Cognitive radios (CRs) use spectrum occupancy prediction (SOP) models to infer the future possible states of a radio spectrum. The use SOP models in CRs provide several benefits such as optimisation of spectrum sensing process, conservation of energy, and interference avoidance with licensed spectrum owners. In literature, there exists various learning-based SOP models, however, their performance is limited due to the dynamic nature of primary user activity models and channel sensing errors. In this context, we propose a deep reinforcement learning (DRL) based SOP model. Different from existing research works where DRL has been used for channel selection, we use DRL to infer future possible states of a wireless channel. Specifically, we use dueling deep Q-network with prioritised experience replay (DDQN-PER) to predict channel occupancy for next k-steps under different sensing conditions. Extensive simulation results show the superior performance of proposed method over widely used learning-based methods for different primary radio activity models.