Traffic congestion (TC) is a major issue in smart cities due to urbanization, leading to financial losses. Existing prediction algorithms fail to predict TC effectively due to dynamic road networks. A Multilevel-Gated Recurrent Unit (MGRU) model is proposed for accurate traffic flow prediction and congestion avoidance. The multilevel softmax layers reduce prediction errors and adapt to dynamic changes. Data including weather and vehicle counts were used, requiring fewer factors than previous methods. Traffic density is estimated for TC prediction, and experimental results show accuracy of 0.887 and MAE of 82.34, outperforming Conv-Bi-LSTM models.