Human-in-the-loop preference calibration for home energy management via LLM-enabled contract-constrained language interfaces

As residential energy systems become increasingly electrified and coordination-intensive, home energy management systems (HEMS) must operate economically while remaining aligned with evolving user-authorized comfort and service boundaries. Existing personalization approaches rely on questionnaire initialization, fixed preference parameters, or indirect behavioral corrections, limiting alignment under changing needs and the use of runtime natural-language feedback for long-horizon calibration. This paper proposes a large language model (LLM)-enabled human-in-the-loop framework that converts user utterances into authorization-aware, contract-constrained, machine-checkable events for downstream control and preference updating. The LLM is a contract-constrained semantic interface, not a direct controller. Explicit intents support bounded execution and direct profile correction, whereas vague feedback requires option negotiation and user confirmation before bounded updating. A Contract-Informed Preference Update (CIPU) mechanism combines structured interaction evidence with projection and trust-region stabilization, while model predictive control remains the safety-constrained execution layer. In simulations, the proposed method reduces normalized preference distance by 22.50% and complaints by 74.46% relative to Survey MPC across three role-conditioned profiles. On a controlled 188-utterance language-interface benchmark, the proposed interface achieves 100% structured-event exact match and a 0% parser-level false direct-execution rate. Closed-loop ablations show 59.40% fewer complaints than the explicit-only variant, while event-log auditing of the no-contract-gate variant yields a 36.90% event-interface mismatch rate after gate removal. Under mid-horizon preference drift, final preference mismatch is reduced by 44.09% on average and up to 71.43% relative to Survey MPC.