Stochastic Model Predictive Control (SMPC) uses a probabilistic approach to take care of disturbances and model uncertainties. One key point is to enforce a prescribed probability for the system to exceed a given constraint. While this probability must not be exceeded, an effective method should approximate it at the best. The inherent mathematical complexity of the probabilistic approach and the related approximations lead often to a violation level far below the prescribed one, a phenomenon called conservativity. This Thesis proposes several approaches to reduce conservativity in a class of SMPC methods called analytical. The first is to reformulate the method in a less abstract form, introducing some reasonable assumption on the noise distribution. Two new reformulated constraints, sharper than the original one and based on these assumptions, are added to the analytical method. The simulations performed show that a significant (77%-83%) reduction of conservativity is achieved, while keeping a still quite arbitrary form of the noise distribution. Then, starting from the observation that the violation level is increased by the fatness of the tails of the noise distribution, the analytical method is adapted to fat-tailed distributions, solving the related problem of infinite moments that hinders the use of these distributions. A reinterpretation in terms of noise with finite and properly selected variance, valid for the use with the analytical method, is derived for these distributions. The simulations performed show that, for every reformulated constraint of the analytical method, a fat-tailed distribution always exists that produces the desired violation level. Finally, an empirical technique is presented, based on the lessening the constraints. Dually with respect to the fat-tailed approach, a distribution reshaped to be thinner than the actual noise one is generated and used to set the constraints in the analytical method. Empirical relationships tune the reshaping to produce optimal results in terms of conservativity. The simulations show that this approach gives optimal results even with the most general constraints, avoiding any further assumption.
Il controllo predittivo stocastico, o Stochastic Model Predictive Control (SMPC) usa un approccio probabilistico per gestire i disturbi e le incertezze del modello. Un punto chiave è imporre al sistema una probabilità prestabilita di non superare un dato vincolo; un metodo efficace dovrebbe approssimarla al meglio. La complessità matematica inerente all’approccio probabilistico e le conseguenti approssimazioni portano spesso a livelli di violazione molto più bassi di quelli desiderati, un fenomeno chiamato conservatività. Questa Tesi propone diversi approcci per ridurre la conservatività per i metodi SMPC chiamati analitici. Il primo è riformulare il metodo in una forma meno astratta, introducendo alcune ragionevoli ipotesi sulla distribuzione del rumore. Sono stati aggiunti al metodo analitico due nuovi vincoli riformulati, basati su queste ipotesi. Le simulazioni eseguite mostrano una significativa riduzione di conservatività (77%-83%), pur mantenendo una forma piuttosto generale della distribuzione del rumore. Quindi, partendo dall’osservazione che il livello di violazione è accresciuto dalla grassezza delle code della distribuzione del rumore, il metodo analitico viene adattato alle distribuzioni fat-tailed, risolvendo il problema connesso dei momenti infiniti, che ostacola l’uso di queste distribuzioni. Per esse, si ricava una reinterpretazione in termini di rumore con varianza finita e opportunamente formulata, valida per l’uso con il metodo analitico. Le simulazioni eseguite mostrano che, per ogni vincolo riformulato del metodo analitico, esiste sempre una distribuzione fat-tailed che produce il livello di violazione desiderato. Infine, viene presentata una tecnica empirica, basata sul rilassamento dei vincoli. Con un approccio duale rispetto a quello fat-tailed, viene generata una distribuzione più fine del rumore reale, poi usata per costruire i vincoli del metodo analitico. Relazioni empiriche calibrano la modifica della distribuzione per produrre risultati ottimali per ridurre la conservatività. Le simulazioni mostrano che questo approccio da risultati ottimali anche con i vincoli più generali, evitando l’uso di ulteriori ipotesi.
Constraint selection in analytical stochastic model predictive control
FICICCHIA, ALBERTO
2015/2016
Abstract
Stochastic Model Predictive Control (SMPC) uses a probabilistic approach to take care of disturbances and model uncertainties. One key point is to enforce a prescribed probability for the system to exceed a given constraint. While this probability must not be exceeded, an effective method should approximate it at the best. The inherent mathematical complexity of the probabilistic approach and the related approximations lead often to a violation level far below the prescribed one, a phenomenon called conservativity. This Thesis proposes several approaches to reduce conservativity in a class of SMPC methods called analytical. The first is to reformulate the method in a less abstract form, introducing some reasonable assumption on the noise distribution. Two new reformulated constraints, sharper than the original one and based on these assumptions, are added to the analytical method. The simulations performed show that a significant (77%-83%) reduction of conservativity is achieved, while keeping a still quite arbitrary form of the noise distribution. Then, starting from the observation that the violation level is increased by the fatness of the tails of the noise distribution, the analytical method is adapted to fat-tailed distributions, solving the related problem of infinite moments that hinders the use of these distributions. A reinterpretation in terms of noise with finite and properly selected variance, valid for the use with the analytical method, is derived for these distributions. The simulations performed show that, for every reformulated constraint of the analytical method, a fat-tailed distribution always exists that produces the desired violation level. Finally, an empirical technique is presented, based on the lessening the constraints. Dually with respect to the fat-tailed approach, a distribution reshaped to be thinner than the actual noise one is generated and used to set the constraints in the analytical method. Empirical relationships tune the reshaping to produce optimal results in terms of conservativity. The simulations show that this approach gives optimal results even with the most general constraints, avoiding any further assumption.File | Dimensione | Formato | |
---|---|---|---|
2016_12_Ficicchia.pdf
accessibile in internet per tutti
Descrizione: Testo della tesi
Dimensione
1.61 MB
Formato
Adobe PDF
|
1.61 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/131992