<p>Traffic crashes significantly contribute to non-recurrent congestion, resulting in lower levels of service in transportation networks. Accurate predictions of crash clearance times could significantly enhance safety and increase efficiency. However, accurate prediction is challenging because of the stochastic nature of crashes. The application of deep learning models in this area has been limited so far. To address this challenge while focusing on novel methodologies, an application of the LRT-CNN-VD model, a deep Bayesian Convolutional Neural Network with local reparameterization trick (LRT) and Variational Inference (VI), was proposed to train the model efficiently. This technique permits optimization without necessitating the tuning of regularization hyperparameters. The model also employs Variational Dropout (VD) to minimize the effects of overfitting. The validity of the model was tested on the Pennsylvania crash dataset and compared it against two other deep learning models (using dropout regularization and without regularization) and two other machine learning algorithms (Random Forest Regression and KNN Regression). According to results, the LRT-CNNVD model outperforms others across all metrics and shows comparable results to regular Dropout’s performance in terms of overfitting. Furthermore, the proposed model significantly reduces the total regularization hyperparameter tuning and training time.</p>