# Reinforcement Learning-Based Management in IoT-Enabled Renewable Energy Communities: An Approach to Optimization for Comfort, Economy, and Sustainable Performance

## TL;DR

This paper introduces a reinforcement learning framework for managing renewable energy communities using IoT sensors to optimize comfort, cost, and sustainability.

## Contribution

A sensor-driven reinforcement learning framework for decentralized energy management in renewable energy communities is proposed.

## Key findings

- The proposed approach outperforms rule-based strategies in energy consumption and thermal comfort.
- Pre-trained agents maintain stable behavior in a community setting with limited exploration sensitivity.
- The framework enables scalable and cooperative energy management without explicit inter-agent communication.

## Abstract

The increasing deployment of Internet of Things (IoT) sensing infrastructures and distributed renewable energy resources is enabling the emergence of Renewable Energy Communities (RECs), which require intelligent, adaptive, and decentralized energy management strategies. This study proposes a sensor-driven reinforcement learning (RL) framework for the coordinated management of residential RECs, aiming to jointly optimize thermal comfort, economic savings, and environmental sustainability. Each household is equipped with a network of IoT sensors monitoring indoor temperature, energy production and consumption, battery state of charge, and user presence, which collectively define a discretized state space for a tabular Q-learning agent controlling heating systems and programmable appliances. A stochastic simulation environment is developed to realistically reproduce weather variability, building thermal dynamics, user activity profiles, and photovoltaic generation. To address the instability typical of multi-agent learning, a two-stage training strategy is adopted: agents are first pre-trained at single-house level using synthetic sensor data and are subsequently deployed within the full community, where coordination is achieved through shared reward components without explicit inter-agent communication. Performance is evaluated on a heterogeneous Renewable Energy Community (REC) composed of eleven households, including both prosumers and consumers. The simulation results show that the proposed approach significantly outperforms rule-based control strategies, achieving lower energy consumption, improved thermal comfort stability, and higher global reward. Moreover, pre-trained agents maintain stable and cooperative behavior when operating concurrently at community level, with limited sensitivity to exploration. These findings demonstrate that sensor-driven, lightweight reinforcement learning represents a viable and scalable solution for decentralized energy management in IoT-enabled Renewable Energy Communities.