探討物聯網感測器空間分布於人工智慧區域淹水預報模式之影響與修正策略

Investigating the Spatial Distribution of IoT Flood Sensor on the Effect of Artificial Intelligence Regional Flood Inundation Forecasting Model and Modified Strategies

楊舜年(Shun-Nien Yang)

指導教授 ： 張麗秋

淡江大學/工學院/水資源及環境工程學系博士班/博士(2020年)

https://doi.org/10.6846/TKU.2020.00887

摘要

全球各地環境與氣候逐漸惡化，發生諸多洪水事件，如2019年美國氣旋、非洲伊達強烈熱帶氣旋、威尼斯大洪水、臺灣0611、0702、0722、0813豪雨與2020年0519豪雨等災害事件，故世界各國投入大量資源，以期透過預警及預報系統，提前部屬相關防災措施與人員，降低極端降雨所造成之災害影響。
若採用二維淹水模擬模式進行即時區域淹水預報，則有許多受限之處，如水文預測具有不確定性、參數設定眾多、運算時間長等，故難以進行即時二維淹水預報，僅適用於災前整備階段使用。故本研究提出使用AI結合二維水理模式與IoT路面淹水感測器資料進行區域淹水預報，將二維水理模擬資料做為AI模式訓練資料，以RNARX模式進行未來1~3小時平均淹水深預報(時間分布)，SOM模式進行各神經元淹水分布(空間分布)，並透過SOM-RNARX模式以平均淹水深串聯RNARX與SOM模式，最終採用IoT感測器資料進行區域淹水修正，預報未來1~3小時區域網格淹水深。
結果顯示RNARX模式在資料傳輸皆為正常情況下，以RNARX 4A為最佳之模式；若感測器無回傳資料，RNARX 4B可做為無感測器資料之備用模式；若雨量站無回傳資料，RNARX 5A可做為無雨量資料之備用模式；若採用IoT虛擬感測器資料，則RNARX 6B為最佳之模式結果。
本研究另提出SOM模式參數設定建議，如初始半徑設定需依據鄰近區域形狀進行設定、迭代次數應避免超過3,000次、以SOM權重展開圖判定拓樸收斂性等，並提出神經元差異性評估指標判斷最佳拓樸架構，結果顯示SOM 5(5×5)為最佳之拓樸神經元架構。
SOM-RNARX模式結果顯示，增加RNARX精確度可提升區域淹水預報精確度，以SOM-RNARX 2模式最佳，但即使RNARX平均淹水深採用SOBEK模擬值，誤差來源仍有一部分為SOM之空間分布誤差，故提出採用IoT感測器資料進行區域淹水修正模式，以改善空間誤差分布。
IoT感測器資料除了實際淹水感測器25點外，另將淹水點進行K-Means分類，透過代表點分析演算法挑選出不同累積代表比例10~100%之虛擬IoT淹水感測器資料，以做為區域淹水修正資料使用，而修正方法分為徐昇氏控制面積修正法與高程差區域淹水修正法，結果顯示因徐昇氏控制面積修正法較適用於高程變化較小之區域，於本研究區域之改善結果較差，改善率為負值；高程差區域淹水修正法亦可分為水平線修正與固定量修正，結果顯示在修正高程為0.1公尺時，有最佳RMSE改善率；感測器資料若使用實際感測器25點，則以固定量修正有最佳結果；若使用虛擬IoT感測器資料，在累積比例為80%(188點)時，皆可達最佳之RMSE改善率。
透過本研究之完整流程，可有效預報未來1~3小時各網格區域淹水深，並使用IoT感測器資料進行修正，使預報結果能夠更加貼近現地淹水情況；另可根據代表性感測器分析法，提供未來淹水感測器設置點位建議，以達平時建置規劃、災前整備與災後評估之效用。

