Neural Network Based Model Predictive Controllers for Modular Multilevel Converters

• Published in: IEEE transactions on energy conversion Vol. 36; no. 2; pp. 1562 - 1571
• Main Authors: Wang, Songda, Dragicevic, Tomislav, Gao, Yuan, Teodorescu, Remus
• Format: Journal Article
• Language: English
• Published: New York IEEE 01.06.2021; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Algorithms; Artificial neural networks; Computation; Computational modeling; control design; Controllers; Converters; Cost function; Data collection; Machine learning; Mathematical model; model predictive control (MPC); Modular multilevel converter (MMC); neural network (NN); Neural networks; Pattern recognition; Predictive control; Predictive models; Real time
• DOI: 10.1109/TEC.2020.3021022
• ISSN: 0885-8969

Modular multilevel converter (MMC) has attracted much attention for years due to its good performance in harmonics reduction and efficiency improvement. Model predictive control (MPC) based controllers are widely adopted for MMC because the control design is straightforward and different control objectives can be simply implemented in a cost function. However, the computational burden of MPC imposes limitations in the control implementation of MMC because of many possible switching states. To solve this, we design machine learning (ML) based controllers for MMC based on the data collection from the MPC algorithm. The ML models are trained to emulate the MPC controllers which can effectively reduce the computation burden of real-time control since the trained models are built with simple math functions that are not correlated with the complexity of the MPC algorithm. The ML method applied in this study is a neural network (NN) and there are two types of establishing ML controllers: NN regression and NN pattern recognition. Both are trained using the sampled data and tested in a real-time MMC system. A comparison of experimental results shows that NN regression has a much better control performance and lower computation burden than the NN pattern recognition.