Physics-informed neural networks (PINNs) have recently emerged and attracted extensive attention as an alternative approach to computational fluid dynamics (CFD) methods, which can provide competitive solutions to a variety of forward and inverse fluid problems. In this study, we reconstruct a 3D wind field around a building model in wind tunnel test with a Reynolds number of 2.4 × 104 by formulating a novel PINN framework, which is the first exploration of PINNs for building wind engineering problems. To surmount the hurdle in multi-objective optimization for PINN training, a dynamic prioritization (dp) self-adaptive loss balance strategy is proposed (termed dpPINN), which adaptively reconciles the loss terms of distinct scales to facilitate convergence in PINN training. A zero-equation turbulence model and the wind velocity data collected in near-wall regions are embedded in dpPINN training. Comparison results indicate that dpPINN predictions show good consistency with observation data, which is superior to two current PINN paradigms in prediction accuracy. Furthermore, the influence of neural network configurations, turbulence models, and the layout arrangements of training points on the dpPINN prediction is investigated. It is demonstrated that the dpPINN could be a powerful auxiliary means for airflow simulation and reconstruction in wind engineering applications.