Physics-Informed Temporal Graph Convolutional Network Model for Multiphase Flow Process Characteristics in Pump Valves

Accurately describing complex flow characteristics of the whole discharge process in check valves is desirable for the structure design of multiphase pumps. A physics-informed temporal graph convolutional network (PITGCN) multistage model is proposed to predict the opening heights and flow rates of check valves with limited samples. First, mechanism models for the lag, mutation and sine stages of the whole discharge process are established for different stage characteristics, respectively. Additionally, to track the opening and boundary points of the discharge process, the physics laws of the lag and mutation stages are embedded into the loss function of TGCN for the recognition and predication of three stages. Subsequently, the mechanism models of the lag and sine stages are used to correct their prediction results. Experimental results demonstrate the proposed multistage model can capture the complex dynamic features by the incorporation of physical knowledge and limited process data.