<p>Traffic congestion in rapidly urbanizing regions poses significant challenges to economic efficiency, environmental sustainability, and public safety. Conventional traffic management approaches often fail to effectively utilize the growing volume of real-time data generated from distributed sources. While recent advances in deep learning, particularly models such as Long Short-Term Memory (LSTM) networks and their variants, have improved traffic prediction capabilities, issues related to data privacy and distributed data availability remain largely unaddressed. To overcome these limitations, this study proposes a hybrid deep learning framework that integrates Bidirectional Long Short-Term Memory (BiLSTM) and Gated Recurrent Unit (GRU) networks with an attention mechanism within a federated learning paradigm. The proposed approach enables decentralized model training across multiple data sources without requiring raw data sharing, thereby preserving privacy while maintaining predictive performance. Experimental evaluation demonstrates the effectiveness of the model in capturing complex temporal dependencies in traffic patterns and generating reliable congestion forecasts</p>