在近年氣候變遷日益加劇的背景下，環境極端事件的發生頻率與嚴重程度持續上升，使得建立可靠的時空預測架構以進行極端事件分析成為重要課題。雖然 PoT-GEV 模型能有效利用區塊極大值資料，同時描述極端事件的發生頻率與強度，但其實際分析效果仍可能受到空間異質性、時間變異性以及閾值設定等因素的影響。為了克服上述問題，本研究基於 Peng et al. (2025) 所提出之空間階層式 PoT-GEV（SH-PGEV）模型，建立一套時空預測架構。此架構透過潛在空間高斯過程描述空間相依性，並進一步結合自我迴歸時間模型與空間平滑技巧，以提升環境極端事件的預測能力。此外，本研究亦透過模擬實驗系統性探討閾值設定對參數估計與預測準確性的影響。模擬結果顯示，適當的閾值選取策略不僅能有效提升參數估計準確性，亦能穩定時空預測表現；同時，所提出之方法能有效捕捉 SH-PGEV 模型中頻率參數與強度參數的時間演變與空間變異特性。在實證分析方面，本研究以台灣年最大降雨資料為例，建立一套針對極端降雨事件的滾動式時空預測流程。透過結合自我迴歸預測與簡單克利金方法，本研究得以對未來期間的重現值與超越機率進行空間預測。將預測結果與實際觀測值比較後發現，納入滾動更新訓練資料能顯著提升預測準確性，並增強模型對氣候變化的適應能力。綜合而言，本研究所提出之方法提供了一套具彈性且計算效率良好的時空極值分析架構。;In recent years， the increasing frequency and severity of environmental extremes under climate change have highlighted the importance of developing reliable spatiotemporal prediction frameworks for extreme-event analysis. Although the PoT-GEV framework provides an effective approach for jointly characterizing the intensity and frequency of extremes using block maxima data， its practical performance can be substantially affected by spatial heterogeneity， temporal variation， and threshold specification. To address these challenges， this study develops a spatiotemporal prediction framework based on the spatial hierarchical PoT-GEV (SH-PGEV) model proposed by Peng et al. (2025). The proposed framework incorporates latent spatial Gaussian processes to account for spatial dependence and further integrates autoregressive temporal modeling with spatial smoothing techniques to improve predictive performance for environmental extremes. In addition， the effects of threshold specification on parameter estimation and forecasting accuracy are systematically investigated through simulation studies. The simulation results demonstrate that an appropriate threshold selection strategy can substantially improve parameter recovery and stabilize spatiotemporal prediction performance. Furthermore， the proposed framework effectively captures the temporal evolution and spatial variability of the frequency and intensity parameters in the SH-PGEV model. For empirical analysis， annual maximum rainfall data in Taiwan are analyzed to construct a rolling spatiotemporal forecasting procedure for extreme rainfall events. By combining autoregressive prediction with simple kriging， the proposed method generates spatial forecasts of return levels and exceedance probabilities over future periods. Comparisons between predicted and observed extremes indicate that incorporating rolling-update training data significantly improves prediction accuracy and enhances the model’s adaptability to evolving climate conditions. Overall， the proposed framework provides a flexible and computationally feasible approach for spatiotemporal extreme-value analysis.
