Day-ahead demand response potential prediction in residential buildings with HITSKAN: A fusion of Kolmogorov-Arnold networks and N-HiTS
Ali Muqtadir, Bin Li, Ying Zhou, Chen Songsong, Sadia Nishat Kazmi
Abstract
Accurate forecasting of Demand Response (DR) is vital for optimizing resource allocation in power systems, especially in markets where Load Aggregators (LAs) bid based on predicted DR potential. Traditional models struggle to capture the nonlinear dependencies of consumer behavior and the temporal patterns in energy consumption. This study aims to overcome these limitations by introducing HITSKAN, a hybrid approach which is a fusion of Kolmogorov-Arnold Networks (KANs) and Neural Hierarchical Interpolation (N-HiTS) to improve day-ahead DR potential forecasting. HITSKAN is able to solve the challenges faced by LAs by integrating the ability of KANs to model complex multivariate functions for nonlinearity together with the strength of N-HiTS in handling temporal dependencies. The methodology employs real-world residential load data from 114 apartments to capture historical demand response potential through thermal response modeling, which does not require appliance-level data and then applies the HITSKAN forecasting model to predict day-ahead DR potential. The performance of model is evaluated on all key metrics which include Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), Root Mean Square Error (RMSE) and systematic Mean Absolute Percentage Error (sMAPE) along with variance, standard deviation and computation time. Results demonstrate that HITSKAN outperforms state-of-the-art forecasting models in both winter and summer seasons. By incorporating KANs into a time series forecasting framework, HITSKAN offers a scalable and effective solution for DR potential forecasting, significantly enhancing grid management and resource optimization in residential settings.