Power system operators evaluate the frequency security of the system by predicting the frequency nadir, which is assumed to indicate the impact of a sudden loss of a generating resource. Recently, frequency nadir prediction has become more challenging because renewables have penetrated and significantly changed the generation portfolio within the system. Conventionally, the frequency nadir is determined using a frequency response model where the features-load damping, system inertia, and effective governor response-are assumed to be known. However, these key features are not easily obtained in a power system that continuously changes during daily operation. This study proposes a supervised learning scheme that traces these key features. It also proposes a new feature-the power gap rate-that better reflects the influence of the load on the system frequency than that of the load damping. Feature importance recognition and the construction of a frequency nadir model (FNM) are realized using the proposed supervised learning scheme. The proposed FNM achieved 54% higher accuracy than the conventional method. Finally, the FNM is implemented in a planning process that quantifies the capacity of the fast responsive reserve (FRR). In two renewable penetration cases, the proposed FRR procurement successfully secured the frequency nadir above the security criterion.