A study of the applicability of the Kolmogorov–Arnold network architecture for time series forecasting

The recently proposed Kolmogorov–Arnold Network (KAN) architecture emerges as a promising alternative to traditional neural networks based on the Multilayer Perceptron (MLP). By leveraging the Kolmogorov–Arnold representation theorem, KAN represents multidimensional functions as combinations of univariate functions, thereby offering potentially higher accuracy and model interpretability through its inherently simpler structure. This paper investigates the applicability of KAN to time series forecasting using the well-known hourly electricity consumption dataset as a benchmark. Meteorological observation data are selected as an additional testbed. A comparative analysis is conducted between KAN networks and traditional MLPs, as well as implementations of recurrent architectures based on KAN (TKAN variants) versus established designs such as Long Short-Term Memory (LSTM) and Gated Recurrent Units (GRU). Experimental results demonstrate the superiority of the KAN architecture over MLPs in temporal prediction tasks. The proposed recurrent architecture, TKAN1, achieves the highest coefficient of determination (R2 = 0.3483) among TKAN variants, with a Root Mean Squared Error (RMSE) of 0.1010 in energy demand forecasting.