Meta-reinforcement learning for multi-energy building microgrids: Curriculum design, rapid adaptation, and scalable deployment

Multi-energy building microgrids meeting local demands is a promising solution for sustainable and self-sufficient energy communities. However, consumption diversity across heterogeneous buildings creates a deployment bottleneck that requires an adaptive and scalable energy-management strategy. A novel meta-reinforcement learning framework was proposed in this study that integrates an enhanced reptile-based meta-learning algorithm with Proximal Policy Optimization (PPO) to learn a generalized policy capable of rapid building-specific personalization. Two curriculum learning approaches were applied to assess the impact of task organization on meta-learning outcomes, focusing on seasonal variations and building diversity. Rigorous hyperparameter optimization was performed to identify high-performance models (three per curriculum) for comprehensive deployment analysis. The trained meta-policies were deployed on five unseen test buildings across four seasons. The adapted policies achieved a mean reward improvement of 23.9% over the trained meta-policy, with constraint violations reducing by 38% and daily operating costs decreasing by 12.2%, statistically validated through a Wilcoxon signed-rank test (p = 1.11 × 10 − 20 ). Curriculum-2 (progressive building diversity) outperformed Curriculum-1 (seasonal progression) by 53% in mean adaptation improvement, demonstrating that learning building heterogeneity is more valuable than mastering seasonal patterns for adaptive energy management. Future research directions include hardware-in-the-loop validation, adaptive curricula, and multi-agent coordination for community-scale deployment.