One model fits all: Individualized household energy demand forecasting with a single deep learning model

Energy demand forecasting at the household level is an important issue in smart energy grids to facilitate applications such as residential Demand Response (DR). However, if a separate machine learning model is trained for each house, the erratic nature of some consumers will lead to significant inaccuracies for the respective models, while predictions for new households with scarce data will not be possible to generate. In this work, we propose an approach with a single deep learning model that is trained on multiple households, which can create hourly energy consumption forecasts for individual households. We present a novel architecture that combines a Recurrent Neural Network (RNN) encoder and a Multilayer Perceptron (MLP). Our approach captures both the impact of past consumption time-series and that of energy profiles on future energy demand. Our model incorporates energy profiles to derive different characteristics between consumers, and it features a "double" clustering procedure that is specially designed for a mixture of time-series and non-time-series data. Experiments with real smart meter data show that the proposed neural network architecture achieves high performance in predicting energy consumption both for known and new consumers not present in the training dataset, with a Mean Absolute Percentage Error (MAPE) of 10.1% and 12.5% respectively.