河川揚塵之空間與時間頻率分析與預測: 以台灣中南部河川為例

Aeolian River Dust in Central and Southern Taiwan Rivers: Spatial-Temporal Characterization and Prediction

陳俊光(Chun-Kuang Chen)

指導教授 ： 蔡宛珊

國立臺灣大學/工學院/土木工程學研究所/碩士(2022年)

https://doi.org/10.6342/NTU202203329

摘要

在台灣西部平原上，有許多湍急的河川發源於高山地形。河川中湍急的水流會挾帶著大量河道中的泥 沉積物，並堆積於下游河口。河川揚塵事件是一種沙塵的極端事件發生於乾燥的河川裸露地。濁水溪 及高屏溪分別為在台灣中、南部河川，河川揚塵事件發生潛在機率最高的兩條河川。當季風盛行的 節，堆積在河口的泥沙沉積物能夠輕易的被帶到空氣中。因此住在河岸附近的居民會容易暴露在高濃 的PM10環境中，吸入過多的高濃度PM10會對於居民的身體健康產生負面的影響。首先，為了瞭解濁 溪以及高屏溪河川揚塵事件在空間與時間上的發生分布。我們利用統計分析法發現了濁水溪河川揚塵事盛行的季節為每年的10月至隔年的4月，而高屏溪盛行的季節則為每年的3月至9月。此外每年冬天濁溪發生的河川揚塵事件的頻率以及事件長度為高屏溪揚塵事件的10至20倍。
接著由於水文氣象因子(例如: 溫度，降雨，相對溼度，以及風速)為主要影響揚塵(PM10)變化的因子。因此本研究探討了揚塵(PM10)與相關水文氣象因子(例如: 溫度，降雨，相對溼度，以及風速)之間的關聯性。然而在自然界中的水文訊號通常具有非線性、非穩態、以及多尺度的特性，限制了傳統時間頻率方法(例如:傅立葉轉換以及小波轉換)應用於分析各水文氣象因子與PM10之間關連性。因此我們引入了基於改進的完全集合經驗模態分解與自適性噪音(Improved Complete Ensemble Empirical Mode Decomposition with Adaptive Noise)的時間依賴性內在關聯性方法(Time dependent intrinsic correlation)分析非線性、非穩態的水文訊號並探討變數間的關聯性。本研究利用 ICEEMDAN方法發現PM10的時間序列與相關水文氣象因子在天與年的時間尺度上是具有顯著性。在年尺度中，只有崙背測站的風速因子與PM10顯示為正相關，其餘因子皆顯示為強烈的負相關; 而在天尺度的分析中觀察到了PM10與溫度、相對溼度、風速之間關連性隨著季節強弱的轉換，正相關通常發生於河川揚塵事件盛行的季節。為了研究風的環流如何影響PM10濃度與河川揚塵事件，本研究提出基於ICEEMDAN方法具有時間尺度的風動能分析。在天尺度上，PM10濃度與海陸風有關; 而在年尺度上，PM10濃度與冬天的季風有著密切的關係。風動能分析在揚塵事件分析上，濁水溪顯示與東北季風有關連，而在高屏溪，河川揚塵事件則與西北季風以及西南季風有關連性。
最後，為了降低河川揚塵事件所帶來對健康危害的風險，本研究提出結合ICEEMDAN-RBFNN的預測模型來預警未來3小時的PM10 濃度。此外本研究中除了ICEEMDAN-RBFNN模型也引入RBFNN和MLP模型進行了在預測未來3小時的PM10的預測能力比較。預測模型的神經網絡建構中，溫度、相對濕度、風速、風向因子和過去的PM10濃度資料被納入模型的輸入因子。在模型測試結果顯示，結合ICEEMDAN-RBFNN模型的精準度在預測未來3小時PM10 高濃度或是低濃度的PM10中都相較於RBFNN以及MLP模型優秀。因此作為河川揚塵的預警系統模型，ICEEMDAN-RBFNN預測模型能夠有效地預測PM10濃度。此外本研究所提出預測模型的穩定性可以由改變放入不同長度的過去PM10濃度資料來證明。最後，風險評估為本研究中所提出的結合預測模型的一個應用。

