高斯过程回归方法在浙江沿海海岛冬春季阵风预报中的应用试验

利用2006—2016年冬春季浙江四个海岛气象站10 m大风观测资料和ERA-interim资料，首先分析了阵风因子和阵风风速的概率分布特征；然后统计阵风与10 m至500 hPa层的气温、风速、散度、涡度、比湿、垂直速度、浮力能等要素的关系，选取高相关的预报因子；最后采用高斯过程回归方法建立阵风概率预报模型，并进行试报。(1)平均风速相同时对应的阵风因子变化较大，导致阵风也出现大的差异，说明阵风数据分布具有混沌性；阵风风速具有正态或准正态分布特点，在自然对数处理后完全符合正态分布，表明采用高斯过程回归方法建立阵风概率预报模型合理可行。(2)阵风与大气低层的动力因子相关较好，而在近中层则与热力因子相关较好。(3)阵风大值样本在大气低层具有更强的下沉速度，有利于上层动量向下输送，且大值样本对应的中层气温和比湿相对大些，说明中层暖湿气流有利于湍流的发展和不稳定能量的交换。(4)试报模型的因子权重尺度分析表明，最佳预报因子绝大多数集中在875 hPa层以下，说明大气低层因子对近地面阵风起主导作用。(5)高斯过程回归模型试报表明，大部分站点阵风预报的50%概率区间上下界跨度约为2.5 m/s，75%概率区间跨度约为4.5 m/s，样本的50%和75%概率区间击中率均符合预期。 

APPLICATION OF GAUSSIAN PROCESS REGRESSION METHOD TO GUST FORECASTING IN WINTER AND SPRING IN ZHEJIANG COASTAL ISLANDS

By the use of the gale data and ERA-interim data from four island meteorological stations in Zhejiang during winter and spring from 2006 to 2016, the climatic probability distribution characteristics of gust factors and gust velocity are analyzed first. Then the relationship between gust and meteorological factors such as temperature, wind speed, divergence, vorticity, specific humidity, vertical velocity and buoyancy energy from 500 hPa to the ground level are statistically analyzed. Finally, the Gaussian process regression method is used to establish a probability forecast model of gust, and the trial and comparison analysis are made. The following conclusions are drawn. (1) When the average wind speed is the same, the corresponding gust factors can change greatly, which leads to the big difference of gust, indicating that there is chaos in gust data distribution. Coastal gust wind speed has normal or quasi-normal distribution characteristics, which fully conforms to normal distribution after natural logarithm processing. Gust data characteristics show that the probability prediction method based on Gaussian process regression is reasonable and feasible. (2) Gust has a good correlation with the dynamic factors in the lower atmospheric layer, whereas it has a good correlation with the thermodynamic factors in the near middle atmospheric layer. (3) The sinking velocity of large gust samples is stronger in the lower layer, which is conducive to the downward transport of momentum in the upper layer. Moreover, the temperature and humidity in the middle layer are relatively higher, which indicates that the warm and humid air flow in the middle layer is more conducive to the turbulent development and the exchange of unstable energy. (4) Weight scale analysis of sample tests shows that most of the selected predictors are concentrated below 875hPa layer, indicating that the lower atmosphere factors play a dominant role in gust near surface. (5) The test results of Gaussian process regression model show that the 75% probability forecast interval span of samples at most stations is about 4.5 m/s, and the 50% probability forecast interval span is about 2.5 m/s. The 50% and 75% probability interval hit rates of samples are in line with expectations.