The escalating volume of data traffic in modern communication networks poses a pressing challenge, driving the need for advanced predictive tools to optimize resource allocation, alleviate congestion, and sustain reliable connectivity. Traditional statistical techniques, such as ARIMA and Kalman filtering, alongside Recurrent Neural Networks (RNNs) and Graph Neural Networks (GNNs) like LSTM, GCN, have laid the groundwork for traffic prediction but falter when addressing the non-Euclidean complexities of network topologies and the limitations of long-term temporal modeling. To overcome these barriers, this paper introduces the SFI-TGNN, a sophisticated model that merges a Temporal Graph Neural Network with a Step Forward Iteration training approach, offering a versatile framework that adeptly captures intricate spatiotemporal dependencies. Rigorous evaluations reveal that the SFI-TGNN surpasses established baselines delivering superior predictive accuracy across a range of forecasting horizons, thus paving the way for transformative advancements in intelligent network management.