Core Concepts
A real-world testbed called HomeLabGym is introduced to ease the deployment and evaluation of innovative Home Energy Management System (HEMS) controllers, particularly those based on reinforcement learning, in a controlled smart home environment.
Abstract
The paper introduces HomeLabGym, a real-world testbed for evaluating Home Energy Management Systems (HEMS) and flexible asset controllers. The key highlights are:
Motivation: Evaluating HEMS solutions typically relies on simulations, which may not capture the full complexity of real-world scenarios. On the other hand, real-world testing is labor-intensive, especially when dealing with diverse assets using different communication protocols.
Approach: HomeLabGym is a Python-based package that interfaces with the various sensors and smart appliances in the IDLab HomeLab, a state-of-the-art smart home testbed. It provides a modular design to abstract away the complexities of communication protocols, enabling researchers to focus on prototyping and deploying HEMS controllers.
Experiment: The authors demonstrate the "plug-and-play" nature of HomeLabGym by deploying a Deep Q-Network (DQN) agent, trained in simulation, to control a home battery for energy arbitrage based on real-time electricity prices. The experiment reveals a 2% lower reward in the real-world deployment compared to simulation, highlighting the value of real-world testing in uncovering complexities not captured in simulations.
Conclusion: HomeLabGym provides an easy-to-use, real-world testbed for HEMS controllers, allowing researchers to prototype, deploy, and analyze their solutions in a controlled smart home environment. The authors invite collaborations to access the IDLab HomeLab testbed facilities.
Stats
The charging power of the home battery often falls below the 1 kW setpoint, resulting in a non-linear charging speed.
The heat pump exhibits a step-like behavior as it approaches the room heating set-point.
Quotes
"Realistic simulation of real-world households is both complex to model and to run, we believe there is a need for such easy-to-use real-world test environments to deploy, validate, and analyze flexible asset controllers."
"Our experiment revealed that the real-world reward is 2% lower than that attained in simulation. This disparity can be attributed to variations between real-world conditions and simulation environments."