Spatio-temporal data modeling is a crucial component of many real-world systems, such as traffic prediction, environmental monitoring, and energy forecasting. Spatio-Temporal Graph Neural Networks (ST-GNNs) have recently emerged as powerful models for learning complex temporal and spatial dependencies simultaneously. However, existing evaluations primarily rely on real-world datasets with fixed characteristics, which limits our understanding of model behavior under varying conditions. This paper proposes a systematic empirical analysis of ST-GNNs using controlled synthetic benchmarks with tunable noise and structural dependencies. Synthetic datasets of sinusoidal time series with known inter-node correlations were designed and used to evaluate a diffusion-based ST-GNN architecture against classical baselines including ARIMA and LSTM. The experiments showed that the ST-GNN model outperformed traditional models in a multi-step forecasting task, particularly when leveraging spatial correlations. Our findings offer insight into the strengths and limitations of ST-GNNs and highlight the value of synthetic benchmarks for analyzing model generalization and robustness.