Abstract: With the proposal of the“dual carbon”goal and the increasing demand for energy conservation and emis?sion reduction, the need for energy-efficient ventilation in highway tunnels has grown substantially, placing higher requirements on the accuracy of traffic volume forecasting. To enable low-carbon regulation of tunnel ventilation systems, this study utilizes traffic volume data recorded at 5-minute intervals from a highway tunnel and compares the performance of several common forecasting models in short-term multi-step prediction tasks. An improved prediction framework is proposed by integrating the dung beetle optimization (DBO) algorithm and the self-attention mechanism into the LSTM neural network, yielding the DBO-A-LSTM model with significantly enhanced predictive performance. To address the issue of high prediction errors during holiday peak periods, the DBO-A-LSTM model is further refined through dataset optimization. Confidence interval analysis, t-tests, and Wilcoxon tests demonstrate that the optimized DBO-A-LSTM model achieves a stable R2 exceeding 0.75 under all scenarios, with accurate trend prediction; the mean absolute percentage error (MAPE) is reduced to below 22.2%, and the mean percentage error remains small. Visual analysis further confirms that the predicted values closely follow the fluctuation patterns of traffic peaks. The proposed model can be applied to real-time tunnel ventilation demand estimation, enabling proactive control of ventilation systems up to 20 minutes in advance based on 5-minute dynamic predictions. This approach significantly improves the accuracy and timeliness of intelligent energy-saving ventilation control in highway tunnels, offering practical implications for advancing low-carbon operation oftransportation infrastructure.