An LSTM neural network prediction model of ultra-short-term transformer winding hotspot temperature

Power transformers are crucial components in power systems, widely used across various applications. The hotspot temperature of transformer oil winding is a key indicator of the transformer’s heat dissipation efficiency. Predicting the top oil temperature accurately is essential for effective monito...

Full description

Saved in:
Bibliographic Details
Main Authors: Kun Yan, Jingfu Gan, Yizhen Sui, Hongzheng Liu, Xincheng Tian, Zehan Lu, Ali Mohammed Ali Abdo
Format: Article
Language:English
Published: AIP Publishing LLC 2025-03-01
Series:AIP Advances
Online Access:http://dx.doi.org/10.1063/5.0213051
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Power transformers are crucial components in power systems, widely used across various applications. The hotspot temperature of transformer oil winding is a key indicator of the transformer’s heat dissipation efficiency. Predicting the top oil temperature accurately is essential for effective monitoring of transformer operations. Traditionally, winding hotspots serve as a reliable alert for the transformer’s operational status. Utilizing principal component analysis (PCA) to reduce data correlation and employing long short-term memory (LSTM) neural networks for predicting winding hotspots can lead to reduced errors and improved prediction accuracy, enabling quicker and more efficient forecasts. In addition, building back propagation neural networks and random forest prediction models to forecast using the same samples allows for a comparison with the LSTM model enhanced by improved PCA. Through parameter training and model development, this approach effectively processes historical time series data. The prediction error margin ranges from −3.5% to 1%, with the trends of predicted and actual values largely aligning, showcasing high accuracy.
ISSN:2158-3226