結合自編碼器及回饋式類神經網路建立區域淹水預測模式之研究

A Study of Building Regional Flood Inundation Forecast Models by Hybrid Autoencoder and Recurrent Neural Network

高毅灃(I-Feng Kao)

指導教授 ： 張斐章

國立臺灣大學/生物資源暨農學院/生物環境系統工程學研究所/博士(2020年)

https://doi.org/10.6342/NTU202003855

摘要

淹水預報為防救災人員之重要參考資訊，惟區域淹水機制非常複雜且數據維度極大，建立區域淹水預報係一個極具挑戰的關鍵議題， 過去有學者使用聚類的方式克服區域淹水數據維度過高的問題，但對區域淹水特徵萃取問題尚少討論。本研究以前人的研究為基礎，思考如何藉由特徵萃取大幅降低高維度之區域淹水資料以減少模式複雜度，提出結合堆疊自編碼器(stacked autoencoder, SAE)及回饋式類神經網路(recurrent neural network, RNN)之區域淹水預報模式。本研究提出之SAE-RNN模式利用SAE大幅降低區域淹水資料之維度，過程中輔以主成分分析(principal components analysis, PCA)決定每次SAE堆疊過程的隱藏層神經元個數及初始化其神經元權重，完成4個SAE模式訓練後使169,797個網格資料降維成2至5維度淹水特徵編碼；淹水預報部分係使用以編碼器-解碼器(encoder-decoder)框架及長短期記憶體(long short-term memory, LSTM)為基礎之回饋式類神經網路(RNN)模式建立淹水特徵編碼預測模式；模式以前24小時雨量為輸入項，輸出未來3小時區域淹水特徵編碼預報結果，再透過SAE中的解碼器(decoder)將預測之特徵編碼還原成區域淹水網格淹水深，完成SAE-RNN模式淹水預測。本研究以宜蘭地區設計降雨及颱風暴雨紀錄所模擬之55場淹水事件數據為SAE-RNN模式訓練、驗證及測試資料，建立4個使用不同維度淹水特徵編碼之模式，預測未來3小時區域淹水深之測試階段RMSE值小於0.09m，R2可達0.95以上；區域網格淹水預測結果MAE分布圖顯示絕對誤差小於0.1m之區域占全區96%以上。根據不同SAE-RNN模式測試結果，發現影響SAE-RNN模式準確度的主因為SAE模式降維還原過程的誤差，次因為RNN產生的預測誤差。綜合考量各模式各項評估指標後以使用4維淹水特徵編碼的SAE-RNN模式為本研究最佳區域淹水預報模式。此外，將不同區域淹水事件資料以SAE降維成二維淹水特徵編碼，依時序繪製連線於平面座標上，再配合將該平面固定間格取樣之網格座標點以SAE還原成區域淹水圖，使每個平面座標網格點可關聯至特定區域淹水分布，完成可於平面圖中掌握多場事件淹水歷程時空變化情形之視覺化圖表，同時也能顯示不同雨型對淹水歷程變化之影響。

關鍵字

區域淹水 ； 降維度 ； 自編碼器 ； 回饋式類神經網路 ； 回饋式類神經網路、長短期記憶體 ； 資料視覺化

並列摘要

Flood forecasting is essential information in disaster management. Because the regional flooding mechanism is very complicated, it is challenging to establish a regional flood forecast. Previous studies have used cluster analysis to overcome the problem that building a high-dimensionality regional flooding data forecasting model is difficult. However, there is an insufficient discussion on the characteristic extraction of regional flooding data. This study focuses on reducing the dimensionality of regional flooding data by feature extraction, and build regional flood inundation forecast models by hybrid Stacked autoencoder (SAE) and Recurrent neural network (RNN). During the SAE training process, this study uses Principal components analysis (PCA) to determine the number of hidden layer neurons and provide initial values of weights. In the case study of 55 flooding event physical simulation data in Yilan, 4 SAE models were completed to reduce the dimension of the regional flooding grid data from 169,797 to 5, 4, 3, or 2 and convert regional flooding data into four different dimensions of flooding characteristic codes. For forecasting the depth of regional flooding, this study uses the encoder-decoder framework and Long short-term memory (LSTM) to establish a forecast RNN model to predict the flooding characteristic codes in the future 3 hours. The input sequence of this RNN model is the rainfall information of the previous 24 hours. Finally, the SAE and RNN models are combined into SAE-RNN model, and the predicted flooding characteristic code is restored to the predicted regional flooding depth. The result of forecasting regional flood inundation shown that the RMSE less than 0.09m and the R2 more than 0.95 in the testing stage. The error distribution map of the forecast area shown that the MAE in 96% of the area is less than 0.1m. According to the results of different SAE-RNN models, this study concludes that the main factor affecting the SAE-RNN model accuracy is the restoration error of SAE, the second factor is the RNN model error of forecasting flooding characteristic codes, and the SAE-RNN model with 4-dimension flooding characteristic codes is the best regional flood inundation forecast model in this case study. In addition, 2-dimension flooding characteristic codes provide a visual effect of the time and space distribution of the flooding process by line chart in the 2-dimensional coordinate plane. In this study, the decoder of SAE is used to restore the flooding characteristic codes of any coordinate as a flood map in 2-dimensional coordinate plane, which correlated the visibility between the flooding process and regional flood inundation depth distribution in the study area.

並列關鍵字

Regional flood ； Dimension reduction ； Autoencoder ； Autoencoder, Recurrent neural network ； Long short-term memory ； Data visualization

參考文獻

1. Audhkhasi, K., Rosenberg, A., Sethy, A., Ramabhadran, B., Kingsbury, B. (2017). End-to-end ASR-free keyword search from speech. IEEE Journal of Selected Topics in Signal Processing, 11(8), 1351-1359.

2. Asadi, A., Safabakhsh, R. (2020). The encoder-decoder framework and its applications. In Deep Learning: Concepts and Architectures (pp. 133-167). Springer, Cham.

3. Badrzadeh, H., Sarukkalige, R., Jayawardena, A. W. (2015). Hourly runoff forecasting for flood risk management: Application of various computational intelligence models. Journal of Hydrology, 529, 1633-1643.

4. Bengio, S., Vinyals, O., Jaitly, N., Shazeer, N. (2015). Scheduled sampling for sequence prediction with recurrent neural networks. In Advances in Neural Information Processing Systems (pp. 1171-1179).

5. Berkhahn, S., Fuchs, L., Neuweiler, I. (2019). An ensemble neural network model for real-time prediction of urban floods. Journal of hydrology, 575, 743-754.