Using machine learning to optimize energy systems

Jan 23, 2020

  • Collaboration between Warwick Business School, the University of Warwick and Fetch.ai explores the use of machine learning in energy systems
  • Optimizing how both heat and power are produced, stored and used across the Warwick campus
  • AI approach that’s model free and data driven: learns the optimal control policy from historical data
  • Initial results from a “virtual twin” reduced energy costs by 13-18% over a two-year period

In September 2019, the University of Warwick declared a climate emergency and set ambitious “net zero” goals for emissions.

Achieving this involves using less energy on campus, and paying attention to energy used in heating, cooling and transport as well as the usual focus on electricity.

Considering the supply, storage and use of energy across multiple energy vectors rapidly becomes complex.  Different, existing control systems model separate energy vectors across different parts of the Campus.

Artificial intelligence and machine learning are enabling a transactive energy approach where key parts of the energy system react and respond in ways not previously possible.

WBS Professors of Practice David Elmes and Mark Skilton have collaborated with Fetch.ai, one of the leading companies in the use of AI to offer a decentralized connectivity platform that enables devices to connect directly with digital agents delivering autonomous solutions to complex tasks. Development of the machine learning approach is led by Dr Yujian Ye of Fetch.ai in collaboration with Chris Conlan, a data science PhD student in the Warwick Institute for the Science of Cities together with Joel Cardinal and Mark Jarvis of the University of Warwick Estates department.

“We need innovative approaches to running multi-vectoral energy systems more efficiently,” comments Professor Elmes.  “Warwick has run our campus as a smart, local energy system for a decade or more and reduced our emissions per unit of income by 48% since 2006.  But we need to do more to reach net zero.”

“This approach uses multi-agent reinforcement learning to give a model-free, data driven approach where AI agents representing key energy assets across the campus learn the optimal control policy from real-world historical data,” adds Professor Mark Skilton.  A virtual twin can then run alongside the real system to demonstrate the different energy management strategies that arise from the AI approach.  “We have started with the combined heat and power engines we have for self-generating energy on campus along with thermal storage and the demand for heat and electricity.  That initial scope is a small part of the campus energy system overall and initial results suggest potential for a 13-18% reduction in energy costs compared with existing control systems.  But we now need to study what decisions the AI twin has made and how they can be brought across into the real world.”

“Fetch.ai uses autonomous software agents to complete useful economic work in a wide range of different markets,” comments Humayun Sheikh, Fetch.ai CEO and co-founder.  “Our collaboration with Warwick Business School shows that energy systems offer a great example of how our decentralized AI capability empowers efficient decision-making.  Enabling autonomous agents to work this way unleashes the true potential for an internet of things”

Two years’ worth of data across 16 sensors on the Warwick campus provided over 200,000 data points for Fetch.ai.  The energy management strategy of each asset was then optimized using an individual actor neural network.

“This is only an initial step in applying machine learning to the more efficient and sustainable management of a smart local energy system,” adds Professor Elmes.  “A model-free, data driven approach offers an exciting alternative for the greater complexity of energy systems that want to include heat, power and transport in a smart local energy system approach.”