PQ prediction by way of parallel computing - benchmark and sensitivity analysis for classical ML approaches

Loading...
Thumbnail Image
Paper number
2049
Working Group Number
Conference name
CIRED 2019
Conference date
3-6 June 2019
Conference location
Madrid, Spain
Peer-reviewed
Yes
Short title
Convener
Authors
Eisenmann, Adrian , University of Stuttgart- Institute of Power Transmission and High Voltage Technology, Germany
Streubel, Tim , University of Stuttgart Institute of Power Transmission and High Voltage Technology, Germany
Rudion, Krzysztof, University of Stuttgart, Germany
Abstract
In this paper, three machine-learning (ML) regression methods for the prediction of current and voltage characteristics in industrial environments are investigated. Apart from the original data, three different feature sets are used for the prediction. In addition, up to three different multi-step prediction methods are used and compared - the direct, the recursive and the multi-output method. The regression methods are the k-Nearest Neighbour Regression (kNNR), the Support Vector Regression (SVR), and the Random Forest Regression (RFR). Eleven features are extracted per metric, from which three feature sets are developed by selection based on Kendall rank, Random Forest Feature import (RFFI), and Sequential Random k-Nearest Neighbour (SRKNN). It could be observed that the prediction of voltage is difficult. In this case the prediction error (the mean absolute error (MAE) was used throughout this investigation) could only be improved by the SVR and just by 1.1 % compared to the primitive prognosis. The comparison was based on the assumption that for the primitive prognosis the measured value would remain unchanged over the forecasting period. In predicting the current signal, significantly better results could be achieved, with the SVR again attaining the lowest error with an improvement of 32 %. The least MAE for the voltage prediction was achieved by the feature set selected by RFFI. In the current prediction, the feature selection using SRKNN proved to be best.
Table of content
Keywords
Publisher
AIM
Date
2019-06-03
Permanent link to this record
https://cired-repository.org/handle/20.500.12455/706
http://dx.doi.org/10.34890/925
ISSN
2032-9644
ISBN
978-2-9602415-0-1