Ray-level dataset for deep-learning direction correction in Concentrating Solar Power (CSP) plants.

This dataset can be considered a necessary tool for the generalization of the training process suggested in https://doi.org/10.5281/zenodo.13941464. Whereas the former dataset was thought to graphically train a neural network for flux density (and hence, power) predictions in flat surfaces, this one can be used to correct the individual simulated ray's directions in flat aperture planes/surfaces. In this case, the correction is numerical and could be based on random forests or deep learning methods, like Feed-Forward Neural Networks, Recurrent Neural Networks, Convolutional Neural Networks or Bayesian Neural Networks, among others. The user can combine both corrections in order to individually modify ray's directions and powers and later project the rays against any geometric receiver surface.Like in the previous case, an Excel file with the 931 tested meteorological conditions is included. The positions of the power plant and the receiver are included. A front view of the Solarturm Juelich (STJ) is also provided in order to identify the aperture plane where the rays are intercepted (2nd level of MFT at 35 metres high). The 931 meteorological conditions are tested with 8 different combinations of focused heliostats also detailed in the dataset complementary files. With all of these cases, a set of 25059 couples of ".txt" files is generated. Each of them contains a column array that define the simulated intercepted rays' directions for two cases: all the simulated heliostats are focused without tracking errors (training input) and the tracking errors are modelled within the simulation environment for all of them (training groundtruth). This dataset is part of the WP1 of TOPCSP european project (HORIZON MSCA Doctoral Network, Project number 101072537).