Adjusted Feature-Aware k-Nearest Neighbors: Utilizing Local Permutation-Based Error for Short-Term Residential Building Load Forecasting


Household load profiles are more fluctuating than higher aggregated load profiles and relative forecast errors are comparatively high. To handle this, adjusted error metric and average concepts have been proposed to be used to obtain more suitable forecasting algorithms. These algorithms have so far only been compared for day-ahead forecasts. They are further not considering external features such as numerical weather data or calendar-based information. We present an extension of an algorithm based on k-nearest neighbors that is capable of incorporating such external features, the Adjusted Feature-Aware k-Nearest Neighbors (AFKNN). We show on 220 households of the Pecan Street dataset that forecast accuracy can be improved for buildings with electrical heating and cooling as well as for intra-day forecasting, at the cost of higher modeling complexity.

In 2018 IEEE International Conference on Communications, Control, and Computing Technologies for Smart Grids (SmartGridComm)
Marcus Voss
Marcus Voss
Intelligence Architect and AI Expert

AI and Sustainability at Birds on Mars and Climate Change AI, Lecturer for AI and Data Science.