In times when computational resources - such as data, code, software tools, libraries, etc. - play a fundamental role in the development of scientific works, it has become evident that transparency regarding all the computational arsenal involved in such type of work is essential for its validation. This concern is the basis of the culture of reproducibility, which aims to add to a work the possibility of it being reproduced by an unknown person or by the author herself/himself in the future. Reproducibility can bring other benefits such as enabling the reuse and continuity of a work, which is associated with other terms such as replicability. This thesis is based on a workflow developed for solar irradiance prediction, and focuses on replicating it and adopting mechanisms to make the new workflow reproducible, as well as better exploiting recurrent neural networks for the prediction task. The prediction of the total solar irradiance at the top of the atmosphere would contribute, for example, in studies of solar variability, or could bring improvements to atmospheric and climate models on Earth; however, it is a service still not much explored by the scientific community in the area of space weather. The new version of the workflow was developed attempting to use free computational resources, such as the Python language and Linux operating system, and performs the prediction task using different recurrent neural network architectures from the Keras library. The work confirms the effec tiveness of recurrent networks in predicting total solar irradiance and for one of the emission lines tested: lyman-α; and suggests that the prediction of other lines of the spectrum need additional parameters to obtain better accuracy. This document reports the replication process, presents the irradiance prediction results, and lists the computational resources employed to try to make the new workflow reproducible.