Medium-term Traffic Flow Forecasting based on a Learning Network of Convolution-based BiLSTM and Transformer

Urbanization has accelerated the need for accurate traffic flow forecasting as an integral part of intelligent transportation systems. Precise medium-term traffic flow forecasting is crucial for alleviating congestion and enhancing road utilization efficiency. However, the technology is fraught with complexities due to the dynamic nature of traffic patterns that vary both temporally and spatially. In this paper, we propose a novel learning network that incorporates a simplified convolutional layer, bidirectional long short-term memory (BiLSTM) layer and an enhanced transformer encoder to predict medium-term traffic flow. Specifically, the convolution operation is employed to extract local features from the traffic data, which helps in identifying the underlying patterns that are crucial for understanding the temporal dynamics of traffic. Combined with the BiLSTM, which captures the longer-term temporal dependencies, and the enhanced Transformer encoder that focuses on capturing the global dependencies and self-attention mechanisms, the proposed model can effectively address the complexities and non-linearities present in traffic flow data. This subtle integration of hybrid structures allows the proposed network to achieve superior performance in handling the intricacies of traffic data. Experimental evaluations indicate that the proposed network outperforms comparative forecasting methods in terms of prediction accuracy and robustness. The results suggest that the network can provide reliable medium-term traffic flow forecasts, which are essential for making informed decisions in transportation systems and urban development.