A deep convolutional generative adversarial network for data imputation: Application to wind speed time series

• Published in: Advances in Wind Engineering Vol. 2; no. 2; p. 100054
• Main Authors: Liu, Kejun, Cai, Yuan
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.06.2025
• Subjects: Convolutional neural networks; Deep learning; Generative adversarial network; Missing data imputation; Wind speed time series; Deep learning; Missing data imputation; Wind speed time series; Convolutional neural networks; Generative adversarial network
• ISSN: 2950-6018; 2950-6018
• DOI: 10.1016/j.awe.2025.100054

Accurate and reliable wind speed data are essential across diverse wind engineering applications, such as maximizing the efficiency and effectiveness of wind energy utilization. However, the continuity of wind speed monitoring is often disrupted by missing data due to sensor malfunctions, adverse environmental conditions, or limited measurement coverage. To address this challenge, this study introduces the CNN-based Wasserstein Generative Adversarial Imputation Network (C-WGAIN), a novel framework designed to impute missing wind speed data. The proposed framework employs two-dimensional convolutional layers to capture temporal features of wind speed dynamics and integrates the Wasserstein distance into the discriminator to enhance the stability and robustness of the imputation process. The framework was tested using in-situ wind speed data from four meteorological stations operated by the Hong Kong Observatory. Comprehensive evaluations were conducted in both single-station and multi-station imputation scenarios. The experimental results demonstrate the exceptional performance of the proposed method, with accurate recovery of missing data even under challenging conditions, including scenarios with a high missing rate of up to 80%.