Recently, In-context learning (ICL) with Large Language Models (LLMs) has achieved promising results on the Aspect-based Sentiment Analysis (ABSA) task. However, existing ICL methods are limited to a single views and lack dynamic adaptability, resulting in low-quality ICL examples. To address this issue, we propose a dynamic multi-views in-context learning LLMs-based framework (DMICL) for ABSA, which includes a multi-views retrieval strategy and a dynamic fusion module. Specifically, the multi-views retrieval strategy retrieves ICL examples by simultaneously considering three views: overall semantic relevance, syntactic structure relevance, and aspect-sentiment semantic relevance. To achieve this goal, we construct positive and negative pairs from these three views and train the retriever using contrastive learning. Since the importance of different views varies across queries, we design a dynamic fusion module in the retriever to evaluate the importance of the three views in different queries. This module adaptively adjusts the weights of each view based on features such as syntactic tree complexity, aspect count, and sentiment diversity. Experimental results on ABSA datasets demonstrate that our framework significantly outperforms baselines.