Smart power consumption forecast model with optimized weighted average ensemble

Abstract

Smart power forecasting enables energy conservation and resource planning. Power estimation through previous utility bills is being replaced with machine intelligence. In this paper, a neural network architecture for demand side power consumption forecasting, called SGtechNet, is proposed. The forecast model applies ConvLSTM-encoder-decoder algorithm designed to enhance the quality of spatial encodings in the input feature to make a 7-day forecast. A weighted average ensemble approach was used, where multiple models were trained but only allow each model’s contribution to the prediction to be weighted proportionally to their level of trust and estimated performance. This model is most suitable for low-powered devices with low processing and storage capabilities like smartphones, tablets and iPads. The power consumption comparison between a manually operated home and a smart home was investigated and the model’s performance was tested on a time-domain household power consumption dataset and further validated using a real time load profile collated from the School of Renewable Energy and Smart Grid Technology, Naresuan University Smart Office. An improved root mean square error (RMSE) of 358 kwh was achieved when validated with holdout validation data from the automated office. Overall performance error, forecast and computational time showed a significant improvement over published research efforts identified in a literature review.