Intelligent transportation systems (ITS) depend on accurate and reliable traffic speed prediction to improve the safety, efficiency, and sustainability of transportation activities. Recently, deep learning approaches have significantly contributed to the development of ITS, but are still facing challenges in cyber-physical context due to the aleatoric uncertainty of increasingly uncertain traffic data and epistemic uncertainty of point-to-point estimation training models. In this work, a Bayesian deep learning model reframing with a universal traffic forecasting framework is devised for traffic speed forecasting with uncertainty quantification. The key idea of proposed network is to introduce time-series features in a latent distribution space. Compared to traditional point estimation neural networks, case studies show that the proposed model can predict more reliable results in cross domain learning tests and is capable of discovering good feature representations in missing traffic data or data-deficient scenarios.