Symbolic regression and multi-parameter optimization for modelling and generalizing chiller performance curves

Cooling systems account for a significant share of global energy consumption and carbon emissions, presenting a key opportunity for energy efficiency improvements through performance optimization. Despite efforts to enhance the Coefficient of Performance (COP), many systems continue to operate inefficiently due to outdated technologies and suboptimal control strategies. This paper proposes a method to model and generalize chiller performance curves by establishing mathematical relationships between COP and key operational parameters (LR, TcdIn and TcwOut), aiming to optimize overall system performance. Initially, Fourier Transform filtering was applied to improve the signal-to-noise ratio. A complexity-augmented symbolic regression approach was then employed to derive mathematical formulas that described the underlying performance relationships, resulting in a mechanistic model. The model was parameterized and optimized using a gradient descent algorithm, yielding a universal model. Model performance was evaluated by comparing Mean Squared Errors (MSE) between train and test sets, and k-fold cross-validation was implemented to assess the extrapolation capabilities and robustness of the model. This study presents a promising approach for modelling COP curve and highlights the potential of symbolic regression in cooling systems research.