基于机器学习的灾害性天气多尺度预测模型

A Multi-Scale Prediction Model for Catastrophic WeatherBased on Machine Learning

为提高传统数值模拟预报结果的准确性,引入机器学习算法构建一种新型灾害性天气多尺度预测模型。通过降尺度时空融合算法实现遥感数据融合,反演推算得到大气气溶胶光学厚度作为天气预测模型的输入变量。利用包含反向解搜索策略的萤火虫优化算法,建立预测模型参数寻优策略,应用机器学习的支持向量机算法,构建包含多项式核函数的复杂多尺度预测模型,在考虑各种不确定因素的情况下进行不断训练,最终得到灾害性天气预测结果。使用该模型对2015年6月23日00—24时成都市灾害性天气进行预测,预测结果的ROC(Receiver Operating Characteristic)曲线的AUC(Area Under the Curve)值为0.88,且龙泉驿、新津和金堂站的预测正确率达90%。基于机器学习的灾害性天气多尺度预测模型可为灾害性天气预测提供一种有效手段。

In order to improve the accuracy of traditional numerical simulation prediction,a new multi-scale prediction model for catastrophic weather was constructed by introducing machine learning algorithms.A downscale spatiotemporal fusion algorithm was applied to achieve remote sensing data fusion,and then the atmospheric Aerosol Optical Depth was inverted as the input variable of the weather prediction model.A prediction model parameter optimization strategy was established using a firefly optimization algorithm that included a reverse solution search strategy.A complex multi-scale prediction model was constructed by using the support vector machine algorithm of machine learning that included a polynomial kernel function.Finally,the disastrous weather prediction results were obtained by continuous training under the conditions of considering various uncertain factors.The catastrophic weather occurred in Chengdu from 00:00 BT to 24:00 BT on June 23,2015 was predicted by using the method.The results showed that the predicted AUC value of the ROC curve was 0.88,and the prediction accuracy of Longquanyi,Xinjin and Jintang stations reached 90%.The multi-scale prediction model for catastrophic weather based on machine learning provided an effective means for predicting catastrophic weather.