Short-term traffic flow prediction with bio-inspired optimization: a case study from China

For short-term traffic flow prediction to support signal timing, this paper follows a three-stage procedure: sensing, predicting, and evaluating. At the sensing stage, an adaptive Gaussian mixture model with a detection zone and morphological filtering reduces repeated counting. A dual closed-loop system improves robustness under illumination changes, achieving 90.28% counting accuracy on Dataset A from Xiangyang, China. At the prediction stage, eight deep learning models are compared: six baselines (LSTM, CNN, CNN-LSTM, CNN-LSTM-Attention, LSTM-Transformer, CNN-LSTM-Transformer) and two optimization-enhanced variants (THRO-LSTM, CPO-LSTM), plus GA, PSO, and Bayesian optimization. The CRITIC-TOPSIS method ranks models by accuracy, efficiency, and robustness. Two datasets are used: Dataset A for fitting and Dataset B for generalization. THRO-LSTM achieves a MAPE of 2.12% and a 20% MSE reduction over baseline LSTM in offline scenarios, while plain LSTM performs well in real-time settings. Sensitivity analysis confirms robustness. This study offers a reference for model selection in traffic prediction.