An online HEMS scheduling method based on deep recurrent neural network

For a daily-basis scheduling of an energy system, energy management system often enough to use not a global optimal scheduling but a near optimal scheduling. The article proposes an online scheduling framework without online optimization. The framework is built from two encoder-decoder architectures to extract features of time series; a multi-layer long short-term memory regression model for multi-step time-series forecasting, and multi-class and single-label classification model for on/off scheduling of a device. The models are estimated at offline, and return scheduling from historical time series as input, at online. We evaluate the accuracy of scheduling from the viewpoint of Kullback-Leibler divergence which measures the dissimilarity between two probability distributions. Through the numerical experiments, we demonstrate the usefulness of the proposed framework.