深度學習法於颱風期間即時水位預報之研究

Deep learning techniques for real time water level forecasting during typhoons

陳郁仁(Yu-Ren Chen)

指導教授 ： 林國峰

國立臺灣大學/工學院/土木工程學研究所/碩士(2018年)

https://doi.org/10.6342/NTU201802361

摘要

台灣位處熱帶與亞熱帶交接處，常受到颱風的侵襲，近年由於氣候極端化也導致每年平均侵襲台灣的颱風數量有上升的趨勢。伴隨颱風而來的大量雨水，常導致淹水災害。本研究建立即時水位預報模式，提供作為防災參考依據，以期達到減災的目的。
研究中提出了以擴展序列卷積網路 (dilated causal convolutional neural network) 為基礎之深度學習模式來建立即時水位預報模式，將傳統卷積類神經網路轉換維度使其適用於一維時序列資料，並透過將輸入資料依照時序列分段傳入網路，訓練時以場次為單位使模式能更全面學習颱風事件所有區段之特性，進而掌握水文反應過程，改善傳統機器學習法無法處理大量資訊之缺點，亦使用了殘留層 (residual) 與門函數 (gated activation units) 之新興網路連結來強化網路整體強健性；另外，也將所建立模式與傳統類神經網路及支援向量機進行比較。
本研究選擇台灣東北部宜蘭河流域作為研究區域。蒐集2012至2017年的颱風雨量資料以及水位資料作為模式輸入因子，建置即時水位預報系統，進行未來1小時至未來6小時水位預報。研究結果顯示，各模式在未來1至3小時之預報系統上都有不錯的表現，但是一旦將預報時間拉長，傳統機器學習法於水位峰值及預報延遲容易出現誤差。傳統類神經於未來3小時即出現明顯誤差，支援向量機則於未來4小時出現明顯誤差。而擴展序列卷積網路利用資料分段處理的特性，能有效處理各時段資料的特性，進而將其傳遞至更深層網路進行計算，並有效學習水文反應，使其相較於其他兩種模式有更高的準確性。未來可根據本研究之模式搭配即時觀測系統，協助相關管理單位擬訂適當的防災策略。

關鍵字

淹水 ； 即時水位預報 ； 深度學習 ； 類神經網路 ； 支援向量機

並列摘要

Taiwan is located between tropical area and subtropical area. Recently, the increased amount of typhoon hitting Taiwan due to extreme climate has caused many disasters. Typhoons always bring considerable rainfall and cause serious floods in the downstream areas. This study develops a hourly water level forecasting model to prevent disasters.
The hourly water level forecasting model is developed based on dilated causal convolution neural network (CNN), which is a novel technology, with rainfall data and water level data during typhoon period as input. There are stacks of convolutions filters and dilated layer applied in the network to access a broad range of history when forecasting. This study also uses residual and skip connections to make the model deeper and more robust. In addition, the proposed model is compared with existing models based on the artificial neural network (ANN) and support vector machine (SVM) to demonstrate the improvements.
To demonstrate the effectiveness of the proposed model, an application to the Yilan river basin in northeastern Taiwan with 16 typhoon events from 2012 to 2017 is presented. The results show that ANN and SVM have reached the limits of forecasting at 3-h and 4-h lead time, respectively. However, dilated causal CNN is able to learn potential dependencies between different time series and get outcomes more precisely than ANN and SVM. This study also trains the dilated causal CNN with different factor as input. The outcomes show that even if there is only rainfall data as input, the model still get precise forecasts. In conclusion, the proposed modeling technique is expected to be useful for supporting disaster warning systems.

並列關鍵字

Flood ； Real time water level forecasting ； Deep learning ； Artificial neural network ； Support vector machine

參考文獻

1. Alipanahi, B., Delong, A., Weirauch, M.T., Frey, B.J., 2015. Predicting the sequence specificities of DNA- and RNA-binding proteins by deep learning. Nature Biotechnology 33, 831–838.

2. Assem, H., Ghariba, S., Makrai, G., Johnston, P., Gill, L., Pilla, F., 2017. Urban Water Flow and Water Level Prediction Based on Deep Learning. Machine Learning and Knowledge Discovery in Databases Lecture Notes in Computer Science 317–329.

3. Bai, Y., Chen, Z., Xie, J., Li, C., 2016. Daily reservoir inflow forecasting using multiscale deep feature learning with hybrid models. Journal of Hydrology 532, 193–206.

4. Bakshi, B.R., Stephanopoulos, G., 1993. Wave-net: a multiresolution, hierarchical neural network with localized learning. AIChE Journal 39, 57–81.

5. Bengio, Y., 2009. Learning Deep Architectures for AI. Foundations and Trends® in Machine Learning 2, 1–127.