Nowadays, system architecture of the fifth generation (5G) cellular system is becoming of increasing interest. To reach the ambitious 5G targets, a dense base station (BS) deployment paradigm is being considered. In this case, the conventional always-on service approach may not be suitable due to the linear energy/density relationship when the BSs are always kept on. This suggests a dynamic on/off BS operation to reduce the energy consumption. However, this approach may create coverage holes and the BS activation delay in terms of hardware transition latency and software reloading could result in service disruption. To tackle these issues, we propose a predictive BS activation scheme under the control/data separation architecture (CDSA). The proposed scheme exploits user context information, network parameters, BS sleep depth and measurement databases to send timely predictive activation requests in advance before the connection is switched to the sleeping BS. An analytical model is developed and closed-form expressions are provided for the predictive activation criteria. Analytical and simulation results show that the proposed scheme achieves a high BS activation accuracy with low errors w.r.t. the optimum activation time.