TY - JOUR
T1 - Affine Neural Network-Based Predictive Control Applied to a Distributed Solar Collector Field
AU - Gil, Paulo José Carrilho de Sousa
AU - Henriques, Jorge H.
AU - Cardoso, Alberto J.L.
AU - Carvalho, Paulo Roberto Antonacci
AU - Dourado, Antonio
N1 - The experiments included in this paper have been carried out within the project Improving Human Potential Program (EC-DGXII), supported by the European Union Program Training and by the Mobility of Researchers. The authors would like to thank the staff at the Plataforma Solar de Almeria.
PY - 2014/2
Y1 - 2014/2
N2 - This paper presents experimental results concerning the control of a distributed solar collector field, where the main objective concerns the regulation of the outlet oil temperature by suitably manipulating the oil flow rate. This is achieved by means of a constrained nonlinear adaptive model-based predictive control framework where the control action sequence is obtained by solving an open-loop optimization problem, subject to a set of constraints. The plant dynamics is approximated by an affine state-space neural network, whose complexity is specified in terms of the cardinality of dominant singular values associated with a subspace oblique projection of data-driven Hankel matrices. The neural network is first trained offline and subsequently improved through a recursive updating of its weights and biases, based on a dual unscented Kalman filter. The control scheme is implemented on the Acurex field of the Plataforma Solar de Almería, Spain. Results from these experiments demonstrate the feasibility of the proposed framework, and highlight the ability to cope with time-varying and unmodeled dynamics, under the form of disturbances, and its inherent capability for accommodating actuation faults.
AB - This paper presents experimental results concerning the control of a distributed solar collector field, where the main objective concerns the regulation of the outlet oil temperature by suitably manipulating the oil flow rate. This is achieved by means of a constrained nonlinear adaptive model-based predictive control framework where the control action sequence is obtained by solving an open-loop optimization problem, subject to a set of constraints. The plant dynamics is approximated by an affine state-space neural network, whose complexity is specified in terms of the cardinality of dominant singular values associated with a subspace oblique projection of data-driven Hankel matrices. The neural network is first trained offline and subsequently improved through a recursive updating of its weights and biases, based on a dual unscented Kalman filter. The control scheme is implemented on the Acurex field of the Plataforma Solar de Almería, Spain. Results from these experiments demonstrate the feasibility of the proposed framework, and highlight the ability to cope with time-varying and unmodeled dynamics, under the form of disturbances, and its inherent capability for accommodating actuation faults.
KW - constrained optimization
KW - affine state-space neural networks
KW - unscented Kalman filter
KW - distributed solar collector field (DSC)
KW - model-based predictive control (MPC)
KW - online training
KW - Adaptive control
KW - Adaptive control
KW - affine state-space neural networks
KW - constrained optimization
KW - distributed solar collector field (DSC)
KW - model-based predictive control (MPC)
KW - online training
KW - unscented Kalman filter
U2 - 10.1109/TCST.2013.2260545
DO - 10.1109/TCST.2013.2260545
M3 - Article
SN - 1063-6536
VL - 22
SP - 585
EP - 596
JO - IEEE Transactions on Control Systems Technology
JF - IEEE Transactions on Control Systems Technology
IS - 2
ER -