Seventeen-class power quality disturbance dataset, repaired, with partition indices and code

Description

Seventeen classes of parametric power quality disturbance records, 1,000 per class and 17,000 in total. Each record is a single cycle of a 50 Hz fundamental, 100 samples at an equivalent 5 kHz, amplitude in per unit, with no measurement noise. Nine classes are single events and eight are compound events formed from two constituent models on one record. Included are the stratified partition indices used in the accompanying article, meaning 12,240 training, 1,360 validation and 3,400 test records with 200 test records per class, a generation specification giving the measured parameter ranges of all seventeen classes, a repair log, and code to verify the data, reproduce the feature analysis and retrain the three compared architectures. The original generation script was not preserved, so the parameter ranges were measured from the records themselves. Four documented defects were repaired, and all 17,000 records are now unique. One artifact remains by design, since the first sample is exactly zero in about 99 per cent of the sag, swell and interruption records, and the article measures its effect with a control run.

Institutions

• University of Asia Pacific

  Dhaka Division, Dhaka

Categories

Licence