The characterization of the nominal power under real operating conditions is crucial for the correct evaluation of a photovoltaic generator. Several earlier studies proposed different methods based on empirical models for the estimation of the nominal power. These methods require experimental data obtained during optimal days under clear sky conditions and are not suitable for days deviating from these optimal conditions and, thus, generating a significant amount of noise in the data. In this sense, we propose a non-parametric statistical approach to filter out this noise to reliably estimate the real nominal power in the latter conditions. The period of study was 107 days. These were divided in two categories, clear and partly cloudy sky conditions. The results show that our statistical method allows to obtain the same nominal power under partly cloudy conditions as under clear sky. This was possible by applying a kernel density estimation to filter the outliers and the noise under cloudy conditions.