Given that the cooling capacity of the absorption refrigeration system depends on both the heat supply and the operating conditions, a data-and-knowledge-driven Wavelet Neural Network (WNN) modeling approach is proposed based on the analysis of its working principle and the coupling relationship between variables. This approach utilizes the Simple Particle Swarm Optimization (SPSO) algorithm instead of the gradient descent method to search the network parameters of WNN, which makes up for the defects of the gradient descent method that is easy to fall into the local optimum. Furthermore, the inequality constraints of energy and temperature are integrated into the loss function of model training as prior knowledge, and the model’s parameters are trained using experimental data, which reduces the training burden of the model and the model’s dependence on the training data, and increases the network’s interpretability and the model’s robustness. The experimental results show that the prediction accuracy and generalization of the WNN model are improved after replacing the optimization algorithm and incorporating the physical knowledge under the same working conditions, and its RMSE and MAE are reduced by 2.08% and 1.63%, respectively, compared with the original WNN model.
Read full abstract