Comparison of Kolmogorov–Arnold Networks and Multi-Layer Perceptron for modelling and optimisation analysis of energy systems

Abstract

Considering the improved interpretable performance of Kolmogorov–Arnold Networks (KAN) algorithm compared to multi-layer perceptron (MLP) algorithm, a fundamental research question arises on how modifying the loss function of KAN affects its modelling performance for energy systems, particularly industrial-scale thermal power plants. In this regard, first, we modify the loss function of both KAN and MLP algorithms and embed Pearson Correlation Coefficient (PCC). Second, the algorithmic configurations built on PCC, i.e., KAN_PCC and MLP_PCC as well as original architecture of KAN and MLP are deployed for modelling and optimisation analyses for two case studies of energy systems: (i) energy efficiency cooling and energy efficiency heating of buildings, and (ii) power generation operation of 660 MW capacity thermal power plant. The analysis reveals superior modelling performance of KAN and KAN_PCC algorithms than those of MLP and MLP_PCC for the two case studies. KAN models are embedded in the optimisation framework of nonlinear programming and feasible optimal solutions are estimated, maximising thermal efficiency up to 42.17 ± 0.88 % and minimising turbine heat rate to 7487 ± 129 kJ/kWh corresponding to power generation of 500 ± 14 MW for the thermal power plant. It is anticipated that the scientific, research and industrial community may benefit from the fundamental insights presented in this paper for the ML algorithm selection and carrying out model-based optimisation analysis for the performance enhancement of energy systems.