PSLSTM-ATT: A hybrid parallel-serial LSTM with self-attention for short-term traffic flow prediction

Abstract

Short-term prediction of traffic flow plays an important role in ITS, allowing various services, such as route planning, signal control, and congestion relief, to be provided in real time. However, the non-stationarity, multiple seasonalities, and complex temporal dependencies within traffic flow data pose significant challenges to prediction accuracy. Existing hybrid deep learning models are mainly based on historical data and cannot effectively consider short-term fluctuations and long-term trends simultaneously. To overcome these drawbacks, we introduce PSLSTM-ATT, a novel hybrid deep learning model based on parallel and serial Long Short-Term Memory (LSTM) networks with a self-attention mechanism. The parallel branch processes the input sequence separately and learns multiscale temporal features, while the serial LSTM encoder captures sequential dependencies among time steps. An important difference in our method is that we use predicted future traffic flow as additional input, giving the model more context for understanding the data. The self-attention module dynamically weights each time step according to its contextual importance, allowing the model to focus on the most relevant temporal information patterns. We test PSLSTM-ATT on four highway datasets from the Amsterdam metropolitan area (A1, A2, A4, and A8). Experimental results show that the developed model performs better than other benchmark techniques, such as KNN, CNN, SVR, SSA-ELM, GSA-KAN, standard LSTM, and LSTM-ATT. The presented PSLSTM-ATT model is viable for short-term traffic prediction in real ITS applications.