Event-driven online adaptation for microclimate control via meta-learning and evolutionary optimization

Ensuring occupant comfort while minimizing energy consumption remains a central challenge in smart residential buildings operating under continuously evolving and unlabeled data streams. This paper proposes and evaluates a simulation framework for residential microclimate control designed to investigate whether online adaptation can improve model-based comfort regulation and estimated energy efficiency under non-stationary conditions. Unlike conventional HVAC controllers based on fixed schedules or static feedback, the proposed approach adapts online without predefined task boundaries. The framework combines Temporal Convolutional Networks to model long-range temporal dependencies, Model-Agnostic Meta-Learning to enable rapid few-step adaptation from recent observations, and a Genetic Algorithm to evolve hyperparameters by maximizing post-adaptation error reduction. The framework is evaluated in a simulated multi-zone residential environment using real weather and occupancy data. Simulation results indicate that, under assumed thermal conductance, HVAC power, runtime reduction, and tariff conditions, Spontaneous Online Learning (SOL) yields an estimated reduction in daily HVAC energy consumption of 19% (approximately 5.7 kWh/day from a 30 kWh baseline) under simplifying assumptions, while maintaining indoor temperature within \(\pm 1,^\circ \textrm{C}\) and Predicted Mean Vote within \(\pm 0.5\) for over 99% of the time. The framework is therefore positioned as a methodological prototype, with real-world performance and deployment feasibility reserved for future validation.