Abstarct: Countering cyber-attacks has become an essential part of modern cybersecurity strategies. Among all the trending solutions to deal with cyber-attacks, early cyber-attack prediction has attracted a lot of attention. This novel approach, which currently is in the early research stages, could prevent much damage and harm if implemented effectively on a real-world scale. Researchers around the world are examining various methods from different perspectives with the goal of predicting cyber-attacks. One of these potential solutions is deep learning, which, according to many researchers, is currently the most promising way to anticipate and counter cyber-attacks. For this purpose, this research has examined the performance of three deep learning models (LSTM, GRU, BiGRU). These models have been studied and compared with different hyperparameters to predict network intrusion attacks. Finally, the prediction errors of these models have been examined and compared with each other. This research explains how raw data from intrusion alerts can be converted into useful information for training deep neural networks. It also introduces features of intrusion alerts that are more important for prediction. In this study, it has been observed that the GRU model generally performs better. However, considering the limitations and conditions of the research, as well as reviewing and comparing the performance of models in the examined hyperparameters, it seems that with more extensive and comprehensive data, the performance of BiGRU would surpass its competitors.