基於深度學習之臺北市緊急醫療服務需求之時空網格化機率預測：以DeepAR與TFT為例

Abstract

緊急醫療服務（Emergency Medical Services, EMS）需求具高度隨機性與時空異質性，若能準確掌握需求於不同時間與空間單元的變化，將有助於救護資源之前置部署與動態調度。本研究以臺北市為研究區，建構「時空網格化 × 機率式多步預測」之 EMS 需求預測框架，目的在於高解析度情境下同時提供點預測與不確定性資訊，以支援風險導向之資源規劃。 本研究以規則網格（1,000m × 1,000m）為空間單元，並以 4 小時為一時段進行時間切分，彙整各網格之需求量形成多序列時間序列資料；整合時間日曆特徵、氣象（降雨、氣溫）與人口結構（總人口、高齡人口）等共變數，其中降雨以 Kriging 內插至網格尺度，人口則由行政區轉換至網格以維持尺度一致。資料期間選取民國 107、108、112、113 年以降低非典型事件干擾。 模型採用 DeepAR 與 Temporal Fusion Transformer（TFT）進行機率預測，並以 MAE、RMSE、預測區間覆蓋率（PICP@80%）衡量準確性與覆蓋表現；另納入容忍誤差率（TRE，τ=±1）評估可操作性，並以分位數校準曲線檢核分位數輸出之校準程度。本研究貢獻在於建立高解析度網格需求之機率預測流程與多源資料整合方法，並比較 DeepAR 與 TFT 之系統性權衡，提供救護部署與決策支援之實證基礎。

Emergency Medical Services (EMS) demand is highly stochastic and spatiotemporally heterogeneous. Accurate forecasting can support ambulance pre-positioning and dynamic dispatching. This study proposes a grid-based probabilistic multi-horizon forecasting framework for EMS demand in Taipei City, providing both point forecasts and uncertainty information for risk-aware planning. Taipei is partitioned into regular 1,000 m × 1,000 m grids, with demand aggregated into 4-hour intervals to form multiple time series. Covariates include calendar/time features, meteorological variables (rainfall and temperature), and demographics (total and older population). Station rainfall is interpolated to grids using Kriging, and administrative-area population is converted to grids to ensure spatial consistency. Data from 2018, 2019, 2023, and 2024 (ROC years 107, 108, 112, 113) are used to reduce atypical effects. We implement DeepAR and the Temporal Fusion Transformer (TFT) for probabilistic multi-horizon forecasting. Performance is evaluated by MAE, RMSE, and the 80% prediction interval coverage probability (PICP@80%), together with the tolerant rate error (TRE, τ=±1) to reflect operational usability. We further assess probabilistic reliability using quantile calibration curves. The results provide a reproducible high-resolution forecasting pipeline and an empirical comparison of DeepAR and TFT to inform EMS deployment and decision support.