隨著太陽能發電在台灣能源結構中的比例迅速攀升,其發電的間歇性與不確定性對電網穩定性與調度作業帶來極大挑戰。為應對氣象預報誤差對太陽能發電預測精度的影響,提出「MTM二階段預測架構」,整合氣象資料修正模組與發電量預測模組,旨在顯著提升預測模型的準確性與穩定性。研究中採用XGBoost、LSTM及Time-GPT三種機器學習演算法,並通過系統化的資料前處理流程,包括資料清洗、異常值剔除與特徵工程,確保輸入資料品質。 本研究以台灣南部太陽能電場的歷史發電資料與氣象資料為基礎,比較Time-GPT、XGBoost及LSTM三種模型在不同訓練資料長度(1個月、3個月、6個月及9個月)下的表現,並採用標準化誤差指標,包括正規化平均絕對誤差(NMAE)與均方根誤差(RMSE),對模型性能進行全面評估。結果顯示,在MTM架構下,LSTM模型於3個月訓練資料時展現最佳性能,NMAE與RMSE分別達到2.72與24.99,相較於傳統單階段預測架構準確度分別提升15.3%與20.0%。特別在短期訓練資料(1個月與3個月)情境下,MTM架構對RMSE的改善尤為顯著,最高提升幅度達24.33%。此結果表明,MTM架構能有效緩解氣象預報不確定性對發電預測的影響,顯著提升預測的穩定性與精確度,尤其在資料量有限的情況下表現更為突出。 本研究提出的MTM二階段預測架構為太陽能發電預測提供了一種高效且穩健的解決方案,對電網調度優化及再生能源的高效利用具有重要的實務價值。然而,本研究目前僅聚焦於單一電場的數據分析,未來可進一步擴展至多地域、多氣候條件的太陽能電場,開發更具通用性的預測模型。;With the rapid increase in solar power generation within Taiwan′s energy structure, its intermittency and uncertainty pose significant challenges to grid stability and scheduling. To address the impact of meteorological forecast errors on solar power forecasting accuracy, this study proposes the "MTM Two-Stage Forecasting Framework," integrating meteorological data correction and power generation forecasting modules to enhance model accuracy and stability. Using XGBoost, LSTM, and Time-GPT algorithms, and systematic data preprocessing, the study evaluates model performance across various training data lengths. Results show that the LSTM model under the MTM framework achieves the best performance with 3 months of training data, significantly improving accuracy compared to traditional methods, especially with limited data. The MTM framework presents an efficient and robust solution for solar power forecasting, offering practical value in enhancing grid scheduling efficiency and maximizing renewable energy utilization. While the current research is based on data from a single solar power plant in southern Taiwan, future work may explore its application across multiple regions and diverse climatic conditions to develop more generalized and scalable forecasting models.