This paper constructs a deep learning model based on the sequential fusion encoder (SFE) for forecasting winter precipitation phase. The model integrates the advantages of convolutional neural network (CNN), convolutional gated recurrent unit (ConvGRU), and Transformer. It can conduct automatic learning and extraction of complex features from meteorological data, handle non-linear relationships, and process large-scale datasets. We utilize hourly precipitation observation data from Wuhan Station in 2010-2024 and ERA5 reanalysis data, select 60-channel grid data (including temperature and geopotential height) from 9 atmospheric layers (1000-500 hPa) as predictors. To address sample imbalance, minute-level data augmentation is employed, involving resampling at intervals of 30 min for rain, 1 min for sleet, and 5 min for snow. Finally, a sample size of 19 932 is obtained. Test results show that this model performs excellently in forecasting solid precipitation (snow and sleet), with F1-scores of 0.92-0.93 in the training set and 0.67-0.68 in the validation set. However, its ability to identify rapid transitions between precipitation phases needs to be improved. Verified by two complex weather processes in February 2024, the model is found to be able to serve as a supplement to numerical prediction and provide an efficient solution for intelligent forecasting of winter precipitation phase, aiding in enhancing the forecasting capabilities of meteorological stations.