Energy performance prediction of the centrifugal pumps by using a hybrid neural network

• Published in: Energy (Oxford) 2020-12, Vol.213, p.119005, Article 119005
• Main Authors: Huang, Renfang, Zhang, Zhen, Zhang, Wei, Mou, Jiegang, Zhou, Peijian, Wang, Yiwei
• Format: Article
• Language: English
• Subjects: Back propagation networks; Centrifugal pump; Centrifugal pumps; Efficiency; Electric power; Electric power systems; Energy; Energy performance; Geometric accuracy; Loss model; Mathematical models; Model accuracy; Neural networks; Performance prediction; Physics-informed neural network; Pumps
• ISSN: 0360-5442; 1873-6785
• DOI: 10.1016/j.energy.2020.119005

It is of great significance to rapidly and accurately predict the energy performance of centrifugal pumps for the macro-control of the entire electric power system. However, some challenges are encountered, for example, the numerical simulation requires huge computing resources and calculating time, the theoretical loss model needs to improve the prediction accuracy, etc. Based on the multiple geometrical parameters and operation conditions, a hybrid neural network is proposed to predict the energy performance (i.e. the head, power and efficiency) of centrifugal pumps, where the theoretical loss model is incorporated into the back propagation neural network and then the neural network structure is optimized by automatically determining the node number of hidden layers. When compared with the experiments, the energy performance is well predicted by using the hybrid neural network with the mean-square-error (MSE) for the head, power and efficiency of 0.0062, 8.4E-4, 0.020, respectively. Besides, by considering the theoretical loss model, the hybrid neural network demonstrates a dramatic decrease in the head MSE and the efficiency MSE when compared with the original neural network. Furthermore, the hybrid neural network performs much better than the traditional linear regression in a wide flow-rate range for multiple centrifugal pumps. •A hybrid neural network is proposed to predict energy performance of pumps.•The hybrid NN is incorporated the theoretical loss model into the classical NN.•The neural network structure is automatically optimized.•Energy performance is predicted well by the hybrid NN.