User Behavior Analytics with Machine Learning for Household Electricity Demand Forecasting

Abstract

This study investigates how smart sensor data that infers user behavior can be applied to machine learning (ML) methods for accurate short-term load forecasting (STLF) in residential buildings. We first collected the appliances energy prediction (AEP) dataset-one of the most popular datasets for use in building-level STLF-and configured four input variables, considering only external factors, external factors and sensor data before one day, external factors and sensor data before one week, and external factors and sensor data before both one day and one week. We then constructed five STLF models for each input variable based on decision tree-based ensemble learning methods. We finally compared the prediction performance of the input variables concerning mean absolute percentage error (MAPE), root mean square error (RMSE), and mean absolute error (MAE). Experimental results showed that the models trained with only external factors outperformed those trained with other input variables on STLF in the AEP dataset. We conclude that when researchers use the AEP dataset to demonstrate the superiority of their proposed model, only external factors should be considered input variables.