The primary aim of this study is to train and evaluate a set of Reduced-Order Models (ROMs) for predicting upper-wing pressure distributions on a civil aircraft configuration. Using wind tunnel data recorded for the Airbus XRF-1 research configuration in the European Transonic Windtunnel (ETW), the ROMs integrate dimensionality reduction and time-evolution components. For dimensionality reduction, both Singular Value Decomposition (SVD) and a convolutional Variational Autoencoder (CNN-VAE) neural network are employed to reduce/encode Instationary Pressure Sensitive Paint (IPSP) data from a transonic buffet flow condition. Fully-Connected (FC) and Long Short-Term Memory (LSTM) neural networks are then applied to predict the evolution of the latent representation, which is subsequently reconstructed/decoded back to its original high-dimensional state. SVD and CNN-VAE are trained on data from five flow conditions and tested on two unseen conditions. When comparing the power spectra of the reconstructed and experimental data, SVD demonstrates marginally better performance. Subsequently, FC and LSTM models are applied for the forward evolution of the reduced/latent representation for one flow condition, resulting in four evaluated ROMs. Among them, the CNN-VAE-LSTM model excels by accurately capturing buffet dynamics, encompassing both transient features and steady-state oscillations. SVD-based models tend to struggle with transient buffet behavior, as they predominantly learn steady-state dynamics. Additionally, the CNN-VAE-LSTM model was employed for an end-to-end training approach. The results suggest that it faces difficulties maintaining dynamics in predictions, demanding further investigation and optimization. GitHub repository