Approximate Predictive Control of Ac Electric Drives

This thesis deals with the development of model predictive control of ac electric drives with long prediction horizons and yet with computational cost of the algorithm comparable to conventional cascade control approaches. The basic idea is to approximate the cost to go function of dynamic programming taking into account long prediction horizon. For specific cases (linear model and quadratic cost function) the optimization of the control problem can be solved analytically which leads to a dramatic computational cost reduction. However the analytical solution does not consider the hard state and input constraints. Therefore proposed control algorithm combines unconstrained solution with constraint manager. Resulting algorithm is very simple and computationally comparable to conventional control schemes, nonetheless it still preserves excellent control performance of MPC solved on long prediction horizon. Proposed control technique has been implemented in DSP and tested in three practical examples of drive control on laboratory prototype of PMSM drive of rated power 10.7 kW. The first case is described in chapter four and it is focused on cascade free speed control of PMSM using PWM. in this case, we aim to achieve comparable performance to existing PWM-based solutions of cascade free control at much lower computational cost. The second case is focused on speed control of PMSM using finite number of admissible control inputs. Existing one step FCS-MPC suffers from high distortion of the stator current. in this case we aim to enhance the current control performance at computational cost comparable to one-step ahead solution of FCS-MPC. The third Example focused on designing MPC for control of the Traction PMSM drive with input LC filter fed from dc centenary. The control problem is divided in two parts.i) stability of the input LC filter and ii) dynamics of the PMSM drive. Both parts are elegantly combined in the cost function of a simple one step FCS-MPC. The term respecting the Input LC filter is designed analytically taking into account long prediction horizons.