New deep reinforcement learning technique sees agents cut home energy costs by nearly 20%

Apr 23, 2020

  • Innovative approach using deep reinforcement learning algorithm sees autonomous agents reduce daily home energy costs by almost 20%
  • Current methods use mathematical models to optimize energy management strategies that rely on external predictions for factors such as renewable energy generation
  • New proposal is model-free and utilizes smart meter data to achieve real-time autonomous energy control and reduce costs

A study by and Imperial College London has demonstrated how’s autonomous agents can use a new deep reinforcement learning technique to lower daily domestic energy costs by nearly 20%.

This cutting edge research, accepted into the highly prestigious International Joint Conference on Artificial Intelligence to be held in Yokohama in Japan in July, uses a model-free and data-driven approach to deliver accurate real-time information. At present, autonomous energy systems rely on generation and consumption models that forecast the amount of energy that will be required by the user the following day. Such an approach is hindered by its lack of flexibility. This inherent flaw prevents it from reacting when circumstances inevitably change. For example, a model-based methodology will have limited success predicting the amount of solar energy generated the following day by panels on the roof of a homeowner’s house. In contrast, the new model-free reinforcement learning method outlined in the paper will help smart energy grids efficiently utilize renewable energy sources.

An overview of the new deep reinforcement learning technique used to optimize energy usage

In the paper,’s machine learning scientist Yujian Ye and head of research Jonathan Ward collaborated with researchers at Imperial College London. Using real-world data, they developed a real-time autonomous energy management system for a smart home equipped with various energy devices including solar panels, a gas boiler, an electric heat pump, a thermal energy storage and an electric energy storage system. The data enabled the agent to be trained using deep reinforcement learning. The agent gradually acquired the best energy management strategies by learning from repeated interactions through the process of trial and error. As such, the agent representing the home wasn’t required to understand how the overall energy system operated, but learned to optimize the usage of each energy resource. Once trained, the agent acted within milliseconds to autonomously respond to changes in the home environment in order to fulfill the homeowner’s energy needs at the lowest possible price.

How the autonomous energy management system operates in the home

This unique deep reinforcement learning method represents a significant improvement on other state-of-the-art approaches to energy management optimization.

For example, the conventional model-based optimization approach called stochastic programming uses a scenario-based method to model potential uncertainties. In order to achieve greater accuracy, a high number of scenarios is required and this quickly becomes computationally very expensive in terms of time and resources. An alternative approach, robust optimization, is less computationally expensive, but the resultant energy management strategies are often too conservative due to the model’s tendency to hedge against the worst case scenario of the uncertain parameters. When the two approaches were compared with the new methodology set out by and Imperial College, stochastic programming was found to be 6.28% less cost-effective and robust optimization was found to be 10.21% less cost-effective.

Moreover, when compared with current deep reinforcement learning techniques, research showed that agents reduced the home’s daily energy costs by 19.1% compared to Deep Q Network and by 7.95% compared to Deep Policy Gradient methodology.

Autonomous, real-time and model-free energy management systems will be at the core of the smart homes of the future. will continue to push the boundaries of AI, machine learning and blockchain technology and will use its platform to deliver optimized energy management systems that both enhance flexibility and reduce the cost for consumers.

In the months ahead, possible areas for research include evaluating the effectiveness of the proposed energy management system for a heating, ventilation and air conditioning system. The most exciting challenge is showcasing the impact the new method of reinforcement learning will have when it is implemented across the energy management systems of several houses. Each property will be represented by an autonomous agent and will provide further evidence of how the multi-agent system works. The research will also demonstrate how agents can incentivize peer-to-peer trading of energy between neighbors in the setting of a local energy market.

Get involved

If you’re a developer, find out more about’s autonomous agents by checking out the documentation section of our website, which includes some demos.

  • Watch videos outlining our technology and follow our tutorials on YouTube
  • If you’re not a developer, you can stay up to date with our progress by joining us on Telegram or by following us on Twitter