Extrapolation model for chillers based on physics-constrained broad learning

In order to realize energy-saving control optimization of the chillers, an extrapolation model for chillers based on physics-constrained broad learning is developed by introducing error backpropagation, Adam optimizer, and custom physics-constrained loss function on the basis of broad learning, which is helpful to accurately obtain the operating state of the chiller under unknown working conditions according to the measured data of the existing working conditions. Comparisons with multilayer perceptron, random forest, convolutional neural network, support vector regression show that, compared with other non-extrapolation models, this extrapolation model reduces MAE by about 22.35% and RMSE by about 25.36%, increases R2by about 19.22%, and reduces energy loss by about 99.45%, and shorts the training time, making it a method that balances model structure complexity, accuracy and training time, suitable for extrapolation scenario modeling of chillers.