Sentiment analysis represents a fundamental task in natural language processing (NLP) critical for extracting and analyzing emotions and opinions represented in textual data. The complexity of the Arabic language presents unique obstacles to accurate sentiment categorization in Arabic sentiment analysis. As cutting-edge NLP models, transformers have demonstrated exceptional performance on various tasks, including sentiment analysis. However, because of the multiplicity of pre-trained transformer models available, fine-tuning for best performance is required, making the procedure time-consuming and demanding significant knowledge. To overcome these challenges, this research paper introduces an ensemble learning approach consolidating multiple transformer models into a unified framework. In the initial stage, a set of baseline transformer models, comprising AraBERT, CAMeIBERT, and Roberta, undergo training and fine-tuning. Subsequently, three distinct ensemble fusion techniques-voting, weighted voting, and meta-learning-are applied to aggregate the predictions generated by these baseline transformer models. Conducting comprehensive experiments on the Egyptian Arabic Corpus, the paper reveals transformer models' superiority over existing methods applied to the corpus. Furthermore, the proposed ensemble approach exhibits enhanced performance for these transformer models, resulting in even more promising outcomes in sentiment classification tasks.