Base station (BS) sleeping operation in transient process has been widely analyzed in current studies to save energy consumption in 5G cellular network. In this paper, we proposed a two-stage BS sleeping scheme with traffic offloading to reduce the energy consumption in inhomogeneous cellular networks with spatio-temporal random arrival of traffic. A Markov decision process (MDP) based algorithm is invoked for analyze the operation and performance of our adaptive sleeping scheme. And the impacts of user aggregation in both temporal and spatial domain are discussed. Numerical results show the proposed BS sleeping scheme can significantly reduce the network energy consumption. However, due to the interference between BSs, the BS sleeping scheme may also bring some loss in user satisfaction when traffic load is high.