GAN-Transformer for Multivariate Power Consumption Imputation

• Huichao Kong,
• Xiaoxia Li,
• Haojun Li,
• Yake Xiao,
• Siqi Ye,
• Changwei Zhang,

Vol. 20, No. 3, March 31, 2026
10.3837/tiis.2026.03.011, Download Paper (Free):

• Time series imputation
• Generative Adversarial Network
• Transformer
• smart meter data
• missing data reconstruction

Abstract

Accurate imputation of missing data in multivariate time series is critical for reliable smart meter analytics and downstream tasks such as demand forecasting and anomaly detection. In this paper, we propose a novel imputation framework that integrates generative adversarial networks (GANs) with Transformer encoders to jointly model temporal dynamics and feature correlations. The generator employs a multi-layer Transformer encoder equipped with sinusoidal positional encoding and inter-feature attention to capture both sequential and cross-sensor dependencies. A lightweight Transformer-based discriminator guides the generator to produce realistic imputations through adversarial learning. We introduce a temporal consistency loss to enforce smooth transitions in the imputed sequences. Extensive experiments on real-world electricity consumption data demonstrate that the proposed GAN-Transformer model achieves an average mean relative error (MRE) of 0.11 and an accuracy within 20% error tolerance (ACC@0.2) of 87% at a 50% missing rate, demonstrating superior robustness and visually coherent reconstructions, and significantly outperforms both GAN-only and Transformer-only baselines across various missing rates, thus offers a scalable and domain-adaptive solution for practical power data imputation scenarios.

Statistics

Cite this article