Network Intrusion Detection Systems (NIDS) are vital in safeguarding computer networks from cyber threats. However, designing an effective NIDS configuration involves optimizing multiple objectives, often leading to suboptimal solutions. This paper presents an innovative approach combining two powerful optimization methods, the Whale Optimization Algorithm (WOA) and Simulated Annealing (SA), for feature selection for NIDS. Our proposed WOA-SA methodology aims to achieve superior results by balancing global exploration and local improvement capabilities. Additionally, Deep Learning (DL) techniques are integrated to enhance feature extraction and classification accuracy for zero-day and new types of attacks with optimal DL models. This paper provides a detailed exposition of WOA-SA for feature selection and its practical application to NIDS optimization. This paper aims to achieve the maximum detection rate for zeroday attacks while reducing the false alarm rate (FAR) and reducing computational complexity. The comprehensive analysis of DL different approaches such as Long Short-term Memory, Convolutional Neural Networks, Recurrent Neural Networks, and Deep Neural Networks was carried out on the original and optimal feature set of the BOT-IOT 2020 dataset. From WOA-SA the feature set was reduced to 13 from 79, these 13 selected feature set performances were tested using the DL approach. Experimental results show that model accuracy improved with optimal features, it was improved to 2.2% and also reduced in FAR of the model to 10%, to show how well the optimal feature-based NIDS model performs in comparison to other well-known DL approaches. The proposed method also shows reduced computational complexity due to a reduced number of features. On the whole, our proposed design outperforms the current approach in terms of computational complexity, zero-day attack detection, accuracy, and FAR.