Learning Disentangled Representation of Residential Power Demand Peak via Convolutional-Recurrent Triplet Network

Abstract

In the time-series models for predicting residential energy consumption, the energy properties collected through multiple sensors usually include irregular and seasonal factors. The irregular pattern resulting from them is called peak demand, which is a major cause of performance degradation. In order to enhance the performance, we propose a convolutional-recurrent triplet network to learn and detect the demand peaks. The proposed model generates the latent space for demand peaks from data, which is transferred into convolutional neural network-long short-term memory (CNN-LSTM) to finally predict the future power demand. Experiments with the dataset of UCI household power consumption composed of a total of 2,075,259 time-series data show that the proposed model reduces the error by 23.63% and outperforms the state-of-the-art deep learning models including the CNN-LSTM. Especially, the proposed model improves the prediction performance by modeling the distribution of demand peaks in Euclidean space.