Day-ahead scheduling of thermal storage systems using Bayesian neural networks

The increased need for energy efficiency in buildings requires sophisticated scheduling strategies. A considerable challenge when developing such strategies is to address the stochasticity of demand appropriately. In this paper, we propose a day-ahead scheduling technique, which aims to minimize electricity costs, as well as power grid congestion. Our method considers energy storage systems with heat pumps and backup resistance heaters under uncertain heat demand. We employ a Bayesian neural network to model the stochastic consumer demand, which takes historical measurements as training inputs, and is able to model complex stochastic patterns. The model is then employed to generate sample demands, which are used to approximate the expected costs. The minimization of the resulting cost function corresponds to a stochastic optimal control problem with quadratic costs and mixed integer constraints. In a numerical simulation of a single-family building, the proposed approach is shown to perform better than a standard neural network-based scheduling scheme.