Technology and networks have improved significantly in recent decades, and Internet services are now available in almost every business. It has become increasingly important to develop information security technology to identify the most recent attack. Many signature-based intrusion detection techniques promise great accuracy and low false alert rates. However, they struggle when confronted with emerging threats. This work employs the autoencoder, a type of neural network, to find possible zero-day threats by looking for differences in the data. This research is significant because, unlike previous efforts, it does not rely on the use of explicit labels during training. The premature convergence they caused was a problem for the optimization strategy used in the real world to identify zero-day attacks and other novel threats. This work proposed a unified ensemble autoencoder based on the Dynamic weighted Cuckoo Search Algorithm (DWCSA) to choose the most practical features for detecting anomalies in network traffic. As a result, the suggested Dynamic weighted based Cuckoo Search Algorithm (DWCSA) is modified to enhance performance while maintaining its fundamental structure. The model uses an improved DWCSA to determine which dataset features are the most important. Then an ensemble Autoencoder is used to improve further the optimal features that the enhanced DWCSA learned. Accuracy was enhanced by the proposed model, which overcame the constraint problems that had arisen throughout the feature selection process. The study demonstrates how to practically design and deploy a suitable approach and procedure, allowing even non-experts to identify the zero-day attack. The investigation of the results indicates that the proposed zero-day detection methods have more excellent results for the highest overall accuracy of 99.93% on CICIDS18 data sets. Overall, the results for identifying zero-day threats using the proposed technique are promising.