应用机器学习技术预测强雨雪天气过程中的积雪

2017年12月～2018年2月冬季,在中国长江中下游流域发生了两次强度强、范围广的强雨雪冰冻天气。在第一次强降雪天气中,由于2018年1月3～4日和5～8日两阶段降雪在中国东部落区高度重叠,导致了较为严重的灾害。为了预测日积雪深度,利用2017年12月～2018年2月和2007年12月～2008年2月这两个时间段上的国家测站日值数据,利用CART决策树算法根据各气象要素生成一个预测当天是否有积雪的二元判别决策树模型。从决策树结构中可以看出,前一日的积雪深度、日最高气温、日平均气温、日最低相对湿度等要素对预测结果的影响重大。且两决策树的结构相似度极高,故该模型对是否有积雪的预测存在普适性。随后利用深度学习方法训练两个时间段上所有预测为有积雪的个例,建立预测积雪深度的回归模型,结果表明,利用该模型训练得到的误差较小,但不足之处在于,预测极端降雪个例的误差大于普通降雪个例。将决策树模型与深度学习模型串接,便能得到预测当天是否有积雪,及积雪深度的模型。相比于前人的研究,该模型能拟合更复杂的特征,得到更精确的预测,使用2018年的数据也能更好地模拟当前的气候背景。

Predicting Snow Depth during Strong Rain and Snowfall Processes using Machine Learning Techniques

During the winter from Dec 2017 to Feb 2018, two strong rain and snow weather processes along with freezing took place in China. During the first process, severe disasters were caused by two periods of snowfall which overlaps mostly. In order to predict the daily snow depth, national weather station observations during the two periods were imported respectively to a CART decision tree model to predict whether there were accumulated snow on that day. Both decision trees show that the snow depth on the previous day, daily maximum temperature, daily average temperature, daily minimum relative humidity and so on have primary influence on prediction result. And the structure of both decision trees resemble each other, which implies that the condition to form snow accumulation is universal, and the two decision trees are representative of the rule of snow accumulation. Next, a regression model using deep learning techniques is trained in order to predict the value of snow depth given other meteorlogical variables. The results show that the model is accurate in snow depth prediction while its performance is limited for instances with a larger snow depth. As a result, by combining the decision tree model and deep learning model, we will get a model which predicts snow depth.