Ensuring the quality of input data is essential for building robust and explainable AI models in critical infrastructure domains such as smart grids. However, in SCADA-based intrusion detection pipelines, structural inconsistencies and latent feature drift are rarely assessed. In this preliminary study, we adapt the DQuaG framework—a graph-based reconstruction model originally developed for general tabular data—to assess data quality in a SCADA dataset based on the DNP3 protocol. Weapply the model in an unsupervised setting, using reconstructionloss to detect potential inconsistencies without labeled errors. Our results show that even within clean datasets, structural outliers can be identified, highlighting the value of structure-aware data validation. We discuss the implications for data-centric AI pipelines in SCADA cybersecurity and outline future directions for improving quality assessment and synthetic data generation.