近年來，再生能源併網比例持續提升，其間歇性與不確定性使電力負載呈現更高的波動性與非平穩性，對電網調度與用戶端能源管理提出更嚴格的預測精度要求。傳統加權最小平方法（WLS）的固定平滑參數無法即時因應不同時段的負載動態，深度學習方法在負載模式劇烈變化時，預測穩定性仍有提升空間。 本研究提出梯度自適應混合負載預測模型（GRU-GA-WLS），以兩階段殘差架構整合統計平滑方法與深度學習的優勢。第一階段引入梯度自適應加權最小平方法（GA-WLS），依負載相鄰時間步的差分調整平滑權重，並結合高斯混合模型（GMM）自動識別場域與季節的作息時段邊界，建立場域與季節雙重自適應的動態週期基線；第二 階段以整合可逆實例正規化（RevIN）與自注意力機制的GRU殘差預測網路，學習基線解釋後剩餘的非線性殘差，採用貝葉斯優化進行模型配置搜尋。本研究以日間高峰規律型場域MT_200（Case1）與夜間高峰動態型場域MT_162（Case2）為驗證對象，涵蓋15分鐘、30分鐘、60分鐘、未來24小時與未來七天的預測視野，並區分夏季與非夏季進行系統性評估。實驗結果顯示，GRU-GA-WLS在兩場域所有預測視野與季節情境下均取得最低MAPE，在動態型場域MT_162（Case2）的實驗結果中，顯示GA-WLS梯度自適應機制對劇烈負載轉折的追蹤能力；在規律型場域MT_200（Case1）的實驗結果中，顯示模型在不同負載特性場域下均具備穩定且優於基準方法的預測能力。;In recent years， the increasing penetration of renewable energy into power grids has introduced greater intermittency and uncertainty， causing electricity load to exhibit higher volatility and non-stationarity， which imposes stricter requirements on forecasting accuracy for both grid dispatch and demand-side energy management. Conventional Weighted Least Squares (WLS) methods rely on fixed smoothing parameters and therefore cannot adapt in real time to varying load dynamics across different periods， while deep learning approaches still face limitations in prediction stability when load patterns change abruptly. This study proposes a Gradient-Adaptive Hybrid Load Forecasting Model (GRU-GA-WLS)， which integrates the strengths of statistical smoothing and deep learning through a two-stage residual forecasting architecture. In the first stage， a Gradient-Adaptive Weighted Least Squares (GA-WLS) method is introduced， which adjusts smoothing weights based on the first-order differences between adjacent time steps and incorporat es a Gaussian Mixture Model (GMM) to automatically identify time-of-day segment boundaries for each site and season， thereby establishing a dynamic periodic baseline with dual adaptivity to site and seasonal characteristics. In the second stage， a Gated Recurrent Unit (GRU) residual forecasting network enhanced with Reversible Instance Normalization (RevIN) and a self-attention mechanism is employed to learn the nonlinea r residuals remaining after baseline explanation， with Bayesian Optimization used to search for optimal model configurations. The proposed model is validated on a daytime-peak regular-pattern site MT_200 (Case1) and a nighttime-peak dynamic-pattern site MT_162 (Case2)， across five forecasting horizons—15 minutes， 30 minutes， 60 minutes， 24 hours ahead， and 7 days ahead—and is systematically evaluated under both summe r and non-summer scenarios. Experimental results show that GRU-GA-WLS achieves the lowest MAPE across both sites for all forecasting horizons and seasonal scenarios. In the dynamic-pattern site MT_162 (Case2)， the results demonstrate the effectiveness of the GA-WLS gradient adaptive mechanism in tracking abrupt load transitions; in the regular-pattern site MT_200 (Case1)， the results confirm that the proposed model delivers stable forecasting performance superior to the baseline methods across sites with markedly different load characteristics.
