Climate change is largely determined by the radiation budget imbalance at the Top Of the Atmosphere (TOA), which is generated by the increasing concentrations of greenhouse gases (GHGs). As a result, the Earth Energy Imbalance (EEI) is considered as an Essential Climate Variable (ECV) that has to be monitored continuously from space. However, accurate TOA radiation measurements remain very challenging. Ideally, EEI monitoring should be performed with a constellation of satellites in order to resolve as much as possible spatio-temporal fluctuations in EEI which contain important information on the underlying mechanisms driving climate change. The monitoring of EEI and its components (incoming solar, reflected solar, and terrestrial infrared fluxes) is the main objective of the UVSQ-SAT pathfinder nanosatellite, the first of its kind in the construction of a future constellation. UVSQ-SAT does not have an active determination system of its orientation with respect to the Sun and the Earth (i.e., the so-called attitude), a prerequisite in the calculation of EEI from the satellite radiation measurements. We present a new effective method to determine the UVSQ-SAT’s in-orbit attitude using its housekeeping and scientific sensors measurements and a well-established deep learning algorithm. One of the goals is to estimate the satellite attitude with a sufficient accuracy for retrieving the radiative fluxes (incoming solar, reflected solar, terrestrial infrared) on each face of the satellite with an uncertainty of less than ±5 Wm−2 (1σ). This new method can be extended to any other satellites with no active attitude determination or control system. To test the accuracy of the method, a ground-based calibration experiment with different attitudes is performed using the Sun as the radiative flux reference. Based on the deep learning estimation of the satellite ground-based attitude, the uncertainty on the solar flux retrieval is about ±16 Wm−2 (1σ). The quality of the retrieval is mainly limited by test conditions and the number of data samples used in training the deep learning system during the ground-based calibration. The expected increase in the number of training data samples will drastically decrease the uncertainty in the retrieved radiative fluxes. A very similar algorithm will be implemented and used in-orbit for UVSQ-SAT.