methods are used. In this context, advances in Machine Learning (ML) make it possible to ex- tract patterns that can be explored without the need to know the entire state of the system. In this research, the aim is to define an efficient and accurate solution for modeling, fore- casting and controlling the flow over a rotating cylinder for a Reynolds number of 100. Instead of making predictions of the complete state of the system, a neural network of the Nonlinear Autoregressive Model with Exogenous Inputs (NARX) type was used, capable of predicting a time series made up of various values of support coefficients. Subsequently, the approach fol- lowed was based on predictive models by incorporating the network into a Predictive Control structure to effectively control it. It was found that the Deep Learning approach stood out not only for its predictive capacity through the NARX network, but also for the integration of the Predictive Control structure into a complex system. It was therefore possible to control the following sets of lift coefficient values, CL= [1, 0, -1]; [0.5; -0.25; -0.5] and [-1]. The search for more efficient techniques to replace computationally expensive simulations is essential to im- prove the design and optimization of complex systems.
In the fields of science and engineering, a significant proportion of systems are governed by non-linear differential equations, giving them substantial complexity. Effective control plays a critical role in several areas, notably fluid mechanics due to its presence in energy production technologies, transportation systems and medical applications. However, the control of unsta- ble flows represents a significant challenge, often becoming an unfeasible task when traditional methods are used. In this context, advances in Machine Learning (ML) make it possible to ex- tract patterns that can be explored without the need to know the entire state of the system. In this research, the aim is to define an efficient and accurate solution for modeling, fore- casting and controlling the flow over a rotating cylinder for a Reynolds number of 100. Instead of making predictions of the complete state of the system, a neural network of the Nonlinear Autoregressive Model with Exogenous Inputs (NARX) type was used, capable of predicting a time series made up of various values of support coefficients. Subsequently, the approach fol- lowed was based on predictive models by incorporating the network into a Predictive Control structure to effectively control it. It was found that the Deep Learning approach stood out not only for its predictive capacity through the NARX network, but also for the integration of the Predictive Control structure into a complex system. It was therefore possible to control the following sets of lift coefficient values, CL= [1, 0, -1]; [0.5; -0.25; -0.5] and [-1]. The search for more efficient techniques to replace computationally expensive simulations is essential to im- prove the design and optimization of complex systems.
Descrição
Palavras-chave
Controlo de Escoamento Machine Learning Nonlinear Autoregressive Model With Exogenous Inputs Controlo Preditivo Coeficiente de sustentação