Master Thesis on What-If Reasoning in LLM Agents for Evaluating HEMS Parameter Effects Using Time-Series Foundation Models via MCP

How can modern Home Energy Management Systems (HEMS) turn complex user requests into concrete and dependable actions? Large Language Models (LLMs) provide powerful capabilities for dialogue and qualitative reasoning, while specialized Time-Series Foundation Models (TSFMs) bring complementary strengths in capturing and predicting complex numerical dynamics. In your thesis, you will explore how these technologies can come together for accurate forecasting and closed-loop what-if reasoning – contribute your ideas to intelligent energy management and apply now!

• During your assignment, you will connect a specialized Time-Series Foundation Model as an independent service to an existing LLM agent ecosystem using the standardized Model Context Protocol (MCP).
• You will investigate how numerical time-series predictions generated by the TSFM can be mathematically abstracted at the server level into concise semantic representations, such as load peaks or solar generation surplus, tailored for LLM reasoning.
• Additionally, you will design an interactive evaluation loop in which the agent explores system parameters, such as charging schedules and setpoints, the model predicts the resulting curve shifts, and the LLM assesses whether the user's objectives are optimally achieved.
• Finally, you will systematically evaluate scenarios and data representations to determine where specialized TSFMs deliver measurable advantages over pure LLM reasoning in terms of computational efficiency, token consumption, latency, and predictive accuracy.