Data-Driven Modeling of Modular Multilevel Converters Based on HHT and CNN-LSTM-AM Neural Network

The mathematical mechanism model of modular multilevel converters (MMC) exhibits high-dimensional complexity, necessitating a trade-off between accuracy and efficiency during the modeling process. To address this issue, this article proposes a nonmechanism modeling approach for MMC, focusing on the mapping relationships between electrical quantities. The proposed method utilizes the convolutional neural network - long short-term memory - attention mechanism (CNN-LSTM-AM) neural network algorithm to drive the data-driven modeling of MMC. Operational data from the MMC system is utilized as the training set. The CNN layer selects input features with strong correlations, while the LSTM layer captures temporal dependencies among the features. The AM layer performs weighted aggregation on the LSTM output to enhance prediction accuracy. In addressing the challenge of fitting frequency, which cannot be directly achieved using electrical quantities, the frequency model incorporates a Hilbert–Huang transform (HHT) decomposition of voltage and current signals. The extracted instantaneous frequencies and amplitudes serve as additional input features to improve the fitting of output features. This approach effectively resolves the issue of inadequate prediction accuracy in the frequency model. Additionally, data-driven models for MMC inverters and rectifiers are trained using simulation data and evaluated for accuracy using a separate test set.