E. C. Biscaia Jr., A. R. Secchi, L. S. Santos Programa de Engenharia Química (PEQ) – COPPE – UFRJ Rio de Janeiro - Brazil Dynamic Optimisation Using Wavelets Bases
Aims of the Contribution Improve numerical methods for solving dynamic optimisation problems: s.t
Sequential Method Control variables are discretized and dynamic model is solved numerically at each iteration of the NLP
Sequential Method Control variables are discretized and dynamic model is solved numerically at each iteration of the NLP
Sequential Method discretization in time domain ns stages Control variables are discretized and dynamic model is solved numerically at each iteration of the NLP
Sequential Method control profile (parameterization)
Sequential Method decision variables
Sequential Method decision variables
Sequential Method decision variables NLP solver Calculates optimal control profile
Sequential Method decision variables NLP solver Calculates optimal control profile Successive Refinement Initial profile NLP solver Refinement NLP solver
Wavelets Sequential Method NLP solver
Wavelets Sequential Method NLP solver Wavelets Improving Adaptation of discrete points at each iteration
Wavelets Sequential Method NLP solver Wavelets Improving Adaptation of discrete points at each iteration
Wavelets Sequential Method NLP solver Wavelets new mesh Improving Adaptation of discrete points at each iteration
Wavelets Sequential Method NLP solver Wavelets new mesh Improving Adaptation of discrete points at each iteration
Wavelets Analysis Considering a function, it can be transformed into wavelet domain as: details
Wavelets Analysis Considering a function, it can be transformed into wavelet domain as: details control variable Inner product
Wavelets Analysis Considering a function, it can be transformed into wavelet domain as: where is the maximum level resolution. details control variable Vector of wavelets details Inner product ResolutionPosition
Wavelets Analysis Haar wavelet has been considered:
Wavelets Analysis Haar wavelet has been considered:
Wavelets Analysis Haar wavelet has been considered: Orthogonal basis
Wavelets Analysis
Control profile NLP solver Wavelets Analysis NLP solver
Control profile NLP solver How Wavelets Work NLP solver Wavelets Iteration 1 Iteration 2
Wavelets Thresholding Analysis details
Wavelets Thresholding Analysis Thresholding: some details are eliminated. details
Wavelets Thresholding Analysis New thresholded control profile Thresholding: some details are eliminated. details
Thresholding strategies Thresholding: (i)decomposition of the data ; (ii)comparing detail coefficients with a given threshold value and shrinking coefficients close to zero, eliminating data noise effects (DONOHO and JOHNTONE, 1995):
Thresholding strategies Thresholding: (i)decomposition of the data ; (ii)comparing detail coefficients with a given threshold value and shrinking coefficients close to zero, eliminating data noise effects (DONOHO and JOHNTONE, 1995): Visushrink (DONOHO, 1992): standard deviation of a white noise details coefficients
Thresholding strategies Thresholding: (i)decomposition of the data ; (ii)comparing detail coefficients with a given threshold value and shrinking coefficients close to zero, eliminating data noise effects (DONOHO and JOHNTONE, 1995): Visushrink (DONOHO, 1992): Fixed user specified (SCHLEGEL and MARQUARDT,2004 and BINDER, 2000): standard deviation of a white noise details coefficients
How Wavelets Work Control profile NLP solver Wavelets Sequential Algorithm Incorporate the Visushrink threshold procedure and compare with other fixed threshold parameters; Observe if the CPU is affected by changes of threshold rule. Improve, at each iteration, the estimate of control profile.
