广东省降水的多尺度时空投影预测方法

采用多尺度时空投影(MSTP)预测思路建立广东月降水和季节降水预测方法。通过EOF分解、小波分析和Lanczos滤波方法进行周期分解, 采用MSTP方法进行预测。借鉴年际增量法, 对预报结果用最小二乘法进行误差订正, 得到降水预测结果。PS预测评分和均方根误差10年独立样本检验(2006—2015年)结果显示:订正后, PS预测评分起伏较小, 68.8%的月降水和季节降水PS预测评分明显提高的年份超过6年, 且有87.5%的月降水和季节降水PS预测平均分达到70以上; 在±0.5个标准差范围内, 订正后均方根误差在40%以上的概率分布明显高于订正前, 订正后的月和季节降水占81.3%, 订正前占31.3%;在±1个标准差范围内, 概率分布在70%以上的月季降水订正前后相差不多, 订正后占56.3%, 订正前占50%。

A Multi-scale Spatial-temporal Projection Method for Monthly and Seasonal Rainfall Prediction in Guangdong

Guangdong Province is located in low latitude areas and affected by both tropical weather systems and high latitude weather systems. A multi-scale spatial-temporal projection (MSTP) method is developed to predict the monthly and seasonal precipitation in Guangdong. Multi-factor and multi-scale forecasting method can be used to seek for the forecast factor by quantity scale separation through scale decomposition conforming to the physical meaning, thus it can reduce climate non-stationary time series and improve the prediction accuracy. The key feature of MSTP mothod is that it considers not only spatially but also temporally varying large-scale field connection between the predictor and predictand. Based on main modes of the empirical orthogonal function (EOF) analysis, periods are gained from the wavelet analysis and decomposed by Lanczos filtering. According to the correlation between the precipitation and Climate Forecast Systems datasets provided by National Centers for Environmental Prediction dynamic model data (CFSv2), significant influencing factors are selected to predict precipitation with MSTP method. Using the least square error correction method, Guangdong monthly and seasonal precipitation predictions are obtained based on inter-annual increment approach. The test of independent samples from 2006 to 2015 shows the correction can improve the performance of prediction, making PS score of operational test change smoothly. After correction PS scores are improved greatly during 6 years of the hindcast period, and the monthly and seasonal rainfall account for 68.8%. For 87.5% of total samples, the forecast average score is over 70. The prediction effect is closely related to period changes of the precipitation main mode. If the inter-annual period is priority to other period, the prediction effect after correction is significantly higher than that before the correction, otherwise it is poor. The root mean square error within 0.5-1 standard deviation rate after correction is higher than that before corrections. Within 0.5 standard deviations, the monthly and seasonal rainfalls, of which the root mean square error of probability is more than 40%, account for 81.3% after correction comparing to 31.3% before correction. Within 1 standard deviation, the monthly and seasonal rainfalls, of which the root mean square error of probability is more than 70%, account for 56.3% comparing to 50% before correction. It suggests that most of rainfall prediction errors are within 1 standard deviation. Therefore, the prediction from MSTP method can offer important reference to operational prediction of short-term climate prediction for monthly and seasonal prediction in Guangdong.