Optical remote sensors are extremely susceptible to clouds. Clouds and their shadows affect remote sensing image processing methods to automatically identify and classify land use and land cover types. To detect cloud and cloud shadows in remote sensing images, many algorithms have been proposed, such as FMask, Sen2Cor and s2cloudless. In image time series analysis, interpolation techniques are used to produce valid values when pixels are covered by clouds or shadows. This paper evaluates the use of deep learning approaches to interpolate cloudy pixels in Sentinel-2 time series. Twelve different model configurations were evaluated and their differences and limitations were highlighted. The model proved to be very promising in dealing with the limitations of the cloud mask interpolating clouds and cloud shadows.