Location
Online
Event Website
https://hicss.hawaii.edu/
Start Date
3-1-2023 12:00 AM
End Date
7-1-2023 12:00 AM
Description
Utilizing data analytics and machine learning (ML) on phasor measurement units (PMUs) data to analyze faults automatically is the focus of this paper. Insufficient labels and natural uneven distribution of different types of line fault events found in field-recorded PMU data make supervised ML model development challenging. To address this issue, we train off-the-shelf Support Vector Machine (SVM) ML models for line fault classification using simulated PMU data obtained from a combination of 12 physical and virtual PMUs placed on a synthetic IEEE 14-bus system as well as using this simulated data unified with field recordings. A conducted sensitivity study is focused on three factors, 1) the number of PMUs used to train the ML model, 2) the voltage level at which the model is trained, and 3) the vicinity of PMUs to transmission line faults. The ML models trained with simulated and field data are evaluated on one-year field-recorded data collected from 38 PMUs sparsely located in the US Western interconnection. We demonstrate that when training ML models with only simulated data, the performance varies significantly with different number of PMUs, voltage level, and PMU placement in separate areas of the synthetic grid (F1 score of 0.78 to 0.92). We obtained an F1 score of 0.94 using the simulated dataset integrated with field recordings. The performance of a ML model developed using simulated data is also evaluated on the three-phase voltage signals extracted from 188 PMUs in the Eastern interconnection accompanied by imprecise labels, where the majority of the labels do not identify the fault type. On this extremely challenging task, we achieved 77% accuracy solely using synthetic data for ML training.
Recommended Citation
Otudi, Hussain; Mohamed, Taif; Kezunovic, Mladen; Hu, Yi; and Obradovic, Zoran, "Training Machine Learning Models with Simulated Data for Improved Line Fault Events Classification From 3-Phase PMU Field Recordings" (2023). Hawaii International Conference on System Sciences 2023 (HICSS-56). 5.
https://aisel.aisnet.org/hicss-56/es/monitoring/5
Training Machine Learning Models with Simulated Data for Improved Line Fault Events Classification From 3-Phase PMU Field Recordings
Online
Utilizing data analytics and machine learning (ML) on phasor measurement units (PMUs) data to analyze faults automatically is the focus of this paper. Insufficient labels and natural uneven distribution of different types of line fault events found in field-recorded PMU data make supervised ML model development challenging. To address this issue, we train off-the-shelf Support Vector Machine (SVM) ML models for line fault classification using simulated PMU data obtained from a combination of 12 physical and virtual PMUs placed on a synthetic IEEE 14-bus system as well as using this simulated data unified with field recordings. A conducted sensitivity study is focused on three factors, 1) the number of PMUs used to train the ML model, 2) the voltage level at which the model is trained, and 3) the vicinity of PMUs to transmission line faults. The ML models trained with simulated and field data are evaluated on one-year field-recorded data collected from 38 PMUs sparsely located in the US Western interconnection. We demonstrate that when training ML models with only simulated data, the performance varies significantly with different number of PMUs, voltage level, and PMU placement in separate areas of the synthetic grid (F1 score of 0.78 to 0.92). We obtained an F1 score of 0.94 using the simulated dataset integrated with field recordings. The performance of a ML model developed using simulated data is also evaluated on the three-phase voltage signals extracted from 188 PMUs in the Eastern interconnection accompanied by imprecise labels, where the majority of the labels do not identify the fault type. On this extremely challenging task, we achieved 77% accuracy solely using synthetic data for ML training.
https://aisel.aisnet.org/hicss-56/es/monitoring/5