A novel spatiotemporal feature fusion-based GAN framework for SCADA data imputation of wind turbines

In the supervisory control and data acquisition (SCADA) system of wind farms, missing data caused by sensor failures and communication disruptions critically undermines the accuracy and operational reliability of downstream prediction models. Existing imputation methods struggle to address the spatiotemporal coupling characteristics of wind farm data while also lacking systematic analysis of how imputation impacts downstream tasks. In this paper, a Spatio-Temporal Fusion Generative Adversarial Imputation Network (STF-GAIN) is proposed to achieve high-precision wind turbine data imputation through single modal imputation and cross-modal fusion. First, differentiated single modal imputation modules are designed to capture the probability distribution characteristics, temporal dynamics, and spatial-physical meaning dependencies of the SCADA data. Second, the discriminator in GAIN framework is utilized to assess the authenticity of these single-modal imputations, yielding corresponding confidence matrices. Finally, a feature integrator is employed to integrate cross-modal features with their confidence scores to produce imputations that are both distribution-consistent and spatiotemporally coherent. Experimental results demonstrate that STF-GAIN significantly outperforms baseline models in imputation accuracy under both random and block missing patterns. Furthermore, this paper systematically investigates the potential impact mechanisms of data imputation on downstream prediction tasks through integrated experiments.