UAS trajectory prediction is stochastic in nature and randomness exists in almost every aspect of UAS Traffic Management. In order to address this challenge, it is critical to ensure a reasonable separation between UAS and obstacles when doing path planning and conflict resolution. In this paper, a novel method to deconflict for rotary-wing UAS traffic management is proposed. The main idea is to integrate a probabilistic dynamic anisotropic operational safety bound as airspace reservation with reinforcement learning method. The operational safety bound is based on UAS performance, weather condition and uncertainties in UAS operations, such as positioning error. Based on the operational safety bound concept, a new reward function in reinforcement learning is developed. The proposed methodology results in a trajectory prediction model under risk-based dynamic separation criterion. The algorithm of Q learning is adopted to find the optimal path planning. Simulations of avoiding static obstacles and multi-UAS conflict resolution are conducted to show the deconflict capability. Comparisons between results with operational safety bound and without operational safety bound are shown and analyzed.