There has been increasing recognition of the importance of utilizing phasor measurement units (PMUs) for power plant model validation (PPMV). In North America, the PMU-based PPMV is typically performed manually by utilities and independent system operators (ISOs) for diagnosing and calibrating power plant model problems. In order to guarantee the accuracy of the calibration results, engineers need to provide manual judgment to the mismatched PPMV cases, so that the type of the modeling problem (such as wrong machine parameters, missing governor models, etc.) can be determined before the model is sent for a detailed calibration. This manual judgment process has become the major bottleneck for automating the PMU-based PPMV. To overcome this difficulty, this paper proposes a feature-based diagnosis framework to determining the types of the power plant model problems. It mimics the human engineering judgment process via a supervised learning approach. Instead of applying purely curve fitting or sensitivity analysis, this approach uses the engineering experience extracted from the labeled historical PPMV cases, establish the critical feature space, and then perform artificial learning to determine the type of a power plant model problem. The proposed framework could serve as a screening tool for the PPMV engineering judgment process, which could potentially help automate the entire PMU-based PPMV applications.