The fast advancement of smart city systems has resulted in a huge quantity of data spread over both time and space, which is marked by intricate spatial relationships and temporal changes. Regular predictive models usually find it difficult to comprehend such interrelated and complex relationships, thus the outcomes are limited in terms of accuracy and stability. On the contrary, a GNN-based spatiotemporal modeling framework for infrastructure analytics is proposed and evaluated in this study, with its performance compared to that of a traditional model. With each model having 60 observations, the study assesses prediction accuracy, Root Mean Square Error (RMSE), inference latency, and spatiotemporal feature utilization. Descriptive statistics show that the GNN spatiotemporal model provides prediction accuracy which is significantly higher (M = 0.890) and RMSE which is substantially lower (M = 9.10) than the traditional approach (accuracy M = 0.715; RMSE M = 18.2). Distributional analyses also reveal the GNN model's predictions as more consistent and with less variability. Extreme value analysis provides a strong point for the proposed framework, indicating the ability to deliver good performance even in difficult settings. Welch's one-way ANOVA confirms that all the models differ significantly in all metrics evaluated (p < .001). To sum up, the results speak for the GNN-based spatiotemporal modeling in being the most predictive, stable, and deep in analytics, hence its adoption in smart city and infrastructure analytics applications being supported.