關鍵字

改進的完全集合經驗模態分解與自適性噪音 ； 時依性本質相關係數 ； 人工神經網絡 ； 河川揚塵事件

並列摘要

In the western plains of Taiwan, there are several rapid rivers that originate from the mountainous terrain. Turbulent flow transports large quantities of river sedimentation that accumulate in the riverbed downstream. River dust episodes are extreme dust events occasionally generated from the dry riverbeds. Choushui and Kaoping rivers are considered to be the two rivers with the highest potential for river dust events in central and southern Taiwan, respectively. When the monsoon prevails, the accumulated river sedimentation in the estuary can easily be lifted into the air. Due to this, vulnerable populations living near riverbanks are exposed to high PM10 concentrations, which may have adverse health effects. The spatial-temporal distribution of river dust episodes is first analyzed through a simple statistical analysis. It has been determined that river dust events will be prevalent along the Choushui river from October to April next year, while along the Kaoping river from March to September.
Furthermore, a correlation analysis is conducted to examine the relationship between river dust ( PM10 ) and relevant hydro-meteorological factors (i.e., temperature, precipitation, relative humidity, and wind speed). However, natural signals are nonlinear, nonstationary, and multiscale, which limits the possibility of determining a correlation between the two factors using traditional time frequency analysis (for example, the Fourier transform and wavelet transform). As a result, we introduce a time-dependent intrinsic correlation (TDIC) based on the Improved Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (ICEEMDAN) algorithm. Based on the ICEEMDAN algorithm, the annual and diurnal scales are identified to be significant among the PM10 and relevant hydro meteorological factors. Only the wind speed at the Lunbei station showed a positive correlation at the annual scale, whereas other factors showed a strong negative correlation. A seasonal switchover of correlation can be observed between PM10 and temperature, relative humidity, and wind speed at the diurnal scale. The positive correlations are attributed to the season of river dust episodes in the Choushui and Kaoping rivers. A scale-dependent wind motion analysis based on the ICEEMDAN algorithm is also presented to examine how wind patterns affect PM10 concentrations and river dust episodes. On a diurnal scale, PM10 concentration is related to land-sea breezes, while on an annual scale, it corresponds to the winter monsoon. The river dust episodes are associated with the northeast monsoon along the Choushui river, while they occur along the Kaoping river both during the northwest and southwest monsoons.
In order to reduce significant health risks associated with river dust episodes, a hybrid Improved Complete Ensemble Empirical Mode Decomposition with Adaptive Noise - Radial Basis Function Neural Network (ICEEMDAN-RBFNN) prediction model is proposed. A comparison is made between the hybrid model and the RBFNN and Multilayer Perceptron (MLP) models for forecasting future three-hour PM10 concentrations. For network modeling, temperatures, relative humidity, wind speeds, wind direction indexes, and previous hours of PM10 concentrations are taken into account. The results showed the accuracy of the hybrid model outperformed the RBFNN and the MLP models in capturing either the high or low PM10 concentrations for the next three hours. Accordingly, it is evident that the hybrid ICEEMDAN-RBFNN model is capable of predicting PM10 efficiently, serving as an early warning system for episodes of river dust. In addition, the stability of the proposed hybrid model can be proven by the changed length of the input previous PM10 concentrations. Finally, risk assessments are shown as an application of the proposed hybrid model in this study.

並列關鍵字

Improved Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (ICEEMDAN) ； Time-dependent Intrinsic correlation (TDIC) ； Radial Basis Function Neural Network (RBFNN) ； river dust events

參考文獻

Al Ameri, I. D., Briant, R. M., Engels, S. (2019). Drought severity and increased dust storm frequency in the Middle East: a case study from the Tigris–Euphrates alluvial plain, central Iraq. Weather, 74(12), 416-426.

Aryal, Y. N., Evans, S. (2021). Global dust variability explained by drought sensitivity in CMIP6 models. Journal of Geophysical Research: Earth Surface, 126(6), e2021JF006073.

ASCE Task Committee on Application of Artificial Neural Networks in Hydrology. (2000). Artificial neural networks in hydrology. II: Hydrologic applications. Journal of Hydrologic Engineering, 5(2), 124-137.

Bracewell, R. N. (1989). The fourier transform. Scientific American, 260(6), 86-95. Bracewell, R. N., Bracewell, R. N. (1986). The Fourier transform and its applications (Vol. 31999, pp. 267-272). New York: McGraw-Hill.

Broomhead, D. S., Lowe, D. (1988). Radial basis functions, multi-variable functional interpolation and adaptive networks. Royal Signals and Radar Establishment Malvern (United Kingdom).