Implement Optimal Vector Control for LCL-Filter-Based Grid-Connected Converters by Using Recurrent Neural Networks

Abstract

This paper proposes a novel vector control method for an LCL-filter-based grid-connected converter (LCL-GCC) by using a recurrent neural network (RNN). The RNN is trained by using the Levenberg-Marquardt (LM) algorithm. A forward accumulation through time (FATT) algorithm for LCL-GCCs is developed to calculate the Jacobian matrix required by the LM algorithm. The objective of the training is to implement optimal control of an LCL-GCC by using an RNN. With the RNN vector control technique, the decoupling of the LCL system is not needed. Simulation study demonstrates that the proposed RNN-based vector control method is a damping-free technique and can tolerate a wide range of system parameter changes. In both simulation- and hardware-based experiments, the RNN vector controller demonstrates much improved performance than that of conventional passive damping (PD) and active damping (AD) vector controllers for LCL-GCCs. In general, the proposed RNN vector controller provides a good solution to overcome the low-efficiency problem associated with conventional PD vector controllers and the problem of sensitivity to system parameter change associated with conventional AD vector controllers. Overall, the RNN vector controller shows strong ability to trace rapidly changing reference commands, tolerate system disturbances, and satisfy various LCL-GCC control needs.