Real-Time Traffic Flow Prediction Method Based on PSO-LSTM-MAEOL

Traffic flow prediction plays a crucial role in mitigating traffic congestion, enhancing road safety, and promoting energy conservation and emission reduction. However, traditional static prediction methods often fail to adapt to the realtime dynamic nature of traffic flow. To address these challenges, this study introduces a real-time traffic flow prediction method that integrates the Particle Swarm Optimization (PSO) algorithm, Long Short-Term Memory (LSTM) neural network, and an online learning strategy. The proposed approach involves three key steps: First, the optimal hyperparameters for the LSTM are determined using PSO based on historical data. Second, an LSTM neural network is constructed with these optimal hyperparameters, and an initial prediction model is trained using historical data. Third, as new monitoring data becomes available, the prediction accuracy is assessed, and the frequency of online learning is dynamically adjusted based on the moving average error. The model is then fine-tuned through online learning, enabling realtime traffic flow prediction based on the continuously updated dynamic model. The proposed method was evaluated using realworld data from the Yangzong Tunnel section of the Kunshi Expressway in China and compared with several existing models. The results demonstrate that the proposed method achieves MAE, MAPE, RMSE, and R2 values of 1.539, 7.879, 2.0125, and 0.9691, respectively, with an average prediction time of 0.3248 seconds per instance. This method not only ensures high prediction accuracy but also maintains efficient computational performance, offering a reliable solution for real-time traffic flow prediction.