The model order reduction and data assimilation techniques have been extensively applied to construct surrogate models as core models for digital twins in different industries. However, in most applications, the online phases rely on a single offline basis construction within the parameter domain, which is computationally expensive and often overlooks local parameter variations due to anisotropic behavior. Consequently, this poses challenges for real-time monitoring where parameters vary dynamically. To address this, we introduce a relative deviation score of reduced basis coefficients, represented by $\mathcal{W},$ to assess the reconstruction capability of basis functions under current parameters. Additionally, we provide a quantitative threshold $\overline{\mathcal{W}}$ for $\mathcal{W}.$ Using the reconstruction of a two-dimensional reactor core field distribution as a case study, we demonstrate that if $\mathcal{W}$ remains below the threshold $\overline{\mathcal{W}}=1,$ the average reconstruction error can be maintained within the same order of magnitude as the fundamental error limit $ε (κ ≤ 10).$ Furthermore, we compare $\mathcal{W}$ with the traditional one-time reduced basis selection method and the metric $C_{R2F}$ [F. Bai and Y. Wang, SN Applied Sciences, 2 (2020), p. 2165], confirming the lower computational cost and enhanced robustness of our proposed algorithm.