This study investigates the feasibility of employing deep learning models, specifically the U-Net3+ and Attention U-Net architectures, to estimate annual solar energy potential maps (ASM) using high resolution LiDAR-derived digital surface models (DSMs). As urban areas increasingly seek sustainable energy solutions, precise localization of rooftop solar panels becomes essential. While accurate, traditional physical models, such as the Area Solar Radiation (ASR) model, are often computationally intensive and time consuming, limiting their application over large areas. This research proposes a novel framework that integrates deep learning techniques to enhance the efficiency and accuracy of solar energy potential estimation. The methodology includes data collection, generation of reference ASMs from LiDAR DSMs using the ASR model, and training various U-Net models on patches of DSM data. The U-Net 3+ model demonstrated the highest correlation with reference ASMs with an RMSE of 94.353 (MWh/m²) and R² of 0.91, indicating its effectiveness in capturing the spatial relationships between topographical features and solar radiation. The results suggest that deep learning models can be a viable alternative to traditional physical models, facilitating quicker and more reliable solar energy potential mapping.