Scholarlink

<p>Tankless (instant) water heaters eliminate standby thermal losses and are rapidly gaining global market share, yet they remain almost entirely absent from the extensive reinforcement learning (RL) literature on smart water-heater control. This paper presents FlowSmart-AI, a multi-agent framework that applies safe reinforcement learning to demand-adaptive, energy-efficient control of electric and gas tankless water heaters within a cloud-edge computing paradigm aligned with AI-driven predictive energy management. The system employs a bifurcated design: (i) an <i>Edge Layer</i> implementing a high-frequency proximal policy optimization (PPO) agent with a recurrent look-ahead flow predictor for sub-second thermal modulation, and (ii) a <i>Cloud Intelligence Layer</i> executing a soft actor-critic (SAC) meta-agent that learns long-term occupancy patterns, schedules Legionella-safe hygiene pulses, and coordinates demand-response participation. A formally defined constrained Markov decision process (CMDP) safety layer ensures that stagnant water at Legionella-growth temperatures (25–45&#xa0;°C) is automatically flushed—embedding formal safety constraints directly in the RL decision loop. A federated averaging (FedAvg) protocol with a household-specific personalization layer enables fleet-wide policy improvement without transmitting raw usage data, thereby preserving user privacy in multi-household deployments. FlowSmart-AI is trained and evaluated in a calibrated simulation environment modelling the rapid thermal dynamics of a 9.6&#xa0;kW electric and 199,000 BTU/h gas tankless unit. Across 12 synthetic household archetypes and 3 tariff structures, the system achieves energy savings of 20–40%, cost reductions of 22–45%, and time-to-temperature improvements of 15–30% compared with thermostat, PID, and rule-based baselines, maintaining outlet temperature within ± 1.0&#xa0;°C of the set-point for 98% of draw events. Additionally, a comparison against a vanilla SAC agent and a DQN baseline isolates the contribution of architectural innovations. Sensitivity analysis of reward weights and statistical significance testing (paired t-tests, <i>p</i> &lt; 0.01) strengthen the validity of reported results. Ablation studies confirm the contribution of each architectural component. The findings demonstrate that safe, privacy-preserving deep RL can extend the energy-saving benefits documented for tank-based controllers to the structurally distinct and commercially significant tankless water-heating domain, positioning cloud-edge AI platforms as viable solutions for building-edge energy management. All simulation code, trained model weights, and environment configurations are publicly available at [repository link] to support reproducibility.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

FlowSmart-AI: Multi-Agent Safe Reinforcement Learning for Demand-Adaptive Control in Tankless Water Heating Systems

  • 期刊论文

  • 发表日期：

    2026年6月15日

查看全文
  • Soujanya Ambala,
  • Manisha Rajendra Dhage,
  • G. Thiraviaselvi,
  • B. Mouleswararao,
  • Kodali Lakshmi Padmavathi,
  • Ruchi Patel

摘要

Tankless (instant) water heaters eliminate standby thermal losses and are rapidly gaining global market share, yet they remain almost entirely absent from the extensive reinforcement learning (RL) literature on smart water-heater control. This paper presents FlowSmart-AI, a multi-agent framework that applies safe reinforcement learning to demand-adaptive, energy-efficient control of electric and gas tankless water heaters within a cloud-edge computing paradigm aligned with AI-driven predictive energy management. The system employs a bifurcated design: (i) an Edge Layer implementing a high-frequency proximal policy optimization (PPO) agent with a recurrent look-ahead flow predictor for sub-second thermal modulation, and (ii) a Cloud Intelligence Layer executing a soft actor-critic (SAC) meta-agent that learns long-term occupancy patterns, schedules Legionella-safe hygiene pulses, and coordinates demand-response participation. A formally defined constrained Markov decision process (CMDP) safety layer ensures that stagnant water at Legionella-growth temperatures (25–45 °C) is automatically flushed—embedding formal safety constraints directly in the RL decision loop. A federated averaging (FedAvg) protocol with a household-specific personalization layer enables fleet-wide policy improvement without transmitting raw usage data, thereby preserving user privacy in multi-household deployments. FlowSmart-AI is trained and evaluated in a calibrated simulation environment modelling the rapid thermal dynamics of a 9.6 kW electric and 199,000 BTU/h gas tankless unit. Across 12 synthetic household archetypes and 3 tariff structures, the system achieves energy savings of 20–40%, cost reductions of 22–45%, and time-to-temperature improvements of 15–30% compared with thermostat, PID, and rule-based baselines, maintaining outlet temperature within ± 1.0 °C of the set-point for 98% of draw events. Additionally, a comparison against a vanilla SAC agent and a DQN baseline isolates the contribution of architectural innovations. Sensitivity analysis of reward weights and statistical significance testing (paired t-tests, p < 0.01) strengthen the validity of reported results. Ablation studies confirm the contribution of each architectural component. The findings demonstrate that safe, privacy-preserving deep RL can extend the energy-saving benefits documented for tank-based controllers to the structurally distinct and commercially significant tankless water-heating domain, positioning cloud-edge AI platforms as viable solutions for building-edge energy management. All simulation code, trained model weights, and environment configurations are publicly available at [repository link] to support reproducibility.