DC Short-Circuit Fault Detection for MMC-HVDC-Grid Based on Improved DBN and DC Fault Current Statistical Features

Abstract

The fault current in a modular multilevel converter-based high voltage direct current grid (MMC-HVDC-Grid) will rise rapidly when DC short-circuit fault (DCSCF) occurs, which poses a great challenge to the prompt and accurate fault protection detection technology. Aiming to solve the issue of (DCSCF) location detection in MMC-HVDC-Grid, a statistical feature-based improved deep belief network (IDBN) method is proposed in this paper. By the proposed method, the three features, such as the standard deviation, information entropy, and kurtosis are extracted from the fault current training samples within a single terminal of MMC-HVDC-Grid and fused as the inputs of IDBN. The IDBN is trained by using the pre-training of contrastive divergence (CD) algorithm and the back fine-tuning of the supervised learning algorithm, and then the DCSCF detection is accomplished with the SOFTMAX as the output layer. In order to improve the detection accuracy and reduce the adverse effect on the converter station, the exponential decay learning rate is used to optimize the network, which can speed up the rate of convergence for the training process. Taking four-terminal MMC-HVDC-Grid as an example, the proposed method is tested and compared with other fault detection methods based on traditional neural networks, such as back propagation, radial basis function, stacked autoencoder (SAE), and support vector machine. The simulation results show that the performances of the overall classification accuracy, kappa coefficient, Jaccard distance, and detection speed by the proposed method are all superior to the other traditional neural network fault detection methods mentioned above. In addition, the online DC fault current data are continuously extracted into the trained IDBN model, and the performance of the proposed method is verified through the real-time digital simulation platform.