Kolmogorov–Arnold Network And Multi-Layer Perceptron for Modelling and Optimisation of Energy Systems

Abstract

Kolmogorov–Arnold Network (KAN) algorithm is recently introduced in the ML space with the improved interpretable performance compared to multi-layer perceptron (MLP). It remains an open research question how integrating the prior data-information into the KAN algorithm design may affect its modelling performance. With this research question, we modify the loss function of both KAN and MLP to introduce Pearson correlation coefficient (PCC) in the loss function of the two algorithms. The new algorithmic configuration, i.e., KAN_PCC and MLP_PCC as well as original configuration of KAN and MLP are deployed for modelling and optimisation analyses. Two case studies from energy systems like energy efficiency cooling (ENC) and energy efficiency heating (ENH) of building as well as power generation operation of 660 MW capacity thermal power plant are taken. The analysis reveals that KAN based algorithmic configuration provides improved modelling performance for case studies of buildings and thermal power plant. KAN based algorithmic configurations are also embedded in the nonlinear programming framework for multi-objective optimisation analysis. Feasible optimal solutions are estimated maximizing thermal efficiency and minimizing turbine heat rate of thermal power plant under three cases of power generation mode. Maximum thermal efficiency of 42.17 % and minimum turbine heat rate of 7487 kJ/kWh is observed corresponding to 500 MW for KAN based multi-objective optimisation analysis. KAN appears to show improved modelling performance for considered case studies that can be applied for improved modelling and optimisation analysis of energy systems.

Keywords: KAN, MLP, energy systems, thermal power plants, energy sustainability