A plug-and-play home energy management algorithm using optimization and machine learning techniques

A smart home is considered as an automated residential house that is provided with distributed energy resources and a home energy management system (HEMS). The distributed energy resources comprise PV solar panels and battery storage unit, in the smart homes in this study. In the literature, HEMSs apply optimization algorithms to efficiently plan and control the PV-storage, for the day ahead, to minimize daily electricity cost. This is a sequential stochastic decision making problem, which is computationally intensive. Thus, it is required to develop a computationally efficient approach. Here, we apply a recurrent neural network (RNN) to deal with the sequential decision-making problem. The RNN is trained offline, on the historical data of end-users' demand, PV generation, time of use tariff and optimal state of charge of the battery storage. Here, optimal state of charge trace is generated by solving a mixed integer linear program, generated from the historical demand and PV traces and tariffs, with the aim of minimizing daily electricity cost. The trained RNN is called policy function approximation (PFA), and its output is filtered by a control policy, to derive efficient and feasible day-ahead state of charge. Furthermore, knowing that there are always new end-users installing PV-storage systems, that don't have historical data of their own, we propose a computationally efficient and close-to-optimal plug-and-play planning and control algorithm for their HEMSs. Performance of the proposed algorithm is then evaluated in comparison with the optimal strategies, through numerical studies.