本研究旨在發展適用於臺灣之深度學習地震預警模型，將既有以單一地動指標為主之預估架構，推進至可同時預估最大地表加速度（PGA）與最大地表速度（PGV）之雙指標預警框架。研究資料採用台灣強地動觀測網（TSMIP）於 1999 至 2019 年間收錄之地震紀錄。本文首先延伸 Taiwan Transformer Shaking Alert Model（TT-SAM）架構，建立以速度波形、低頻速度波形與物理特徵為輸入之 PGV 預估模型；進一步提出 Seismic Acceleration and Velocity Alerting Neuralnet Transformer Model（SAVANT），透過並行卷積分支擷取加速度波形、速度相關波形與物理特徵資訊，結合測站位置、場址效應參數與 Transformer Encoder 建立多測站空間關係，並以雙混合密度網路同時輸出 PGA 與 PGV。為改善高震度樣本稀缺造成之低估問題，本研究導入以震度 4 級為門檻之 MSFE 損失函數，使高、低震度樣本於訓練過程中具有較均衡之影響力。結果顯示，SAVANT 在 7 秒輸入下之 PGA 與 PGV 預估 R² 分別達 0.708 與 0.805；隨著輸入波形增加至 13 秒，R² 進一步提升至 0.752 與 0.828。；於震度 4 級警報門檻下，7 秒輸入之 PGA 與 PGV 判定 Recall 分別為 0.81 與 0.69，顯示模型於地震早期已具備辨識潛在警報區域之能力。2016 年美濃地震與台東外海地震案例亦顯示，SAVANT 可於 P 波抵達第一個測站後 3 秒掌握主要強地動分布，並透過滾動式更新逐步修正預估結果與補足警報範圍。綜合而言，SAVANT 透過雙指標預估、物理特徵整合與不平衡資料處理，提升深度學習模型於臺灣地震預警應用中之穩定性與實務潛力。;This study develops a deep learning-based earthquake early warning model for Taiwan， extending existing single-output frameworks toward a dual-output system that simultaneously predicts peak ground acceleration (PGA) and peak ground velocity (PGV). Using seismic records from the Taiwan Strong Motion Instrumentation Program (TSMIP) from 1999 to 2019， this study first extends the Taiwan Transformer Shaking Alert Model (TT-SAM) to construct a PGV prediction model， and further proposes SAVANT (Seismic Acceleration and Velocity Alerting Neuralnet Transformer Model). SAVANT uses parallel convolutional branches to extract features from acceleration waveforms， velocity-related waveforms， and physical features， and integrates station information， VS30， and a Transformer Encoder to model multi-station spatial relationships. Dual mixture density networks are used to simultaneously output PGA and PGV predictions. To mitigate underestimation caused by scarce high-intensity samples， an MSFE loss function with intensity level 4 as the threshold is introduced. With a 7-s input window， SAVANT achieves R² values of 0.708 and 0.805 for PGA and PGV， respectively. At the intensity level 4 warning threshold， the recall values are 0.81 and 0.69 for PGA and PGV， respectively. Case studies of the 2016 Meinong and Taitung offshore earthquakes show that SAVANT can capture the main strong-motion distribution at an early stage and progressively refine warning areas through rolling-update predictions. Overall， SAVANT improves the stability and practical potential of deep learning-based earthquake early warning by integrating dual-output prediction， physical features， and imbalance-aware training.
