The Instrumentation and Control (I&C) system in nuclear power plants (NPPs) comprises a hierarchical architecture of closed-loop controllers. It generate multivariate time-series operational data, including control setpoints, measured process variables, and control actions. Given its critical role in NPP operation, developing advanced surveillance and diagnostic methodologies for I&C systems becomes imperative to assist operator monitoring and decision-making. This study proposes a novel spatio-temporal feature extraction framework integrating Graph Convolutional Networks (GCN) with Long Short-Term Memory (LSTM) networks. Leveraging inter-variable relationships encoded in adjacency matrices derived from priori system knowledge, the developed GCN-LSTM architecture employs a Masked Autoencoder (MAE) framework to simultaneously capture spatial dependencies and temporal dynamics of process variables. Validation through missing variable imputation tasks demonstrates the model's capability to reconstruct arbitrarily selected process variables in the Nuclear Steam Supply System (NSSS) of the HTR-PM600 with satisfactory reconstruction accuracy.