Forecasting short-term residential electricity consumption using a deep fusion model

Electricity consumption forecasting is practically significant for either detecting abnormal power usage pattern or resource-conserving purpose. Indeed, it is a non-trivial task since electricity consumption is related to multiple complex factors, including historical amount of consumption, calendar dates and holidays, as well as residential power consumption habits. To this end, we propose an end-to-end structure to collectively forecast short-term power consumption of private households, called RCFNet (Residual Conventional Fusion Network). Specifically, our RCFNet uses (1) three branches of residual convolutional units to model the temporal proximity, periodicity and tendency properties of electricity consumption, (2) one fully connected neural network to model the weekday or weekend property, and (3) a residual convolution network to fuse the above output to produce short-term prediction. All the convolutions used here are one-dimensional. Through experimental studies on residential electricity consumption dataset in Australia, it is validated that the proposed RCFNet outperforms several well-known methods. Besides, we demonstrate that residential power consumption is closely related to the living characteristics of residents.