Telecommunication networks play a critical role in our society. They make it possible to share a massive amount of data across the globe. While networks are complex systems in terms of size and technological diversity, a failure results in numerous alarms from the series of devices. This makes monitoring and maintenance activities challenging due to the growing complexity of alarm management systems and the need for highly educated experts to deploy. Alarms are generated in vast quantities every day by today's large and complicated telecom networks. The alarm sequence offers significant information on the network's activity, but most of it is fragmented and hidden in the massive amount of data. Alarm regularities can be utilized in fault management systems, for example, to filter redundant alarms, locate network problems, and even anticipate catastrophic faults. In the presence of flooding alarms, alarms that are inadequately configured and maintained, and a large number of nuisance alarms, operators are expected to make vital judgments. If the incoming alarms can be correctly predicted before they occur, the operators may be able to address and possibly avoid anomalous behaviors by taking corrective actions in a timely manner. This paper presents an alarm prediction method based on data mining to generate patterns from historical alarm data, and use such patterns to train three deep learning approaches, namely long short-term memory (LSTM), bidirectional LSTM (Bi-LSTM) and gated recurrent unit (GRU). The prediction performance of the three deep learning approaches has been compared. Domain trained word embedding and pretrained word embedding (Word2vec) are used to feed embeddings to neural networks. The relevance of applying different word embedding is to explore the effect of the data preparation on the model performance. The Frequent pattern growth (FP) algorithm implemented in Rapidminer studio has been used to mine five months' worth of alarm logs. Finally, the best performing model is selected based on the accuracy of the model. The models are tested with a sequence of alarms and Bi-LSTM with domain trained word embedding achieves 93% in predicting the target alarms. However, from the results, we can also say that all three deep learning approaches can be used for predicting telecom alarms.