• Weather-informed adaptive reinforcement learning is evaluated under sustained heatwave stress at neighbourhood scale. • Decentralized control redistributes flexibility non-uniformly across buildings, shaping demand rather than only clipping peaks. • Forecasting stabilizes adaptive learning, suppressing extreme demand tails while preserving thermal comfort. • Resilience emerges from coherent coupling of anticipation, policy adaptation, and action smoothness. • Stress-oriented indicators reveal how decentralized learning enhances climate-resilient urban energy operation. Heatwaves increasingly challenge urban electricity networks by intensifying and reshaping cooling demand over extended periods. This study investigates how weather-informed Adaptive Reinforcement Learning for Energy Management (ARLEM), a decentralized model-free control framework, redistributes cooling demand and manages systems stress during sustained heatwaves, focusing on the roles of forecast horizon, policy-update frequency and action-selection constraints in determining climate-resilient control performance. ARLEM is evaluated for a representative residential neighbourhood in Madrid consisting of four building archetypes and simulated under near-future summer climate conditions derived from an ensemble of regional climate projections. A broad range of control configurations is explored, including predictive and non-predictive learning, different forecast horizons, policy-update schedules, and action-selection constraints. System behaviour is assessed using stress-oriented indicators that capture demand flexibility, adaptive responsiveness, and amplification of demand under extreme conditions. Results show that predictive ARLEM operates as a distribution-shaping controller rather than a peak-only optimizer. During heatwaves, predictive ARLEM reduces neighbourhood-scale mean cooling demand by approximately 10–15% while lowering high-quantile demand levels by 80–130 kWh, without destabilizing indoor thermal conditions. Flexibility is allocated unevenly across buildings, allowing heterogeneous responses that collectively reduce system stress without synchronized rebound effects. The findings demonstrate that resilience in decentralized adaptive control emerges from coherent alignment between anticipation, adaptation speed, and action constraints. By managing sustained stress rather than merely suppressing short-lived peaks, weather-informed ARLEM provides a scalable, privacy-preserving, and context-aware framework for neighbourhood-scale energy management under increasingly severe climate conditions.
Vahid M. Nik (Wed,) studied this question.