Multi-Scale Conditional Generative Adversarial Networks for Wind Speed Data Imputation in Earthen Ruins Protection

Abstract

Time-series data are vital for preserving earthen ruins and evaluating wind erosion effects. Harsh conditions at these sites often lead to sensor degradation and significant data gaps. To tackle wind speed data imputation for such environments, we introduce a Multi-Scale Conditional Generative Adversarial Network (MSC-GAN) with a Transformer-based generator. This model integrates features across hourly, daily, and weekly scales, combined with real-time wind direction data and random noise. Utilizing the Transformer’s ability to model long-range dependencies and multi-scale information, MSC-GAN adeptly manages complex missing data patterns. We validate MSC-GAN using nearly two years of near-surface wind speed data from the Suoyang City earthen ruins. Our experimental results reveal that MSC-GAN substantially improves imputation accuracy—by 40.2% for short gaps and 6.7% for long gaps—over traditional methods. Code is available at https://github.com/zizhou001/msc-gan.