Feature Selection for Fault Prediction in Distribution Systems

Abstract

While conventional power system protection isolates faulty components only after a fault has occurred, fault prediction approaches try to detect faults before they can cause significant damage. Although initial studies have demonstrated successful proofs of concept, development is hindered by scarce field data and ineffective feature selection. To address these limitations, this paper proposes a surrogate task that uses simulation data for feature selection. This task exhibits a strong correlation (r = 0.92) with real-world fault prediction performance. We generate a large dataset containing 20000 simulations with 34 event classes and diverse grid configurations. From 1556 candidate features, we identify 374 optimal features. A case study on three substations demonstrates the effectiveness of the selected features, achieving an F1-score of 0.80 and outperforming baseline approaches that use frequency-domain and wavelet-based features.

Figures (10)

• Figure 1: Zero-sequence current measurements from a 20 grid showing normal operation (a) and exemplary pre-fault precursor symptoms (b-d). The fault was caused by a conductor falling onto a wooden pole, which then ignited (e).
• Figure 2: Fault prediction pipeline during training (gray) and deployment (black). Training is executed offline on recorded data, while deployment processes data in real-time. The pipeline utilizes voltage and current measurements could also incorporate additional data from different sources.
• Figure 3: Identification of the most relevant windows for an exemplary signal, where the selected continuous window is marked in yellow and the transient window in red.
• Figure 4: Distribution grid model used for simulation data generation. Event buses are dynamically connected to random network nodes during simulation.
• Figure 5: Performance of randomly selected sets of features on the simulation task and the real-world task.