Abstract: Aiming at the problems that the training data of distributed photovoltaic power prediction is polluted by dynamic line loss noise, the model accuracy is insufficient and it is difficult to promote on a large scale, a photovoltaic power prediction method based on dynamic line loss modeling and cross-region parameter migration is proposed. A dynamic line loss model is constructed in the source-domain demonstration station with dual monitoring of inverters and point of common coupling (PCC), and the migratable line loss characteristic parameters are extracted. The parameters are migrated to the target stations only equipped with PCC metering equipment, and the pure theoretical gross power sequence is reconstructed inversely through the measured net power and estimated line loss to eliminate the nonlinear interference of line loss. A dual-path decoupled architecture of gross power prediction and line loss estimation is adopted to learn the power generation law and loss law respectively, and the PCC power prediction results are synthesized. Experiments show that the proposed method can effectively eliminate the interference of line loss noise, improve the training quality and generalization ability of the prediction model, and significantly improve the prediction accuracy compared with the traditional fixed line loss method. It is suitable for low-cost and large-scale power prediction of massive distributed photovoltaic stations。