Global network infrastructures are increasing with the development of new technologies andrngrowth in Internet traffic. As network infrastructures increases, maintaining and monitoring themrnwill become very challenging since thousands of alarms are generated every day. Clearing thosernalarms by corrective maintenance activities require considerable effort and resources (car, labor,rnand budget). rnIn mobile networks, a Base Transceiver Station (BTS) is one key infrastructure elementrnperforming the task of connecting customer equipment with the cellular network. BTS servicesrnmay be interrupted due to transmission, optical fiber cut, power system failure, natural disaster orrnmany more. In the case of Ethio Telecom (ET), the sole telecom service provider in Ethiopia,rnpower system failure takes the biggest share for interruption of BTS services. Minimizing powerrnsystem failure will reduce downtime of the BTS thereby, guarantee customer satisfaction andrnmaximize revenue. Recently, machine learning algorithms are used to predict failure in variousrnareas like power distribution, hydropower generation plants, solar power generation plants, highrnvoltage transmission grid and many more. rnThis thesis investigates predicting BTSs power system failure using a recurrent neural networkrn(RNN) types namely, long short term memory (LSTM) and gated recurrent unit (GRU) with linearrnand sigmoid activation function applied for the output. In parallel, the prediction performance ofrnLSTM and GRU has been compared. Data collected from five BTS sites for twenty weeks ofrnobservations are used to train and test the model. The data are prepared with two different datarnarrangements, which are a single site and multiple sites. The relevance of using different data sizernis, to check the impact of increasing data size with different arrangements on the prediction results.rnMean squared error (MSE) and number of epoch are used to evaluate the performance of thernmodels with different configurations. Based on the results found, GRU using sigmoid activationrnfunction with feature reduction achieves better performance than LSTM. In addition, both LSTMrnand GRU can be used for predicting BTS power system failure.