Improved 1DCNN arc detection algorithm for DC microgrid based on PSO-MRFO

Abstract Arc faults in DC microgrids can easily escalate into serious incidents, such as fires and other dangerous disasters. Therefore, it is extremely necessary to conduct in-depth research on DC arc faults and carry out timely detection. In arc fault detection, artificial intelligence-based detection technology is usually used. However, there are two main limitations in the technology, namely the lack of diverse arc current data and the poor optimization of hyperparameters. The limitations could result in insufficient arc detection accuracy. Ultimately, arc faults in DC microgrids can not be timely and accurately detected. To solve this problem, this paper proposes an arc fault detection method using one-dimensional convolutional neural network (1DCNN) and optimizing it through the hybrid particle swarm optimization and manta-ray foraging optimisation (PSO-MRFO) algorithm. First, raw DC arc-fault data were acquired under a wide range of operating conditions to diversify the training samples for the 1DCNN, and arc characteristics extracted by the Tsfresh library were employed to enrich the input feature space. Second, to substantially enhance the performance of the 1DCNN, the rapid local-search capability of particle swarm optimisation (PSO) was fused with the powerful global-exploration capacity of the manta-ray foraging optimisation (MRFO), augmented by a local-optimum-escape mechanism, to construct the hybrid PSO–MRFO algorithm. A probabilistic mechanism dynamically selects PSO or MRFO to update particle positions, while a rolling-foraging strategy actively avoids local optima. Experimental results demonstrate that the PSO–MRFO algorithm outperforms either constituent algorithm in both global-search ability and convergence speed. Moreover, the 1DCNN optimised by PSO–MRFO significantly surpasses alternative models in detection performance.