<p>Continuous aggravation of the greenhouse effect and the increase of energy costs have pressurized the manufacturing industry to reduce energy consumption and improve energy efficiency in manufacturing systems. Motivated by the poor inference ability and hallucination trap suffered in the conventional energy management methods, this paper proposed an energy knowledge graph (EKG)-embedded Large Language Models (LLMs) for the energy management of the manufacturing systems. Firstly, the Bidirectional Encoder Representation from Transformers Improved Transformer-Improved Conditional Random Field is proposed to construct an energy knowledge graph for manufacturing systems. Then, this developed EKG is embedded into the LLM using pre-training. Finally, this embedded LLM is compensated using the subgraph generation-retrieval method. By using the EKG-embedded LLMs, the final developed model can be constructed in a guided way and assigned for a good inference ability over different energy management questions. Experiments have been done on the actual manufacturing industry, and they have been evaluated subjectively and objectively. Results indicate that the proposed methods have attained an impressive 98.33% in terms of precision, recall, F-1 score and accuracy. Besides, the overload test demonstrates a correctness rate exceeding 95%, and the proposed method received scores of over 9 out of 10 across various aspects in the subjective evaluation segment, suggesting its capability to effectively address the diverse questions encountered in manufacturing operations.</p>