Power Load Data Cleaning Method Based on DBSCAN Clustering and WGAN Algorithm

The various processes of acquisition and transmission of measurement data are subject to malfunction or interference, resulting in missing data. Traditional data restoration methods ignore the historical load change pattern and have low reconstruction accuracy. In this paper, we first use the DBSCAN...

Full description

Saved in:
Bibliographic Details
Published in:IEEE International Conference on Power, Intelligent Computing and Systems (Online) pp. 652 - 657
Main Authors: Wei, Liyong, Ding, Yi, Wang, En, Liu, Lixin
Format: Conference Proceeding
Language:English
Published: IEEE 14.07.2023
Subjects:
cluster analysis
Clustering algorithms
Data reconstruction
electricity load fluctuation
Generative adversarial networks
neural network learning
Neural networks
Power measurement
Real-time systems
Time measurement
Training
WGAN
ISSN:2834-8567
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:The various processes of acquisition and transmission of measurement data are subject to malfunction or interference, resulting in missing data. Traditional data restoration methods ignore the historical load change pattern and have low reconstruction accuracy. In this paper, we first use the DBSCAN clustering algorithm to detect outliers in the original load dataset and remove the outliers with large deviations to form a dataset with vacant values. Then, the Wasserstein distance is used to improve on the original GAN network. Through non-supervised training of WGAN, the neural network will automatically learn complex spatio-temporal relationships that are difficult to model explicitly, such as correlations between measurements and load fluctuation patterns. Finally, the authenticity constraint and contextual similarity constraint are used to optimize the hidden variables, so that the trained generator will be able to generate highly accurate reconstructed data. The algorithm analysis proves the performance stability of the proposed method, and the reconstructed data can reflect the real time characteristics of the measured data.
ISSN:2834-8567
DOI:10.1109/ICPICS58376.2023.10235595