The use of full model representation in power system studies may lead to undesirable levels of computational burden and inaccuracy due to modern system complexities and uncertainties. To address the tractability issue, network reduction methods aim to create a simplified model, with reduced dimension, of a given power system. Current techniques consider only one operating point in their reduction process, falling short in properly performing for a wide range of operating conditions. Additionally, a nonlinear AC power flow solution features worse computation performance, but better accuracy when compared against its linearized counterpart (DC power flow solution). Unfortunately, the DC power flow approximation disregards the line losses and nonlinear effects due to changes in voltage levels and reactive power. In this context, we propose a novel optimization–based framework to create equivalent power flow models. Thus, to overcome the computational performance limitations and imprecision for multiple operating scenarios, we use the proposed framework to produce a DC–based network reduction method that performs well in many operating points. The solution of a linear optimization problem, which considers multiple AC power flow scenarios or network measurements, determines the equivalent network parameters. To ensure modeling accuracy, we consider a set of artificial dynamic loads to represent the mismatch between observed scenarios and the response of the equivalent. These artificial loads are polynomial functions of the operating point, and their coefficients are co-optimized with the reduced network parameters. Principal Component Analysis (PCA) is used to extract the relevant components of the load vector defining the operating point, reducing the equivalent model dimensionality, and improving out–of–sample performance. We test the methodology against traditional Ward equivalent for different operating conditions. We present case studies with generated data to investigate the model generalization capability for different noise levels. Finally, we conduct a case study based on realistic load profiles from a Brazilian distribution company within the IEEE 118–Bus test system.