Bayesian-optimized CNN-LSTM neural network for predicting road construction dust concentrations

High dust concentrations from road construction degrade air quality, threaten human health, and increase machinery wear and fuel use. Accurate prediction of dust concentrations is therefore critical for proactive environmental control and low-carbon construction. This study proposes a Bayesian-optimized neural network model that integrates spatial, temporal, and environmental information from multi-source data, including particulate sensors, meteorological parameters, and construction records. The convolutional neural network (CNN) captures spatial features, the long short-term memory (LSTM) learns temporal dependencies, and Bayesian optimization (BO) automatically tunes hyperparameters to enhance prediction performance. The proposed model achieves high accuracy (R2 = 0.884) and exhibits superior short-term and long-term robustness compared with conventional models. These results demonstrate that the BO-CNN-LSTM framework effectively improves dust prediction accuracy and stability, providing a practical and intelligent tool for dust mitigation, energy-efficient scheduling, and carbon reduction in road construction projects.