RHEA: Residential Home Energy Advisor

Large language models (LLMs) can learn how to solve many common knowledge tasks such as language translation, code generation or dialogue generation. To provide a response, the LLMs combine the knowledge that was acquired from training along with the user’s input. Although this approach covers many cases, it would be valuable if the models could also use data from different sources, such as meter measurements. Incorporating such data could open up new avenues for using LLMs in data-centric applications like energy management systems. One way towards this direction is with the use of external tools such as APIs, which can retrieve and process relevant data for each task. This scenario is explored in current research with the creation of RHEA, an autonomous home energy advisor. RHEA is designed to monitor, control, and provide insight into the energy usage of electrical home appliances. This is accomplished by enabling a pre-trained LLM to generate calls to external energy analytics APIs, which process energy measurements obtained by the user’s home. The assistant is built on top of a known pretrained LLM that was fine-tuned to learn how to generate calls to three specific energy-related external tools only when needed while maintaining its capabilities gained from the original training. Additionally, the fine-tuning enables the assistant to generalize and learn to generate calls to similar APIs with minimal examples and without requiring further training.