- 學位論文
整合智慧都市防洪系統水位預報及操作策略-以臺北市為案例
Integrating Water Level Forecasting and Operational Strategies of Intelligent Urban Flood Control Systems - A Case Study of Taipei City
若您是本文的作者，可授權文章由華藝線上圖書館中協助推廣。
摘要
近年來受氣候變遷影響，極端降雨事件明顯增加，提高都市排水系統負荷及淹水風險，因此如何有效管理都市下水道排水系統、降低淹水風險，並提高災害應變能力，已成為亟需重視問題。本研究選定臺北市中山抽水站集水區為研究區域，蒐集中山抽水站集水區之雨量、下水道水位、抽水站前池水位、外水水位及歷史操作紀錄資料，透過隨機搜尋(Random search)調整模式參數以建置10分鐘單位之水位預測模式，結合最佳化操作策略，整合為抽水機組操作策略模式。本研究分為水位預測及操作策略兩大部分，水位預測方面，建置卷積類神經網路混合倒傳遞類神經網路(CNN-BP)與長短期記憶類神經網路(LSTM)，預測未來10至40分鐘(T+1~T+4)之下水道水位，從兩種模式結果顯示CNN-BP為較佳之模式，故將CNN-BP模式預測未來一小時(T+1~T+6)之抽水站前池水位，同時亦使用CNN-BP模式預測未來一小時(T+1~T+6)之外水水位；於操作策略方面，建置非支配排序遺傳演算法Ⅲ(NSGA-Ⅲ)模式進行抽水站操作規則最佳化，以Entropy-TOPSIS選擇最佳之操作規則，並進行最佳化操作模擬；最後整合水位預測與NSGA-Ⅲ優選之操作規則模擬歷程作為調適性網路模糊推論系統(ANFIS)訓練樣本，以建立抽水機組操作策略模式，預測抽水機操作數量。結果顯示，CNN-BP模式預測全區域下水道水位監測站水位結果皆優於LSTM模式；透過相關性分析可降低CNN-BP模式預測前池水位之輸入項，於高水位事件預測中，T+1至T+3大致可良好掌握水位趨勢，足以作為抽水站操作參考依據；於預測外水水位方面，因水位振幅高度變動小，CNN-BP模式對於掌握外水水位變動性具有極高能力；透過NSGA-Ⅲ模式搜尋最小化風險值(Obj1)、最小化前後水位差總和(Obj2)、最小化抽水機開關次數(Obj3)三種目標函數，並進行最佳化操作模擬，結果顯示三種目標函數皆優於抽水站現有規則模擬，改善率分別為3.07%、47.15%及55.14%，最佳化操作模擬結果可大幅提升抽水站在風險、水位平穩性及抽水機操作次數三大面向結果；最後建置ANFIS模式依現實情況預測最佳化操作模擬之抽水機數量，透過相關性分析篩選ANFIS模式之輸入項，由預測結果得知ANFIS模式可大致掌握抽水機組操作趨勢。 本研究藉由結合水位預測模式及抽水站最佳化操作，透過ANFIS建置抽水機組操作策略模式，此模式具備即時性與可變動性，於颱風或豪雨期間，可迅速掌握雨量及整體防洪排水系統水位情況，即早提供抽水站管理單位有效之抽水機組操作參考，以增加災害預警與應變能力，實現智慧防洪管理。
並列摘要
In recent years, the impact of climate change has led to an increase in extreme rainfall events, significantly raising the load on urban drainage systems and the risk of flooding. Consequently, effectively managing urban sewer systems, reducing flood risks, and enhancing disaster response capabilities have become urgent issues. This study selects the Zhongshan Pumping Station catchment area in Taipei City as the research area, collecting data on rainfall, sewer water levels, internal and external water levels of the pumping station, and historical operation records of the Zhongshan Pumping Station. Multiple water level prediction models with a 10-minute interval are established by adjusting model parameters using Random Search. Combined with an optimized operation strategy, these models are integrated into a pumping unit operation strategy model. This study is divided into two main parts: water level prediction and operation strategy. For water level prediction, Convolutional Neural Networks combined with Back Propagation Neural Networks (CNN-BP) and Long Short-Term Memory Neural Networks (LSTM) are established to predict sewer water level for the next 10 to 40 minutes (T+1 to T+4). Results from both models show that CNN-BP is the superior model. Therefore, the CNN-BP model is used to predict the internal water level for the next hour (T+1 to T+6) and the external water level for the next hour (T+1 to T+6). For operation strategies, the Non-dominated Sorting Genetic Algorithm Ⅲ (NSGA-Ⅲ) is constructed to optimize the pumping station operation rules. The best operation rules are selected using Entropy-TOPSIS, and the optimized operation is simulated. Finally, the water level prediction and NSGA-Ⅲ optimized operation rules simulation process are integrated as training samples for the Adaptive Network-Based Fuzzy Inference System (ANFIS) to establish a pumping unit operation strategy model, predicting the number of pump operations. Results indicate that the CNN-BP model's prediction of sewer water level across all monitoring stations outperforms the LSTM model. Correlation analysis can reduce the input items of the CNN-BP model for predicting internal water levels. T+1 to T+3 can generally grasp the water level trends well enough to serve as a reference for pumping station operations in predicting high water level events. Due to the small amplitude of water level changes in predicting external water levels, the CNN-BP model can accurately capture external water level variability. Through NSGA-Ⅲ, three objective functions are optimized: minimizing risk value (Obj1), minimizing the total difference in front and back water level (Obj2), and minimizing the number of pump switches (Obj3). The optimized operation simulation results show that all three objective functions perform better than the existing pumping station rules simulation, with improvement rates of 3.07%, 47.15%, and 55.14%, respectively. The optimized operation simulation results significantly enhance the pumping station's performance in terms of risk reduction, water level stability, and pump operation frequency. Finally, the ANFIS model predicts the number of optimized pump operations based on real conditions, with input items selected through correlation analysis. The prediction results show that the ANFIS model can generally grasp the trend of pumping unit operations. This study combines water level prediction models and optimized pumping station operations to establish a pumping unit operation strategy model through ANFIS. This real-time and adaptive model enables quick monitoring of rainfall and overall flood drainage system water level during typhoons or heavy rain. It provides effective pump operation references to the pumping station management unit, enhancing disaster warning and response capabilities, and realizing smart flood management.
參考文獻
1. Ambadekar, P. K., & Choudhari, C. M. (2020). CNN based tool monitoring system to predict life of cutting tool. SN Applied Sciences, 2, 1-11.
2. Awodi, N. J., Liu, Y. K., Ayo-Imoru, R. M., & Ayodeji, A. (2023). Fuzzy TOPSIS-based risk assessment model for effective nuclear decommissioning risk management. Progress in Nuclear Energy, 155, 104524.
3. Bhandari, H. N., Rimal, B., Pokhrel, N. R., Rimal, R., Dahal, K. R., & Khatri, R. K. (2022). Predicting stock market index using LSTM. Machine Learning with Applications, 9, 100320.
4. Bi, H., Huang, L., Zhang, H., Xie, S., Zhou, L., & Liu, L. (2024). A deep learning-based model for tropospheric wet delay prediction based on multi-layer 1D convolution neural network. Advances in Space Research.
5. Cai, B., Pan, G. L., & Fu, F. (2020). Prediction of the postfire flexural capacity of RC beam using GA-BPNN machine learning. Journal of Performance of Constructed Facilities, 34(6), 04020105.