Distributed Physics-Informed Neural Networks for Temperature Field Reconstruction in the Regenerator of Hot Blast Stove Considering Parametric Spatial-Temporal Dynamics

Hot blast stove as an important heating equipment in the iron and steel production process, reconstructing its regenerator temperature field is of great significance to provide guidance of structural parameter design. In this paper, a distributed physics-informed neural network (DPINN) method considering spatial-temporal dynamics of physical parameters is proposed to reconstruct the two-dimensional transient temperature field of the regenerator with small amount of measurement data. Specifically, the regenerator is partitioned into several non-overlapping subsystems according to spatially-varied physical parameters, in which physical parameters in each local subsystem are assumed to be consistent. Then, PINN is applied to reconstruct local temporally-varied temperature fields. The global consensus strategy is designed for exchanging boundary information with neighbouring subsystems, in order to ensure the global temperature field satisfying model centralized constraints. The effectiveness of the proposed method is verified by comparing with the temperature field from ANSYS FLUENT simulation model, the results demonstrate the proposed DPINN improves the modeling accuracy compared with PINN.