enhancing short-term traffic flow forecasting by hybrid deep learning architectures and attention mechanisms (case study: high-density karaj-chalous road, iran)

The main tool to mitigate congestion and improve travel experiences effectively in intelligent traffic management is to predict the accurate and timely short-term traffic flow on high-volume roads. we present the performances of different deep learning models, such as lstm, gru, cnn, their hybrids cnn-lstm and cnn-gru, and versions with an attention mechanism for one-hour-ahead traffic flow prediction on mountainous and high-density karaj-chalous road. the input data include the traffic data from two traffic counters. the cited data were derived for a period ranging from 01/01/1401 to 01/01/1403. besides, the synoptic meteorological data were acquired within three-hour intervals, while the models are compared based on various quantitative accuracy and error metrics. the results showed that the cnn-lstm model was the best among the rest, with an r² value of 0.83, because it captured complex traffic patterns and temporal dependencies effectively. the other models ranked next were lstm, gru, cnn-lstm-gru, and cnn-gru, with r2 values of 0.82, 0.81, 0.80, and 0.80, respectively. while the weakest models, cnn and cnn-multihead-attention, yielded an r² of 0.60 and 0.62, respectively, this is due to a lack of consideration in these models regarding the nature of traffic data as a time series. employing attention mechanisms improved prediction accuracy in some model architectures. this effect was highly varied based on the model structure itself. the results depict that deep, hybrid models with the integration of attention mechanisms can give more reliable and valuable forecasts to the intelligent transportation management systems for better travel planning and congestion reduction in similar roadways.