Inaccurate topology and line parameters in low-voltage distribution networks(LVDNs) render traditional power flow calculation methods ineffective. While data-driven approaches reduce reliance on physical parameters, they often lack interpretability. To address this, we propose a hybrid method integrating physical knowledge with data-driven techniques. The input-output feature vectors of the deep learning model are constructed based on the DistFlow model, where the head-end node voltage, photovoltaic(PV) output, and load power at user nodes serve as input features, and the voltage magnitude at user nodes is the output feature. A multi-channel convolutional network is designed by incorporating a three-phase linearized power flow model. This network processes voltage, active power, and reactive power through independent channels, with convolutional kernel weights initialized using resistance(R) and reactance(X) parameters. Finally, aiming at the problem that gray data(data containing measurement errors and outliers) used for training will affect model performance, an improved denoising autoencoder(DAE) is proposed to filter and eliminate anomalous samples. Experimental results demonstrate that the proposed method outperforms conventional data-driven approaches in accuracy and generalization capability, while significantly reducing the influence of gray data on model performance.