A winter precipitation-type prediction is a challenging problem due to the complexity in the physical mechanisms and computability in numerical modeling. In this study, we introduce a new method of precipitation-type prediction based on the machine learning approach LightGBM. The precipitation-type records of the in situ observations collected from 32 national weather stations in northern China during 1997–2018 are used as the labels. The features are selected from the conventional meteorological data of the corresponding hourly reanalysis data ERA5. The evaluation results of the model performance reflect that randomly sampled validation data will lead to an illusion of a better model performance. Extreme climate background conditions will reduce the prediction accuracy of the predictive model. A feature importance analysis illustrates that the features of the surrounding area with a –12 h offset time have a higher impact on the ground precipitation types. The exploration of the predictability of our model reveals the feasibility of using the analysis data to predict future precipitation types. We use the ECMWF precipitation-type (ECPT) forecast products as the benchmark to compare with our machine learning precipitation-type (MLPT) predictions. The overall accuracy (ACC) and Heidke skill score (HSS) of the MLPT are 0.83 and 0.69, respectively, which are considerably higher than the 0.78 and 0.59 of the ECPT. For stations at elevations below 800 m, the overall performance of the MLPT is even better.

Based on daily 500-hPa geopotential height from ERA-Interim reanalysis data, this study analyzed the day-to-day circulation variance in cold season (October–March) by composite and correlation analysis. Two same-length time periods were compared, namely, the hiatus period (1999–2013) and the rapid warming period (1984–1998). Spectral analysis revealed that over the mid–high latitudes of the Northern Hemisphere, the most outstanding peak in the daily 500-hPa geopotential height variance was of quasi-biweekly timescale (about 10–20 days), accounting for about 32% of the total variance. During the warming hiatus, quasibiweekly disturbance (QBD) changed remarkably in Northeast Asia. On average, within the domain 42°–50 °N, 128°–142 °E, the QBD variance changed from 1860 m² in the rapid warming period to 2475 m² in the hiatus period, increasing by about 33% and statistically significant at the 95% confidence level. Lead–lag analysis showed that the QBD signal could be traced back by about 14 days, with an origin around the Ural Mountains. Then, the signal developed and propogated southeastward, with its location about 10 days prior to its peak in West Siberia, and about 6 days prior to its peak in the Sayan Mountains, and finally moving to Northeast Asia. By comparing the propagation process between the two periods, we found that the propagation paths were basically the same, but there were evident differences in the intensity of the signals. The intensification of QBD may have been related to the increased energy conversion from mean flow to QBD transients. The frequency of low-temperature extremes in negative QBD phases was much higher than under normal conditions or in positive phases. Associated with the enhanced QBD, the probability of extreme low temperature increased from 19% during the rapid warming period to 27% during the warming hiatus.