A Novel Short-Term Residential Load Forecasting Methodology using Two-Stage Stacked LSTM and Hampel Filter

Abstract

Various deep learning-based algorithms have been developed recently for accurate residential load forecasts. The deep learning models in the literature are trained using a few years of data to learn patterns over all the seasons. The authors of some of the work have appended appliance level data to train for the outliers in the load data caused by the non-periodic usage of medium-duty appliances. These methodologies are data-intensive and challenging to implement for a large number of customers. Hence, this work proposes a novel technique of disaggregating the residential load data set into a base/general load pattern and a non-periodic outlier load pattern using a Hampel filter and training two stacked long short term memory (LSTM) models. The proposed methodology also uses less data to train the LSTM model. The model achieved comparable results just by training over four months of residential load data. The model is trained using actual residential load data acquired using a WiFi energy monitor installed at one of the households in Fairbanks, Alaska.