Emotion prediction seeks to automatically identify and analyze emotional tendencies in text. Incorporating user profiles enables a more precise and personalized approach by leveraging individual characteristics and behavioral data. Recently, Large Language Models (LLMs) have demonstrated remarkable generalization capabilities, often surpassing traditional methods in emotion prediction tasks. However, achieving effective personalization remains a critical challenge. In this paper, we propose PIE-E (Personalized Instruction-tuned Emotion prediction with LLMsEnsemble), a novel framework that fine-tunes LLMs using user profile information to enhance personalized emotion prediction. Our approach retrieves relevant historical user data to provide supplementary context for a more comprehensive understanding of emotions in text. We also introduce emotional stimuli prompts inspired by emotion regulation theories to enrich the model’s emotional comprehension. To ensure robustness, we adopt a cross-consistency mechanism to ensemble predictions from multiple LLMs, including Baichuan2, Qwen2, and glm4. Extensive experiments on a Weibo dataset demonstrate that PIE-E effectively integrates user profiles and psychological insights, achieving significant improvements in both accuracy and personalization for emotion prediction.