PI-MBKAN: Physics-Informed Multi Branch Kolmogorov–Arnold Network for high-precision Chiller Power Prediction

Accurate chiller power prediction is critical for building energy optimization and operational cost control. This paper defines the Chiller Power Prediction (CPP) problem, aiming to minimize the error between predicted and actual power. Existing methods suffer from limited nonlinear fitting efficiency and poor extrapolation under unknown operating conditions. To tackle these issues, this study reveals a key insight: inherent physical decoupling exists between the evaporator and condenser. Mixed modeling of these two subsystems causes feature interference and degrades prediction accuracy. Then, a Physics-Informed Multi-Branch Kolmogorov–Arnold Network(PI-MBKAN) is proposed for CPP problem. The core design of PI-MBKAN aligns with chiller physical characteristics: (1) A dual-branch modular architecture is built based on the decoupling of chilled and condenser water sides, with each branch using a Kolmogorov–Arnold Network(KAN) to independently extract nonlinear features; (2) A weighted fusion mechanism dynamically adjusts branch weights to adapt to varying operating conditions; (3) Constraints from the first and second laws of thermodynamics are embedded into a hybrid loss function, ensuring the model adheres to physical laws while learning statistical patterns. Experiments on a real-world dataset and an open-source simulation dataset validate PI-MBKAN’s superiority: In interpolation tests, MBKAN outperforms KAN, MLP, etc, reducing the Root Mean Square Error(RMSE) by up to 16.4%/32.6% vs. the suboptimal KAN on the two datasets, respectively. In extrapolation tests with unknown data distribution, PI-MBKAN outperforms data-driven models and existing physics-informed neural network.