This paper presents STGA-Net (Spatio-Temporal Graph Attention Network), a deep learning framework designed for accurate short-term traffic flow prediction in complex Indian urban environments. Standard models such as ARIMA, CNN, and LSTM often fail to effectively represent dynamic spatial dependencies among interconnected road segments. The proposed model employs Graph Attention Networks (GAT) for adaptive spatial learning and Long Short-Term Memory (LSTM) units for modeling temporal sequences to address this issue. One big thing that makes this work is the use of correlation-based adaptive adjacency learning, which lets the model figure out how road segments are related to each other in real time, not just on static road maps. Embeddings of day-of-week and other time-based features can also help find traffic patterns that happen over and over again.The model was tested on a real-world traffic dataset from Bangalore and showed stable convergence and better predictive performance. The results of the experiment show that STGA-Net has a Mean Absolute Error (MAE) of 0.6536 and a Root Mean Squared Error (RMSE) of 0.8544, which means that it can reliably make short-term predictions. The model also got a (R2) score of 0.2390, which shows that it can capture important changes in space and time. These results show that STGA-Net works well for smart city and intelligent traffic management.