In this work, we aim to enhance the robustness of conventional adaptive beamforming schemes against both signal steering vector uncertainties and interference nonstationarity here. For this purpose, a random model with Watson prior distribution is applied to the steering vector of interest, and cluster structured prior belief about the interferer motion is incorporated into the statistical analysis. However, the difficulty in handling the normalization factor of the Watson distribution makes the Bayesian inference of this probabilistic model prohibitive. Towards this end, an efficient algorithm using variational Bayesian expectation-maximization is proposed, by iteratively majorizing a lower bound on the variational free energy to refine the latent parameters and circumvent intractable moment computation. A specific beamformer encompasses a trade-off among accuracy, speed and ease of implementation is thus obtained. Further, an extension of the algorithm to the case of moving interferences is presented by assigning a Dirichlet process prior to the hidden parametric space. By doing so, the underlying time-aware clusters embedded in the interference-plus-noise vectors can be automatically revealed. Both simulated and experimental results validate the superiority of the algorithms over the existing algorithms.