Algorithm and Parameters 1.Integrator: Runge Kutta fourth order (ode45 Matlab); 2.Optimisation: Interior Point (Matlab) was used as NLP solver; 3.Wavelets: Routines of Matlab 7.6; 4.Stop Criteria
Flowsheet of Wavelet Refinement Algorithm Example: Semi-batch Isothermal Reactor (SRINIVASAN et al.,2003) Constant by parts interpolation
Flowsheet of Wavelet Refinement Algorithm Example: Semi-batch Isothermal Reactor (SRINIVASAN et al.,2003)
Flowsheet of Wavelet Refinement Algorithm Example: Semi-batch Isothermal Reactor (SRINIVASAN et al.,2003)
Flowsheet of Wavelet Refinement Algorithm Visushrink ThresholdFixed Threshold Example: Semi-batch Isothermal Reactor (SRINIVASAN et al.,2003)
Flowsheet of Wavelet Refinement Algorithm Visushrink ThresholdFixed Threshold Example: Semi-batch Isothermal Reactor (SRINIVASAN et al.,2003)
Flowsheet of Wavelet Refinement Algorithm Visushrink ThresholdFixed Threshold Example: Semi-batch Isothermal Reactor (SRINIVASAN et al.,2003)
Flowsheet of Wavelet Refinement Algorithm Visushrink ThresholdFixed Threshold Locations of discontinuity points ~ large details coefficients Example: Semi-batch Isothermal Reactor (SRINIVASAN et al.,2003)
Flowsheet of Wavelet Refinement Algorithm Visushrink ThresholdFixed Threshold Example: Semi-batch Isothermal Reactor (SRINIVASAN et al.,2003)
Case Studies
Semi-batch Isothermal Reactor (SRINIVASAN et al., 2003)) Optimal Control Profile: 128 stages
Control profile evolution
Results: Semi-batch Isothermal Reactor Uniform mesh with 128 stages Fixed Threshold Fixed Threshold Fixed Threshold Visushirink Iter. Total CPU ns Total CPU ns Total CPU ns Total CPU ns Total CPU ns Reference CPU time: Uniform mesh
Bioreactor problem (CANTO et al., 2001)) Optimal Control Profile: 128 stages M: monomer S: substrate
Control profile evolution
Results: Bioreactor problem ) Uniform mesh with 128 stages Fixed Threshold Fixed Threshold Fixed Threshold Visushirink Iter. Total CPU ns Total CPU ns Total CPU ns Total CPU ns Total CPU ns Reference CPU time: Uniform mesh
Mixture of Catalysts (BELL and SARGENT, 2000) Optimal Control Profile: 64 stages
Control profile evolution
Results: Bioreactor problem Uniform mesh with 64 stages Fixed Threshold Fixed Threshold Visushirink Iter. Total CPU ns Total CPU ns Total CPU ns Total CPU ns Reference CPU time: Uniform mesh
Conclusions 1.Two threshold procedures were analyzed here: Fixed and Visuhrink (level dependent threshold). According to the results, Visushrink strategy is able to denoise the details in a more efficient way than Fixed strategy, that is more conservative. We have shown that the choice of a threshold procedure can improve the wavelet adaptation algorithm;
Conclusions 1.Two threshold procedures were analyzed here: Fixed and Visuhrink (level dependent threshold). According to the results, Visushrink strategy is able to denoise the details in a more efficient way than Fixed strategy, that is more conservative. We have shown that the choice of a threshold procedure can improve the wavelet adaptation algorithm; 2.Other examples from literature ( BINDER et al., 2000; SRINIVASAN et al., 2003; SCHLEGEL, 2004) was solved and have been presented similar results: a considerable improvement of CPU time when Visushrink is used.
Conclusions 1.Two threshold procedures were analyzed here: Fixed and Visuhrink (level dependent threshold). According to the results, Visushrink strategy is able to denoise the details in a more efficient way than Fixed strategy, that is more conservative. We have shown that the choice of a threshold procedure can improve the wavelet adaptation algorithm; 2.Other examples from literature ( BINDER et al., 2000; SRINIVASAN et al., 2003; SCHLEGEL, 2004) was solved and have been presented similar results: a considerable improvement of CPU time when Visushrink is used. 3.Other wavelets thresholding strategies (as Sureshrink and Minimaxi) has been investigated, however in some cased these strategies have undesirable results, with worse performance than Fixed strategy.
Future Works 1.This algorithm will be used to solve more complex problems with several control variables in order to improve the sequential adaptation of each control profile. Our expectation is to observe the intensification of threshold influence for these problems.
Future Works 1.This algorithm will be used to solve more complex problems with several control variables in order to improve the sequential adaptation of each control profile. Our expectation is to observe the intensification of threshold influence for these problems. 2.As observed here, wavelets are able to detect discontinuity points and therefore the location of different control arcs. A more sophisticated interpolation of control profile will be implemented in these regions with aims to reduce he number of stages and consequently decision variables.
Thank You