Estrategias de optimización del control predictivo de un convertidor multinivel-NPC y su implementación en FPGA

Resumo

Multilevel converters are well-established alternative used in high-power and high or medium voltage systems as wind energy conversion, generation, transmission and distribution of energy to industrial applications. Several control strategies have been developed for these converters as Space Vector Modulation (SVM) or Selective Harmonic Elimination (SHE) to name the most commonly used. However, Model Predictive Control (MPC) techniques are generating considerable interest to implement different controllers for multilevel converters, due to its accuracy, fast dynamic response and multi-target simultaneous control. The basic idea of this concept is to predict the future behavior of the converter based on its discrete model or mathematical descriptions of the system. Against, the major drawback of this technique is the high computational load required, because the model is repetitive evaluated at each sampling time for each switching state. Finite Control Set Model Predictive Control (FCS-MPC) has been developed as a particular case of MPC where the algorithm is only applied for a finite number of switching states, reducing the computational burden. Then, the controller configures the converter with the optimal switching state that minimizes a given cost function over a sampling interval. In order to establish the importance of one controlled target in relation to the others, a weighting factor is included for each term in the cost function. Adjusting the optimal values of these weighting factors is a crucial stage in the specification of a FCS-MPC and it is a difficult challenge too. Due to this, in all works currently presented, once the weighting factors have been determined, they remain unchanged. In this thesis, a FCS-MPC has been applied for a Three-Level Neutral Point Clamped (3L-NPC) voltage source converter, including an automatic tuning process of the weighting factors. Different strategies have been studied, and finally, an artificial neural network-based approach has been considered as the better solution to implement a novel controller strategy: Adaptive Model Predictive Control (A-MPC) as an optimized version of a traditional MPC. The cost function includes predictions of the grid currents, the switching frequency of the Insulated Gate Bipolar Transistors (IGBT) and the balance of the DC-link bus voltages. Merit figures such as the current tracking error, Total Harmonic Distortion (THD), average switching frequency and DC-link voltage balancing along with the active and reactive power references, are the inputs of the Artificial Neural Network (ANN) whose outputs constitute the weighting factors. In order to reduce the computational burden, alpha-beta components transformation and a FPGA-based implementation have been used. Finally, in order to evaluate the proposed method, numerous simulations and experiments have been carried out on a platform located in the research laboratory of the GEISER group.