With commercial deployment of the Citizens Band Radio Service commencing in the last quarter of 2018, efforts are in progress to improve the efficiency of the Spectrum Access System (SAS) functions. An area of concern as identified in recent field trials is the timebound evacuation of unlicensed secondary users from a frequency band by the SAS on the arrival of an incumbent user. In this thesis, we propose a way to optimize the evacuation process by reducing the number of secondary spectrum grant revocations to be performed. The proposed work leverages knowledge of incumbent user spectrum occupancy pattern obtained from historical spectrum usage data. Using an example model trained on 48 hours of an incumbent user transmission information, we demonstrate prediction of future incumbent user spectrum occupancy for the next 15 hours with 94.4% accuracy. The SAS uses this information to set the time validity of the secondary spectrum grants appropriately. In comparison to a case where spectrum grants are issued with no prior knowledge, the number of revocations declines by 87.5% with a 7.6% reduction in channel utilization. Further, the proposed technique provides a way for the SAS to plan ahead and prepare a backup channel to which secondary users can be redirected which can reduce the evacuation time significantly. / Master of Science / Studies on spectrum occupancy show that, in certain bands, licensed incumbent users use the spectrum only for some time or only within certain geographical limits. The dynamic spectrum access paradigm proposes to reclaim the underutilized spectrum by allowing unlicensed secondary users to access the spectrum opportunistically in the absence of the licensed users. In the United States, the Federal Communications Commission (FCC) has identified 150 MHz of spectrum space from 3550-3700 MHz to implement a dynamic spectrum sharing service called the Citizens Broadband Radio Service (CBRS). The guiding principle of this service is to maximize secondary user channel utilization while ensuring minimal incumbent user disruption. In this study, we propose that these conflicting requirements can be best balanced in the Spectrum Access System (SAS) by programming it to set the time validity of the secondary license grants by taking into consideration the incumbent spectrum occupancy pattern. In order to enable the SAS to learn incumbent spectrum occupancy in a privacy-preserving manner, we propose the use of a deep learning model, specifically the long-short term memory (LSTM). This model can be trained by federal agencies on historical incumbent spectrum occupancy information and then shared with the SAS in a secure manner to obtain prediction information about possible incumbent activity. Then, using the incumbent spectrum occupancy information from the LSTM model, the SAS could issue license grants that would expire before expected arrival time of incumbent user, thus minimizing the number of revocations on incumbent arrival. The scheme was validated using simulations that demonstrated the effectiveness of this approach in minimizing revocation complexity.