Short-term traffic flow prediction based on multi-model combination: PSO-SAO-CNN-FuzzyBiLSTM

In order to overcome the shortcomings of existing methods in characterizing complex spatio-temporal features and data uncertainty, this paper proposes a Convolutional Neural Network-Fuzzy Bidirectional Long Short-Term Memory (CNN-FuzzyBiLSTM) fusion prediction model based on Particle Swarm Optimization-Snow Ablation Optimizer (PSO-SAO) optimization. Based on the high-speed traffic flow data of M25 in the UK from August to September 2019, Standard Z-score (Z-score), Interquartile Range (IQR) and Median Absolute Deviation (MAD) methods were used to detect anomalies, and Min-Max normalization preprocessing was used. In terms of the model architecture, the three-layer Convolutional Neural Network (CNN) extracts the spatial correlation between adjacent monitoring stations, the fuzzy logic system is combined with Bidirectional Long Short-Term Memory (BiLSTM), and the forgetting gate bias is dynamically adjusted by the activation rate of the input gate and the forgetting gate to deal with data uncertainty. In addition, this paper proposes a PSO-SAO hybrid optimization strategy, which integrates the global exploration of Partricle Swarm Optimization (PSO) and the local fine-tuning advantages of Snow Ablation Optimizer (SAO) through an adaptive interaction mechanism, and optimizes the model hyperparameters (learning rate, number of Long Short-Term Memory (LSTM) units, L2 regularization coefficient, Dropout rate, convolution kernel siz). The experimental results show that the Mean Absolute Error (MAE) of the model is 0.6633 and the Root Mean Square Error (RMSE) is 0.9361. Compared with CNN-BiLSTM, Transformer and Graph Neural Network (GNN), MAE is reduced by 98.66%, 99.04% and 99.25%, respectively. Main contributions: 1) CNN-FuzzyBiLSTM fusion framework; 2) PSO-SAO hybrid optimization strategy. Limitations include not considering external factors such as weather, and real-time computing efficiency needs to be improved.