Parallel multi-level learning and optimization algorithms for control of cyclic processes on embedded systems

点击翻译按钮获取中文摘要

The goal of this subproject is to use innovative algorithmic ideas to enable the control of fast, cyclical processes in real time. The challenges associated with the real operation of the PCCI and GCAI engines in Aachen and Zurich in particular require new algorithmic developments and numerically favourable problem formulations that go beyond the results achieved in the first funding period. Relevant features of the realistic operation are, first of all, even tougher requirements on computation times. The higher rotational speeds (over 3000 rpm) lead to significantly shorter maximum computation times for the preparation and - with the same actuator delay - especially for the feedback phase (< 1 ms). Furthermore, the inner-cyclic control which is indispensable for real motor operation requires parallel calculations on an even smaller time scale.Secondly, real operation is characterized by time-dependent load profiles. For the GCAI process, an optimization-based reference generator should prevent controller convergence from being impaired in transient operation. The generation of feasible periodic references requires its own tailored algorithms and should run in parallel to control and estimation.Thirdly, it has been shown that the stochastic properties of the GCAI process vary across the engine characteristic map. Especially in case of late ignitions, the stochastic system behavior cannot be suppressed sufficiently with the control system designed so far. In order to guarantee a robust satisfaction of the system limitations, a robust variant of the optimization-based control needs to be developed. Since robust problem formulations are more challenging to solve, innovative algorithms and problem formulations have to be developed to meet the real-time requirements also with the robust MPC.