Please use this identifier to cite or link to this item:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/71341
Title: | 應用三維卷積神經網路在濕對流的參數化 The representation of moist convection using 3D Convolutional Neural Networks |
Authors: | Shih-Wen Tsou 鄒適文 |
Advisor: | 吳健銘 |
Keyword: | 機器學習,深度學習,三維卷積神經網路,濕對流參數化,雲解析模式,VVM, Machine Learning,Deep Learning,3D-Convolutional Neural Networks,representation of moist convection,cloud-resolving model,VVM, |
Publication Year : | 2019 |
Degree: | 碩士 |
Abstract: | 濕對流過程參數化是造成氣候模式和數值天氣模式不確定性的主要原因。隨著硬體和技術的發展,前人嘗試使用機器學習藉由高解析雲解析模式的結果求得特定網格解析度下的濕對流效應(機器學習濕對流參數化),並期待未來能夠取代現存模式中由積雲參數化所診斷的濕對流效應。
而另一方面,氣候模式及數值天氣模式的發展潮流為增加水平解析度,以獲得更局部的氣象資訊並期望更準確的預測濕對流的發展過程。然而現存的積雲參數化在模式解析度逐步上升至~O(10km)尺度時,濕對流過程該由積雲參數化處理或是由網格尺度的物理過程處理變得模糊不清(grey zone resolution)。在這樣的水平解析度下,網格內可能不再同時包含對流以及其環境,意即決定網格對流效應的環境將延伸至相鄰網格。 然而,至今仍未有研究評估在grey zone resolution下以相鄰網格作為決定對流效應之環境對於模擬結果的影響。因此本研究嘗試利用高解析雲解析模式(VVM)的資料,並使用一種可掃描區域內參數梯度的3D卷積神經網路,在Grey zone resolution利用各個網格及其相鄰網格的資訊,預測各個網格的對流效應。為了證實在grey zone resolution鄰近網格資訊對於預測對流效應的助益,本研究呈現多元線性回歸(Multiple Linear Regression)及深度神經網路(Deep Neural Networks, DNN)的結果作為對照組,它們分別是機器學習模型的最低標準,及最基礎的深度學習模型。研究結果指出,由3D卷積神經網路所預測的對流效應在對流位置、對流強度高度分布及對流效應對可降水量的反應皆較另兩種模型有更好的表現,其中在對流效應對可降水量的反應的預測中,3D卷積神經網路甚至在環境濕度較高的環境下的預測,可以捕捉到5公里處freezing level非常細節的對流特徵,這是其他模型較難掌握的現象。 The parameterization of moist convection contributes to the uncertainty in climate modeling and numerical weather prediction. As the improvement of computational resources and the theory of machine learning, building parameterization of moist convection based on machine learning method becomes feasible nowadays. In the meanwhile, the grid size used by GCMs become finer in order to represent processes at finer scale, in which more accurate representation of moist convection is expected. However, the ambiguity between subgrid-scale and resolved scale processes is introduced which arouses the grey zone problem. In such resolution ~O(10km), the relationship between moist convection and its environment may not be distinguishable upon the same grid box. This study tries to apply the machine learning method to evaluate the model that considers the adjacent grid boxes in grey zone resolution. We use 3-Dimension Convolutional Neural Networks (3D-CNN) as our model to predict the effect of moist convection according to the meteorology parameters from both the target grid box and its surroundings. The high-resolution data is obtained from the results of a vector vorticity equation cloud-resolving model (VVM) under various environmental conditions. The unique aspect of 3D-CNN is that it can detect the gradient of each fields between adjacent grids within a certain domain. The result based on two other models, which are Multiple Linear Regression and Deep Neural Network is also shown in this study to manifest our findings. The result shows that 3D-CNN predicts more reasonable spatial distribution and vertical distribution of the effect of moist convection, and the response of the effect of moist convection to total precipitable water compared to the other models. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/71341 |
DOI: | 10.6342/NTU201900572 |
Fulltext Rights: | 有償授權 |
Appears in Collections: | 大氣科學系 |
Files in This Item:
File | Size | Format | |
---|---|---|---|
ntu-108-1.pdf Restricted Access | 4.27 MB | Adobe PDF |
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.