Recurrent Neural Network-based Base Transceiver Station Power Supply System Failure Prediction

In mobile telecom networks, Base Transceiver Station (BTS) is a key infrastructure that connects customers with the mobile network. BTSs are geographically scattered across the networks service area and thousands of fault indicating alarms are generated by a typical BTS on a daily basis. Thus, proactively maintaining the BTS before faults happen is beneficial to guarantee proper operation of the network and reduce operational costs. In the mobile networks installed in the city of Addis Ababa, Ethiopia, failure in power supply system of BTSs takes the biggest share for mobile services interruption. This work investigates the early prediction of BTS failures due to power system and environmental abnormalities using recurrent neural networks (RNN) with long short term memory (LSTM) and gated recurrent unit (GRU) cells. Eleven power-supply system related features and alarms were collected from a real-time power and environmental monitoring system installed in each BTS. Moreover, the experiments were performed on five BTS sites with sixteen weeks of observations. The experimental results show that GRU using sigmoid activation function with feature reduction achieves better performance than using LSTM with the different configurations investigated. Unlike the prevailing practice in the network operator, which is taking corrective actions in response to alarms, the results in this experimental provide a new insight to the mobile operator to exploit patterns inherent in the daily measured data and predict the well-being of its infrastructure.