Building Integrated Photovoltaics (BIPV) presents a viable strategy for enhancing energy sustainability in urban environments. Accurate detection of usable façade surfaces is critical for promoting PV integration. This study proposes a deep learning approach that applies a Cycle Consistent Generative Adversarial Network (CycleGAN) to segment areas suitable for PV installations from building façade images. Initially, we implemented a U-Net-based Pix2Pix GAN model, but it exhibited inadequate segmentation performance. To address this, we replace the Pix2Pix with the CycleGAN model, which allows for unpaired image translation between real-world façade images and their corresponding segmentation masks. The outputs of segmentation were post-processed to estimate the area available for BIPV. These segmented masks served as input data for PVGIS to predict energy yield. A case study in Edinburgh (Lat/Lon 55.933, −3.213) was conducted for south orientation, considering the PV orientation losses (19.96%) due to installation in vertical tilt angle (90°). The model successfully identified 81 m2 with a potential yield of 629.9 (kWh/kWp/year) of usable PV area, which represents approximately 41.51% of the total 195 m2 façade. These findings demonstrate the effectiveness of integrating AI with energy modelling in predicting BIPV potential at the facade level, providing a scalable and efficient solution for assessing urban solar potential.