Neural networks applied to forecasting the astroclimatic and meteorological parameters in an operational context
Abstract
Forecasting optical turbulence and atmospheric parameters is crucial for science operations of present and new generation ground-based telescopes (ELT) as well as for free-space optical communication (FSOC). These forecasts are also useful to optimize the AO use, particularly in the ELT framework too. Recent research has demonstrated that the AR technique currently provides the best performances in predicting atmospheric and astroclimatic parameters on a time scale of 1 hour (e.g., RMSE 0.1" for the seeing), however deep learning offers some intriguing features that include the high adaptability, the correlation that can be found between different parameters treated simultaneously and different architectures can explore the capability for possible improvements. For this reason, it is important to investigate on the possibility to improve forecast performances beyond the score achieved by AR or simply to apply deep learning in those context in which it is not possible to apply the AR. In this contribution we investigate a set of algorithms belonging to the family of the neural networks that in principle are more suitable than the Random Forest (RF) model for this kind of application. We address specific questions related to the forecast performances in operational configurations. Besides that, we also investigate the impacts on gaps on the trained data set due to the lack of measurements of the optical turbulence during the day time. We show evidences that with measurements covering the whole 24-hours forecast performances might improve. Our test has been done on the wind speed in this contribution. This work is done in the framework of the operational forecast system for LBT (ALTA Center) and for the VLT (FATE project) that are leaded by our team.
Related
- Characterisation of the NewAthena WFI's DEPFET Flight Production's Operational Parameters
- Influence of local and global solar magnetic fields on medium-term forecasting of the number and parameters of coronal mass ejections
- Forecasting Coupled Dark Energy Parameters with the One-Loop Galaxy Power Spectrum
- Stochastic analysis of finite-temperature effects on cosmological parameters by artificial neural networks
- CCC Aurora Threshold Program: Realizability-Gated Residual Coherence Projection-Compressed Observability and Operational Closure
BibTeX
@misc{weinberger2026neural,
title = {{Neural networks applied to forecasting the astroclimatic and meteorological parameters in an operational context}},
author = {C. Weinberger and E. Masciadri and M. De Sepibus and A. Otarola and S. Miesske and C. Veillet and S. Ragland},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.19070},
eprint = {2610.19070},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.19070},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Neural networks applied to forecasting the astroclimatic and meteorological parameters in an operational context
AU - C. Weinberger
AU - E. Masciadri
AU - M. De Sepibus
AU - A. Otarola
AU - S. Miesske
AU - C. Veillet
AU - S. Ragland
PY - 2026
DA - 2026/10/13
PB - arXiv
DO - 10.48550/arxiv.2610.19070
UR - https://arxiv.org/abs/2610.19070
AB - Forecasting optical turbulence and atmospheric parameters is crucial for science operations of present and new generation ground-based telescopes (ELT) as well as for free-space optical communication (FSOC). These forecasts are also useful to optimize the AO use, particularly in the ELT framework too. Recent research has demonstrated that the AR technique currently provides the best performances in predicting atmospheric and astroclimatic parameters on a time scale of 1 hour (e.g., RMSE 0.1" for the seeing), however deep learning offers some intriguing features that include the high adaptability, the correlation that can be found between different parameters treated simultaneously and different architectures can explore the capability for possible improvements. For this reason, it is important to investigate on the possibility to improve forecast performances beyond the score achieved by AR or simply to apply deep learning in those context in which it is not possible to apply the AR. In this contribution we investigate a set of algorithms belonging to the family of the neural networks that in principle are more suitable than the Random Forest (RF) model for this kind of application. We address specific questions related to the forecast performances in operational configurations. Besides that, we also investigate the impacts on gaps on the trained data set due to the lack of measurements of the optical turbulence during the day time. We show evidences that with measurements covering the whole 24-hours forecast performances might improve. Our test has been done on the wind speed in this contribution. This work is done in the framework of the operational forecast system for LBT (ALTA Center) and for the VLT (FATE project) that are leaded by our team.
ER -