Abstract:

In order to improve the fine defense ability of late frost disaster in the vineyards in the eastern foothills region of the Helan Mountain, the minimum temperature observation data of the vineyards from April to May during 2020-2023 were used to analyze the variation characteristics of minimum temperature, the occurrence frequency and regional distribution characteristics of late frost in the vineyards. Based on the European Centre for Medium-Range Weather Forecasts (ECMWF) model forecast products and the actual temperature of grid points in Ningxia, the radial basis function (RBF) neural network algorithm was used to construct the minimum temperature and frost prediction model in the vineyards in the eastern foot of the Helan Mountain. The results show that the light frost was the most common in the vineyards in the eastern foothills region of the Helan Mountain, followed by the medium frost. April was the main month for frost occurrence. The frost in D.F. Yuxing winery appeared most frequently, and the frost in Guanlan winery was the least. The verification results of the minimum temperature and frost forecast show that compared with the ECMWF model, the RBF model has improved the accuracy of the minimum temperature forecast in the Helan, Yongning and Hongsipu production areas, with the highest increase of 33.8%, and the average absolute error reduced by 0.20-1.50 °C. For the single station frost forecast, the RBF model has obvious advantages, the accuracy rate generally increased by 1.0%-14.0%, and the average absolute error reduced by 0.04-0.37 °C. For the average of the production areas, the RBF model has the highest accuracy of frost prediction in the Hongsipu production area, up to 13.0%. In the case analysis of frost, the RBF model has better prediction effect, especially for moderate frost prediction. Compared with the ECMWF model, the accuracy rate increased by 25.0%-50.0%, and the average absolute error reduced by 1.80-2.10 °C.

Key words: the eastern foothills region of the Helan Mountain, microclimate stations, radial basis function, late frost

摘要：

为提升贺兰山东麓葡萄园晚霜冻灾害精细化防御能力，利用2020—2023年4—5月贺兰山东麓葡萄园农田小气候站最低气温观测数据，分析葡萄园最低气温变化特征、晚霜冻发生频率和区域分布特征，并基于欧洲中期天气预报中心（European Centre for Medium-Range Weather Forecasts，ECMWF）模式预报产品和宁夏地区格点气温实况，采用径向基函数（Radial Basis Function，RBF）神经网络算法，构建贺兰山东麓葡萄园最低气温和霜冻预报模型。结果表明：贺兰山东麓葡萄园轻霜冻最为普遍，其次是中霜冻，4月是霜冻发生的主要月份，东方裕兴酒庄霜冻出现最频繁，观兰酒庄霜冻最少，红寺堡产区是霜冻易发区。最低气温和霜冻预报检验结果显示，与ECMWF模式相比，RBF模型对贺兰、永宁和红寺堡产区的最低气温预报准确率提高，最高提升幅度达33.8%，平均绝对误差降低0.20~1.50 ℃。从单站霜冻预报看，RBF模型有明显优势，准确率普遍提升1.0%~14.0%，平均绝对误差降低0.04~0.37 ℃；从产区平均看，RBF模型对红寺堡产区霜冻预报准确率提高最多，达13.0%。在针对霜冻的实例分析中，RBF模型预报效果更优，特别是对中霜冻预报优势明显，相比ECMWF模式准确率提升25.0%~50.0%，平均绝对误差降低1.80~2.10 ℃。

关键词: 贺兰山东麓, 农田小气候站, 径向基函数, 晚霜冻