Accurate and efficient calculation of a transformer's magnetic field is fundamental for the rapid calculation of its losses, temperature rise, and structural forces. However, existing numerical methods for calculating the harmonic magnetic field of a product-level transformer are time-consuming and fail to meet the rapid requirements of digital operations and maintenance. To address this, this paper first utilises the harmonic field method to obtain the snapshot matrix of the transformer's magnetic field. Subsequently, a response surface model of the magnetic field is constructed using intrinsic quadrature theory and radial basis functions in the augmented form. To enhance the efficiency of constructing the reduced-order model, an adaptive Latin hypercube sampling method, integrating the additive rule and leave-one-out cross-validation, is introduced, significantly improving the efficiency of sample space construction. The effectiveness of the proposed method is validated by applying the proper orthogonal decomposition-radial basis function including linear polynomial (POD-RBFLP) method to calculate the harmonic magnetic field of a three-phase power transformer in reduced order. The results are compared with those from COMSOL calculations, showing that the reduced-order model maintains the calculation error within a reasonable range, thereby confirming the accuracy of the proposed method. Additionally, the reduced-order model demonstrates a significant advantage in computation time compared to COMSOL simulations, enabling the calculation of the transformer's magnetic field in seconds.