Scientists at South Korea’s Chung-Ang University have developed a framework for smart-home energy management systems.

The steady advance into the era of smart grids and smart homes presents opportunities for using energy more efficiently. Fortunately, most decisions and actions needed to optimise energy use can be delegated to home energy management systems (HEMSs), which manage the energy consumption of home appliances by scheduling when washing machines should start and turning air conditioners on and off.

In general, a HEMS works to reduce electricity bills while taking the user’s preferences and comfort into account.

Usually, hand-crafted models that use abstract equations to represent appliances and distributed energy resources are used to programme HEMSs. But these models and optimisation methods are not very versatile and give suboptimal schedules. An alternative is to use centralised machine learning, where data from thousands of users are collected, sent to a central server, and used to train a model from the ground up. This strategy, however, may prove costly, computationally complex, and susceptible to hacking.

To tackle these problems simultaneously, associate professor Dae-Hyun Choi and PhD student Sangyoon Lee from Chung-Ang University have proposed a novel data-driven strategy. The researchers have developed a framework for HEMS based on federated deep reinforcement learning (F-DRL), combining the advantages of various machine learning techniques.

Their study was published in IEEE Transactions on Industrial Informatic.

The key word to note in F-DRL is federated, which indicates a decentralised form of machine learning. In the proposed framework, each home has a HEMS connected to various appliances and devices, which collects data about its users’ energy consumption and tries to optimise a schedule for the appliances by creating a local model.

These local models are all uploaded to a global server, which averages them to produce a global model. Afterwards, each HEMS replaces its local model with the global model and proceeds to train it once again using local data. This process is repeated several times, progressively improving the accuracy of both global and local models.

“In a conventional centralised DRL model, the global server must have access to the data of all local devices to generate the model of the global system,” said Choi. “This results in data privacy concerns for local devices. However, in our federated DRL method, the system does not require the sharing of user data because only the parameters of local and global model are exchanged. In turn, this helps prevent local data leakages and protects the users’ privacy.”

The researchers tested their approach through simulations, showcasing its optimal performance when scheduling the operation of various appliances at different homes.

“To the best of our knowledge, this is the first HEMS framework based on federated DRL that can manage the energy consumption of multiple smart homes and ensure the comfort of the consumers while taking their preferences into account in a distributed manner,” said Choi.

In addition to its low computational complexity and relatively fast training process, the proposed framework can easily support the addition of more appliances in each house.

Choi envisions a more comprehensive version of this framework that also considers electric cars and energy trading between households.

“We certainly have our fingers crossed for the secure optimisation of energy consumption in the future,” Choi said.