A Particle Swarm Optimization-neural Network Ensemble Prediction Model for Persistent Freezing Rain and Snow Storm in Southern China

Based on daily minimum temperature, maximum temperature and precipitation data of 756 stations in China, National Center for Environmental Prediction (NCEP)/National Center for Atmospheric Research (NCAR) reanalysis data during 1951-2013 and NCEP 24 h forecast data, a nonlinear statistical ensemble prediction model based on the particle swarm optimization-neural network (PSONN-EPM) is developed for predicting and verifying the regional persistent freezing rain and snow storm process in southern China by analyzing and extracting significant predictors. Results show that model performance can be effectively improved when dividing low-temperature processes into the general process and severe process which are constructed based on cold extents, humidity and influence ranges of the freezing rain and snow storm processes. In 10-day independent forecast test, the average relative errors for the general process and the severe process are 2.04 and 0.6 using stepwise regression equation forecast method, while those are 1.33 and 0.30 by using PSONN-EPM technique. It means forecast errors are reduced by 0.71 and 0.3 as compared with the stepwise regression method. In addition, the predication result for the severe freezing rain and snow storm process is better than that for the general process. The PSONN-EPM integrates predictions of multiple ensemble members, thus the prediction accuracy and stability are higher than those of the traditional linear regression method. Furthermore, such method does not contain any tunable parameters, and is applicable for practical operational weather prediction.