- 學位論文
應用多變數迴歸與類神經網路模擬機場能見度
Using Multiple Regression and Artificial Neural Network Approach for Modeling Airport Visibility
摘要
能見度變化起因於天氣條件。據飛航安全統計，能見度低乃肇生飛機失事因之一。產生低能見度之因素很多，在地面為濃霧、濃霾或降雨，在高空為雲中或高層霾。低能見度除影響起飛降落及地面作業外，惡劣之垂直能見度將阻礙投彈、偵照、船艦行動及三棲作戰配合之進行，進而影響戰演訓任務順遂。天氣守視室，為觀測及記錄機場天氣現象之重要場所，因其設置地點限於機場跑道起飛或降落區，提供氣象觀測資料予飛行員及飛航管制人員執行飛行任務。由於低能見度發生時機大部分在於清晨，因此，本研究主要以晨間6至10時連續逐時能見度於3,600公尺以下之觀測天氣資料進行訓練模擬，以建置完整推估模式，尋求以氣象資料透過科學計算來判斷能見度的可能性。 研究中蒐集本省南部機場及離島機場天氣守視室自1984年至2004年間共21年之氣象電報資料做為研究材料，資料計有以風向、風速、陣風、視障、雲組、總雲量、溫度、露點、相對溼度、高度表撥定值、氣壓變量、降雨量及降雨時數等13項氣象因子。本研究以統計學中相關係數及F 值檢定進行能見度影響因子篩選，並據以進行建置多變數迴歸與倒傳遞類神經網路模型。由影響因子篩選結果，在本島機場資料不分季:相對溼度、視障、溫度、天空狀況為能見度的影響因子。而分季後春季的影響因子為相對溼度、視障及天空狀況，夏季的影響因子為相對溼度、溫度、天空狀況、降雨量及視障，影響秋季能見度的因子為相對溼度及溫度，冬季的影響因子為相對濕度及視障。對離島而言，在資料不分季的條件下，相對溼度、天空狀況、視障、溫度、風速為影響能見度的因子。比較研究中就資料不分季情況下所完成之倒傳遞類神經網路與多變數迴歸模型之判定係數顯示以倒傳遞類神經網路之表現較佳。又以倒傳遞類神經網路預測階段所得結果透過分級預測在1,800公尺以下之能見度準確率分別為92.6％與77.4％，而1,801公尺至3,600公尺之準確率則分別達93.4％與91.4％。整體而言，以類神經網路來協助本研究中的二個機場在低能見度的判斷是可行的。
並列摘要
The fluctuation in visibility is influenced by climatic condition. According to flight safety statistics, low airport visibility is one of the reasons that cause flight accidents. Low airport visibility can be originated from many factors, including thick fog, rain or thick haze for on ground situation; midst of cloud or high altitude for the condition of up in the air. Apart from affecting the taking off or landing of planes and ground works, appalling vertical visibility will also obstruct weapons launch, surveillance assignments, vessel activities and the cooperation among the air force, navy and army divisions, thus disturbing the success of warfare missions. Weather forecast division is an important place where monitors and records the weather condition, therefore its location is restricted in the area of taking off and landing strips only. Due to low airport visibility usually occurs during early morning, thus this study emphasized on the airport visibility below 3,600 meters from 6am to 10am in the morning. The main objective of the study is to construct a model which has the capability for forecasting airport visibility by using the weather data collected from weather division. This research utilized weather data from two airports located in the southern region of mainland Taiwan and the other one located on an offshore island of Taiwan dating from 1984 to 2004, a total of 21 years. The data includes wind direction, wind speed, gust wind speed, visual hazard, cloud, total cloud mass, temperature, dew point, relative humidity ,altimeter setting, air pressure mass, rainfall, the hours of rainfall and 13 more weather related factors. This research uses correlation coefficient and F test to study the possible influence factor to airport visibility. Following the determined influence factor, multiple linear regression and back-propagation artificial neural network are used for modeling airport visibility. From the results of the influence factor determination, without considering seasonal effects, humidity, visual hazard, temperature, and atmospheric condition are statistically significant for the airport located at southern Taiwan studied herein. As for the airport on offshore island researched in the study, humidity, atmospheric condition, visual hazard, temperature and wind speed are the influence factors to airport visibility. Under non-seasonal consideration, by comparing the coefficients of determination, one can say that the back-propagation artificial neural network approach is superior to the conventional regression method. Finally, applying the results from artificial neural network approach through class predictive accuracy test, the accuracy are 92.6% and 77.4% for airport visibility under 1,800 meters and 93.4% and 91.4% for airport visibility between 1,801 meters and 3,600 meters respectively for back-propogation artificial neural network and multiple linear regression models built in this study.