This article introduces two methods designed for wide-range parametric modeling of microwave components with neural networks. These innovative techniques address high nonlinearity and low training efficiency challenges. A parallel decomposition method decomposes the parameter domain into two subdomains for modeling, training and correction, and integrates the submodels back into one refined model. A sensitivity-driven stepwise method uses the Pearson correlation coefficient technique to classify geometric parameters and incorporates transfer learning (TL) for step-by-step training to construct the final model. The methods are applied to two stripline structures to demonstrate the significant improvements in testing error, training error and training time.