Site-specific phosphorylation events affect nearly all the cellular processes and correct phosphosite localization plays an important role in biological or medical health studies. However, direct false localization rate (FLR) control remains challenging in phosphoproteomics. Here, we propose DeepFLR, a deep learning-based framework utilizing spectrum prediction and the target-decoy method for FLR estimation. We demonstrate that the similarity between predicted and experimental phosphopeptide spectra is comparable to the measurement reproducibility. We further benchmark our method with four synthetic datasets and three real biological sample datasets, showcasing its ability for sensitive phosphosite localization with accurate FLR estimation.