K-Nearest patterns for electrical demand forecasting in residential and small commercial buildings

Abstract

This work presents a case study of Big Data and Machine Learning whose objective is to improve energy Demand Response (DR) programs by providing accurate energy demand forecasts. Given the present state of the art, this research work introduces the proposed methodology for Time Series Forecasting based on two variants of the K-neighbours method (KNN): K-Nearest Features in Time Series (KNFTS) and K-Nearest Patterns in Time Series (KNPTS) algorithms. These algorithms are valuable in this field since only a historical data set consisting the time and energy consumption variables are used to find similar patterns of electricity consumption and then make future forecasts. Furthermore, the proposal to use elastic similarity measures such as DTW and EDR shows to have advantages over the use of common error metrics. It has been proven on data from 122 houses and small commercial buildings located on the island of Lanzarote that the KNPTS achieves minor errors in 89% of cases. Therefore, it shows that the KNPTS algorithm provides a good accuracy, more efficient in prediction than the KNFTS algorithm, to improve DR programs.