關鍵字

人工智慧 ； 類神經網路 ； RNARX ； SOM ； IoT即時淹水感測器 ； 區域淹水預報模式 ； 區域淹水修正法 ； 最具代表性感測器挑選

並列摘要

With the environment and the climate gradually worsen, a great number of flooding happened around the world, such as Hurricane Dorian in the USA, Tropical Cyclone Idai in the southeast Africa, Venice flooding and the torrential rain happening on June 11, July 2, July 22, August 13 2019 and May 24 2020 in Taiwan. Therefore, in the hope of deploying a precaution and crew against a disaster via forewarnings and forecasts in advance, every country devotes vast resources to reducing the influence caused by the extremely rainfall.
To forewarn and forecast floods, the official in Taiwan set out to produce the flood potential map in 2000. Since then, although the flood potential map has been updated to the third generation, it is merely applicable to pre-disaster preparedness due to the following reasons. Because of the uncertainty in hydrological prediction, the computational complexity and the enormous data used for the flood potential analysis, it is difficult to simulate and analyze the map online in real time. However, this study indicates that using artificial intelligence(AI) combined with the data from two-dimensional hydrological model and Internet of Things(IoT) flood inundation sensors forecasts regional flood. The AI model, including RNARX (Recurrent Nonlinear Autoregressive with Eogenous inputs), SOM (Self-Organizing Map) and SOM-RNARX model, are fed by the 2-D simulation data as the training data. The RNARX model is built to predict average inundation depth in 1 to 3-h-ahead (Time distribution) whereas the SOM network produces a regional flood topology map of the study area (Spatial distribution) through neurons. SOM-RNARX model connects SOM and RNARX with the average inundation depth. Finally, IoT sensors to revise the regional flood are adopted to forecast regional flood inundation depth of every grid.
The result shows that RNARX 4A is the optimum RNARX model under the condition which the data is transmitted normally. If the sensors do not return the data, RNARX 4B can be a standby model when there is no sensor data; If rain stations do not return the data, RNARX 5A can be a standby model when there is no rainfall data; If the data of IoT virtual sensors are adopted, RNARX 6B can achieve optimum result.
In addition, this study indicates the assessment index of difference of neurons, which can judge which size of neuron topology is the best architecture, and several suggestions to set parameters in the SOM network. For example, the initial radius has to be set according to the shape of neighbor area, the epochs should be avoided over 3,000 times, and SOM weights extension figure to judge the topological convergence. The result shows that SOM 5(5×5) is the best topology architecture of neuron.
The result of SOM-RNARX model shows that better RNARX accuracy can lead to better forecast accuracy of regional flood. Under this condition, SOM-RNARX 2 is the optimum model. Even if the RNARX model which predicts the average inundation depth achieves 100% accuracy, it still exists bias, some of which comes from bias of spatial distribution from the SOM network. Thus, IoT sensors to revise the regional flood are adopted to improve the bias of spatial distribution.
Except the 25 real flood inundation sensor points in the IoT sensor data, K-Means clustering is applied to classify all the other inundation points. The ratios of cumulative representative of the virtual IoT flood inundation sensor data ranging from 10 to 100% are selected by the representative point analysis algorithm, which is applied to revise the regional flood data by Thiessen's Control Area Revision Method and Regional Inundation Depth Revision Method of Elevation Difference. In this case, the improvement rate of Thiessen's Control Area Revision Method is negative in this study area, which means that it gets worse result, because it is more applicable to the area whose elevation hardly varies. As for Regional Inundation Depth Revision Method of Elevation Difference, including horizontal revision and certain amount revison, the result shows that it can achieve the best RMSE of the improvement rate under the condition that the revised elevation equals to 0.1 meter. If the real-sensor 25 points are utilized as the sensor data, certain amount revison can achieve the optimum result. Similarly, if the virtual IoT sensor data are used, it can also achieve the optimum RMSE of the improvement rate when the cumulative ratio equals to 80% (188 points).
In conclusion, it is effective to forecast the regional flood inundation depth of every grid with the compete process of this study. With the IoT sensor data revising the model, the forecast can be more close to real flooding situations. Moreover, this study also suggests the location where flood inundation sensors should be set in the future based on the representative sensor analysis algorithm, which is extremely effective on construction project in non-disaster period, pre-disaster preparedness and post-disaster needs assessment.

並列關鍵字

Artificial neural networks ； Recurrent nonlinear autoregressive with exogenous inputs (RNARX) ； Self-organizing map(SOM) ； Internet of Things(IoT) realtime flood inundation sensors ； Regional flood inundation forecasting model ； Regional flood inundation revision methd ； Representative sensors selected

參考文獻

1.Abou Rjeily, Y., Abbas, O., Sadek, M., Shahrour, I., Hage Chehade, F., 2017. Flood forecasting within urban drainage systems using NARX neural network. Water Science and Technology, 76(9), 2401-2412.

2.Ahmad, N., 2019. IOT Based Flood Alert Management System.

3.Amale, O., Patil, R., 2019. IOT Based Rainfall Monitoring System Using WSN Enabled Architecture. In 2019 3rd International Conference on Computing Methodologies and Communication (ICCMC) (pp. 789-791). IEEE.

4.Aziz, A., 2019. Coastal alerting IoT system in response to high tides and turbulent weather. In 2019 10th International Conference on Computing, Communication and Networking Technologies (ICCCNT) (pp. 1-7). IEEE.

5.Babu, V., Rajan, V., 2019. Flood and Earthquake Detection and Rescue Using IoT Technology. In 2019 International Conference on Communication and Electronics Systems (ICCES) (pp. 1256-1260). IEEE.