To address the issues of missing data caused by inconsistent sampling rates and the difficulty in fully characterizing complex spatio-temporal coupling relationships among variables in practical industrial processes, this paper proposes a multi-rate soft sensor modeling method based on a Generative Spatio-Temporal Graph Network. First, targeting the periodic sparsity of data induced by sampling discrepancies, a generative adversarial imputation network integrating a time decay mechanism is proposed. This enables the missing sequence imputation to follow the dynamic evolution laws of process variables, thereby achieving high-quality reconstruction of the periodically sparse data. Second, to extract spatio-temporal features, a spatial topology module is designed. By introducing the maximal information coefficient (MIC) into a graph attention network, a spatial topology graph is constructed to decouple the nonlinear spatial correlations among variables. Furthermore, a temporal evolution module is developed to mine the temporal dependencies of the process data, realizing the effective prediction of key quality variables. Finally, the proposed method is applied to a continuous stirred tank reactorand a wastewater treatment plant process. Experimental results verify the effectiveness and superior prediction accuracy of the proposed model.□