The first part of the talk deals with Stochastic Model Predictive Control (SMPC) for discrete-time linear systems subject to additive disturbances with chance constraints on the states and hard constraints on the inputs. Current chance constrained MPC methods tend to be conservative partly because they fail to exploit the predefined violation level in closed-loop. For many practical applications, this conservatism can lead to a loss in performance. We propose an adaptive SMPC scheme that starts with a standard conservative chance constrained formulation and then on-line adapts the formulation of constraints based on the empirical violation level. Using martingale theory we establish guarantees of convergence to the desired level of constraint violation in closed-loop for a special class of linear systems.

In the second part of the talk grid frequency control with buildings is considered. The increasing penetration of fluctuating renewable energy sources leads to more demand for frequency control reserves. Frequency control reserves are traditionally provided by generators in order to ensure that supply-demand mismatches in the electricity grid can be mitigated at all times. It is commonly argued that more demand-side participation for reserve provision would be desirable. In this work we focus on frequency control reserve provision by heating, ventilation, and air conditioning systems of an office building aggregation. We first propose a hierarchical model predictive control (MPC) formulation to robustly provide frequency control reserves while satisfying comfort constraints in the buildings. We then investigate the aggregator's problem of bidding day-ahead in the frequency control reserve market, which we formulate as bilevel optimization problem. The bilevel problem is reformulated and solved as mixed-integer linear program (MILP).

 

Unabhängig von GPS kann und wird der Pilot eines Luftfahrzeugs sich auch an auffälligen Landmarken orientieren. Dabei werden in dicht besiedelten Gebieten wie Deutschland große künstliche Strukturen verwendet, wie etwa Autobahnkreuze. Da stellt sich dann die Frage nach der Automatisierung der visuellen Landmarkenerkennung. Neben den Ansätzen der lernenden Objekterkennung kommen gerade für vom Menschen gemachte Infrastrukturbauwerke wissensbasierte Ansätze in Frage, die sich statt auf (sehr viele) Lernbeispiele auf (wenige) Regeln stützen, die die Eigenschaften solcher Strukturen und ihres Erscheinens in Luftbildern beschreiben. Vor allen Dingen ist dann die Semantik solcher Systeme zu diskutieren, also inwieweit korrekte Bedeutungszuweisungen zustande kommen und wie oft Fehlinterpretationen vorkommen. Dazu allerdings sind realistische Beispieldaten erforderlich.

Virtuelle Globus Systeme wie GoogleEarth können hierzu als Datenquelle, also als Kamerasimulator, dienen. Sie werden eingebunden in eine Active Vision Schleife. Dabei wird die Abdrift des UAVs mit dem Zufallsgenerator bestimmt, und das wissensbasierte Erkennungssystem dient dann wieder der Kurskorrektur. Der Pfad, eine Kette von Landmarken, wird vorher festgelegt. Am Ende kann man abzählen, wie oft man nicht am Ziel angekommen ist und analysieren, was schief gegangen ist. Man kann also eine recht repräsentative semantisch/-statistische Analyse durchführen, ohne tatsächlich ein UAV zu fliegen.

 

The lecture will deal with the fault detection problem, which is a very important part of automatic control. Passive and active fault detection approaches that make it possible to increase safety and reliability and to reduce maintenance costs will be explained. The active fault detector, contrary to the passive one, generates not only decisions but also the input signal to the observed or controlled system. The stress will be laid on Active Fault Detection and Control (AFDC), which represents a challenging research topic and provides higher quality of detection and control compared to passive approaches.

The talk will start with a brief introduction to passive and active fault detection systems. A unified formulation of the active fault detection and control problem for discrete-time stochastic systems and its optimal solution will be proposed. The design of the active fault detector and controller is formulated similarly as in the stochastic optimal control problem, where controller design can be carried out using three different assumptions on measurements availability at individual time steps. These three assumptions, denoted as the Open Loop (OL), Open Loop Feedback (OLF) and Closed Loop (CL) information processing strategies, can be used equally well in AFDC. The general formulation includes several important special cases depending on how the detection and control aims are preferred. These special cases will be introduced as well.

 

The lecture will deal with numerical integrations methods with a focus on their utilization in local nonlinear filters. Traditionally, the integrals corresponding to state and measurement predictive moments are numerically solved by deterministic rules such as the Gauss-Hermit rule or the cubature rule or stochastic rules such as the perfect Monte Carlo or the importance sampling.

In the lecture, a promising alternative to these rules will be introduced, which represents a compromise between deterministic and stochastic rules and takes advantage from both groups. Utilization of this stochastic rule in the Gaussian filter leads to the stochastic integration filter. Its usage in nonlinear state estimation will be shown and illustrated in a numerical example.