Real-time thermal evaluation on three-phase enclosure gas insulated bus by PSO-PINN approach

Real-time thermal evaluation including temperature field calculation and internal heat source inversion is crucial for power equipment ampacity and safety. This study presents a particle swarm optimized physics-informed neural network (PSO-PINN) for thermal evaluation on three-phase enclosure gas insulated bus (GIB) unit. Computational fluid dynamics (CFD) simulations are conducted to generate two-dimensional sampling points for model training and testing. Fluid flow velocity and temperature at the sampling points are predicted by fully connected neural network (FCNN) with spatial coordinates as inputs, in which PSO strategy is utilized to determine the optimal width and depth of network architecture. Navier-Stokes (N-S) equations are incorporated into loss functions via differential operations of network outputs; besides multi-loss function balancing strategy is adopted to balance weights of different loss terms. The trained PSO-PINN model predicts the temperature field within 0.11 seconds, achieving a maximum relative error of less than 2.7% when compared against physical experiments and reference CFD simulations. The proposed model can also invert the internal heat source with an error of 2.83% using tank surface observation data contaminated with 10% noise. With natural incorporation of field-observed data, instantaneous inference time and measurement noise resistance, this work proves to be a promising tool for real-time digital twin (DT) thermal management of GIB units. Objective Real-time temperature field computation and internal heat-source inversion for three-phase gas-insulated bus (GIB). Method A particle-swarm-optimized physics-informed neural network (PSO-PINN) is trained with CFD-generated data; Navier–Stokes and energy equations are embedded as residual loss terms. Network depth/width are automatically optimized by PSO (8 layers × 72 neurons best). Conclusion The PSO-PINN surrogate provides an accurate, instantaneous and noise-robust thermal–fluid digital twin for in-service GIB ampacity management and overheating early warning. • Innovative PSO-PINN framework optimizes neural architecture for real-time thermal evaluation of gas-insulated buses. • A ccurate forward thermal modeling with ≤2.7% error validated by CFD and experiments, enabling 0.11 s predictions. • Robust inverse heating-source identification achieves 2.83% error under 10% noise, critical for fault detection. • Computational efficiency surpasses traditional CFD by 600, ideal for digital twin applications. • Field-deployable solution integrates physical laws and noisy measurements for reliable GIB health management.