Energy-domain tables often face significant challenges in semantic interoperability because semantically related concepts may be represented with inconsistent terminology, units, and structural formats across data sources. Large language models provide a promising solution due to their strong contextual reasoning ability, but their predictions can vary across models and simple aggregation strategies are often insufficient to ensure robust ontology-guided annotation. In this paper, we propose a multi-LLM framework for ontology-based semantic annotation of energy-domain tables. The framework combines ontology preprocessing, multi-model candidate generation, and a confidence-aware aggregation mechanism, termed Multi-LLM Soft Voting (MSV), to improve the reliability of semantic prediction. It supports both CTA-only setting and joint CTA+CPA setting, and further investigates how different ontology representations affect annotation quality and computational efficiency. We evaluate the proposed framework on real energy-domain tables using multiple commercial LLMs. The experimental results show that MSV consistently improves annotation performance over the baseline, while structured ontology input further reduces both computational time and LLM token cost.