The rapid proliferation of social media platforms has generated vast amounts of user-generated content, offering valuable data for sentiment analysis. While significant progress has been made in English sentiment analysis, Arabic sentiment analysis remains challenging due to linguistic complexities and dialectal variations. This study addresses these challenges by employing a novel approach that combines multiple deep learning models with stacking ensemble learning to analyze sentiment in Modern Standard Arabic (MSA) and Saudi dialects. The proposed methodology involves deep learning models, including Convolutional Neural Networks (CNN), Bidirectional Long Short-Term Memory (Bi-STM), and Recurrent Neural Networks (RNN), utilizing AraVec pre-trained word embeddings. To further enhance performance, we propose stacking ensemble learning, combining the strengths of these base models with meta-models. We investigated the effectiveness of three meta-models (Logistic Regression, Random Forest, and Gradient Boosting) to determine the optimal combination for Arabic sentiment analysis. Experiments demonstrated that individual models achieved strong baseline performances, with RNN achieving the highest accuracy of 91.83%. Stacking ensemble models significantly outperformed individual models, with the Logistic Regression meta-model achieving the highest accuracy of 97.22%, representing an improvement of 5.39%. The results highlight the effectiveness of our proposed model in understanding Arabic sentiment, demonstrating that stacking ensemble learning can leverage the unique strengths of different deep learning models to improve predictive performance.