水庫下游水位預報之研究

Abstract

近年來短延時降雨頻率和強度增加，而水庫常採用高水操作，導致水庫及下游河岸預警系統應變時間縮短。現行的水位預報時間僅未來6小時，若在傍晚水位逼近警戒值，主管機關無法決策河川堤防橫移門內是否移車，導致作業困難。本研究提出一長延時水庫下游水位預報模式，提供下游水位預報給防災單位參考。 本研究以台北橋水位作為預報目標，蒐集2014年至2021年之颱風和暴雨事件資料。台北橋上游集水區降雨量、石門水庫放水量、石門水庫入流量和翡翠水庫放水量和淡水河口潮位作為模式之備選因子。以四種人工智慧 (Artificial intelligence, AI)方法：支援向量機 (Support vector machine, SVM)、長短期記憶網路 (Long Short-Term Memory, LSTM)、雙向長短期記憶網路 (Bidirectional Long Short-Term Memory, BiLSTM)及序列到序列 (Sequence-to-Sequence, Seq2Seq)，預報未來1小時水位，並以網格搜尋法篩選出各因子之輸入步長和各模式之超參數。為達到長延時預報，將四種模式分別搭配多步階預報，建立長延時水位預報模式，採用評鑑指標評估模式表現。結果顯示Seq2Seq預報模式能準確預報至未來24小時，其CC值皆大於0.9，最高可達0.98，RMSE值於0.16 公尺至0.33 公尺之間，MAE值皆小於0.3公尺，而CE值大部分達0.85。 為驗證本研究預報模式於實際應用之可行性，採上述最佳模式Seq2Seq，介接氣象局定量降水預報，以及經入流量預報轉換之出流量預報，以多步階預報產出未來24小時之台北橋水位預報。結果顯示本研究所提出之水位預報模式介接雨量預報和出流量預報後，以預報未來12小時水位最符合預警的應用；模式能提供準確之長延時預報，其CC值皆大於0.94，最高可達0.97，RMSE值於0.17 公尺至0.32 公尺之間，MAE值皆小於0.3公尺，而CE值最高達0.89。 本研究提出之Seq2Seq預報模式，在後續的實際應用，搭配即時定量降水預報和警戒水位，可提供未來12小時準確的警戒時段以及最大峰值到達時間，大幅提升即時水位預警系統之準確度。能更有效地操作橫移門和疏散門，以利大台北地區能夠於警戒水位前提早疏散附近民眾及停駐於高灘地的車輛，降低損失生命財產的風險。

The response time for the early warning system of reservoirs and downstream riverbanks has been shortened due to higher frequency and greater intensity of short-duration rainfall events in recent years. The existing water level forecasting methods are limited to 6-hour ahead forecasts. Therefore, this study proposes a long-term water level forecasting model for downstream areas. In this study, the water level of Taipei Bridge has been chosen as the target variable. Data from typhoon and storm events between 2014 and 2021 have been collected. The dataset includes various factors such as precipitation in the upstream watershed of Taipei Bridge, outflow discharge of Shimen Reservoir, inflow discharge of Shimen Reservoir, outflow discharge of Feitsui Reservoir, and tidal measurements at the Tamsui River estuary. . Subsequently, this study uses four artificial intelligence methods, namely Support Vector Machine (SVM), Long Short-Term Memory (LSTM), Bidirectional Long Short-Term Memory (BiLSTM), and Sequence-to-Sequence (Seq2Seq), for predicting 1-hour-ahead water levels. The grid search method is employed to determine the optimal input step size for each factor and to tune the hyperparameters of each model. In order to achieve long-term forecasting, the four models are combined with multi-step forecasting to establish long-term water level forecasting models, and the evaluation indexes are used to evaluate the model performance. The results demonstrate that the Seq2Seq model outperforms the other models, accurately forecasting the water levels for the next 24 hours, with correlation coefficient (CC) exceeding 0.9 (up to 0.984), root mean square error (RMSE) ranging from 0.161 m to 0.332 m, mean absolute error (MAE) below 0.3 meters, and coefficient of efficiency (CE) reach 0.85. These indexes indicate the high accuracy and reliability of the Seq2Seq model in long-term water level forecasting. This study integrates the Seq2Seq model with quantitative precipitation and converted outflow forecasts to predict 12-hour ahead water levels at Taipei Bridge. The results show strong accuracy, with CC exceeding 0.94 (up to 0.97), RMSE values ranging from 0.17 m to 0.32 m, MAE values below 0.3 m, and a high CE value of 0.89. This highlights the feasibility and effectiveness of the proposed forecasting model in practical applications.