台灣在夏秋季常受颱風侵襲，且因颱風風力強力吹襲海面其造成的波浪比一般季節風所造成波浪大、破壞亦大，因此民眾對颱風風力和近岸風浪預報資訊的需求亦日益提升。傳統上，颱風風場和風浪模擬預測為採用數值氣候模式和數值波浪模式。數值模式優點是預報過程為由物理關係解釋，然而模擬過程較費時無法提供即時性；而近來廣受提及的人工智慧則可讓計算效率提升，然而缺點是較少物理關係描述，因此，本計畫構想為希望能結合兩者優點以建立颱風浪模擬模式。本計畫目的為發展一新型數值智慧風浪預測模式，此整合模式技術上包含數值天氣預測模式(WRF)、波浪數值模式(SWAN/WWIII)和人工智慧的機器/深度學習法(如DNN、TDRNN、MLP)。許多研究指出風場初始值對數值模式模擬結果的好壞影響甚鉅，特別是模擬颱風時期台灣鄰近海域風浪時，颱風環流因受陸地地形影響，使得風場狀況變的複雜。模組化過程有四個階段，第一階段採用WRF模式進行風場模擬；第二階段將WRF風場模擬結果成為機器學習的學習目標並建立風場預測模式(NUM-AI_WIND)；第三階段使用WRF風場模擬作為初始值，進行SWAN/WWIII模式模擬近岸風浪；第四階段使用SWAN/WWIII的波浪模擬作為機器學習的學習目標並建立波浪預測模式(NUM-AI_WAVE)。研究區域為台灣東北海域，模擬期距為1小時，預測長度以1至12小時為目標。本計畫在整合了上述各階段模組後，將設計了一套即時風浪預測的流程，並測試幾場颱風事件以驗證成效。即時風浪預測過程中，主要是以上述所研發的NUM-AI_WIND、NUM-AI_WAVE模式和即時獲得的觀測數據進行推算，以此期望加速運算效率並獲致合理可接受的解答。特別說明的是，WRF和SWAN/WWIII模式並不會在即時預測過程中需要進行模擬運算，而是在平常時期進行推算波浪以供給機器/深度學習時訓練之用。本計畫將以近十年的颱風場次作為建模數據。在資料部分，本計畫首先收集NOAA National Center for Environmental Prediction (NCEP)的FNL全球模式分析場作為WRF模式的初始場。另外，將收集JTWC的颱風預測路徑、氣象局的地面測站的氣候參數和浮標的海氣象參數等。本計畫期望以實際案例進行分析和探討，使學術理論能進一步活用於實務面上，以開拓解決颱風風浪即時短時距預測上的新技能。 Strong typhoon winds blow over the sea and land, and produce large waves with enough energy to influence marine structures, erode beaches, and buildings. Therefore, a useful scheme for wind field and wind-wave forecasts during typhoons is highly desirable for Taiwan. Traditionally, numerical models, based on the physical meanings, are applied to simulate the wave filed in the ocean. However, the numerical model, which involve time-consuming calculations, might not be able to satisfy the needs of real-time operations. By contrast, artificial intelligence (AI), powerful machine learning techniques, can be used to forecast the behavior of complex systems. Then, they achieve accurate and fast forecasts with low modeling costs. However, these techniques might be lack of physical meanings.The purpose of this study is to develop an integrated numerical-AI-based model to predict wind field and wind-wave field offshore of Taiwan during typhoons. The proposed integrated models employ a combined wind and wave numerical models and AI data-driven techniques. First, the WRF (Weather Research and Forecasting) numerical model is employed as the wind simulation model to generate the wind fields. Then, the high reliability wave field simulations using WRF-based wind field results are obtained by SWAN or WAVEWATCH III (WWIII) simulation models. Then, the both wind and wave field simulation results will be served as inputs and targets to generate the proposed AI-based wind-wave model. The main four stages of the proposed integrated model use models, techniques and their relationships are briefly described as: 1) at 1st stage, the WRF numerical model is employed as the numerical wind simulation model; 2) at 2nd stage, the numerical-AI-based wind prediction model (shorten as NUM-AI_WIND) is developed and used to prediction the wind field of the studied region; 3) at 3rd stage, the SWAN or WWIII numerical models are used to simulate the wave during typhoons; and 4) at 4th stage, we develop the numerical-AI-based wave prediction model (shorten as NUM-AI_WAVE), which is used to prediction the wave field of the studied region.The experimental area near Northeast Coast of Taiwan is used as a study case. This study will include typhoon events over the past 10 years or more. The time interval is one hour and forecasted horizon is 12 hours. We design a real-time prediction processes of the wind-wave predictions using the proposed integrated model. Note that, when the “on-line” predict the wind-wave amounts, we employ the two AI-based models (i.e., NUM-AI_WIND and NUM-AI_WAVE) in order to speed up the computational time during real-time operational procedure (e.g., within 1 hour). That is to say, the simulations regarding the WRF wind numerical model and SWAN/WWIII wave numerical model are not executed in the stage of processing the real-time prediction procedures (because these numerical simulations are executed when modeling the NUM-AI_WIND and NUM-AI_WAVE models in the peace time). Regarding a complex typhoon system, the collected data comprise: 1) FNL (Final) operational global analysis data for WRF model, 2) typhoon tracks and their characteristics from JTWC, and 3) the buoy atmospheric properties and ground weather data from CWB.