Computationally efficient direct model predictive control with long prediction horizons for medium voltage drives.
Final Report Abstract
In the scope of this project a long-horizon finite control set model predictive control (FCS-MPC) algorithm for medium voltage drives was investigated. During steady-state operation this algorithm operates the motor at low switching frequencies for reduced switching losses while the total harmonic distortion (THD) of the current is kept as low as possible for reduced iron and copper, and thus thermal, losses in the machine. To address the computational issues that arise from the implemented long prediction horizon, modifications to a discrete search approach called “sphere decoder” were suggested. It was shown that this smart branch-and-bound technique can be further optimized. In doing so, its computational complexity can be reduced by more than 75% when long horizons are considered. In addition, the sphere decoding algorithm (SDA) was further refined to consider output constraints by translating them into input constraints. Thanks to this modification, the sphere decoder can still perform effectively, while it manages to find the feasible optimal solution that guarantees safe operation since, e.g., any overcurrents that can damage the power electronic converter are avoided. As a result, the imposed constraints are fully respected allowing the plant to operate at its physical limits without deteriorating its performance. Additional modifications to the SDA were proposed to include nonlinearities—in the form of balancing the neutral point potential of a three-level inverter neutral point clamped inverter—in the control problem. Further, due to the direct control principle of the designed controller, which implies variable switching frequency, a method to adjust the switching frequency of the drive system was suggested, so as to keep it constant over the whole operating regime. This method which does not require recalculation of the offline matrices, thus keeping the associated computational load at bay. Subsequently a field programmable gate array (FPGA) was used to implement the algorithm in real time. Several optimization methods were proposed to reduce the latency and resource usage of the implemented algorithm. As a result, the real-time implementation of the control algorithm in question was rendered feasible. An FPGA in the loop simulation validated that the computational burden of the algorithm can be managed in real time.
Publications
-
“Computationally efficient sphere decoding for long-horizon direct model predictive control”, in Proc. IEEE Energy Convers. Congr. Expo., Milwaukee, WI, Sep. 2016
P. Karamanakos, T. Geyer, and R. Kennel
-
“Constrained long-horizon direct model predictive control for power electronics”, in Proc. IEEE Energy Convers. Congr. Expo., Milwaukee, WI, Sep. 2016
P. Karamanakos, T. Geyer, and R. Kennel
-
“Long-horizon direct model predictive control with active balancing of the neutral point potential”, in Proc. Workshop on Pred. Control of Elect. Drives and Power Electron., Pilsen, Czech Republic, Sep. 2017
E. Liegmann, P. Karamanakos, T. Geyer, T. Mouton, and R. Kennel
-
“A modified sphere decoder for online adjustment of the switching frequency”, in Proc. IEEE Ind. Electron. Conf., 2019
Q. Yang, E. Liegmann, P. Karamanakos, and R. Kennel
-
“Ultrazohm—A powerful real-time computation platform for MPC and multi-level inverters”, in Proc. Workshop on Pred. Control of Elect. Drives and Power Electron., Quanzhou, China, May 2019, pp. 1–6
S. Wendel, A. Geiger, E. Liegmann, et al.
-
“High-level synthesis of a long horizon model predictive control algorithm for an FPGA”, PCIM Europe digital days 2020; International Exhibition and Conference for Power Electronics, Intelligent Motion, Renewable Energy and Energy Management, Germany, 2020, pp. 1-8
S. A. B. Khalid, E. Liegmann, P. Karamanakos, and R. Kennel
