本研究聚焦於智慧電網場域中「長短時段用電趨勢預測」與「異常風險預警」的同步需求，提出一套雙分支深度架構。第一分支將一維負載序列透過 Gramian Angular Summation Field（GASF）轉為影像特徵，並與原始序列在 Time-Aware Transformer 中以門控機制動態融合（Gated Feature Fusion），成功降低極短時間與長時間預測的誤差累積；第二分支利用重取樣的資料處理與 Focal Loss 訓練之 Bi-LSTM，在保持低誤報率前提下，對未來每一分鐘異常事件給出機率預警。兩分支模型推論後，並以 TimeSHAP 生成時間步與特徵層面的可解釋性。實驗顯示，所提方法在 ETT 系列四個資料集的 96、192、336 步視窗皆取得最低 MSE；異常分支於 Server Machine Dataset 測試集達到 0.885 PR-AUC 與 0.953 F1，證明本研究模型設計能兼顧精準度、即時性與可解釋性，為能源管理與工業監控提供一體化解決方案。;This study addresses the dual need in smart-grid operations for both short- and long-horizon load-trend forecasting and minute-level anomaly‐risk early warning. We propose a two-branch deep architecture. The forecasting branch converts one-dimensional load sequences into Gramian Angular Summation Field (GASF) images and fuses them with the raw series through a Gated Feature Fusion mechanism inside a Time-Aware Transformer, markedly reducing error accumulation over very short to very long prediction windows. The warning branch applies data re-sampling and Focal-Loss-trained Bi-LSTM to output minute-ahead anomaly probabilities while maintaining a low false-alarm rate. After inference, both branches are explained with TimeSHAP, providing step-wise and feature-level interpretability. Experiments on four ETT benchmark datasets show the proposed model achieves the lowest MSE at 96-, 192- and 336-step horizons; on the Server Machine dataset the anomaly branch attains a PR-AUC of 0.885 and an F1 score of 0.953. These results demonstrate that the framework delivers accuracy, real-time capability and interpretability in a single solution, making it well-suited to energy management and industrial monitoring tasks.
