Security of data is the most important concern in the present day. It is important to perceive the vulnerability of data from a variety of intrusion attacks that can negatively affect the functionality of any network or system. Prevailing intrusion detection systems are currently not able to cope with the evolving and complex nature of intrusion operations on computer networks. To prevent intrusion, we propose a new feature selection technique that takes advantage of deep learning techniques. Initially, intrusion data is collected from two different datasets. Then, optimal features are selected using a Stacked Autoencoder and subsequently the selected features are passed to an Adaptive Bi-directional Long Short-Term Memory (ABi-LSTM) classifier to classify the data as normal or intrusive. Furthermore, the parameters of ABi-LSTM model are optimally selected using the Enhanced Pelican Optimization Algorithm technique. The performance of the suggested method is examined using the suggested system and several metrics.