Accurate ultra-short-term power prediction for distributed photovoltaic (PV) systems plays a crucial supporting role in providing essential information for distributed PV electricity sellers participating in the electricity spot market. Most existing methods employ cluster-based information-sharing modeling approaches to improve prediction accuracy. However, these methods either rely on direct data sharing, leading to serious data privacy issues, or depend solely on model interaction sharing.Although data privacy is ensured, the limited amount of shared information leads to poor prediction results. To address these challenges, a privacy-preserving, deep federated learning-based approach is proposed for ultra-short-term power prediction in distributed PV systems. Firstly, a federated learning framework based on dual-layer information interaction is constructed to enable richer information sharing through global model and global feature interactions among stations, avoiding direct data exchange to protect data privacy. Secondly, an attention-based autoencoder is adopted as the local prediction model to extract temporal features of power sequences, which are then coupled with global features to form spatiotemporal correlation features. By effectively mining and utilizing spatiotemporal correlation information, the power prediction accuracy is enhanced. Finally, simulation experiments using actual distributed PV data from a region in Hebei Province demonstrate the effectiveness of the proposed method.