圖書典藏及影音數位平台

北區水資源分署自2013年起委辦「石門水庫防洪、排淤及供水運轉系統建置及運轉操作諮詢」迄今，因應氣候變遷下極端降雨頻發之挑戰，以石門水庫逾六十年累積之豐富歷史資料為基礎，在AI大數據分析之創新概念與傳統水文模式結合的架構上，導入機器學習與深度學習之長時距入流預報模式（如RNARX-Storage混合模式），並規劃運用生成式AI 大型語言模型（LLM）建置自然語言決策問答與智慧化操作介面，形成「AI 入流預報＋傳統水文模式＋生成式 AI 決策輔助」之整體架構。本系統採用SQL Server關聯式資料庫管理系統，資料庫中已完整建置166場颱風事件與66場豪雨事件之水庫操作與水文資料；通過擴充LINE服務，新增歷史颱風事件各出水工放流量資訊之查詢功能，便利使用者取得石門集水區歷史颱風事件相關資訊；為強化石門水庫在防洪應變上的即時應對能力與決策效率，「石門水庫防洪運轉操作決策支援系統」導入智慧導引式介面設計與AI模式推薦機制，結合微互動視覺設計與模組化功能整合，提升操作效率與決策產製速度，同時預留AI加值應用擴充空間，並保有新舊介面自由切換之彈性。在網頁端呈現，本計畫將原「PHP＋資料庫直讀」之監看頁重構為前端決策儀表板，採用 JAMstack 概念強化系統安全性與維護性，並於頁面導入淺色／深色雙主題、卡片化資訊展示、一鍵更新與畫面更新時間標示，以兼顧實務操作便利性、後續稽核與檢討需求；預報模式面則針對上升段與退水段之全洪程入流量預報模組，提出RNARX 機器學習與貯蓄函數模式之參數建議與適用條件，提供自動化判定與切換退水模組與既有模組使用時機；並依據豪雨事件特性建置長時距入流預報混合模式，作為水庫操作策略擬定之參考依據。今(2025)年度提供202518樺加沙颱風、202526鳳凰與20251020豪雨等3場事件之專職小組操作協勤與諮詢作業，並依據協勤過程之回饋持續精進石門水庫防洪決策支援系統相關功能，另於執行期間完成1場計畫內容與系統操作之教育訓練，強化分署同仁水庫防洪操作與AI應用知能。

Since 2013, the Northern Region Water Resources Branch (Water Resources Agency) has continuously commissioned the “Shihmen Reservoir Flood Control, Sediment Flushing, and Water Supply Operation System Development and Operational Consulting” project. In response to the growing challenges of extreme rainfall under climate change, this project builds on over sixty years of historical data from Shihmen Reservoir and integrates innovative AI and big-data analytics with conventional hydrological models. Machine learning and deep learning–based long-lead inflow forecasting models, such as the RNARX–Storage hybrid model, are introduced, and the use of generative AI large language models (LLMs) is planned to develop a natural-language decision-support interface. Together, these components form an integrated framework of “AI-based inflow forecasting + traditional hydrological modeling + generative AI–assisted decision support.”The system adopts Microsoft SQL Server as the relational database management system and currently stores complete operational and hydrometeorological records for 166 typhoon events and 66 heavy-rain events. LINE services have been enhanced to include on-demand queries of historical typhoon events, including discharge hydrographs for each outlet structure, enabling users to quickly retrieve past event information for the Shihmen Reservoir watershed. To further strengthen the timeliness and effectiveness of flood-response decision making, the Shihmen Reservoir Flood Control Operation Decision Support System incorporates an intelligent guided user interface and an AI-based model recommendation mechanism. Stepwise guidance reduces the risk of misoperation, while micro-interaction visual design and modular function integration enhance operational efficiency and the speed of generating decision products. The interface retains the flexibility to switch between the legacy and new layouts and reserves space for future AI-added-value applications.On the web front end, the original “PHP + direct database access” monitoring page has been refactored into a decision dashboard based on the JAMstack concept, improving system security and maintainability. The pages adopt dual light/dark themes, card-style information presentation, one-click data refresh, and explicit timestamps to balance operational convenience with future auditing and review needs. In terms of modeling, the project refines full-hydrograph inflow forecasting by providing parameter recommendations and applicability conditions for combining RNARX machine-learning inflow forecasting with storage-function models, enabling automated determination of when to switch between rising-limb and recession-limb modules. For heavy-rain events, a long-lead hybrid inflow forecasting model is developed to support the formulation of reservoir operation strategies.In 2025, the project team provided real-time operational support and consulting for three events—Typhoons Ragasa (202518) and Fung-Wong (202526) and a 20251020 heavy-rain event—supplying comprehensive inflow, rainfall, and operation information through the decision-support system. Based on feedback from these support operations, the flood control decision-support functions were further refined. In addition, one training session on project content and system operation was conducted to strengthen the Northern Region Water Resources Branch (Water Resources Agency) personnel’s capabilities in the reservoir flood-operations and AI-applications.