【作者】 曾静；

【导师】 马雷鸣； 张长江；

【作者基本信息】 浙江师范大学 ， 物理电子学， 2020， 硕士

【摘要】 地面气温是人们日常生活最关注的气象要素之一,对农业、工业、服务业等领域产生重要的影响。为了提高现有日常预报业务中模式预报温度的准确率,我们基于欧洲中期天气预报中心(ECMWF)的经度为120～130°E,纬度为20～40°N的温度集合控制预报历史资料以及华东地区实况站点温度资料对模式预报温度进行订正。主要包括以下三个方面的研究工作:(1)分析和处理实验数据。在本文中利用均方根误差,绝对误差和温度预报准确率构造的综合评价指标对插值性能进行评价,最终采用克里金插值法将模式预报数据与实际站点温度进行空间上的对应,然后采用小提琴图对插值后的数据进行异常值分析并清洗,得到可靠的实验数据。接着为了充分挖掘模式预报数据与实际温度之间潜在关系,将最高温和最低温相关模式预报特征分别进行多项式扩充构造后续温度订正建模的输入特征;(2)基于传统机器学习的多气象因子模式预报温度订正模型。在第一步多项式特征扩充的基础上,针对最高温和最低温分别采用线性回归,支持向量机,决策树和随机森林分别建立温度订正模型。根据不同相关系数的设置对二、三、四次多项式扩充后的数据进行筛选,分别选取二次扩充后与实际温度相关性在0.2～0.8以上的特征以及全部特征进行建模,三次扩充后与实际温度相关性在0.2～0.8以上的特征以及四次扩充后与实际温度相关性在0.4～0.8以上的特征进行建模比较,从中选择最低温和最高温订正性能最佳的模型;(3)基于深度学习的多气象因子模式预报温度订正模型。采用K-means分别对最高温和最低温扩充后数据进行聚类。聚类能够对不同样本有针对性分析和处理,为了保证聚类的准确性采用手肘法则和轮廓系数对类别的选择进行选择,将最低温样本聚类为三类,最高温样本聚类为四类。聚类后使用长短时记忆网络(LONG SHORT-TERM MEMEORY,LSTM)结合注意力机制对模式预报的最高温和最低温建立订正模型。聚类后分别针对性地建模能够充分利用样本的多样性以获得最佳模型。最后将所提出的深度学习温度订正模型和机器学习温度订正模型与一般用于温度订正的卡尔曼滤波和神经网络模型订正后性能进行对比分析。更多还原

【Abstract】 Surface temperature is one of the most concerned meteorological factors in people’s daily life,which has an important impact on agriculture,industry,service industry and other fields.In order to improve the accuracy of the model forecast temperature in the current daily forecast operation,we correct the model forecast temperature based on the ensemble control forecast historical data with longitude of 120-130 ° E and latitude of 20-40 ° N of European Centre for Medium-Range Weather Forecasts(ECMWF)and the temperature data of the actual stations in East China.It mainly includes the following three modules:(1)Analyze and process experimental data.The interpolation performance is evaluated by using the comprehensive evaluation index constructed by root mean square error,absolute error and temperature prediction accuracy.Finally,Kriging interpolation method is used to spatially correspond the model prediction data with the actual station temperature,and then the violin chart is used to analyze and clean the abnormal value of the interpolated data to obtain reliable experimental data.Then,in order to fully mine the potential relationship between the model prediction data and the actual temperature,the prediction features of the highest temperature and the lowest temperature are polynomial extended to obtain the input features of the subsequent modeling;(2)Multi-meterological factors model temperature prediction correction model based on traditional machine learning.On the basis of the first step polynomial feature expansion,linear regression,support vector machine,decision tree and random forest are used to establish the temperature correction model respectively for the highest temperature and the lowest temperature.According to the setting of different correlation coefficients,the data after the quadratic,cubic and quartic polynomials expansion are selected,and the features with the correlation between the quadratic expansion and the actual temperature above 0.2-0.8 and all features are selectedrespectively for modeling.The features with the correlation between the cubic expansion and the actual temperature above 0.2-0.8 and the features with the correlation between the quartic expansion and the actual temperature above 0.4-0.8 are built The most suitable model for the lowest and highest temperature correction was selected;(3)Multi-meterological factors model temperature prediction correction model based on deep learning.K-means is used to cluster the extended data of the highest temperature and the lowest temperature.In order to ensure the accuracy of clustering,elbow rule and contour coefficient are used to select the categories.The lowest temperature samples are classified into three categories,and the highest temperature samples are classified into four categories.After clustering,long-term memory network(LONG SHORT-TERM MEMEORY,LSTM)combined with attention mechanism is used to build a correction model of the highest temperature and the lowest temperature.After clustering,respectively targeted modeling can make full use of the diversity of samples to obtain the best model.Finally,the performance of the proposed deep learning temperature correction model and machine learning temperature correction model is compared with that of Kalman filter and neural network model.更多还原