Online Fault Line Detection in Small-Sample and Streaming Data Environments

This article focuses on medium- and low-voltage networks with neutral grounded through arc-suppression coils. Standard data-driven fault line detection (FLD) approaches assume that the training sample is sufficient, static, and reusable. In practical scenarios, such approaches may be infeasible due to the sporadic and temporary nature of single-phase-to-ground faults, which provide insufficient fault samples and due to large-scale high-speed dynamic data streams associated with measurements. To tackle these issues, this article proposes a novel FLD scheme based on personalized federated learning (PFL) and incremental stochastic configuration networks (SCNs) for small-sample and streaming data environments. Concretely, the SCN, a concise noniterative neural network, is exploited as the FLD classifier. To adapt effectively to dynamic and nonreusable environments, an incremental SCN is proposed that can learn fault features without experiencing forgetting when dealing with streaming data. The proposed FLD scheme based on PFL selectively aggregates fault features from multiple substations. This approach addresses the challenge of limited sample sizes while preserving the personalization of each local model. Extensive experimental results using real data show that the proposed method can significantly improve accuracy when dealing with small samples and continuously learn fault features in streaming data.