As the gold standard for phase retrieval, phase-shifting algorithm (PS) has been widely used in optical interferometry, fringe projection profilometry, etc. However, capturing multiple fringe patterns in PS limits the algorithm to only a narrow range of application. To this end, a deep learning (DL) model based digital PS algorithm from only a single fringe image is proposed. By training on a simulated dataset of PS fringe patterns, the learnt model, denoted PSNet, can predict fringe patterns with other PS steps when given a pattern with the first PS step. Simulation and experiment results demonstrate the PSNet's promising performance on accurate prediction of digital PS patterns, and robustness to complex scenarios such as surfaces with varying curvature and reflectance.