Stochastic home energy management system via approximate dynamic programming

Abstract

This study proposes an approximate dynamic programming (ADP) method for a stochastic home energy management system (HEMS) that aims to minimise the electricity cost and discomfort of a household under uncertainties. The study focuses on a HEMS that optimally schedules heating, ventilation, and air conditioning, water heater, and electric vehicle, while accounting for uncertainties in outside temperature, hot water usage, and non-controllable net load. The authors approach the ADP-based HEMS via an effective combination of Sobol sampling backward induction and a K–D tree nearest neighbour techniques for the value function approximation. A subset of possible states is sampled and used to create an approximation of the value of being in aggregated states. They compare the ADP approach with other prevailing HEMS methods, including dynamic programming (DP) and mixed-integer linear programming (MILP), in a model predictive control framework. Simulation results show that the proposed ADP approach can yield near-optimal appliance schedules under uncertainties when finely discretised. Merits and drawbacks of the proposed ADP method in comparison with DP and MILP are also revealed.