Home energy management using deep reinforcement learning

November 30, 2022

With the increase of global energy consumption, the interest in demand-side management (DSM) is growing because of its ability to reduce the total electricity cost. With the increasing role of renewable energy resources in residential areas, especially solar power production, a key element for energy usage optimization may be the implementation of energy storage systems. This work focuses on applying deep reinforcement learning (DRL) to a home energy management problem that can meet the needs of real-life conditions with the objective of reducing the total energy cost. An environment integrating an energy storage system, photovoltaic (PV) solar energy generation, residential household load demand with a time of use (TOU) pricing scheme is presented. A mathematical formulation of the environment with a discrete action space is proposed and solved with Double Deep Q-Networks (DDQN), Duelling DDQN, Rainbow and Proximal Policy Optimization (PPO) algorithms. A state-of-the-art CNN-LSTM load forecasting model, which combines convolutional neural network (CNN) and long short-term memory (LSTM), is proposed and integrated with the DRL algorithms. A deterministic optimization mixed-integer linear programming (MILP) model is used as benchmark. The suggested methodology produces consistent and employable energy management strategies for real-time scenarios. The most promising results show that PPO algorithm is able to achieve a 35% reduction in total cost when compared with the MILP model, for a testing house with previously unseen data

Authors in the community:

• 

  A

  António Maria Geraldo Corte Real

  ist425731

Supervisors of this institution:

• 

  S

  Susana Margarida da Silva Vieira

  ist45346

mechanical-engineering - Mechanical engineering

por - Portuguese

Instituto Superior Técnico