The characteristic curve of a photovoltaic array affected by partial shadowing presents more than one maximum power points, having the rightmost one a shape more squared than a curve under uniform irradiance. Different machine learning tools take advance of this change of curvature to distinguish between the uniform case and the mismatched case using as input a small percentage of samples around this maximum power point, in such a way a full scan is avoided, reducing energy losses. In addition, the influence of the amount of points used as input on the classification accuracy is studied. Moreover, aimed to apply this approach to the real-world scenarios, an analysis in terms of noise is also presented. Finally, the proposed method has been implemented in an embedded low-cost board that can be integrated in a grid inverter, providing execution times for each approach.