Urban Traffic State Prediction Based on SA-LSTM

Abstract

Abstract Traffic prediction is the key to intelligent traffic guidance, management and control. Aiming at the spatiotemporal characteristics of urban traffic and the complexity of nonlinear variation of traffic parameters, a long short-term memory network (LSTM) urban traffic condition prediction model (SA-LSTM) based on self-attention mechanism is proposed. SA-LSTM uses the self-attention mechanism to assign different weights to traffic state information in different space and time, reflecting the spatio-temporal correlation of traffic data prediction. It can also avoid the vanishing gradient problem faced by traditional RNN and improve the defect that LSTM cannot accurately express the different importance and spacio-temporal characteristic of traffic information. Based on SA-LSTM, experiments were conducted on the Shenzhen road network data and floating car data. LSTM and SA-LSTM were selected for Comparative verification experiments, and the results confirmed that SA-LSTM is superior to LSTM in multiple evaluation indicators. Moreover, the road spatio-temporal correlations obtained by traffic data analysis and obtained by model learning are highly consistent, which proves that SA-LSTM can precisely learn and express the spatio-temporal characteristics and changing trend of the traffic.