Improved Dung Beetle Optimizer–Tuned Convolutional–Recurrent Temporal Self-Attention Network for Multi-Horizon Traffic Speed Forecasting

Reliable traffic-speed forecasting requires local and temporal dependence to be modeled under consistent evaluation. This study presents an Improved Dung Beetle Optimizer (IDBO)-selected convolutional–recurrent temporal self-attention network for direct three-location forecasting. Separate models predict 5, 15, and 30 min speeds from synchronized 60 min histories in the Caltrans Performance Measurement System (PeMS) freeway and Harbin urban-road case studies. IDBO combines stagnation-triggered Lévy perturbation and adaptive weighting to select convolutional filters, recurrent units, and attention heads using a shared six-task validation objective. Chronological evaluation includes nine forecasting methods, five final runs, paired inference, matched-budget search comparisons, context controls, and ablations. The proposed configuration achieves the lowest three-location macro root mean square error (RMSE), mean absolute error, and denominator-bounded mean absolute percentage error (ε-MAPE, with ε = 5 km/h) among the nine benchmark methods in all six tasks. Relative to Transformer, RMSE reductions are 13.97–14.42% on PeMS and 12.79–13.08% on Harbin; reductions relative to its capacity-matched counterpart are 7.55–7.92% and 6.66–7.24%, respectively. All twelve paired comparisons remain significant after separate six-task Holm adjustments for the two comparators. IDBO achieves the lowest mean normalized search endpoint and area under the curve under the common query budget. The network uses 240,387 parameters and 2,804,672 analytical multiply–accumulate operations per sample, documenting the accuracy–complexity trade-off.