Versatile and self-adapting smart home control

The research focuses on the control of a residential photovoltaic and battery system, designed to mitigate uncertainties from the distribution grid perspective while following prescheduled power profiles at the residential meter level. Specific attention is paid to reduced communication requirements, contrary to the traditional energy management strategy. This autonomy is achieved here using reinforcement learning, which has proven effective in the dynamic management of energy systems. The controller’s effectiveness is compared to other methods, such as rule-based controls and optimization approaches. In the use case, the objective is to follow in real-time a power profile based on the moving average of the net load over a week. The controller’s decision-making process is driven by a proximal policy optimization agent, which uses sequential temporal inputs such as the target power profile, the state of charge, and the net load data from the previous time step. To enhance the performance of the controller, the research explores the integration of a neural network architecture inspired by image processing techniques. Performance evaluation criteria include the daily energy deviation and the daily avoidable energy deviation from the target profile.